text
stringlengths 5
1.89M
| meta
dict | domain
stringclasses 1
value |
|---|---|---|
---
abstract: 'We study the influence of a thermal environment on a non-adiabatic spin-flip driving protocol of spin-orbit qubits. The driving protocol operates by moving the qubit, trapped in a harmonic potential, along a nanowire in the presence of a time-dependent spin-orbit interaction. We consider the harmonic degrees of freedom to be weakly coupled to a thermal bath. We find an analytical expression for the Floquet states and derive the Lindblad equation for a strongly non-adiabatically driven qubit. The Lindblad equation corrects the dynamics of an isolated qubit with Lamb shift terms and a dissipative behaviour. Using the Lindblad equation, the influence of a thermal environment on the spin-flip protocol is analysed.'
author:
- Brecht Donvil
- Lara Ulčakar
- Tomaž Rejec
- Anton Ramšak
bibliography:
- 'bibliography\_qubit\_list.bib'
title: 'Thermal effects on a nonadiabatic spin-flip protocol of spin-orbit qubits'
---
Introduction
============
Electron-spin qubits are promising candidates as building blocks of quantum computers. They can be realized in gated semiconductor devices based on quantum dots and quantum wires [@wolf01; @hanson07] and their state can be manipulated via magnetic fields [@dresselhaus55; @bychkov84] or the spin-orbit interaction, which is easily controlled with electrostatic gates [@stepanenko04; @flindt06; @coish06; @sanjose08; @golovach10; @bednarek08; @fan16; @gomez12; @pawlowski16; @pawlowski16b; @Pawlowski2017; @Pawlowski2018]. Such systems were already experimentally realized in various semiconducting devices [@nadjperge12; @nadjperge10; @fasth05; @fasth07; @shin12].
Recent studies proposed non-adiabatic protocols, where spin-qubit manipulation is achieved by translating a spin-qubit in one dimension [@cadez13; @cadez14; @Veszeli2018] in the presence of a time-dependent Rashba interaction [@nitta97; @liang12; @Yang2015]. Refs. [@cadez13; @cadez14] give exact analytical solutions for the time-dependent Schrödinger equation of such a driven spin-qubit, confined in a harmonic potential, and express the spin rotation in terms of the non-adiabatic non-Abelian Anandan phase [@anandan88]. While qubit transformations in linear systems are limited to spin rotations around a fixed axis, this limitation can be eliminated on a ring structure [@kregar16; @kregar16b].
Manipulations of quantum systems are inevitably accompanied by external noise, coming from fluctuating electric fields [created]{} by the piezoelectric phonons [@sanjose08; @sanjose06; @huang13; @echeveria13], for example, or due to phonon-mediated instabilities in molecular systems with phonon-assisted potential barriers [@mravlje06; @mravlje08]. Flying qubits could be carried by surface acoustic waves, where the noise can [arise]{} due to time dependence in the electron-electron interaction effects [@giavaras06; @rejec00; @jefferson06]. Recent related studies [@Gefen05; @lara17; @lara18; @Pyshkin2018; @Li2018; @Lu2018] considered the effects of additive noise present in the driving functions of the qubit.
In order to study effects of a thermal environment on the qubit manipulation, we aim to derive an effective dynamics for the spin-qubit from a full microscopic model of the qubit weakly interacting with a thermal bath. In the case of weak interactions, there exists a well-known approximation scheme to integrate out the dynamics of the bath and obtain the effective dynamics for the system, which is given by the Lindblad equation [@BreuerBook]. For adiabatic and weakly non-adiabatic driving of the system, the aforementioned weak-coupling scheme still holds and leads to a slightly modified Lindblad equation. However, for strongly non-adiabatic driving, which we consider in the present work, the necessary assumptions for this approximation scheme break down.
In a recent work[@Dann2018], the authors showed how a modified weak-coupling scheme can be performed in order to derive the Lindblad equation for an arbitrarily driven weakly-coupled system. In the case of periodic driving, this result reduces to the earlier derived Floquet-Lindblad form [@BlBu1991; @BPfloquet]. Although these results show how in principle a Lindblad equation can be derived, employing these methods essentially requires one to solve the evolution of the driven system in absence of the bath. In general, obtaining an analytic expression for the evolution of a driven system is non trivial and one has to resort to numerical approximations. In the present work, however, we consider periodic driving for which the analytic solution is known [@cadez13], giving us access to the exact Floquet-Lindblad equation.
The paper is structured as follows: in Section \[sec:model\] we introduce the model and the coupling to the thermal bath. In Section \[sec:Floquet\] the Floquet formalism and the corresponding exact solutions are given, which, in Section \[sec:Lind\], serves as a basis for an exact derivation of the Lindblad operators and the Floquet-Lindblad equation. The formalism is then applied to a simple example of a non-adiabatic driving and spin rotation in Section \[sec:example\]. In Section \[sec:conclusion\] we give conclusions and in Appendices A and B we present derivations of individual terms of the Floquet-Lindblad equation.
Model {#sec:model}
=====
Our system of interest is a spin-qubit represented as an electron confined in a quantum wire with a harmonic potential.[@cadez13; @cadez14] The centre of the trap, $\xi(t)$, can be arbitrarily translated along the wire by means of time-dependent external electric fields.
The spin-orbit Rashba interaction couples the electron’s spin with its orbital motion, resulting in the system Hamiltonian $$\label{H}
H(t)=\frac{p^{2}}{2m^{*}}+\frac{m^{*}\omega_0^{2}}{2}\big(x-\xi(t)\big)^{2}+\alpha(t)\,p\, \sigma_y,$$ where $m^{*}$ is the effective electron mass, $\omega_0$ is the frequency of the harmonic trap, and $p$ and $\sigma_y$ are the momentum and spin operators, respectively. The strength of the spin-orbit interaction $\alpha(t)$ is time dependent due to time dependent external electric fields and the spin rotation axis is fixed along the $y$-direction [@nadjperge12]. Throughout the paper we set $\hbar=1$ and initial time $t=0$. The exact time dependent solution of the Schr[ö]{}dinger equation corresponding to the Hamiltonian Eq. is given by[@cadez14] $$\label{eq:psi}
|\Psi_{}(t)\rangle=U(t,0)|\Psi_{}(0)\rangle,$$ where, in the time evolution operator $$\label{eq:ev}
U(t,0)=\mathcal{U}^\dagger(t)e^{-i H_{0} t}\mathcal{U}(0),$$ $H_0$ represents the time independent harmonic oscillator, i.e., Eq. (\[H\]) with $\xi(t)=\alpha(t)=0$, and
$$\begin{aligned}
\label{psi}\label{eq:transf}
\mathcal{U}^\dagger(t)&=&e^{-{i }(\varphi_0(t)+\varphi(t) \sigma_y)}\mathcal{A}_{\alpha}(t)\mathcal{X}_{\xi}(t),\\
\mathcal{A}_{\alpha}(t)&=&e^{-i\dot{a}_{c}(t)p\sigma_y/\omega_0^2}
e^{-im^{*}a_{c}(t){x}\sigma_y},\\
\mathcal{X}_{\xi}(t)&=&e^{im^{*}\big(x-x_{c}(t)\big)\dot{x}_{c}(t)}e^{-ix_{c}(t)p}.\end{aligned}$$
The unitary transformations $\mathcal{X}_{\xi}(t)$ and $\mathcal{A}_{\alpha}(t)$ are completely determined by the classical responses $x_c(t)$ and $a_c(t)$ to the driving. The responses solve the differential equations
\[xcac\] $$\begin{aligned}
\ddot{x}_c(t)+\omega_0^{2}x_{c}(t)&=&\omega_0^{2}\xi(t),\\
\ddot{a}_c(t)+\omega_0^{2}a_{c}(t)&=&\omega_0^{2}\alpha(t).\end{aligned}$$
$\varphi(t)=-m^{*}\int_{0}^{t}\dot{a}_{c}(\tau)\xi(\tau)\mathrm{d}\tau$ while the time dependent phase $\varphi_0(t)$, given in Ref. , is irrelevant for the time evolution operator Eq. (\[eq:ev\]) and will be omitted. Also note that the time evolution operator is invariant with respect to the gauge transformation $$\label{gauge}
\mathcal{U}^\dagger(t) \to e^{i(\delta_1+\delta_2 \sigma_y)} \mathcal{U}^\dagger(t),$$ where $\delta_1$ and $\delta_2$ are real constants.
The system described by the time dependent $H(t)$ is coupled to a bosonic thermal bath. The total Hamiltonian of the spin-qubit interacting with the bath is $$\label{eq:Hamiltonian}
H_{\rm tot}(t)=H(t)+H_B+H_I(t),$$ where the bath Hamiltonian $H_B$ represents a set of oscillators $$H_B=\sum_{\mathbf{k}} \omega_{\mathbf{k}}b^\dagger_{\mathbf{k}} b_{\mathbf{k}},$$ where $b^\dagger_{\mathbf{k}}$ ($ b_{\mathbf{k}}$) are creation (annihilation) operators and the oscillators have a linear dispersion relation $\omega_{\mathbf{k}}=c|\mathbf{k}|$, $\mathbf{k}\in \mathbb{R}^3$. We consider only states with energies below a cut-off energy $\omega_c$.
The spin-qubit is coupled to the bath through the interaction Hamiltonian $H_I(t)\propto\left(x-\xi(t)\right)\sum_{\mathbf{k}} x_{\mathbf{k}}$ which couples the respective position operators, $$\label{eq:intHam}
H_I(t)=g\sqrt{\frac{\pi c^3}{\omega_0 v}}\left(a^\dagger+a-\sqrt{2m^*\omega_0}\xi(t)\right)\sum_{\mathbf{k}}(b_{\mathbf{k}}+b^\dagger_{\mathbf{k}}),$$ where $g$ is a dimensionless coupling strength, $a^\dagger$ ($ a$) is a bosonic creation (annihilation) operator of the harmonic trap and $v$ is the volume of the bath.
Floquet Theory for quantum systems {#sec:Floquet}
==================================
Before deriving the Lindblad equation for the driven spin-qubit, it is instructive to briefly discuss the Floquet theory for periodically driven quantum systems. For systems with a Hamiltonian periodic in time, $H(t+T_d)=H(t)$, it is possible to describe the time-evolution in terms of periodic eigenvectors of the Schrödinger equation called the Floquet states [@Shirley; @Zeldovich]. The Floquet states $|\phi_q(t)\rangle$ form a complete basis and are defined as solutions of the eigenvalue problem
\[eq:Flset\] $$\begin{aligned}
&(H(t)-i\partial_t)|\phi_q(t)\rangle=\epsilon_q|\phi_q(t)\rangle,\\
&|\phi_q(t+T_d)\rangle=|\phi_q(t)\rangle,\end{aligned}$$
where $\epsilon_q$ is called a quasi-energy. The evolution of an arbitrary initial state $|\psi_0\rangle$ from an initial time $0$ to time $t$, expressed in terms of the Floquet states, is $$|\psi(t)\rangle=\sum_q e^{-i\epsilon_q t}|\phi_q(t)\rangle \langle\phi(0)|\psi_0\rangle.$$
In practice, solving Eq. proves to be non-trivial. However, with a gauge transformation $V(t)$ such that $$\label{eq:flfl}
G=V(t)H(t)V(t)^\dagger+i(\partial_t V(t))V(t)^\dagger$$ is time independent, the Floquet states and quasi-energies can be found in terms of eigenvectors and eigenvalues of $G$. Let $|q\rangle$ be an eigenvector of $G$ with the eigenvalue $\epsilon_q$. One can check that the state $$\label{eq:fl}
|\phi_q(t)\rangle= V(t)^\dagger |q\rangle$$ is a Floquet state with the quasi-energy $\epsilon_q$. Note that $\mathcal{U}(t)$, as defined in Eq. , has exactly the property Eq. with $G=H_0$. Therefore, the Floquet states of the driven qubit are $|\phi_q(t)\rangle= \mathcal{U}^\dagger(t) |\psi_q\rangle$, where $|\psi_q\rangle$ is an eigenstate of $H_0$ with the energy $\epsilon_q=\omega_0(q+\frac{1}{2})$.
Derivation of the Lindblad equation {#sec:Lind}
===================================
The reduced density matrix of the spin-qubit at time $t$ in the Schrödinger picture is $$\bar{\rho}(t)=\operatorname{tr}_B\left( U_{\textrm{tot}}(t,0)\rho_{\textrm{tot}}(0)U_{\textrm{tot}}^\dagger(t,0)\right)$$ where $\mathrm{tr}_B$ denotes the trace over bath degrees of freedom, $U_\textrm{tot}(t,0)$ the time evolution operator of the whole system, and $$\rho_\textrm{tot}(0)=\rho(0)\otimes \frac{e^{-\beta H_B}}{\operatorname{tr}_B e^{-\beta H_B}},
\label{rhotot}$$ is an initially separable density matrix consisting of the qubit in the state $\rho(0)$ and the bath at the inverse temperature $\beta=\frac{1}{k_BT}$. In the interaction picture, $$\rho(t)=U^\dagger(t,0)\bar{\rho}(t)U(t,0)$$ where $U(t,0)$ is the time evolution operator of the qubit, the Floquet-Lindblad equation for the qubit interacting with the bath via Eq. is of the form[@BreuerBook] $$\begin{aligned}
\label{eq:lind}
\frac{{\mathop{}\!\mathrm{d}}}{{\mathop{}\!\mathrm{d}}t}\rho(t)=-{i}\big[H_{LS},\rho(t)\big]+\mathcal{D}\big(\rho(t)\big).\end{aligned}$$
In what follows we assume that the driving frequency $\omega_d=\frac{2\pi}{T_d}$ is $\omega_d=\frac{\omega_0}{n_d}$ with $n_d\in\mathbb{N}$, as appropriate for the spin-flip protocol studied in this paper. The first term on the right hand side of Eq. (\[eq:lind\]) contains the Lamb shift Hamiltonian $$\label{Lambshift}
H_{LS}=\sum_{n\in \mathbb{Z}} S(n\omega_d) A_n^\dagger A_n,$$ where $A_n$ are the Lindblad operators (to be defined below) and $$S(\omega)=\frac{g^2}{2\pi\omega_0}{\cal P}\!\!\int_0^{\omega_c} {\mathop{}\!\mathrm{d}}\omega' \omega'^2\left(\frac{1+N(\omega')}{\omega-\omega'}+\frac{N(\omega')}{\omega+\omega'}\right)$$ with $\cal P $ denoting the principal value and $$N(\omega)=\frac{1}{\exp (\beta\omega)-1}$$ being the Bose occupation numbers of the bath degrees of freedom. Throughout the paper we express the temperature in terms of the average occupation of the system oscillator $n_T=N(\omega_0)$. In Appendix \[secapp:ladder1\] we show the Lamb shift Hamiltonian in an explicit form.
The second term on the right hand side of Eq. is the dissipator term $$\label{eq:diss}
\mathcal{D}\big(\rho(t)\big)=\sum_{n\in \mathbb{Z}} \gamma(n\omega_d)\left(A_n\rho(t)A_n^\dagger -\tfrac{1}{2}\{A_n^\dagger A_n,\rho(t)\}\right),$$ with rates $$\gamma(\omega)=\begin{cases}g^2\frac{\omega^2}{\omega_0}(1+N(\omega)), &\omega\ge0,\\g^2\frac{\omega^2}{\omega_0}N(|\omega|),&\omega<0.\end{cases}$$
The Lindblad operators are obtained by finding $A_n$ such that [@BPfloquet; @BreuerBook] $$\begin{aligned}
\label{eq:ladders}
U^\dagger(t,0)\left(a^\dagger+a-\sqrt{2m^*\omega_0}\xi(t)\right)U(t,0)\nonumber\\
=\sum_{n\in\mathbb{Z}} A_n e^{-i n\omega_d t}.\end{aligned}$$ We present the actual calculation of the Lindblad operators in Appendix \[secapp:ladder\]. The result for $n\ge0$ is $$\begin{aligned}
\label{eq:ladder1}
A_n&=&\delta_{n,n_d}\,a\nonumber
+\sqrt{{ 2m^* \omega_0}}\bigg(\hat{x}_{c,n}-\hat{\xi}_n+\frac{\hat{\dot{a}}_{c,n}}{\omega^2_0}\sigma_y+
\nonumber\\
& &+\delta_{n,n_d}\frac{1}{2}\left(-\frac{\dot{a}_c(0)}{\omega_0^2}+i\frac{a_c(0)}{\omega_0}\right)\sigma_y+
\nonumber\\
& &+\delta_{n,n_d}\frac{1}{2}\left(-x_c(0)-i\frac{\dot{x}_c(0)}{\omega_0}\right) \bigg),\end{aligned}$$ and $A_{-n}=A_n^\dagger$. Here by $\hat{f}_n$ we denote a Fourier component of the function $f$. The Lindblad operators are completely determined by the solutions of the response Eqs. .
We can greatly simplify the form of the dissipator Eq. by absorbing the constant terms in the Lindblad operators into the Lamb shift. This procedure is outlined in Appendix \[secapp:ladder1\]. Let us define a rate $$\bar{\gamma}=\frac{2 m^* }{\omega_0^3}\sum_{\substack{n\in \mathbb{Z}\\ n\ne\pm n_d}} \gamma(n\omega_d)|\hat{\dot{a}}_{c,n}|^2,$$ and a new Lindblad operator $$\bar{A}=a+\sqrt{\frac{m^* }{2\omega_0}}\bigg(2\frac{\hat{\dot{a}}_{c,n_d}}{\omega_0}-\frac{\dot{a}_c(0)}{\omega_0}+ia_c(0)\bigg)\sigma_y.$$ With these definitions, Eq. reduces to a much simpler form $$\begin{aligned}
\label{eq:Lindf}
\frac{{\mathop{}\!\mathrm{d}}}{{\mathop{}\!\mathrm{d}}t}\rho(t)=&-{i}\left[\bar{H}_{LS},\rho(t)\right]+\bar{\gamma} \left(\sigma_y\rho(t)\sigma_y-\rho(t)\right)+
\nonumber\\&+ \gamma(\omega_0)\left(\bar{A}\rho(t)\bar{A}^\dagger-\tfrac{1}{2}\left\{\bar{A}^\dagger\bar{A},\rho(t)\right\}\right)+\nonumber\\&+ \gamma(-\omega_0)\left(\bar{A}^\dagger\rho(t)\bar{A}-\tfrac{1}{2}\left\{\bar{A}\bar{A}^\dagger,\rho(t)\right\}\right).\end{aligned}$$ The redefined Lamb shift Hamiltonian $\bar{H}_{LS}$ is shown explicitly in Appendix \[secapp:ladder1\]. The dissipator in the above equation consists of two types of terms: the term proportional to $\bar{\gamma}$ is a dephasing term and causes a decay of spin size. The other terms lead to thermal activation in the oscillator component.
Example {#sec:example}
=======
At the initial time $t=0$, let the electron be in the ground state manifold of $H(0)$, spanned by a Kramers doublet. In particular, we choose the qubit to be in the spin-up state, i.e., with $\langle\Psi(0)| \sigma_x |\Psi(0)\rangle=\langle\Psi(0)| \sigma_y |\Psi(0)\rangle=0$. Using Eq. (\[gauge\]), such a state can be constructed as $$|\Psi(0)\rangle=e^{i m^*\big({a}_{c}(0){x}_{c}(0)+\frac{\dot{a}_{c}(0)\dot{x}_{c}(0)}{\omega_0^2}\big)\sigma_y}
\mathcal{U}^\dagger(0)|\psi_{0}\rangle|\chi_{\uparrow}\rangle.$$ Here $|\psi_{0}\rangle$ is the ground state of the harmonic oscillator Hamiltonian $H_0$ and $|\chi_{\uparrow}\rangle$ is the up-state spinor in the eigenbasis of $\sigma_z$. The initial density matrix of the qubit Eq. (\[rhotot\]) is $\rho(0)=|\Psi(0)\rangle\langle\Psi(0)|$.
As a simple example of the theory outlined in the previous section, we consider periodic spin transformations following an elliptic path in the parametric space $[\xi(t),\alpha(t)]$ with
\[driving\] $$\begin{aligned}
&\xi(t) = \xi_0 \cos\frac{\omega_0 t}{2},\\
&\alpha(t)=\alpha_0-\alpha_0 \sin\frac{\omega_0 t}{2},
\end{aligned}$$
i.e, $n_d=2$. This choice of driving, together with initial conditions $x_c(0)=\xi(0)$, $a_c(0)=\alpha(0)$ and $\dot{x}_c(0)=0$, $\dot{a}_c(0)=0$ for differential equations , leads to classical responses
\[eq:exresp\] $$\begin{aligned}
x_c(t)&=&\frac{\xi_0}{3} \left(4 \cos\frac{ \omega_0 t}{2}-\cos\omega_0 t \right),\\
a_c(t)&=&\alpha_0-\frac{\alpha_0}{3} \left(4 \sin\frac{\omega_0 t}{2} -2\sin \omega_0 t\right).
\label{response}
\end{aligned}$$
The driving guarantees that after a completed cycle the state $|\Psi(T_d)\rangle$ returns to the ground state – with spin rotated around the $y$-axis by the Anandan quantum phase $\phi_A=2\varphi(T_d)$ determined solely by the contour $\mathcal{C}_{\xi}$ in the parametric space $[\xi(t),a_c(t)]$ or, equivalently, $\mathcal{C}_{\alpha}$ in the space $[x_c(t),\alpha(t)]$, $$\label{anandan}
\phi_A=2m^{*}\oint_{\mathcal{C}_{\xi}} a_{c}[\xi]{\rm d}\xi=2m^{*}\oint_{\mathcal{C}_{\alpha}}x_{c}[\alpha]{\rm d}\alpha.$$ The corresponding contour ${\cal{C}}_{\alpha}$ is shown in Fig. \[fig:Bloch\](a) with a full black line and the dashed line indicates the driving protocol $[\xi(t),\alpha(t)]$. Note that the area of the shaded region equals $\phi_A/(2m^*)$.
During the motion of the system, the electron’s spatial wave function is a superposition containing also the oscillator’s excited states and the spin of the electron is rotated around the $y$-axis by the angle $\phi(t)$. After a completed driving cycle the electron’s spin has, according to Eq. (\[anandan\]), rotated by the angle $\phi(T_d)=\phi_A =\frac{8}{3} \pi m^*\alpha_0\xi_0$.
In the upcoming numerical studies we take $\gamma/\omega_0\lesssim \frac{1}{10}$. This is to ensure that the effect of the bath remains a perturbation to the free dynamics of the spin-qubit, which is the underlying assumption for the Lindblad equation Eq. . We also fix the Rashba coupling $\alpha_0=\frac{3}{16} (\omega_0/m^*)^{1/2}$ and $\xi_0=(m^*\omega_0)^{-1/2}$ such that $\phi_A={\pi \over 2}$, i.e., two driving cycles are needed for a spin-flip. The cut-off frequency is set to $\omega_c=2 \omega_0$ and all numerical calculations were carried out using the QuTiP framework [@qutip12; @qutip13].
Spatial position and spin properties
------------------------------------
In Fig. \[fig:Bloch\](b) we show the time dependence of the expectation value of the spin. Since we are considering rotations around the $y$-axis, $\langle \sigma_y \rangle=0$, i.e., the expectation value is confined to the $x$-$z$ plane and [*within*]{} the Bloch sphere. Bullets represent the values at equal time steps for the total time duration of two cycles. Note that in the absence of the interaction, $g=0$, the spin transformation is exactly one spin-flip (black). The red bullets represent the corresponding result for $g=0.2$, $n_T=1$ and $\omega_c=2\omega_0$. The orange square indicates the spin at $t=T_d$.
![(a) Contour ${\cal{C}}_{\alpha}$ of the driving and the response in the parametric space spanned by the spatial position and the spin-orbit driving. The red bullet represents the starting point and the orange square the response $[\langle x\rangle,\alpha]$ after one cycle for $g=0.3$, $n_T=1$ and $\omega_c=2\omega_0$. (b) Two cycles of the qubit rotation within the Bloch sphere: the black bullets and the black arrow represent the noninteracting result, $g=0$, with spin rotation around the $y$-axis for an angle $\phi=\pi$. The red bullets and the red arrow show the rotation for $g=0.2$, $n_T=1$ and $\omega_c=2\omega_0$. Bullets show positions at equal time intervals.[]{data-label="fig:Bloch"}](Fig1.pdf){width="5"}
The interaction with the bath influences the spin-flip protocol, namely: the angle of the spin rotation, the size of the spin, the expectation value of the position $\langle x \rangle$ and the oscillator part of the wave function, which does not return to the ground state at the end of the transformation. Below we discuss these effects.
The influence of the interaction with the bath on the expected position of the oscillator $\langle x\rangle$ can be observed from Figs. \[fig:Bloch\](a) and \[fig:x\](a) by comparing the result for the non-interacting pure dynamics (black line) to the interacting result (red line). In the non-interacting case, the expected position is equal to the classical response, $\langle x\rangle=x_c(t)$, as can easily be shown from Eq. . In the presence of the bath $\langle x\rangle$ deviates from $x_c(t)$ and is shifted from the starting point (red bullet) after one cycle (orange square). Fig. \[fig:x\](a) shows that the shift of $\langle x\rangle$ to larger values is increased after the second cycle, $t=2T_d$, and is further increased until the driven system reaches a steady state. This is in agreement with the classical result for a driven weakly damped harmonic oscillator, $x_c \to {4 \over 3}\xi_0$, after a large number of cycles \[the first term in Eq. (\[eq:exresp\]a)\]. The approach to the steady state can be observed in Fig. \[fig:x\](b) which shows $\langle x\rangle$ in terms of the number of driving cycles (needed to half-flip the spin), for different values of the coupling strength $g$ at $T=0$. Due to the low frequencies introduced by the Lamb shift in Eq. , $\langle x\rangle$ exhibits slow damped oscillations. The inset of Fig. \[fig:x\](b) shows $\langle x\rangle$ after one cycle as a function of $g$ for various values of the cut-off energy $\omega_c$. Note the absence of oscillations for $\omega_c=1.16\, \omega_0$, dotted line. At this cut-off energy $S(\omega_0)+S(-\omega_0)\to 0$ and consequently spatial parts of $H_{LS}\to 0$, see Eqs. (\[HLS\]) and (\[Spm\]), which leads to results qualitatively close to the classical result for a damped oscillator. At higher cut-off energies quantum terms in $H_{LS}$ leads to a much richer dynamics. The orange squares and circles indicate the spatial position after the first cycle, Figs. \[fig:Bloch\](a) and \[fig:x\]. The shift of the position and consequently the changed contour in the parametric space affect the spin behaviour, as will be discussed below.
![(a) Driving $\xi(t)$ as a function of time for two cycles (dotted line) and the expected position of the electron $\langle x \rangle$ in the case of no interaction (black) and for $g=0.3$, $n_T=1$ (red). The orange square corresponds to the square in Fig. \[fig:Bloch\]. (b) $\langle x \rangle$ as a function of the number of qubit transformation cycles, for different values of the coupling $g$ at $T=0$. Inset: $\langle x \rangle$ after one cycle as a function of $g$ for $\omega_c/\omega_0=1.16$, 2, and 3. The orange circle corresponds to $g=0.3$.[]{data-label="fig:x"}](Fig2.pdf){width="6.5"}
The spin response of the system is analysed in Fig. \[fig:spin\]. First we concentrate on the size of the spin, which for $g=0$ does not depend on the position coordinates regardless of the driving, as can be checked by the application of the unitary transformation Eq. (\[psi\]) and a direct evaluation of the spin expectation values. The exact result is $$\label{eq:sigmana}
\sqrt{\langle \sigma_x \rangle^2+\langle \sigma_z \rangle^2}=e^{- {m^*} \big(a^2_c(t)/\omega_0 +{\dot a}_c^2(t)/\omega_0^3\big)},$$ which reduces to $e^{- {m^*} \alpha(t)^2/\omega_0 }$ for the case of slow (adiabatic) driving. In the case of a time independent Rashba coupling this is a known result for Kramers doublets [@cadez13]. Fig. \[fig:spin\](a) shows the size of the spin during the spin-flip protocol in absence of the environment (full black line), the zero temperature result at $g=0.3$ (red line) and the result at a finite temperature $n_T=1$ (dashed line). The orange square indicates the result after the first cycle, as in previous figures. Effects of the coupling to the bath are more pronounced at an elevated temperature.
![(a) Time dependence of the spin size $\sqrt{\langle \sigma_z \rangle^2+\langle \sigma_x \rangle^2}$ during two cycles for $g=0$ (black line), $g=0.3$, $T=0$ (red) and $g=0.3$, $n_T=1$ (dashed). (b) Angle $\phi$ as a function of time. Inset: Angle error $\delta \phi=\phi-{\pi \over 2}$ at the final time of one cycle $t=T_d$ as a function of $g$. In (a) and (b) the red bullets and the orange squares indicate the one cycle starting and final values, correspondingly.[]{data-label="fig:spin"}](Fig3.pdf){width="8.5"}
The spin rotation, the most relevant property for the qubit manipulation, is shown in Fig. \[fig:spin\](b) for $g=0$ (black) and $g=0.3$, $n_T=0$ (red). Due to the interaction with the bath the rotation angle is slightly increased with respect to the non-interacting value $\phi={\pi \over 2}$ \[the orange square marks the value after one completed cycle\]. The deviation $\delta\phi=\phi-{\pi \over 2}$ is additionally presented as a function of $g$ in the inset. Comparing the finite temperature result $n_T=1$ (blue dashed line) with $n_T=0$ (full red) shows that the effect is further increased at finite temperatures.
![(a) Deviation of the angle $\phi$ from $\frac{\pi}{2}$ after $N_{\rm c}$ cycles using renormalised Rashba coupling $\frac{\alpha_0}{N_c}$ for various $g$ at $T=0$. (b) Number of cycles $N_g$ from (a) as a function of $g$. Note the orange symbol relating (a) and (b). The dotted line represents the approximation $N_g\approx 0.36 \,g^{-2}$.[]{data-label="fig:angle"}](Fig4.pdf){width="8"}
Error analysis and fidelity
---------------------------
Deviations of $\phi$ from the target value ${\pi \over 2}$ are additionally explored and presented in Fig. \[fig:angle\]. Here we performed $N_{\rm c}$ consequent cycles, with renormalised value of the Rashba coupling $\alpha_0(N_{\rm c}) = \alpha_0 / N_{\rm c}$ such that after $N_{\rm c}$ cycles at $g=0$ exactly one half spin-flip is performed. Fig. \[fig:angle\](a) shows that initially the error decreases with an increasing number of cycles. For a large number of cycles, however, the error approaches a constant value. The typical transition number of cycles $N_g$ is for $g=0.3$ indicated by an orange square and an arrow. In Fig. \[fig:angle\](b), $N_g$ is shown as a function of $g$. It clearly exhibits a $g^{-2}$ scaling as expected for the typical relaxation time scale (measured by the number of cycles) for error generating Lamb shift Hamiltonian and Lindblad terms, Eq. (\[eq:lind\]).
Let us discuss the effects of the environment also in the framework of fidelity of the spin-qubit transformation. In particular, we consider the Uhlman-Josza fidelity $$F=\operatorname{tr}\sqrt{\sqrt{\rho_0}\rho_g\sqrt{\rho_0}},$$ where $\rho_0$ and $\rho_g$ represent the density matrix for the non-interacting and the interacting regimes of the model, respectively. At the initial time $t=0$ (or in the absence of interaction) $F=1$, but with increasing time $F$ progressively diminishes due to error generating processes in the Lindblad equation. Fig. \[fig:fidelity\] shows the fidelity calculated at the end of each cycle (red dots) as a function of the number of cycles $N_{\rm c}$ for $g=0.2$ and $T=0$. Here, the Rashba coupling $\alpha_0$ is independent of the number of cycles $N_{\rm c}$.
The structure of $F$ exhibits different short time and long time behaviours. To analyse these behaviours, we make a simple estimate of fidelity at the end of the driving cycle, $$F\approx |\langle\psi_0|\psi_{\delta x}\rangle|^2 |\langle\chi_0|\chi_{\delta \phi}\rangle|^2,
\label{F}$$ where $|\psi_0\rangle$ is the target final harmonic oscillator ground state and $|\psi_{\delta x}\rangle$ is the ground state of the harmonic oscillator with the potential displaced by $\langle x\rangle-\xi_0=\delta x\xi_0$, giving $$\label{eq:ov1}
|\langle\psi_0|\psi_{\delta x}\rangle|^2=e^{-{1 \over 2}\delta x^2}.$$ Similarly, $|\chi_{0}\rangle$ is the target final spin state and $|\chi_{\delta \phi}\rangle$ is the spin state with the angle $\delta \phi=\phi-N_{\rm c} {\pi \over 2}$ off from the target state, so that $$\label{eq:ov2}
|\langle\chi_0|\chi_{\delta \phi}\rangle|^2=\cos^2\frac{\delta \phi}{2}.$$ The fidelity (red line and bullets) and its estimate Eq. (black), shown in Fig. \[fig:fidelity\] for $g=0.2$, behave qualitatively similar. At zero temperature considered here the two overlaps Eqs. and represent major sources of the fidelity reduction. The remaining contributions, much more pronounced at finite temperatures (not shown), are mainly due to the fact that the system is not in a pure state and $F$ simply cannot be expressed solely in terms of wave function overlaps.
The separate curves $|\langle\psi_0|\psi_{\delta x}\rangle|^2$ (green dots) and $\langle\chi_0|\chi_{\delta \phi}\rangle|^2$ (blue dots) allow us to analyse the short and long-time behaviour of the fidelity. For small number of cycles, $N_{\rm c} \lesssim 10 $, the fidelity is mainly reduced due to the shift of the electron position $\delta x>0$ after a completed cycle. This affects the fidelity due to the reduced overlap of the target spatial wave function and the actual result in the presence of interaction. As discussed before, $\delta x$ exhibits oscillatory behaviour which is damped out (see Fig. \[fig:x\]) and $\delta x \to {1 \over 3}$ at larger times $t \gtrsim 10 T_d$, hence the overlap approaches $|\langle\psi_0|\psi_{\delta x}\rangle|^2=e^{-1/18} = 0.95$.
The spin contribution to the fidelity reduction due to the error in the angle of rotation, $\cos^2\frac{\delta \phi}{2}$, is at short times also generated due to oscillations of the orbit $[\langle x \rangle,\alpha]$ and the corresponding deviations from the noninteracting contour ${\cal C}_\alpha$, see Fig. \[fig:Bloch\](a). For $N_{\rm c} \gtrsim 10 $ the orbit in the parametric space progressively relaxes to the steady state contour and then the error $\delta\phi$ increases monotonously, similar to the recent study of adiabatic non-Abelian dephasing [@gefen19a; @gefen19b]. There are several competing error-generating sources also in the Lamb shift Hamiltonian $H_{LS}$, Eq. (\[HLS\]): the spin rotation terms are of the Rashba coupling form $\propto p\, \sigma_y$, a space dependent magnetic field $\propto x\,\sigma_y$ and a constant magnetic field term $\propto \sigma_y$. Dissipative terms in the Lindblad equation are another important source of the fidelity reduction at larger times. They additionally contribute to spin errors and most importantly, to the size of the spin, shown in Fig. \[fig:spin\](a). At elevated temperatures dephasing effects discussed above amplify due to the increase of coupling factors $S(\omega)$ and $\gamma(\omega)$.
![Uhlmann-Josza fidelity $F$ as a function of the number of cycles for $g=0.2$ and $T=0$ (red). The green squares represent $|\langle\psi_0|\psi_{\delta x}\rangle|^2$, the overlap of two ground states of a harmonic oscillator, one at the origin and the other displaced by $\delta x$. The blue squares represent the overlap $|\langle \chi_0|\chi_{\delta\phi} \rangle |^2$ of two spin states with relative spin angle $\delta \phi$. The black bullets represent the combined overlaps as an estimate of the fidelity.[]{data-label="fig:fidelity"}](Fig5.pdf){width="6"}
Conclusions {#sec:conclusion}
===========
In this paper we have studied the effects of a thermal environment on a non-adiabatic spin-flip protocol. The protocol is based on confining an electron in a harmonic trap and simultaneously manipulating the position of the centre of the trap and the Rashba interaction. For arbitrary driving protocols, assuming weak coupling between the system and its environment, the effective dynamics of the system can be obtained in terms of the Lindblad equation [@Dann2018]. In the case of periodic driving it reduces to the Floquet-Lindblad equation[@BlBu1991; @BPfloquet]. However, to obtain the explicit form of the Floquet-Lindblad equation one needs to solve the closed system dynamics, which in the case of periodic driving translates to solving the eigenvalue problem Eqs. . The protocol we are considering is exactly solvable[@cadez13], and the resulting Floquet-Lindblad equation is fully determined by two classical responses Eqs. to the driving.
Access to the Floquet-Lindblad equation allows us to study the effects of the thermal environment on the driving protocol. The Lindblad equation modifies the free spin-qubit Hamiltonian by introducing low frequencies, an effect called the Lamb shift. Additionally, there are dissipative terms giving rise to spin-dephasing effects as well as thermal activation of the oscillator.
As an example we consider a specific driving protocol with classical response functions Eq. . Interestingly, we find that the low frequencies and other terms introduced by the Lamb shift Hamiltonian result in an optimal number of driving cycles to complete the protocol. Figure \[fig:angle\] displays this result, where depending on the coupling strength to the bath, the optimal number of driving cycles allows us to minimise the error in the final angle of the spin. The Lamb shift Hamiltonian is a time independent shifted harmonic oscillator with spin dependent terms, and the exact solution is a coherent state analogous to Eq. (\[eq:psi\]). This allows for an analytical analysis and a deeper understanding of the decoherence dynamics, enabling further possibilities of protocol optimisation at zero and finite temperatures.
The exact study of the considered spin-qubit interacting with the environment can be extended to any driving protocol and any kind of a bath as long as the interaction is weak.
Calculation of the Lindblad operators {#secapp:ladder}
=====================================
The time evolution operator Eq. of the free spin-qubit consists of three terms. We calculate $U^\dagger (t,0) (a+a^\dagger-\sqrt{{2m^*\omega_0}}\xi(t) )U(t,0)$ by first applying $\mathcal{U}^\dagger(t)$, resulting in the expression $$a+a^\dagger+\sqrt{{2m^* \omega_0 }}\left(x_c(t)-\xi(t)+\frac{\dot{a}_c(t)}{\omega^2_0}\sigma_y\right).$$ Applying the time independent harmonic oscillator term $e^{-iH_0 t}$ transforms this into $$a e^{-i\omega_0 t}+a^\dagger e^{i\omega_0 t}+\sqrt{{2m^* \omega_0 }}\left(x_c(t)-\xi(t)+\frac{\dot{a}_c(t)}{\omega^2_0}\sigma_y\right).$$ Finally, we apply $\mathcal{U}(0)$ and thus obtain $$\begin{split}
&a e^{-i\omega_0 t}+a^\dagger e^{i\omega_0 t}+\sqrt{{2m^* \omega_0 }}\bigg(x_c(t)-\xi(t)+\frac{\dot{a}_c(t)}{\omega^2_0}\sigma_y-\\
&-x_c(0)\cos(\omega_0t)-\frac{\dot{a}_c(0)}{\omega_0^2}\cos(\omega_0t)\sigma_y+\frac{a_c(0)}{\omega_0}\sin(\omega_0t)\sigma_y-\frac{\dot{x}_c(0)}{\omega_0}\sin(\omega_0t)\bigg).
\end{split}$$ Expressing $x_c(t)$ and $a_c(t)$ in terms of their Fourier components $\hat{x}_{c,n}$ and $\hat{a}_{c,n}$ results in the form Eq. , $$\begin{split}
&a e^{-i\omega_0 t}+a^\dagger e^{i\omega_0 t}+\sqrt{{2m^* \omega_0 }}\sum_{\substack{n\in \mathbb{Z}\\n\ne\pm n_d}}
\left(\hat{x}_{c,n}-\hat{\xi}_n+\frac{\hat{\dot{a}}_{c,n}}{\omega^2_0}\sigma_y\right)e^{-in \omega_d t}+\\
&+\sqrt{{\frac{m^* \omega_0}{2} }}\left(-x_c(0)-i\frac{\dot{x}_c(0)}{\omega_0}-\frac{\dot{a}_c(0)}{\omega_0^2}\sigma_y+i\frac{a_c(0)}{\omega_0}\sigma_y\right){e^{-i\omega_0 t}}+\\
&+\sqrt{{\frac{m^* \omega_0}{2} }}\left(-x_c(0)+i\frac{\dot{x}_c(0)}{\omega_0}-\frac{\dot{a}_c(0)}{\omega_0^2}\sigma_y-i\frac{a_c(0)}{\omega_0}\sigma_y\right){e^{i\omega_0 t}},
\end{split}$$ from which we read the jump operators of Eq. . The Fourier components $\hat {f}_n$ are defined by $f(t)=\sum_{n\in\mathbb{Z}} {\hat f}_n e^{-i n \omega_d t}$ and are interconnected by useful relations $\hat{\dot{a}}_{c,n}=-i {n \over n_d} \omega_0 \hat{a}_{c,n}$ and, for $n\ne\pm n_d$, $\hat{a}_{c,n}={n_d^2 /( {n_d^2-n^2})}\hat{\alpha}_n$ and $\hat{x}_{c,n}-\hat{\xi}_n={n^2 /( {n_d^2-n^2})}\hat{\xi}_n$.
Lamb shift Hamiltonian {#secapp:ladder1}
======================
The Lindblad equation is invariant under inhomogeneous transformations
\[B1\] $$\begin{aligned}
& A_n\rightarrow \bar{A}_n=A_n+z_n,\\
& H_{LS}\rightarrow \bar{H}_{LS}=H_{LS} +\frac{1}{2i}\sum_{n\in\mathbb{Z}} \gamma(n\omega_d)\left(z_n^\ast A_n-z_nA_n^\dagger\right)+c\end{aligned}$$
where $z_n\in \mathbb{C}$ and $c\in \mathbb{R}$. The Lindblad operators Eq. consist of two parts, one proportional to $\sigma_y$, $a$ and $a^\dagger$ and the other proportional to the identity. The latter can be eliminated using the above transformation, leading to Eq. with the transformed Lamb shift Hamiltonian $$\label{hlsbar}
\bar{H}_{LS}=H_{LS}
-\frac{1}{2i}g^2\omega_0\sqrt{2 m^* \omega_0}\left(\big(\hat{x}_{c,-n_d}-\hat{\xi}_{-n_d}-\frac{1}{2}x_c(0)+i{\dot{x}_c(0) \over 2\omega_0}\big)A_{n_d} - \mathrm{h.c.}\right)+\bar{B}\sigma_y$$ where $$\bar{B}=2g^2 m^*\sum_{\substack{n\in \mathbb{N}\\ n\ne n_d}}
{n^5 \omega_0\over n_d\left( {n_d^2}-n^2\right)^2} \mathrm{Re}\left\{ \hat{\alpha}_n \hat{\xi}_n^\ast\right\}.
$$
In the particular case of an even periodic driving function $\xi(t)$ and an odd periodic driving function $\alpha(t)$, i.e., when $\xi(-t)=\xi(t)$ and $\alpha(-t)-\alpha(0)=-\left(\alpha(t)-\alpha(0)\right)$ \[as is the case in the example studied in Section \[sec:example\]\], $x_c(-t)=x_c(t)$ and $\dot{a}_c(-t)=\dot{a}_c(t)$ and the Lamb shift Hamiltonian Eq. (\[Lambshift\]) simplifies. It can be represented as a shifted harmonic oscillator in the presence of the Rashba interaction and an inhomogeneous magnetic field,
\[HLS\] $$\begin{aligned}
H_{LS}&=&\frac{p^2}{2m_{LS}} + \frac{m_{LS}\omega_{LS}^2}{2}(x-x_{LS})^2
+\big(\alpha_{LS}p +b_{LS}x+B_{LS}\big)\sigma_y,\\
B_{LS}&=&-b_{LS} x_{LS}-
4\zeta m_{LS}\sum_{\substack{n\in \mathbb{Z}\\n\ne\pm n_d}}
{n_d n^3 \over ({n_d^2-n^2})^2}S(n\omega_d)
\;i\hat{{\alpha}}_n\hat{\xi}_n.\end{aligned}$$
Here $\zeta=\left(S(\omega_0)+S(-\omega_0)\right)/\omega_0$, $m_{LS}=\zeta^{-1}m^*$, $\omega_{LS}=\zeta \omega_0$, $x_{LS}=x_c(0)-2\big(\hat{x}_{c,n_d}-\hat{\xi}_{n_d}\big)$, $\alpha_{LS}=\zeta a_c(0)$, $b_{LS}=\zeta^2m_{LS}\big(2\hat{\dot{a}}_{c,n_d}-\dot{a}_c(0)\big)$ and, at $T=0$, $$\label{Spm}
\zeta=-\frac{g^2}{2\pi}\left(\frac{\omega_c (\omega_c+2\omega_0)}{2\omega_0^2}+\log \frac{\left|\omega_c-\omega_0\right|}{\omega_0}\right),$$ which is zero at $\omega_c=1.16\,\omega_0$ and at $\omega_c=2\omega_0$, as used throughout the paper, $\zeta=-{1 \over \pi}g^2$. Note that the resonant frequency components of driving, $\hat{\xi}_{\pm n_d}$, should vanish if the steady state regime of driving and response is to be studied. Note also that $\hat{\xi}_{\pm n_d}=0$ does not imply $\hat{x}_{c,\pm n_d}=0$, respectively. Applying the transformation Eqs. (\[B1\]) results in $\bar{B}=0$ and $$\bar{H}_{LS}=H_{LS}+\frac{1}{2}g^2\omega_0x_{LS}\left(p+m_{LS}\alpha_{LS}\sigma_y\right).$$
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The year 2019 marks the 30$^{\rm th}$ anniversary of BES and the 100$^{\rm th}$ anniversary of Rutherford’s discovery of the proton. In spite of the fact that when BES operations started the proton was already 70 years old and the strange hyperons were all over 25, BES continues to make important and unique measurements of nucleon and hyperon properties, including some interesting discoveries.'
address: |
University of Chinese Academy of Science,\
Beijing 100039, CHINA\
$^*$E-mail: [email protected]\
author:
- Stephen Lars Olsen
title: 'Nucleon/Hyperon Physics at BES'
---
Introduction {#sec1:intro}
============
In 1919, Ernest Rutherford reported the observation of “swift” charged particles produced by $\alpha$-particles impinging on $^{14}$N atoms (in modern parlence the $^{14}$N($\alpha ,p$)$^{17}$O reaction). The swift particles were positively charged and had a range in a zinc sulphide screen that was “far beyond the range of the $\alpha$ particles.” He concluded that these were hydrogen ions that had been “constituent parts of the nitrogen nucleus,” confirming the hypothesis of the chemist Willian Prout that the nuclei of all elements contained hydrogen ions. The neutron, the isotopic spin partner of the proton, was discovered 13 years later by James Chadwick in $\alpha$-particle beryllium collisions. Since that time, the proton has been the prime subject of a huge amount of research, including its anomalously large (compared to expectations for a Dirac point-like particle) magnetic moment, $\mu_p$, first measured by Otto Stern in 1933, and its non-zero charge radius first measured by Robert Hofstadter in 1956. The Lambda hyperon ($\Lambda$) was discovered via its production in high energy cosmic rays in 1947. The $\Sigma^+,\ \Sigma^0,\ \Sigma^-$ and $\Xi^0,\ \Xi^-$ hyperons were discovered at Brookhaven and Berkely in the 1950s, and the $\Omega^-$ at Brookhaven in 1964. These particles and their properties provided essential clues for Gell-Mann’s discovery of flavor-$SU(3)$ and the Gell-Mann/Zweig quark model (see Fig. \[fig:octet\]). Most of these particles decay via weak interactions, and measurements of their parity-violating weak decay parameters were major experimental activities in the 1960s. By 1989, when the BES experimental program started, this was already ancient history. Nevertheless BES, an experiment specialized for $\tau$-charm physics, has managed to contribute substantially to our understanding of the well established nucleons and strange hyperons.
![The baryon octet (left) and decuplet (right). The horizontal axis is isospin, the vertical axis is strangeness, and the quark contents are indicated. All octet baryons have spin-parity $J^P=1/2^+$, nd the decuplet baryons have $J^P=3/^+$. Other than the stable proton and the $\Sigma^0$, which decays electromagnetically to $\gamma\Lambda$, the octet baryons decay via weak interactions. The $\Omega^-$ is the only weakly decaying decuplet baryon. []{data-label="fig:octet"}](octet-decuplet.pdf){width="4in"}
Space-like baryon form-factors
==============================
Most of what we know about the internal structure of baryons comes from a huge number of experiments that measured the elastic scattering of high energy electrons from proton and, to a lesser extent, neutron targets. For single photon exchange, the laboratory differential cross section is given by
$$\frac{d\sigma}{d\Omega} = \frac{\alpha^2}{4E^2_e \sin^4\frac{\theta}{2}}\frac{E_e^{\prime}}{E_e}
\Big(\frac{G^2_E(Q^2) +\tau G^2_M(Q^2)}{1+\tau}\cos^2\frac{\theta}{2}+2\tau G^2_M(Q^2)\sin^2\frac{\theta}{2}\Big),
\label{eqn:space-like}$$
where $E_e$ ($E_e^{\prime}$) is the incident (scattered) electron energy, $\theta$ is the scattering angle, $Q^2=-q^2=(p_e-p_e^{\prime})^2$ and $\tau=Q^2/4m^2_p$; $q^2$ is the exchanged photon’s squared invariant mass, which in scattering experiments is always [*space-like*]{} (i.e., $q^2<0 $).
The functions $G_E(Q^2)$ and $G_M(Q^2)$ are the baryon’s electric and magnetic form-factors that characterize the proton’s charge and current densities. These are real, analytic functions: at $Q^2 {\rightarrow}0$, $G_E {\rightarrow}1$ and $G_M {\rightarrow}\mu_p$. The proton’s electric and magnetic form-factors have been measured in hundreds of experiments; averaged $G_E(Q^2)$ and $|G_M(Q^2)|$ measurements from recent experiments are shown on the space-like side of Fig. \[fig:form-factors\]a. These measurements were used to infer the proton’s rms charge-radius with impressive precision: $r^{\rm rms}_{ff}=0.8775\pm 0.0005$ fm. However, this value disagrees rather violently with recent charge-radius determinations based on the Lamb shift of muonium (i.e., $\mu^-$p atoms): $r^{\rm rms}_{\mu^- p}=0.842\pm 0.001$ fm (Fig. \[fig:form-factors\]b). Thus, in spite of 60 years of measurements, we still do not understand the way charge and currents are configured in the proton.
In its most naïve form, the quark model has the proton comprised of three spin-1/2 quarks in $S$-waves, where two of them have $S=0$ with the third one providing the proton’s $J=\hbar/2$ net angular momentum (Fig. \[fig:form-factors\]c). This can be tested with asymmetry measurements for polarized electron beams scattering from polarized proton targets. The green points on Fig. \[fig:form-factors\]a’s space-like side show the resulting polarization form-factor $|G_{E(pol)}|$ values determined from such measurements. In somewhat of a surprize, these polarization measurements show that the net quark-spin contribution to the proton’s total spin is about $0.15\hbar$, or only a third of the proton’s total angular momentum; the rest must come from gluons and/or orbital angular momentum. This was totally unanticipated and is further evidence of the paucity of our understanding of the proton.
![[**a)**]{} Proton form-factor measurements (from ref. [@pacetti]). [**b)**]{} Comparison of $r^{\rm rms}_{ff}$ (blue band) and $r^{\rm rms}_{\mu^- p}$ (red stripe) measurements. [**c)**]{} The proton’s spin in the naïve quark model. []{data-label="fig:form-factors"}](form-factors.pdf){width="4.0in"}
Time-like baryon form-factors
=============================
When faced with a challenging puzzle, what can an experimenter do? ...the BESIII answer is do more experiments. While electron-positron colliders cannot access space-like baryon form-factors, they are extremely well suited to measure the analytic continuations of $G_E$ and $G_M$ into the time-like region ($Q^2=q^2>0$) via ${e^+e^-}{\rightarrow}B\bar{B}$ interactions. This is especially true for BESIII, since the BEPCII CM energy range includes the thresholds for the pair production of all of the stable baryons. The time-like equivalent of eqn. \[eqn:space-like\] is $$\frac{d\sigma}{d\Omega} = \frac{\alpha^2\beta {\mathcal C}}{4E^2_{CM}}
\Big((1+\cos^2\Theta)|G_M(Q^2)|^2 +\frac{\sin^2\Theta}{\tau}|G^2_E(Q^2)|^2\Big),~~~\beta=\sqrt{1-\frac{1}{\tau}},
\label{eqn:time-like}$$ where $\beta$ and $\Theta$ are the baryon’s CM velocity and production angle, and ${\mathcal C}$ accounts for the effects of Coulomb interactions between the baryons: for point-like charged particles, ${\mathcal C}=\pi\alpha/[\beta(1-\exp(-\pi\alpha/\beta)]$; for point-like neutral particles ${\mathcal C}=1$.
It is evident from eqn. \[eqn:time-like\] that separate measurements of $|G_E|$ and $|G_M|$ can be extracted from the $\cos\Theta$ dependence of the differential cross section. However, in contrast to space-like form-factor measurements, existing data on time-like form-factors are very limited and, prior to recent BaBar and BESIII measurements, separate determinations of $G_E$ and $G_M$ of any significance had not been possible. Instead measurements have focused on the magnitude of the “effective” form-factor $|G_{\rm eff}|^2 = (|G_E|^2+2\tau|G_M|^2)/(2\tau +1)$, and this is what is shown on the time-like side of Fig. \[fig:form-factors\]a. Analyticity requires that at threshold $R\equiv|G_E(4m^2_B)/G_E(4m^2_B)|=1$, and separate $|G_E|$ and $|G_M|$ measurements near threshold could be used to test this relation. Unlike space-like form-factors, which both have to be real, time-like form-factors can be complex. A non-zero relative complex phase between $G_E$ and $G_M$ ($\Delta\Phi$) would result in polarization of the final-state baryons. Measurements of this polarization for ${e^+e^-}{\rightarrow}p\bar{p}$ are impractical with BESIII. However, the large parity-violating weak decay asymmetry parameters for the $\Lambda$, $\Sigma^+$ and $\Xi$ hyperons make BESIII uniquely well suited for $\Delta \Phi$ measurements for these baryons.
Near-threshold ${e^+e^-}{\rightarrow}B\bar{B}$ cross section measurements
-------------------------------------------------------------------------
For point-like charged particles at and just above threshold ($E_{CM} = 2m_B$), ${\mathcal C}\approx \pi\alpha/\beta$; the $\beta$ in the denominator of ${\mathcal C}$ cancels the $\beta$ in the numerator of eqn. \[eqn:time-like\] and the cross section is $\pi^2\alpha^3/2m^2_B $, and non-zero right at threshold. Thus, for example, the cross section for ${e^+e^-}{\rightarrow}\tau^+\tau^-$ jumps from zero to $\delta\sigma(m_\tau)=\pi^2\alpha^3/2m^2_\tau = 237$ pb right at threshold. (This cross-section relation, with its 237 pb threshold jump, is used for all of the BES $\tau$-mass measurements.) This abrupt jump reflects the influence of the infinite number of Bohr-like Coulomb bound $\tau^+\tau^-$ states that are just below the $2m_{\tau}$ threshold. For neutral point-like particles, ${\mathcal C}=1$ and there is no cancelation, and the cross section starts at zero and increases as $\sigma\propto\beta$.
In the ${e^+e^-}{\rightarrow}p\bar{p}$ reaction at threshold, the $p\bar{p}$ pair is produced via a $Q^2\approx 3.5$ GeV$^2$ virtual photon, with a Compton wavelength that is $\approx 0.1$ fm – one-tenth the proton’s rms size. On this distance scale, the proton is definitely not point-like and the expression in eqn. \[eqn:time-like\] should not apply. Thus it was a big surprize that near-threshold $\sigma({e^+e^-}{\rightarrow}p\bar{p})$ measurements from BaBar [@BaBar-ppbar], subsequently confirmed by CMD3 [@CMD3-ppbar] and BESIII [@BESIII-ppbar; @BESIII-ppbar-isr], show a dramatic jump at threshold that is consistent with the $\delta\sigma(m_p)=\pi^2\alpha^3/2m^2_p = 850$ pb value that would be expected for a point-like proton (see Fig. \[fig:sigma-ppbar-LcLcbar\]a). The CMD3 measurements are particularly notable in that they span the $Q^2=4m^2_p$ threshold in small $E_{\rm CM}$ increments and find that the “jump” occurs across less than a 1 MeV change in CM energy. BESIII recently reported a similar behavior for $\sigma({e^+e^-}{\rightarrow}\Lambda^+_c \bar{\Lambda}^-_c)$ near the $E_{\rm CM}=2m_{\Lambda_c}$ threshold [@BESIII-LcLcbar], shown in Fig. \[fig:sigma-ppbar-LcLcbar\]b, where a measured $230\pm 50$ pb threshold cross section jump is even larger than expectations for a charged point-like particle, i.e., $\delta\sigma(m_{\Lambda_c})=140$ pb.
![Near-threshold behavior of [**a)**]{} $\sigma({e^+e^-}{\rightarrow}p\bar{p})$ and [**b)**]{} $\sigma({e^+e^-}{\rightarrow}\Lambda^+_c \bar{\Lambda}_c^-)$. []{data-label="fig:sigma-ppbar-LcLcbar"}](sigma_ppbar_LcLcbar-2box.pdf){width="3.8in"}
For neutral point-like particles, the threshold cross-section is expected to be zero and the increase as $\sigma\propto\beta_\Lambda$. However, while measured $p\bar{p}$ and $\Lambda_c^+\bar{\Lambda}_c^-$ cross-sections seem to reflect point-like charged-particle behavior, BESIII results on $\sigma( {e^+e^-}{\rightarrow}\Lambda\bar{\Lambda})$, see Fig. \[fig:sigma-LLbar-SpSmbar\]a, do not follow expectations for a neutral point-like particle. These measurements, which start at $E_{\rm CM}=2m_\Lambda+1.0$ MeV ($\beta\approx 0.03$), find a non-zero threshold value of $305\pm 50$ pb that decreases with increasing CM energy [@BESIII-LLbar]. This jump is about half of the $\delta\sigma(m_\Lambda)=600$ pb expectation for a point-like charged particle with the $\Lambda$ mass.
![Near-threshold behavior of [**a)**]{} $\sigma({e^+e^-}{\rightarrow}\Lambda \bar{\Lambda})$ and [**b)**]{} BESIII and BaBar measurements of $\sigma({e^+e^-}{\rightarrow}K^+K^-K^+K^-)$. []{data-label="fig:sigma-LLbar-SpSmbar"}](sigma_LLbar_4K-2box.pdf){width="3.8in"}
### Influence of sub-threshold baryonium states?
Theoretical predictions for $\sigma({e^+e^-}{\rightarrow}p\bar{p})$ [@Meissner-ppbar] and $\sigma({e^+e^-}{\rightarrow}\Lambda\bar{\Lambda})$ [@Meissner-LLbar] based on chiral effective theory fail to reproduce the very sharp threshold jumps, especially those with less than a 1 MeV turn-on as seen for $p\bar{p}$ by CMD3 and $\Lambda\bar{\Lambda}$ by BESIII. For the $\Lambda\bar{\Lambda}$ case, the authors of ref. [@Meissner-LLbar] suggested that this anomalous threshold behavior could be the influence of a narrow $^3S_1$ $\Lambda\bar{\Lambda}$ bound state with mass very near the $2m_\Lambda$ threshold. Intriguingly, measurements of $\sigma({e^+e^-}{\rightarrow}K^+K^-K^+K^-)$ by both BESIII and BaBar, shown in Fig. \[fig:sigma-LLbar-SpSmbar\]c, have, respectively, $\sim 3\sigma$ and $\sim 2\sigma$ enhancements in the CM energy bin that includes $E_{\rm CM}= 2m_{\Lambda}$.
Perhaps the observed threshold behavior in other channels are also due to near-threshold $^3S_1$ $B\bar{B}$ bound states. In 2003, BESIII reported the discovery of a strong candidate for a sub-threshold $^1S_0$ $p\bar{p}$ bound state in ${J/\psi}{\rightarrow}\gamma p\bar{p}$ decays [@BESII-X1860], as discussed in detail at this symposium by Prof. Shan Jin. Thus, it does not seem unreasonable to think that a corresponding $^3S_1$ state might exist. These results suggest that extended BESIII operations in the vicinities of all of the stable baryon-antibaryon thresholds might yield very interesting results.
Time-like form-factor measurements at higher $Q^2$ values
---------------------------------------------------------
In addition to the near-threshold measurements described in the previous subsection, BESIII has performed above-threshold measurements of ${e^+e^-}{\rightarrow}p\bar{p}$ and ${e^+e^-}{\rightarrow}\Lambda\bar{\Lambda}$. The former have yielded precise separate measurements of $|G_E|$ and $|G_M|$ for the proton; the latter have resulted in the first-ever measurement of a non-zero $\Delta\Phi$ value for any baryon.
### ${e^+e^-}{\rightarrow}p\bar{p}$
BESIII measured the proton’s time-like proton form-factor using ${e^+e^-}{\rightarrow}p\bar{p}$ events collected at 12 CM energy points ranging from 2.232 GeV to 3.671 GeV [@BESIII-ppbar] and ${e^+e^-}{\rightarrow}\gamma p\bar{p}$ initial-state-radiation (isr) events in large data sets with $E_{\rm CM}$ between 3.773 and 4.600 GeV [@BESIII-ppbar-isr]. In the isr events, the $\gamma$-rays are preferentially emitted at small angles and do not register in the detector; their presence and four-momenta are inferred from energy-momentum conservation. Figure \[fig:ppbarff-Lambda-Pol\]a (top) shows BESIII $|G_{\rm eff}|$ measurements along with those from previous experiments. The black curve shows the result of a fit to a dipole shape ($|G_{\rm eff}|\propto 1/[(1+Q^2/m^2_a)(1-Q^2/Q^2_0)]$) that is commonly used to characterize form-factors. Oscillatory deviations from the dipole shape, first noted by BaBar [@BaBar-ppbar], are pronounced, as indicated by the fit residuals shown in the lower panel of Fig. \[fig:ppbarff-Lambda-Pol\]a. These deviations may be the indication of broad $J^{PC}=1^{- -}$ resonances in the $p\bar{p}$ channel or some unanticipated dynamics in baryon-antibaryon systems [@ppbar-oscill]. In either case, they require further study both in the nucleon-antinucleon and hyperon-antihyperon channels.
![ [**a)**]{} The top panel shows measurements of $|G_{\rm eff}(Q^2)|$ of the proton. The black curve is the result of a fit with a simple dipole form. The bottom panel shows the residuals from the dipole fit. [**b)**]{} The red points show BESIII $|G_E(Q^2)|$ (top) and $|G_M(Q^2)|$ (bottom) measurements. [**c)**]{} BESIII $d\sigma /d\cos\Theta$ measurements for ${e^+e^-}{\rightarrow}\Lambda\bar{\Lambda}$ at $E_{\rm CM}=2.396$ GeV (top) and the $\cos\Theta$ dependence of the $\Lambda$ polarization (bottom). The red curves are fits to the data. []{data-label="fig:ppbarff-Lambda-Pol"}](ppbarff-Lambda-Pol-3box.pdf){width="5.0in"}
The isr measurements have the advantage of providing access to a broad range of $Q^2$ at once, and can be done with the large samples of BESIII data that are taken at $E_{\rm CM}=3.77$ GeV for $D$-meson decay measurements and $E_{\rm CM}>4.0$ GeV for $XYZ$-meson studies. However, the required emission of a hard isr $\gamma$-ray incurs a luminosity penalty of order $\alpha_{\rm QCD}/\pi$. Thus, for BESIII at least, detailed form-factor measurements, including precise separate determinations of $|G_E|$ & $|G_M|$, and measurements of $\Delta \Phi$, their relative phase, are best done with direct ${e^+e^-}{\rightarrow}B\bar{B}$ production events and this reqires machine-time-consuming dedicated runs.
The BESIII direct ${e^+e^-}{\rightarrow}p\bar{p}$ measurements [@BESIII-ppbar] were used to make the precison extractions of $|G_E$ and $|G_M|$ shown, respectively, in the top and bottom panels of Fig. \[fig:ppbarff-Lambda-Pol\]b. Here, for the point nearest threshold (at $E_{\rm CM}=2.0$ GeV), $|G_E|\neq|G_M|$; BESIII measures $R(Q^2=4.0\ {\rm GeV}^2)=1.38\pm 0.11$, in agreement with a previous BaBar measurement ($1.48\pm 0.16$) [@BaBar-ppbar], with better precision. Analyticity requires $R=1$ at $Q^2=4m_p^2=3.52$ GeV$^2$, and this may be in trouble unless there are rather dramatic changes in the $|G_E|$ and $|G_M|$ trends for $Q^2<4.0$ GeV$^2$.
With data collected at $E_{\rm CM}=2.396$ GeV, 165 MeV above the $\Lambda\bar{\Lambda}$ threshold, BESIII used a low background sample of 555 ${e^+e^-}{\rightarrow}\Lambda\bar{\Lambda}$ events to make first “complete” time-like form-factor measurements for any baryon: $G_{\rm eff}=0.123\pm 0.004$; the $d\Gamma/d\cos\Theta$, distribution, where $\Theta$ is the $\Lambda$ production angle, is shown in Fig. \[fig:ppbarff-Lambda-Pol\]c (top), is consistent with being flat and $R=|G_E/G_M|=0.96\pm 0.14$; the bottom panel of the figure shows strong evidence for a non-zero $\cos\Theta$-dependent polarization corresponding to a $|G_E|$-$|G_M|$ phase difference of $\Delta\Phi=37^o\pm 11^o$.
Search for $CP$ violation in hyperon decays
===========================================
The CKM mechanism for $CP$ violation in the Standard Model (SM) fails to explain the matter-antimatter asymmetry of the universe by more than 10 orders-of-magnitude. This suggests that additional, heretofore undiscovered, $CP$ violating processes occur, and motivates aggressive searches for new, non-SM sources of $CPV$. The LHCb experiment recently reported the first observation of $CPV$ in charmed $D$-meson decays at the $10^{-3}$ level [@Aaij], which lies in the upper range of SM expectations. To date, $CP$ violation in hyperon decays have never been observed. Standard Model $CP$ violations in hyperon decays are expected to be $\sim 10^{-5}$; any value higher than this level would be a signature of new, beyond the SM physics [@pakvasa].
BESIII limit on hyperon $CPV$ using ${J/\psi}{\rightarrow}\Lambda\bar{\Lambda}$ events
--------------------------------------------------------------------------------------
The branching fraction for ${J/\psi}{\rightarrow}\Lambda\bar{\Lambda}$ is $1.94\pm 0.3 \times 10^{-3}$, which translates into about $3.2\times 10^{6}$ fully reconstructed events with $\Lambda{\rightarrow}p\pi^-$ and $\bar{\Lambda}{\rightarrow}\bar{p}\pi^+$ in BESIII’s existing 10B ${J/\psi}$ event data sample. These events are jewels: they are essentially background-free and the $\Lambda$ and $\bar{\Lambda}$ are quantum-entangled with completely correlated spin orientations. In the limit of strict $CP$ symmetry, the decay asymmetry parameters for $\Lambda{\rightarrow}p\pi^-$ $(\alpha_-)$ and $\bar{\Lambda}{\rightarrow}\bar{p}\pi^+$ $(\alpha_+ )$ are equal in magnitude but opposite in sign, i.e. $\alpha_- = -\alpha_+$; deviations from this equality would be a clear sign of $CP$ violation in hyperon decays.
In ${J/\psi}{\rightarrow}\Lambda\bar{\Lambda}$ decay, the $\Lambda\bar{\Lambda}$ pair is produced with either parallel or antiparallel helicities. Prior to the BESIII discovery of non-zero $\Lambda$ polarization in the form-factor measurement discussed in the previous section [@Lambda-ff], it was thought that for ${J/\psi}$ decay, the amplitudes for the two helicity configurations were both real and the $\Lambda$ and $\bar{\Lambda}$ unpolarized. In this case, only the product $\alpha_- \alpha_+$ would be measureable, precluding the possibilty of a test of $CP$ symmetry. However the ref. [@Lambda-ff] results motivated BESIII to reexamine the case for $\Lambda$ polarization in ${J/\psi}$ decays [@faldt].
With 1.3B ${J/\psi}$ events collected in 2009 and 2012, a total of 420K fully reconstructed ${J/\psi}\to {\Lambda\bar{\Lambda}}$ events with $\Lambda\to p\pi^-$ and $\bar{\Lambda}\to \bar{p}\pi^+$ were found. Fits to the data require a large relative phase between the two helicity amplitudes: $\Delta\Phi=42.4^o\pm 0.8^o$ [@Lambda-CPV], and a $\cos\Theta$-dependent $\Lambda$ ($\bar{\Lambda}$) transverse polarization, shown in Fig. \[fig:Xi\_2\_Lambda\]a, that varies between $-$25% and $+$25%. Although the $\Lambda$ polarization averaged over $\cos\Theta$ is zero, the BESIII analysis uses the event-by-event polarization, which has the average value $\langle |\vec{\mathcal P}_\Lambda| \rangle \approx 13\%$. This polarization enables independent determinations of $\alpha_-$ and $\alpha_+$ and a test of the $CP$ relation.
![ [**a)**]{} The $\cos\Theta_\Lambda$-dependent $\Lambda$ polarization in ${e^+e^-}{\rightarrow}{J/\psi}{\rightarrow}\Lambda\bar{\Lambda}$, where $\Theta_\Lambda$ is the $\Lambda$ production angle. The red histogram is a fit to the data. [**b)**]{} The $\Lambda$ angular distribution and polarization direction relative to the $\Xi$ polarization direction for $\Xi{\rightarrow}\Lambda\pi$ decay. Here $\alpha, \beta, \gamma$ refer to the $\Xi$ weak decay parameters and $\theta$ is the angle between the $\Lambda$ momentum and $\Xi$ polarization. []{data-label="fig:Xi_2_Lambda"}](Xi_2_Lambda-2box.pdf){width="4.7in"}
For $\Lambda\to p\pi^-$, BESIII measured $\alpha_-=0.750\pm 0.010$, which is more than five standard deviations higher than the previous world average value of $\alpha_-=0.642\pm 0.013$ [@pdg2016] that was based entirely on pre-1974 measurements. The measured $\bar{\Lambda}\to \bar{p}\pi^+$ asymmetry parameter, $\alpha_+=-0.758\pm 0.012$, is also high, and consistent, within uncertainties, with the $\alpha_-$ result. The $CP$ asymmetry, ${A_{\rm CP}}\equiv {(\alpha_-+\alpha_+)}/{(\alpha_--\alpha_+) } =-0.006\pm 0.014$, is compatible with zero and a factor of two more sensitive than the best previous $\Lambda$-based measurement, ${A_{\rm CP}}=+0.013 \pm 0.022$ [@barnes]. These results were based on only $\sim 1/8^{\rm th}$ of BESIII’s current 10B ${J/\psi}$ event sample. Extending the ref. [@Lambda-CPV] analysis to the full data sample should improve BESIII’s $CPV$ sensitivity by at least a factor of three.
Prospects for hyperon $CPV$ searches using ${J/\psi}{\rightarrow}\Xi\bar{\Xi}$ decays
-------------------------------------------------------------------------------------
BESIII’s $CPV$ sensitivity will be further improved, potentially rather substantially, by using polarized $\Lambda$ daughters from $\Xi{\rightarrow}\Lambda\pi$ decays, where the $\Xi$s are produced via ${J/\psi}{\rightarrow}\Xi\bar{\Xi}$ ($\Xi=\Xi^-\ {\rm or}\ \Xi^0$). Figure \[fig:Xi\_2\_Lambda\]b shows the $\Lambda$ polarization $\vec{\mathcal P}$ in $\Xi{\rightarrow}\Lambda\pi$ decay. Of particular note is its component along $\hat{z}^\prime$, the $\Lambda$ flight direction: ${\mathcal P}_{z^\prime}=\alpha + {\mathcal P}_\Xi\cos\theta$. Here $ {\mathcal P}_\Xi$ is the $\Xi$ polarization, $\alpha$ is the $\Xi{\rightarrow}\Lambda\pi$ weak decay parameter and $\theta$ is the $\Lambda$ decay angle. When averaged over $\Theta_\Xi$, $\langle{\mathcal P}_{z^\prime}\rangle = \alpha$. For both $\Xi^-$ and $\Xi^0$, the $\alpha$ values are substantial: for $\Xi^-{\rightarrow}\Lambda\pi^-$ $\alpha_- =-0.458\pm 0.012$; for $\Xi^0{\rightarrow}\Lambda\pi^0$ $\alpha_0 =-0.406\pm 0.013$. In both cases, the average $\Lambda$ polarization is more than three times that for $\Lambda$s from ${J/\psi}{\rightarrow}{\Lambda\bar{\Lambda}}$, as mentioned in the previous subsection.
The $A_{CP}$ sensitivity for $\Lambda$ decay is proportional to $\langle{\mathcal P}\rangle\times \sqrt{N_{\rm evts}}$. The three times higher $\langle{\mathcal P}\rangle$ value means that a $\Lambda$ originating from ${J/\psi}{\rightarrow}\Xi\bar{\Xi}$ is [*nine times*]{} more valuable than one from ${J/\psi}{\rightarrow}{\Lambda\bar{\Lambda}}$ decays. Thus, even though the ${J/\psi}{\rightarrow}\Xi\bar{\Xi}$ branching fractions are about half that for ${J/\psi}{\rightarrow}{\Lambda\bar{\Lambda}}$, and the $\Xi\bar{\Xi}$ reconstruction efficiencies are also only about half as large, the overall $A_{CP}$ sensitivity is almost three times better. This improved sensititivity has been confirmed by detailed simulations [@kupsc] and it is expected that BESIII’s ultimate $A_{CP}$ sensitivity for the existing 10B ${J/\psi}$ event sample will reach the $10^{-3}$ level.
Three comments
==============
[**1.**]{} Most of the physics program described above was not mentioned in the 850 page “Physics at BESIII” tome that we prepared prior to BESIII operation. These topics are just a few examples of the rich variety of physics contained in the BESIII data. After ten years of operation, we still discover new and unexpected research opportunities with ${e^+e^-}$ data in the 2-5 GeV energy range.
[**2.**]{} After 100 years of research, there is still lots of physics to learn from the nucleons and (after 50 years of research) hyperons. The time-like form-factor measurements discussed here are just scratching the surface of what appears to be interesting and mostly unexplored physics.
[**3.**]{} The search for matter-antimatter asymmetries with hyperons is the last frontier for SM $CP$ physics. For this, exclusive ${e^+e^-}{\rightarrow}{J/\psi}{\rightarrow}\Xi \bar{\Xi}$ processes are uniquely well suited.
Personal remarks
================
While this symposium celebrates the 30$^{\rm th}$ anniversary of BES, it also marks the 35$^{\rm th}$ year of my close connection to BES and IHEP. I first visited IHEP in 1984 to attend the first BES physics workshop. During that visit I arranged with then Director-General Ming-Han Ye and researcher Zhi-Peng Zheng the participation of Zhi-Peng’s group in the AMY experiment at Tristan. In 1992, after eight fruitful years of collaboration on AMY, Zhi-Peng invited me to join BES, which I did. (I then recruited Fred Harris, which turned out to be my biggest contribution to BES.) Our work together over these 35 years has been a long, often difficult, but always interesting trip. In 1984, IHEP and BES were poor and struggling. Today IHEP is a world premier laboratory for particle physics, and BESIII is a highly regarded, world-class experiment. The whole story is much too complex to summarize succinctly in a report like this. Maybe some day I’ll write a book.
![ [**Left**]{} Dinner with Zhi-Peng Zheng, an unidentified DESY physicist and Yu-Can Zhu at Zhi-Peng’s apartment in 1984. [**Right**]{} A 35$^{\rm th}$ year reunion at this symposium.[]{data-label="fig:photo"}](photo.pdf){width="5.0in"}
[10]{}
S. Pacetti, R. Baldini and E. Tomasi-Gustafsson, Phys. Rep., [**550-551**]{}, 1 (2015).
J.P. Lees [*et al.*]{} (BaBar Collab.), Phys. Rev. [**D87**]{}, 092005 (2013).
E.P. Solodov [*et al.*]{} (CMD-3 Collab.) EPJ Web Conf. [**212**]{}, 07002 (2019).
M. Ablikim [*et al.*]{} (BESIII Collab.), Phys. Rev. [**D91**]{}, 112004 (2015) and M. Ablikim [*et al.*]{} (BESIII Collab.), arXiv:1905.09001.
M. Ablikim [*et al.*]{} (BESIII Collab.), Phys.Rev. [**D99**]{}, 092002 (2019).
M. Ablikim [*et al.*]{} (BESIII Collab.), Phys. Rev. Lett. [**120**]{}, 132001 (2018).
M. Ablikim [*et al.*]{} (BESIII Collab.), Phys. Rev. [**D97**]{}, 032013 (2018).
J. Haidenbauer, X.-W. Kang and U.-G. Meissner, Nucl. Phys. [**A929**]{}, 102 (2014).
J.Z. Bai [*et al.*]{} (BES Collab.), Phys.Rev.Lett. [**91**]{}, 022001 (2003).
J. Haidenbauer and U.-G. Meissner, Phys. Lett. [**B761**]{}, 456 (2016).
M. Ablikim [*et al.*]{} (BESIII Collab.), Phys. Rev. [**D100**]{}, 032009 (2019).
A. Bianconi and E. Tomasi-Gustafsson, Phys. Rev. [**C93**]{}, 035201 (2016).
M. Ablikim [*et al.*]{} (BESIII Collab.), Phys. Rev. Lett. [**123**]{}, 122003 (2019).
R. Aaij [*et al*]{} (LHCb Collab.), Phys. Rev. Lett., [**122**]{}, 211803 (2019).
J.F. Donoghue, X.-G. He, and S. Pakvasa, Phys. Rev. [**D34**]{}, 833 (1986).
G. Fäldt and A. Kupsc, Phys. Lett. [**B772**]{}, 16 (2017).
M. Ablikim [*et al.*]{} (BESIII Collab.), Nature Physics [**15**]{}, 631 (2019).
C. Patrignani [*et al.*]{} (Particle Data Group), Chin. Phys. [**C40**]{}, 100001 (2016)
P.D. Barnes [*et al.*]{}, Phys. Rev. [**C54**]{}, 1877 (1996).
P. Adlarson and A. Kupsc, arXiv:1908.03102.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Within the ground-state auxiliary-field quantum Monte Carlo technique, we introduce discrete Hubbard-Stratonovich transformations (HSTs) that are suitable also for spatially inhomogeneous trial functions. The discrete auxiliary fields introduced here are coupled to local spin or charge operators fluctuating around their Hartree-Fock values. The formalism can be considered as a generalization of the discrete HSTs by Hirsch [\[J. E. Hirsch, Phys. Rev. B [**28**]{}, 4059 (1983)\]](https://link.aps.org/doi/10.1103/PhysRevB.28.4059) or a compactification of the shifted-contour auxiliary-field Monte Carlo formalism by Rom [*et al.*]{} [\[N. Rom [*et al.*]{}, Chem. Phys. Lett. [**270**]{}, 382 (1997)\]](http://www.sciencedirect.com/science/article/pii/S0009261497003709). An improvement of the acceptance ratio is found for a real auxiliary field, while an improvement of the average sign is found for a pure-imaginary auxiliary field. Efficiencies of the different HSTs are tested in the single-band Hubbard model at and away from half filling by studying the staggered magnetization and energy expectation values, respectively.'
author:
- Kazuhiro Seki
- Sandro Sorella
bibliography:
- 'biball.bib'
title: ' Benchmark study of an auxiliary-field quantum Monte Carlo technique for the Hubbard model with shifted-discrete Hubbard-Stratonovich transformations '
---
Introduction
============
The numerical solution of the Hubbard model with strong correlations is one of the most challenging issues in the theory of strongly correlated electron systems [@LeBlanc2015; @Zheng2017]. Attempts to determine the ground state are often based on iterative techniques based on a repeated application of a short imaginary time propagator, or by using the simple power method and more advanced Krylov-subspace techniques, as, for instance the Lanczos algorithm, where a Hamiltonian operator is repeatedly applied to a properly chosen trial state. In both cases the ground-state component of the trial state is filtered out after several iterations.
Among these projection techniques, the auxiliary-field quantum Monte Carlo (AFQMC) [@Sugiyama1986; @Sorella1989; @Imada1989; @Becca_Sorella_book] is one of the most powerful schemes, as it allows us to study, for example, the ground-state properties of the Hubbard model with several thousands electrons and lattice sites, when the negative-sign problem is absent [@Sorella2012; @Otsuka2016; @Otsuka2018]. In the ground-state AFQMC, even if the Hamiltonian is the same, there exists some arbitrariness in choosing the trial wave function and the type of auxiliary fields (e.g., real, complex, continuous, or discrete). Experience has shown that an appropriate choice of these ingredients may significantly improve the efficiency of the Monte Carlo simulations [@Shi2013].
It has been demonstrated [@Qin2016; @Zheng2017] that a Slater-determinant obtained from an unrestricted Hartree-Fock (UHF) approximation [@Xu2011] provides a good trial wave function for the doped Hubbard model in the constrained-path AFQMC [@Zhang1997]. Recently, for a particular parameter set at doping $\delta=1/8$ and electron-electron repulsion $U/t=8$, the ground state of the Hubbard model on the square lattice has been predicted to exhibit a vertical stripe order [@Zheng2017], where the stripe states with periods $\lambda=5,6,7$ and $8$ in units of the lattice constant are nearly degenerate, while a spatially homogeneous $d$-wave superconducting state should have, according to their study, a higher energy. Recent variational Monte Carlo (VMC) calculations [@Zhao2017; @Ido2018; @Darmawan2018] have also shown that various vertical-stripe orders with different periods appear depending on the doping and the hopping parameter. In most of the calculations in Ref. [@Zheng2017], the symmetry of finite-size clusters is broken due to the use of UHF trial wave functions or by applying pinning magnetic fields, and the results are extrapolated to the thermodynamic limit. The success of utilizing symmetry-broken wave functions is rather surprising, because symmetry breakings do not occur in the exact ground state of finite-size systems. The similar issue is known as the symmetry dilemma in first-principles calculations for molecules [@Perdew1995; @Carrascal2015]. Recently, it has been shown that the quality of the trial wave function can be improved by restoring the symmetries that are once broken by UHF or mean-field treatments [@Tahara2008; @Rodriguez2012; @Shi2014]. However, in the present work, we do not enter into the issue on symmetry breakings of trial wave functions, and rather focus on the arbitrariness of the auxiliary field to improve the efficiency of AFQMC simulations with such symmetry-broken trial wave functions.
The way of transforming a quartic interaction term into a quadratic one via the Hubbard-Stratonovich transformation (HST) [@Hubbard1959] is not unique and affects the efficiency of simulations [@Hirsch1983; @Motome1997; @Held1998; @Sakai2004; @Han2004; @Broecker2016]. Recently the popularity of this technique is substantially increased, because it has been realized that, with continuous auxiliary fields, one can treat interaction terms beyond the on-site Hubbard interaction, up to the complete treatment of the long-range Coulomb interaction [@Buividovich2012; @Ulybyshev2013; @Hohenadler2014; @Tang2015; @Tang2018], or of the long-range electron-phonon interaction [@Batrouni2019], and even both of them on the same footing [@Karakuzu2018], without being vexed by the sign problem in a certain parameter region on bipartitle lattices. Interestingly, such a parameter region coincides with the one where rigorous statements on the ground state of an extended Hubbard-Holstein model are available [@Lieb1989; @Miyao2018]. It is also noteworthy that, even when the sign problem cannot be eliminated completely, continuous auxiliary fields with a proper shift [@Rom1997; @Rom1998; @Baer1998] can improve the efficiency of simulations compared to the one without the shift. A similar idea has been employed also in the AFQMC [@Zhang2003; @Motta2014] within the constrained-path approximation [@Zhang1997].
In this paper, we introduce shifted-discrete HSTs, where auxiliary fields are coupled to the fluctuation of local spin or charge. The method is applied for AFQMC simulations of the Hubbard model on the square lattice. It is shown that the shifted-discrete HSTs can improve the efficiency of the AFQMC simulations. Moreover we present results on the magnetic order parameter as a function of $U/t$ with high statistical accuracy, that represents important benchmark, useful also for comparison with experiments.
The rest of the paper is organized as follows. In Sec. \[sec:2\], the Hubbard model is defined and the AFQMC method is described. In Sec. \[sec:3\], the shifted-discrete HSTs are introduced. In Sec. \[sec:4\], numerical results of the AFQMC simulations for the Hubbard model are presented. Section \[sec:5\] is devoted to conclusions and discussions.
Model and method {#sec:2}
================
We consider the Hubbard model whose Hamiltonian is defined by $\hat{H} = \hat{K} + \hat{V}$, where $$\begin{aligned}
\label{Ham0}
\hat{K} &=& -t\sum_{\langle ij \rangle, {\sigma}}
\left( \hat{c}_{i{\sigma}}^\dag \hat{c}_{j{\sigma}} + {\rm H.c.} \right), \\
\hat{V} &=& U\sum_{i} \hat{n}_{i {\uparrow}} \hat{n}_{i {\downarrow}}, \label{HamU}\end{aligned}$$ $\hat{c}_{i {\sigma}}^\dag$ ($\hat{c}_{i {\sigma}}$) creates (annihilates) a fermion with site index $i$ and spin index ${\sigma}(={\uparrow},{\downarrow})$, $\hat{n}_{i{\sigma}} = \hat{c}_{i{\sigma}}^\dag \hat{c}_{i{\sigma}}$, $t$ is the hopping parameter between the nearest-neighbor sites on the square lattice and $U > 0$ is the on-site electron-electron repulsion. We consider the Hubbard model on $N=L \times L$-site clusters. Boundary conditions will be specified for each calculation in Sec. \[sec:4\]. The lattice constant is set to be unity.
In the AFQMC, the expectation value of an operator $\hat{O}$ is calculated as $$\langle \hat{O} \rangle_{\beta} =
\frac{
\langle \Psi_{\rm T} |
{\mathrm{e}}^{-\frac{\beta}{2} \hat{H}} \hat{O}
{\mathrm{e}}^{-\frac{\beta}{2} \hat{H}}| \Psi_{\rm T} \rangle }
{\langle \Psi_{\rm T} | {\mathrm{e}}^{-\beta \hat{H}} |\Psi_{\rm T} \rangle },
\label{afqmc}$$ where $\beta$ is the projection time and $ |\Psi_{\rm T} \rangle $ is a trial wave function. If $\beta$ is infinitely large, one can obtain the ground-state expectation value as long as $|\Psi_{\rm T} \rangle$ has a finite overlap with the ground state [@Horn1984]. If $\beta$ is finite, the results depend on the trial wave function (see for example Ref. [@Weinberg2017]). If $\beta=0$, Eq. (\[afqmc\]) reduces to the expectation value of $\hat{O}$ with respect to the trial wave function.
At finite dopings, $|\Psi_{\rm T} \rangle$ is obtained by solving the eigenvalue problem of the following UHF Hamiltonian self consistently: $$\hat{H}_{\rm UHF} = \hat{K}
+ U_{\rm eff} \sum_{i}\left(
\langle \hat{n}_{i{\uparrow}} \rangle_0 \hat{n}_{i{\downarrow}}
+ \hat{n}_{i{\uparrow}} \langle \hat{n}_{i{\downarrow}} \rangle_0
-
\langle \hat{n}_{i{\uparrow}} \rangle_0
\langle \hat{n}_{i{\downarrow}} \rangle_0
\right),
\label{eq:uhf}$$ where $U_{\rm eff}$ is an arbitrary parameter and the expectation value $\langle \cdots \rangle_0$ in Eq. (\[eq:uhf\]) is defined in Eq. (\[afqmc\]) with $\beta=0$. A fine tuning of $U_{\rm eff}$ can improve the quality of the trial wave function [@Qin2016]. We set $U_{\rm eff}/t=2.5$ which has turned out to provide a good trial wave function for the doped cases studied here, in the sense that the energy expectation value decreases quickly with increasing $\beta$. By adding a small bias in the initial condition for the self consistent UHF loop to pin the direction of the stripe, $|\Psi_{\rm T} \rangle$ shows a vertical stripe order with period $\lambda=8$ around $\delta=1/8$ doping on the $16 \times 16$ cluster.
At half filling, $|\Psi_{\rm T} \rangle$ is obtained as a ground state of non-interacting electrons on the square lattice under a staggered magnetic field along the spin-quantized axis ($z$ direction): $$\hat{H}_{\rm MF} = \hat{K} - \Delta_{\rm AF} \sum_{i} (-1)^{i} (\hat{n}_{i{\uparrow}} - \hat{n}_{i{\downarrow}}),
\label{eq:Hmf}$$ where $(-1)^i = 1 (-1)$ if the site $i$ belongs to the $A$ ($B$) sublattice and $\Delta_{\rm AF}$ can be chosen arbitrarily. The value of $\Delta_{\rm AF}$ will be specified with the numerical results in Sec. \[sec:4\].
By using the second-order Suzuki-Trotter decomposition [@Trotter1959; @Suzuki1976], the imaginary-time propagator can be expressed as $${\mathrm{e}}^{-\beta \hat{H}} = \prod_{n=1}^{N_\tau}
\left(
{\mathrm{e}}^{-\frac{\Delta_\tau}{2} \hat{K}}
{\mathrm{e}}^{-\Delta_\tau \hat{V}}
{\mathrm{e}}^{-\frac{\Delta_\tau}{2} \hat{K}}
\right) + O\left(\Delta_\tau^2\right),
\label{Trotter}$$ where the projection time $\beta$ is discretized into $N_\tau$ time slices and $\Delta_\tau=\beta/N_\tau$. For the doped cases, we set $\Delta_\tau t=0.05$ so that the discretization error is within statistical errors. For the half filled case, we perform extrapolations of $\Delta_\tau \to 0$ to eliminate the discretization error, which becomes non negligible for large $U/t$ as compared to statistical or extrapolation errors for the results shown in Sec. \[sec:result\_half\]. An HST is applied to ${\mathrm{e}}^{-\Delta_\tau \hat{V}}$ and the summation over the auxiliary fields is performed by the Monte Carlo method with the importance sampling, where a proposed auxiliary-field configuration is accepted or rejected according to the Metropolis algorithm. In the next Section, we introduce shifted-discrete HSTs for ${\mathrm{e}}^{-\Delta_\tau \hat{V}}$.
Shifted-discrete Hubbard-Stratonovich transformations {#sec:3}
=====================================================
In this Section we derive shifted-discrete HSTs which couple the auxiliary field to the local spin fluctuation in Sec. \[hst:spin\] and to the local charge fluctuation in Sec. \[hst:charge\]. Although the two HSTs can be formulated almost in parallel, we provide both of them separately for completeness.
Auxiliary field coupled to spin fluctuation {#hst:spin}
-------------------------------------------
The Hubbard interaction in Eq. (\[HamU\]) can be written as $$\begin{aligned}
\hat{V} =
&-&\frac{U}{2} \sum_{i}
\left[\left(\hat{n}_{i{\uparrow}} -\hat{n}_{i{\downarrow}} - \tilde{m}_{i} \right)^2
-\tilde{m}_{i}^2\right]\notag \\
&+&\frac{U}{2} \sum_{i} \left[
(1-2 \tilde{m}_{i}) \hat{n}_{i{\uparrow}} +
(1+2 \tilde{m}_{i}) \hat{n}_{i{\downarrow}} \right]
\label{eq:HamUspin}\end{aligned}$$ where $\tilde{m}_i$ is an arbitrary number. Then ${\mathrm{e}}^{-\Delta_\tau V}$ can be written as $$\begin{aligned}
{\mathrm{e}}^{-\Delta_\tau \hat{V}} &=&
{\mathrm{e}}^{ \frac{\Delta_\tau U}{2}\sum_{i} \left[(\hat{n}_{i{\uparrow}} - \hat{n}_{i{\downarrow}} - \tilde{m}_{i})^2 - \tilde{m}_{i}^2\right]} \notag \\
&\times&
{\mathrm{e}}^{-\frac{\Delta_\tau U}{2}\sum_{i} (1-2\tilde{m}_i)\hat{n}_{i{\uparrow}}}
{\mathrm{e}}^{-\frac{\Delta_\tau U}{2}\sum_{i} (1+2\tilde{m}_i)\hat{n}_{i{\downarrow}}}.
\label{propUspin}\end{aligned}$$ Let us consider the first exponential factor in the right-hand side of Eq. (\[propUspin\]). For each site $i$, we consider the following HST: $$C_i {\mathrm{e}}^{-\frac{\Delta_\tau U}{2}\tilde{m}_i^2} {\mathrm{e}}^{\frac{\Delta_\tau U}{2}(\hat{n}_{i{\uparrow}} - \hat{n}_{i{\downarrow}} - \tilde{m}_i)^2}
= \frac{1}{2} \sum_{s_i=\pm1} {\mathrm{e}}^{ \alpha_i s_i (\hat{n}_{i{\uparrow}} - \hat{n}_{i{\downarrow}} - m_{i})},
\label{assumption}$$ where $s_i=\pm 1$ is the discrete auxiliary field, and the undetermined four parameters $\alpha_i$, $m_i$, $\tilde{m}_i$ and $C_i$ are related through the following three equations (see Appendix \[appA\] for derivation): $$\begin{aligned}
&&\frac{\cosh{\alpha_i (1-m_i)} \cosh{\alpha_i(1+m_i)}}{\cosh^2{\alpha_i m_i}}= {\mathrm{e}}^{\Delta_\tau U}, \label{alpha_spin}\\
&&\tilde{m_i} = \frac{1}{2\Delta_\tau U} \ln \frac{\cosh{\alpha_i (1+m_i)}}{\cosh{\alpha_i (1-m_i)}}, \label{mtilde_spin} \\
&& C_i = {\mathrm{e}}^{\Delta_\tau U\tilde{m}^2/2}\cosh{\alpha_i m_i} \label{C_spin}.\end{aligned}$$ Therefore, if say $m_i$ is given, $\alpha_i$, $\tilde{m}_i$, and $C_i$ are determined from Eqs. (\[alpha\_spin\])-(\[C\_spin\]). Finally we obtain $${\mathrm{e}}^{-\Delta_\tau \hat{V}} \propto \prod_i \sum_{s_i=\pm 1}
{\mathrm{e}}^{
\left[ \alpha_i s_i -\frac{\Delta_\tau U}{2} (1-2\tilde{m}_i)\right] \hat{n}_{i{\uparrow}}
+ \left[-\alpha_i s_i -\frac{\Delta_\tau U}{2} (1+2\tilde{m}_i)\right] \hat{n}_{i{\downarrow}}
-\alpha s_{i} m_i},
\label{propUspin2}$$ Note that in general $m_{i} \not = \tilde{m}_i$ and $C_i$’s are irrelevant for results of simulations because they cancel out from the numerator and the denominator in Eq. (\[afqmc\]). If $m_i=0$, the HST reduces to the one introduced by Hirsch [@Hirsch1983]. However, the arbitrariness of $m_i$ can be utilized to improve the efficiency of AFQMC simulations as shown in Sec. \[sec:4\].
In the right-hand side of Eq. (\[propUspin2\]), the auxiliary field $\alpha_i s_i$ is shifted by $\Delta_\tau U \tilde{m}_i$ as compared to the case of $m_{i}=\tilde{m}_i=0$. To obtain more physical intuitions for $m_i$, we rewrite the exponent of the right-hand side of Eq. (\[propUspin2\]) as $$\alpha_i s_i \left(\hat{n}_{i\uparrow} - \hat{n}_{i\downarrow} - m_i \right)
+ \Delta_\tau U \tilde{m}_i \left(\hat{n}_{i\uparrow} - \hat{n}_{i\downarrow}\right)
- \frac{\Delta_\tau U}{2} \left(\hat{n}_{i\uparrow} + \hat{n}_{i\downarrow}\right).
\label{eq:exponent}$$ In the first term, the auxiliary field $\alpha_i s_{i}$ is coupled to the fluctuation of the local magnetization $(\hat{n}_{i\uparrow} - \hat{n}_{i\downarrow} - m_i)$, while the shift of the local magnetization by $-m_i$ in the first term is compensated by the spatially inhomogeneous magnetic field $\Delta_\tau U \tilde{m}_i$ in the second term.
We set the parameter $m_{i}$ as the local magnetization in the trial wave function $$m_{i} =
\langle \hat{n}_{i{\uparrow}} - \hat{n}_{i{\downarrow}} \rangle_0.
\label{eq:mi}$$ This $m_i$ can be easily calculated and is expected to stabilize the simulation by keeping the first term in Eq. (\[eq:exponent\]) “small” during the imaginary-time evolution. For a given $m_i$, $\alpha_i$ can be determined from Eq (\[alpha\_spin\]), $\tilde{m}_i$ from Eq. (\[mtilde\_spin\]), and $C_i$ from Eq. (\[C\_spin\]). The solution $\alpha_i$ of Eq. (\[alpha\_spin\]) can be found by the Newton method with an initial guess $\alpha_{i,{\rm initial}}=\cosh^{-1} {\mathrm{e}}^{\Delta_\tau U/2}$, for example.
Auxiliary field coupled to charge fluctuation {#hst:charge}
---------------------------------------------
In this subsection, $\alpha_i$ and $C_i$ will be re-defined. The Hubbard interaction in Eq. (\[HamU\]) can be written as $$\begin{aligned}
\hat{V}
&=&\frac{U}{2} \sum_{i}
\left[\left(\hat{n}_{i{\uparrow}} + \hat{n}_{i{\downarrow}} - \tilde{n}_{i} \right)^2
-\tilde{n}_{i}^2\right]\notag \\
&-&\frac{U}{2} \sum_{i} \left[
(1 - 2 \tilde{n}_{i}) \hat{n}_{i{\uparrow}} +
(1 - 2 \tilde{n}_{i}) \hat{n}_{i{\downarrow}} \right]
\label{eq:HamUcharge}\end{aligned}$$ where $\tilde{n}_i$ is an arbitrary number. Then ${\mathrm{e}}^{-\Delta_\tau V}$ can be written as $$\begin{aligned}
{\mathrm{e}}^{-\Delta_\tau \hat{V}} &=&
{\mathrm{e}}^{-\frac{\Delta_\tau U}{2}\sum_{i} \left[(\hat{n}_{i{\uparrow}} + \hat{n}_{i{\downarrow}} - \tilde{n}_{i})^2 - \tilde{n}_{i}^2\right]} \notag \\
&\times&
{\mathrm{e}}^{ \frac{\Delta_\tau U}{2}\sum_{i} (1-2\tilde{n}_i)\hat{n}_{i{\uparrow}}}
{\mathrm{e}}^{ \frac{\Delta_\tau U}{2}\sum_{i} (1-2\tilde{n}_i)\hat{n}_{i{\downarrow}}}.
\label{propUcharge}\end{aligned}$$ Let us consider the first exponential factor in the right-hand side of Eq. (\[propUcharge\]). For each site $i$, we consider the following HST: $$C_i {\mathrm{e}}^{\frac{\Delta_\tau U}{2}\tilde{n}_i^2}
{\mathrm{e}}^{-\frac{\Delta_\tau U}{2}(\hat{n}_{i{\uparrow}} + \hat{n}_{i{\downarrow}} - \tilde{n}_i)^2} =
\frac{1}{2} \sum_{s=\pm1} {\mathrm{e}}^{i \alpha_i s_i (\hat{n}_{i{\uparrow}} + \hat{n}_{i{\downarrow}} - n_i)},
\label{assumption_charge}$$ where $s_i=\pm 1$ is the discrete auxiliary field, and the undetermined four parameters $\alpha_i$, $n_i$, $\tilde{n}_i$ and $C_i$ are related through the following three equations (see Appendix \[appA\] for derivation): $$\begin{aligned}
&&\frac{\cos{\alpha_i(2-n_i)} \cos{\alpha_i n_i}}{\cos^2{\alpha_i(1-n_i)}} = {\mathrm{e}}^{-\Delta_\tau U}\label{alpha_charge}, \\
&&\tilde{n}_i = 1 - \frac{1}{2 \Delta_\tau U} \ln \frac{\cos{\alpha_i n_i }}{\cos{\alpha_i (2-n_i)}}\label{ntilde_charge}, \\
&&C_i = {\mathrm{e}}^{-\Delta_\tau U \tilde{n}_i^2/2}\cos{\alpha_i n_i} \label{C_charge}.\end{aligned}$$ Therefore, if say $n_i$ is given, $\alpha_i$, $\tilde{n}_i$, and $C_i$ are determined from Eqs. (\[alpha\_charge\])-(\[C\_charge\]). Finally we obtain $${\mathrm{e}}^{-\Delta_\tau \hat{V}} \propto \prod_i \sum_{s_i=\pm 1}
{\mathrm{e}}^{
\left[i \alpha_i s_i +\frac{\Delta_\tau U}{2} (1-2\tilde{n}_i)\right] \hat{n}_{i{\uparrow}}
+ \left[i \alpha_i s_i +\frac{\Delta_\tau U}{2} (1-2\tilde{n}_i)\right] \hat{n}_{i{\downarrow}}
-i \alpha s_{i} n_i},
\label{propUcharge2}$$ Note that in general $n_{i} \not = \tilde{n}_i$ and $C_i$’s are irrelevant for results of simulations because they cancel out between the numerator and the denominator in Eq. (\[afqmc\]). If $n_i=1$, the HST reduces to the one introduced by Hirsch [@Hirsch1983]. However, the arbitrariness of $n_i$ can be utilized to improve the efficiency of AFQMC simulations as shown in Sec. \[sec:4\].
In the right-hand side of Eq. (\[propUcharge2\]), the auxiliary field $i \alpha_i s_i$ is shifted by $\Delta_\tau U (1-\tilde{n}_i)$ as compared to the case of $n_{i}=\tilde{n}_i=1$. To obtain more physical intuitions for $n_i$, we rewrite the exponent of the right-hand side of Eq. (\[propUcharge2\]) as $$i \alpha_i s_i \left(\hat{n}_{i\uparrow} + \hat{n}_{i\downarrow} - n_i \right)
+ \Delta_\tau U (1-\tilde{n}_i) \left(\hat{n}_{i\uparrow} + \hat{n}_{i\downarrow}\right)
- \frac{\Delta_\tau U}{2} \left(\hat{n}_{i\uparrow} + \hat{n}_{i\downarrow}\right).
\label{eq:exponent_charge}$$ In the first term, the auxiliary field $ i \alpha_i s_{i}$ is coupled to the fluctuation of the local density $(\hat{n}_{i\uparrow} + \hat{n}_{i\downarrow} - n_i)$, while the shift of the local density by $-(1-n_i)$ in the first term is compensated by the spatially inhomogeneous chemical potential $\Delta_\tau U (1-\tilde{n}_i)$ in the second term.
We set the parameter $n_{i}$ as the local charge density in the trial wave function $$n_{i}
= \langle \hat{n}_{i{\uparrow}} + \hat{n}_{i{\downarrow}} \rangle_0.
\label{eq:ni}$$ This $n_i$ can be easily calculated and is expected to stabilize the simulation by keeping the first term in Eq. (\[eq:exponent\_charge\]) “small” during the imaginary-time evolution. For a given $n_i$, $\alpha_i$ can be determined from Eq (\[alpha\_charge\]), $\tilde{n}_i$ from Eq. (\[ntilde\_charge\]), and $C_i$ from Eq. (\[C\_charge\]). The solution $\alpha_i$ of Eq. (\[alpha\_charge\]) can be found by the Newton method with an initial guess $\alpha_{i,{\rm initial}}= \cos^{-1} {\mathrm{e}}^{-\Delta_\tau U/2}$, for example.
Numerical results {#sec:4}
=================
Finite dopings
--------------
At finite dopings, the sign problem occurs [@Hirsch1985; @Loh1990]. In the presence of the sign problem, the projection time $\beta$ cannot be taken as large as that for the half filled case because the average sign (of the statistical weight) decreases exponentially in $\beta$ [@Loh1990], otherwise the number of statistical samplings has be increased exponentially to keep the statistical error small. We set the maximum $\beta$ at which the average sign is $\sim 0.05$. It will be shown that even in the presence of the sign problem, the AFQMC can still provide a good upper bound of the ground-state energy.
Figure \[fig:energyU8\] shows the energy per site $E(\beta) = \langle \hat{H} \rangle_\beta /N$, the average sign, and the acceptance rate as a function of $\beta$ at $U/t=8$ for the $16 \times 16$ cluster with $224$ electrons, corresponding to $\delta = 1/8 = 0.125$. Note that since $$\frac{{\mathrm{d}}E(\beta)}{{\mathrm{d}}\beta} = -
\frac{1}{N}
\left( \langle \hat{H}^2 \rangle_\beta - \langle \hat{H} \rangle^2_\beta \right) \leqslant 0,
\label{eq:slope}$$ $E(\beta)$ is a decreasing function of $\beta$ and its slope ${\mathrm{d}}E(\beta)/{\mathrm{d}}\beta$ is proportional to the energy variance [@Horn1984]. The energies calculated by different HSTs coincide within the statistical errors, though their reachable $\beta$ is different, as shown in Fig. \[fig:energyU8\](a). In Fig. \[fig:energyU8\](b), the acceptance rate of the real auxiliary field with the shift (HST spin with shift) is increased the one from without the shift (HST spin without shift). The reason can be attributed to that since the first term of Eq. (\[eq:exponent\]) with a relevant $m_{i}$ is expected to be “smaller” than that with $m_{i}=0$, the factor ${\mathrm{e}}^{\pm 2 \alpha_i (\hat{n}_{i\uparrow} - \hat{n}_{i\downarrow} - m_{i})}$ is closer to unity, so that the fluctuation of the norm of the determinant ratio is stabilized. On the other hand, the shift of the real auxiliary field does not affect the average sign significantly because the shift does not affect the sign of the determinant ratio, as can be seen in Fig. \[fig:energyU8\](c). The situation is different for the pure-imaginary auxiliary fields. Without the shift (HST charge without shift), the average sign diminishes significantly, even at $\beta t=0.1$. By introducing the shift (HST charge with shift), the average sign is improved significantly. The reason can be attributed to that since the first term of Eq. (\[eq:exponent\_charge\]) with a relevant $n_{i}$ is expected to be “smaller” than that with $n_{i}=1$, the factor ${\mathrm{e}}^{\pm 2 i \alpha_i (\hat{n}_{i\uparrow} + \hat{n}_{i\downarrow} - n_{i})}$ is closer to unity, so that the fluctuation of the phase of the determinant ratio is stabilized. However, the average sign is still quite smaller than that with the real auxiliary fields. Although the acceptance rate is higher than the real auxiliary fields, the pure-imaginary fields may not be practical in the presence of the sign problem.
![ \[fig:energyU8\] (a) The energy per site, (b) the average sign, and (c) the acceptance rate as a function of the projection time $\beta$ with different HSTs. Calculations are done on the $16 \times 16$ cluster with $224$ electrons ($\delta=0.125$) at $U/t=8$. ](energy_sign_acceptance_beta.pdf){width=".95\columnwidth"}
To show the usefulness of the AFQMC with a short imaginary-time propagation, we make a comparison with the state-of-the-art variational wave functions for the Hubbard model [@Zhao2017; @Ido2018]. To this end, we move to the smaller doping with the larger $U/t$, where the more severe sign problem is expected. Figure \[fig:energy\] shows the energy per and the average sign as a function of $\beta$ at $U/t=10$ for the $16 \times 16$ cluster with $228$ electrons, corresponding to $\delta = 0.109375$. Here, only the shifted real auxiliary field is employed because it turned out to be the most efficient, as shown in Fig. \[fig:energyU8\] for $U/t=8$ and $\delta=0.125$. We use periodic- (antiperiodic-) boundary condition in the $x$ ($y$) direction to compare directly with the reference VMC results [@Zhao2017; @Ido2018]. Notice that our AFQMC energy, computed at finite projection time $\beta$ when the average sign is sufficiently large, respects the Ritz’s variational principle \[see Eq. (\[afqmc\])\], because it corresponds to the variational expectation value of $\hat{H}$ over the state ${\mathrm{e}}^{-\frac{\beta}{2} \hat{H}} | \Psi_{\rm T} \rangle/
\langle \Psi_{\rm T}| {\mathrm{e}}^{-\beta \hat{H}} | \Psi_{\rm T}\rangle^{\frac{1}{2}}$. This is a useful property of an approximate technique that is not always satisfied, as for instance for the constrained-path AFQMC [@Zhang1997; @Zhang2003]. At $\beta t=0.7$, where the average sign remains $\sim 0.99$, the AFQMC energy is already lower than the VMC energy without variance extrapolation. At $\beta t = 1.1$, the AFQMC energy almost coincides with the VMC variance-extrapolated one, while the slope ${\mathrm{d}}E(\beta)/{\mathrm{d}}\beta$ is still finite, indicating that the AFQMC energy variance is nonzero \[see Eq. (\[eq:slope\])\]. At $\beta t = 1.5$, the AFQMC energy is $E/t=-0.6552(4)$, which is lower than the variance-extrapolated VMC energy $E/t=0.6538(9)$ [@Zhao2017] which may be compatible with our number within two standard deviations. This result suggest that the ground-state AFQMC method remains very useful for providing upper bound values of the ground-state energy even in the presence of the negative-sign problem.
![ \[fig:energy\] (a) The energy per site and (b) the average sign as a function of the projection time $\beta$. Calculations refer to the $16 \times 16$ cluster with $228$ electrons ($\delta = 0.109375$) at $U/t=10$. In (a), the horizontal lines and the shaded regions are the VMC energies and their error bars taken from Refs. [@Zhao2017; @Ido2018]. FTTN stands for fat-tree tensor network and Var. ext. for variance extrapolation. ](energy_sign_vs_beta.pdf){width=".95\columnwidth"}
Half filling {#sec:result_half}
------------
At half filling, the sign problem is absent. Therefore the AFQMC can provide exact results which often serve as a reference benchmark for other numerical techniques. An excellent agreement in the ground-state energies of the two-dimensional Hubbard model between the AFQMC and other many-body techniques has been reported in Ref. [@LeBlanc2015]. Moreover, within the AFQMC, the staggered magnetization $m$, i.e., the order parameter at half filling, can be estimated accurately by using the twist-averaged boundary condition for small $U/t$, e.g., $U\lesssim 4$ [@Qin2016twist; @Karakuzu2018twist]. However, for large $U/t$, AFQMC simulations still face a difficulty of large fluctuations of the magnetization, which often lead to a relatively large error bar in $m$ [@Qin2016twist; @LeBlanc2015]. The same difficulty arises also in finite-temperature determinant QMC simulations [@Hirsch1989; @Varney2009]. In previous works, in order to overcome the difficulty, a pinning-field method has been proposed with a clear improvement for the determination of $m$ in the thermodynamic limit [@Assaad2013; @Wang2014]. In the following, we report an accurate estimate $m$ especially for large $U/t$ by making use of a symmetry-broken trial wave function.
Figure \[fig:maf\] shows the staggered magnetization along the $z$ direction $$m(\beta) = \frac{1}{2N} \sum_{i} (-1)^i
\langle \hat{n}_{i {\uparrow}} - \hat{n}_{i {\downarrow}} \rangle_\beta
\label{eq:mag}$$ as a function of the Monte Carlo sweep with different HSTs. The calculations are done for $U/t=8$, $\beta t=24$, and $\Delta_\tau t = 0.1$ on the $L=16$ cluster with periodic-boundary conditions. We use $\Delta_{\rm AF}/t=0.001$ to give a finite staggered magnetization in the trial wave function. This small value of $\Delta_{\rm AF}$ is effective to pin a sizable value of the finite size order parameter $m(0)$ because the single-particle states at $U/t=0$ have a large degeneracy ($\propto L$) at the Fermi level,and are therefore strongly renormalized upon an arbitrary small $\Delta_{\rm AF}$.
Since $m(0)$ is finite \[see Eq. (\[eq:Hmf\])\], $m(\beta)$ remains finite even for finite $L$. Note that, at half filling, the HST in the charge channel with shift is equivalent to the one without shift because $n_i=1$. A very large equilibration time of $\sim 5000$ Monte Carlo sweeps is found for $m$ with the standard real HST coupled to the on-site electron spins. In this case, our shifted HST improves the equilibration time, allowing also higher acceptance rate (not shown) as in the doped cases, but the improvement is not really important. Amazingly, $m$ is equilibrated almost immediately for the complex HST coupled to the on-site electron charges. This result implies that this pure-imaginary auxiliary field, which was first introduced by Hirsch [@Hirsch1983], is very useful to estimate $m$ at half filling for large $U/t$. Here we emphasize also that, not only the correlation time is highly reduced with this technique, but also fluctuations, thanks to this pinning strategy in the trial wave function, do not show any problem of large fluctuations, even at very large $U/t$ and values.
![ \[fig:maf\] The staggered magnetization $m$ as a function of the Monte-Carlo sweep for the half-filled Hubbard model on the square lattice at $U/t=8$ with different HSTs. Calculations are done on the $16 \times 16$ cluster with $\beta t = 24$. ](maf_hstspin_vs_hstcharge.pdf){width=".95\columnwidth"}
Figure \[fig:mvsl\] shows the finite-size scaling of $m$ for $U/t=10$. The cluster sizes used are $L=6, 8, 10, 12, 14, 16, 18, 20, 22,$ and $24$. Here, the projection time $\beta$ is chosen proportional to $L$, i.e., $\beta t = \alpha L$, with $\alpha=0.5, 1$ and $1.5$. These $\beta t$ values are an order of magnitude smaller than those used with the pinning-field method [@Assaad2013; @Wang2014], because in our approach we can reach the thermodynamic limit consistently without unnecessarily large values of $\beta$. Indeed, the extrapolated values at $1/\beta=1/L=0$ are consistent for all values of $\alpha$, which validates our approach. Our best estimate is obtained from the $\beta t = 1.5L$ set of data, yielding $m=0.3046(1)$ in the $\Delta \to 0$ limit, where the number in the parentheses indicates the extrapolation error in the last digit. Calculations are done for $\Delta_\tau t = 0.2$, $0.1$, and $0.05$ and the extrapolations to $\Delta_\tau \to 0$ are obtained by a linear fit in $(\Delta_\tau t)^2$, determined by the least-squares method. The ground-state expectation value $m$ in the thermodynamic limit is obtained by extrapolating the results to $L\to \infty$. In this case we fit the data in the range $6 \leqslant L \leqslant 24$ with quadratic polynomials in $1/L$. As it can be seen in Fig. \[fig:mvsl\], the time-discretization error is not negligible for $U/t=10$. The extrapolated value is certainly smaller than the one in the Heisenberg model [@Anderson1952; @Reger1988; @Sandvic1997; @Calandra1998], where the latest Monte Carlo estimate is $m=0.30743(1)$ [@Sandvic2010; @Jiang2011]. In Table \[tab:m\] and Fig. \[fig:mvsU\] we show the values for $m$ in the thermodynamic limit for $U/t=2,4,6,8,10$ and $12$ and compare them with the ones available in the literature [@Karakuzu2018; @Sorella2015; @Qin2016twist; @Karakuzu2018twist]. The main outcome of this work is the estimated value of $m$ for $U/t\geqslant 8$, which is usually the accepted value for cuprates. Here, our error bar at $U/t=8$ is two orders of magnitude smaller than the previous AFQMC estimate [@LeBlanc2015; @Qin2016twist]. Thanks to this high statistical accuracy, our results clearly show that $m$ increases monotonically in $U/t$. This is consistent with a strong-coupling expansion around the Heisenberg limit [@Delannoy2005]. Here, finite-size scaling analyses are performed as follows. For $U/t \geqslant 6$, the scheme of finite-size scaling analyses is the same as that for $U/t = 10$ which has been described before. For $U/t=4$ ($U/t=2$), cluster sizes up to $L=32$ ($L=50$) with twist-averaged boundary conditions [@gros_exact_diag; @gros_gtabc; @Koretsune2007; @Qin2016twist; @Karakuzu2017; @Karakuzu2018twist] are used because the finite-size effects are more important than those we have found at larger $U/t$ values. A much larger value of $\Delta_{\rm AF}/t=10$ is used for $U/t \leqslant 4$ because the twists remove the degeneracy of the single-particle states at $U/t=0$, as discussed before. All the results are obtained in the $\Delta_\tau \to 0$ limit using data at $\Delta_\tau t = 0.2$, $0.1$, and $0.05$ for $U/t \geqslant 4$ and $\Delta_\tau t = 0.25$, $0.2$, and $0.1$ for $U/t = 2$, respectively.
[width=1]{}
$U/t$ $2$ $4$ $6$ $8$ $10$ $12$ $\infty$ (Heisenberg antiferromagnet)
------------------------------------------------- ---------- ----------- ----------- ----------- ----------- ----------- ---------------------------------------
AFQMC (this work) 0.120(1) 0.2340(2) 0.2815(2) 0.2991(2) 0.3046(1) 0.3067(2) –
AFQMC TABC [@Qin2016twist] 0.119(4) 0.236(1) 0.280(5) 0.26(3) – – –
AFQMC TABC [@Karakuzu2018twist] 0.122(1) 0.2347(4) – – – – –
AFQMC PBC [@Karakuzu2018] – 0.238(3) – – – – –
AFQMC MBC [@Sorella2015] 0.120(5) – – – – – –
QMC Heisenberg model [@Sandvic2010; @Jiang2011] – – – – – – 0.30743(1)
![ \[fig:mvsl\] Finite-size scaling of the staggered magnetization $m$ of the half-filled Hubbard model at $U/t=10$ with $\beta t =0.5L, L$, and $1.5L$ for (a) $\Delta_\tau t \to 0$, (b) $\Delta_\tau t = 0.05$, (c) $\Delta_\tau t = 0.1$, and (d) $\Delta_\tau t = 0.2$. The dashed horizontal lines indicate $m$ of the Heisenberg model in the thermodynamic limit taken from Ref. [@Sandvic2010]. ](mvsl_U10_dtall.pdf){width=".95\columnwidth"}
![ \[fig:mvsU\] The staggered magnetization $m$ in the thermodynamic limit as a function of $U$. For comparison, previous AFQMC results are taken from Ref. [@Qin2016]. The dashed horizontal line indicates $m$ of the Heisenberg model in the thermodynamic limit taken from Ref. [@Sandvic2010]. ](mvsU.pdf){width=".95\columnwidth"}
Conclusions and Discussions {#sec:5}
===========================
In this work we have shown that, within the ground-state AFQMC technique, the choice of the trial function and the one for the auxiliary field are extremely important. In particular we have improved the efficiency of the method, by introducing shifted-discrete HSTs, that are useful for performing the imaginary-time evolution of symmetry-broken trial wave functions. The formalism can be considered as a generalization of the discrete HSTs in Ref. [@Hirsch1983] or a compactification of the shifted-contour auxiliary-field Monte Carlo formalism in Ref. [@Rom1997; @Rom1998] specialized to the on-site Hubbard interaction.
Properly chosen auxiliary fields can improve the efficiency of AFQMC simulations. The shifted real auxiliary fields can improve the acceptance ratio, while the shifted pure-imaginary auxiliary fields can improve the average sign. The reason is that the shift in the real auxiliary field can stabilize the fluctuations of the norm of the determinant ratio, while the shift in the pure-imaginary auxiliary field can stabilize the fluctuations of the phase of the determinant ratio. However, even after the improvement, the average sign with the pure-imaginary auxiliary field remains worse than the one obtained with real auxiliary field for the doped cases. Therefore, in the presence of the sign problem, the real auxiliary field is still recommended for achieving longer imaginary-time propagations. On the other hand, at half filling with large $U/t$, the pure-imaginary auxiliary field is dramatically more efficient than the real-auxiliary fields for evaluating the staggered magnetization $m$.
In our approach, $m_{i}$ or $n_{i}$ in Eqs. (\[eq:HamUspin\]) or (\[eq:HamUcharge\]) are arbitrary parameters, that do not have to be necessarily chosen as in Eq. (\[eq:mi\]) or in Eq. (\[eq:ni\]). For example, $m_i$ ($n_i$) can be updated iteratively by the AFQMC expectation value of $\hat{n}_{i \uparrow} - \hat{n}_{i \downarrow}$ ($\hat{n}_{i \uparrow} + \hat{n}_{i \downarrow}$) with iterative simulations. This kind of scheme has already been employed to construct self-consistently an optimized trial wave function in the AFQMC [@Qin2016]. Obviously, shifted-discrete HSTs can be used straightforwardly also in this case. Moreover, we expect that imaginary-time dependent $m_i$ or $n_i$ could further improve the efficiency of the AFQMC, especially within the constrained path formalism. A study along this line is in progress [@SorellaAAD].
Finally, we remark on the $d$-wave superconducting order which has not been considered in the present study. It is noteworthy that an early study on a $t$-$t'$-$J$ model [@Himeda2002] has shown that a stripe state with spatially oscillating $d$-wave superconductivity is favored around $1/8$ hole doping. Considering such an inhomogeneous superconductivity in a trial wave function might be of interest for a possible improvement of AFQMC simulations for doped Hubbard models with large $U/t$.
The authors would like to thank Seher Karakuzu, Federico Becca, Luca Fausto Tocchio, and Tomonori Shirakawa for helpful discussions. K.S. acknowledges Emine K[[ü]{}]{}[[ç]{}]{}[[ü]{}]{}kbenli and Stefano de Gironcoli for bringing his attention to Refs. [@Perdew1995; @Carrascal2015]. Computations have been done by using the HOKUSAI GreatWave and HOKUSAI BigWaterfall supercomputers at RIKEN under the Projects No. G18007 and No. G18025. K.S. acknowledges support from the JSPS Overseas Research Fellowships. S.S. acknowledges support by the Simons foundation.
Derivation of shifted-discrete HSTs {#appA}
===================================
In this Appendix, we derive Eqs. (\[alpha\_spin\])-(\[C\_spin\]) and Eqs. (\[alpha\_charge\])-(\[C\_charge\]). First we derive Eqs. (\[alpha\_spin\])-(\[C\_spin\]), i.e., the shifted-discrete HST in the spin channel. Since the fermion density operator $\hat{n}_{i{\sigma}}$ is idempotent, i.e., $\hat{n}_{i {\sigma}}^2 = \hat{n}_{i {\sigma}}$, its exponential function is written as $${\mathrm{e}}^{\alpha s \hat{n}_{{\sigma}}} = 1 + \left({\mathrm{e}}^{ \alpha s} - 1\right) \hat{n}_{{\sigma}},
\label{idem}$$ where, and hereafter, the site index $i$ is dropped for brevity. Then the right-hand side of Eq. (\[assumption\]) is given as $$\begin{aligned}
&&
\frac{1}{2}\sum_{s=\pm1}
\left[
1 + \left({\mathrm{e}}^{\alpha s} -1 \right) \hat{n}_{{\uparrow}}
\right]
\left[
1 + \left({\mathrm{e}}^{-\alpha s} -1 \right) \hat{n}_{{\downarrow}}
\right]
{\mathrm{e}}^{-s \alpha m}\notag \\
&=&
\cosh{\alpha m} \notag \\
&+& \left[\cosh{\alpha(1-m)} - \cosh{\alpha m} \right] \ \hat{n}_{\uparrow}\notag \\
&+& \left[\cosh{\alpha(1+m)} - \cosh{\alpha m} \right] \ \hat{n}_{\downarrow}\notag \\
&+& \left[2\cosh{\alpha m} - \cosh{\alpha(1-m)} - \cosh{\alpha(1+m)} \right] \hat{n}_{\uparrow}\hat{n}_{\downarrow}.
\label{hst_right}\end{aligned}$$ The left-hand side of Eq. (\[assumption\]) is given as $$\begin{aligned}
&&C {\mathrm{e}}^{- \Delta_\tau U \tilde{m}^2/2}
{\mathrm{e}}^{\frac{\Delta_\tau U}{2}(1-2\tilde{m}) \hat{n}_{\uparrow}}
{\mathrm{e}}^{\frac{\Delta_\tau U}{2}(1+2\tilde{m}) \hat{n}_{\downarrow}}
{\mathrm{e}}^{-\Delta_\tau U \hat{n}_{\uparrow}\hat{n}_{\downarrow}} \notag \\
&=& C {\mathrm{e}}^{- \Delta_\tau U \tilde{m}^2/2} \notag \\
&+& C {\mathrm{e}}^{- \Delta_\tau U \tilde{m}^2/2} \left[{\mathrm{e}}^{\frac{\Delta_\tau U}{2}(1-2\tilde{m})}-1 \right] \hat{n}_{\uparrow}\notag \\
&+& C {\mathrm{e}}^{- \Delta_\tau U \tilde{m}^2/2} \left[{\mathrm{e}}^{\frac{\Delta_\tau U}{2}(1+2\tilde{m})}-1 \right] \hat{n}_{\downarrow}\notag \\
&+& C {\mathrm{e}}^{- \Delta_\tau U \tilde{m}^2/2} \left[2
- {\mathrm{e}}^{\frac{\Delta_\tau U}{2}(1-2\tilde{m})}
- {\mathrm{e}}^{\frac{\Delta_\tau U}{2}(1+2\tilde{m})} \right]
\hat{n}_{\uparrow}\hat{n}_{\downarrow}.
\label{hst_left}\end{aligned}$$ By comparing Eq. (\[hst\_right\]) with Eq. (\[hst\_left\]), we obtain Eqs. (\[alpha\_spin\])-(\[C\_spin\]).
Next, we derive Eqs. (\[alpha\_charge\])-(\[C\_charge\]), i.e., the shifted-discrete HST in the charge channel. As in Eq. (\[idem\]), we have $${\mathrm{e}}^{i \alpha s \hat{n}_{{\sigma}}} = 1 + \left({\mathrm{e}}^{i \alpha s} - 1\right) \hat{n}_{{\sigma}}.
\label{idem:charge}$$ Then the right-hand side of Eq. (\[assumption\_charge\]) is given as $$\begin{aligned}
&&
\frac{1}{2}\sum_{s=\pm1}
\left[
1 + \left({\mathrm{e}}^{i \alpha s} -1 \right) \hat{n}_{{\uparrow}}
\right]
\left[
1 + \left({\mathrm{e}}^{i \alpha s} -1 \right) \hat{n}_{{\downarrow}}
\right]
{\mathrm{e}}^{-i s \alpha n}\notag \\
&=&
\cos{\alpha n} \notag \\ &+&
\left[\cos{\alpha(1-n)} - \cos{\alpha n})\right]
(\hat{n}_{\uparrow}+ \hat{n}_{\downarrow}) \notag \\
&+& \left[\cos{\alpha(2-n)} - 2\cos{\alpha(1-n)} + \cos{\alpha n} \right] \hat{n}_{\uparrow}\hat{n}_{\downarrow}.
\label{hstc_right}\end{aligned}$$ The left-hand side of Eq. (\[assumption\_charge\]) is given as $$\begin{aligned}
&&
C
{\mathrm{e}}^{\Delta_\tau U \tilde{n}^2/2}
{\mathrm{e}}^{-\frac{\Delta_\tau U}{2}(1-2\tilde{n})(\hat{n}_{\uparrow}+ \hat{n}_{\downarrow})}
{\mathrm{e}}^{-\Delta_\tau U \hat{n}_{\uparrow}\hat{n}_{\downarrow}} \notag \\
&=&
C{\mathrm{e}}^{\Delta_\tau U \tilde{n}^2/2} \notag \\ &+&
C{\mathrm{e}}^{\Delta_\tau U \tilde{n}^2/2}
\left[{\mathrm{e}}^{-\frac{\Delta_\tau U}{2}(1-2\tilde{n})}-1 \right]
\left( \hat{n}_{\uparrow}+ \hat{n}_{\downarrow}\right) \notag \\ &+&
C{\mathrm{e}}^{\Delta_\tau U \tilde{n}^2/2} \left[
{\mathrm{e}}^{-\Delta_\tau U}
{\mathrm{e}}^{-\Delta_\tau U(1-2\tilde{n})}
- 2 {\mathrm{e}}^{-\frac{\Delta_\tau U}{2}(1-2\tilde{n})} + 1\right] \hat{n}_{\uparrow}\hat{n}_{\downarrow}.
\label{hstc_left}\end{aligned}$$ By comparing Eq. (\[hstc\_right\]) with Eq. (\[hstc\_left\]), we obtain Eqs. (\[alpha\_charge\])-(\[C\_charge\]).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The physical results of quantum field theory are *independent* of the various specializations of Dirac’s gamma-matrices, that are employed in given problems. Accordingly, the physical meaning of Majorana’s equation is very dubious, considering that it is a consequence of *ad hoc* matrix representations of the gamma-operators. Therefore, it seems to us that this equation cannot give the equation of motion of the neutral WIMPs (weakly interacting massive particles), the hypothesized constitutive elements of the Dark Matter.'
address:
- 'A.L. – Dipartimento di Fisica, Università di Milano, Via Celoria, 16 - 20133 Milano (Italy)'
- 'T.M. – Liceo Classico “G. Berchet”, Via della Commenda, 26 - 20122 Milano (Italy)'
author:
- Angelo Loinger
- Tiziana Marsico
title: '[**On Majorana’s equation**]{}'
---
0.50cm
0.80cm Key words: SUSY-particles; Dark Matter.\
PACS 11.10.Qr – Relativistic wave-equations.
1.20cm **1.** – Let us consider the Dirac equation for the field operator $\psi$ in the absence of any interaction with other fields:
$$\label{eq:one}
i \, \gamma^{\mu} \frac{\partial \psi}{\partial x^{\mu}} - m\psi
=0 \quad, \quad (\hbar=c=1; \quad \mu=0,1,2,3) \quad ;$$
$$\label{eq:oneprime}
\gamma^{\mu} \, \gamma^{\nu} + \gamma^{\nu} \, \gamma^{\mu} = 2 \,
\eta^{\, \mu \nu}\tag{1$'$}\quad,$$
where $\eta^{\, \mu \nu}$ is the customary Minkowskian tensor.
The corresponding charge-conjugate operator $\psi_C$ satisfies the *same* Dirac equation (\[eq:one\]) [@1].
(We recall that with any matrix representation of the $\gamma^{\mu}$’s, the operator $\psi_C$ is a simple function of the adjoint of $\psi$).
It is clear that Dirac equation (\[eq:one\]) describes perfectly also the *neutral* particles with their *identical* antiparticles.
1.20cm **2.** – In quantum field theory the physical concepts (*e.g.*, the energy eigenvalues, the four-current $j^{\mu}$, *etc.*) and the physical conclusions are *independent* of any particular matrix representation of Dirac’s $\gamma^{\mu}$–operators. As it is known.
1.20cm **3.** – Majorana’s equation (see [@2] and [@3]) is a specialization of Dirac equation [@1] such that the elements of the $\iota^{\mu}$–matrices are all *imaginary*. Consequently, the expression $[i\,
\gamma^{\mu}_{im} \, \partial_{\mu}-m]$is a *real* expression, and we can put, with Majorana, $\psi=\varphi + i\chi$, with a “real” $\varphi$ and a ‘real’’ $\chi$ – two *selfadjoint* operators –, that satisfy the *same* Dirac equation. And it seems that Majorana’s equation
$$\label{eq:two}
i \, \gamma^{\mu}_{im} \frac{\partial \varphi}{\partial x^{\mu}} -
m\varphi =0$$
describes particles which coincide with their antiparticles, *i.e.* neutral objects.
However, this conclusion depends on a particular choice of the $\gamma^{\mu}$–matrices, and is consequently very problematic. Moreover eq. (\[eq:two\]) is *not* invariant under the phase (*gauge*) transformations of $\varphi$:
$$\label{eq:three}
\left\{ \begin{array}{l} \varphi(x) \rightarrow \varphi'(x)\equiv
\varphi(x) \exp[\pm iF(x)] \quad, \\ \\
\frac{\partial}{\partial x^{\mu}} \rightarrow
\frac{\partial}{\partial x^{\mu}}\mp \frac{\partial F(x)}{\partial
x^{\mu}} \quad. \tag{3}
\end{array} \right.$$
Of course, if we write
$$\label{eq:four}
i \, \gamma^{\mu}_{im} \frac{\partial \psi}{\partial x^{\mu}} -
m\psi =0 \quad, \quad (\psi=\varphi+i\chi) \quad,$$
we have a standard instance of Dirac equation (\[eq:one\]), which is invariant under the phase transformations of $\psi$.
1.20cm **4.** – According to many astrophysicists, Dark Matter is composed of WIMPs (weakly interacting massive particles) [@4]. And there is a widespread belief that these neutral particles are described by Majorana’s equation (\[eq:two\]. Now, the above considerations show that this conviction is not well founded.
The motions of the hypothetical WIMPs can be properly described by means of Dirac equation (\[eq:one\]).
2.00cm
***APPENDIX A***
0.40cm The condition $\gamma^{\mu}=\gamma^{\mu}_{im}$, $(\mu=0,1,2,3)$, is necessary and sufficient for the formal validity of Majorana’s equation (\[eq:two\]). Indeed, let us assume that the $\gamma^{\mu}$’s are different from the $\gamma^{\mu}_{im}$’s, but – for simplicity – with the same transposition on properties of the $\gamma^{\mu}_{im}$’s *i.e.*: $\gamma^{0 T}= - \gamma^{0}$; $\gamma^{kT}=\gamma^{k}$, $(k=1,2,3)$ [@3]. We know that, quite generally, $\gamma^{0}$ is Hermitian and the $\gamma^{k}$’s are anti-Hermitian.
Denoting the adjointness with an asterisk, from the equation
$$\label{eq:A1}
\left( i \, \gamma^{0} \partial\,_{0} + i \, \gamma^{k}
\partial\,_{k} - m \right) \psi = 0$$
we get easily:
$$\label{eq:A2}
\left( i \, \gamma^{0} \partial\,_{0} + i \, \gamma^{k}
\partial\,_{k} - m \right) \psi^{*T} = 0 \quad.$$
If $\psi_C$ is the charge-conjugate operator of $\psi$, we have:
$$\label{eq:A3}
\left( i \, \gamma^{0} \partial\,_{0} + i \, \gamma^{k}
\partial\,_{k} - m \right) \psi_{C} = 0 \quad.$$
Now, the assumption
$$\label{eq:A4}
\psi_{C} = \psi^{*T}$$
shows that eq. (\[eq:A3\]) coincides with eq. (\[eq:A2\]); and we see that Majorana’s condition that $\psi$ is selfadjoint:
$$\label{eq:A5}
\psi_{C} = \psi^{*T} = \psi$$
implies $\gamma^{\mu} = \gamma^{\mu}_{im}$, *i.e.* the assumption that the expression $(\ldots)$ of eq. (\[eq:A1\]) is *real*.
Remak that the restriction $\gamma^{\mu} = \gamma^{\mu}_{im}$ *without* the assumption (\[eq:A4\]) tells us that $\psi =
\varphi + i \chi$, with $\varphi$ and $\chi$ selfadjoint operators.
It is now obvious to conclude for the validity of Majorana’s equation (\[eq:two\]). It is not difficult to generalize the above reasoning for any transposition property of $\gamma^{0},
\gamma^{1}, \gamma^{2}, \gamma^{3}$.
2.00cm
***APPENDIX B***
0.40cm We have considered Majorana’s equation, with Majorana [@2] and Wilczek [@3], from the standpoint of quantum field theory (“second” quantization of Dirac theory). Since *ubi maius minus cessat*, the previous treatment implies that a consideration of Majorana’s equation as a mathematical object of a “first” quantization, or of a classical field theory (wave-picture, *Wellenbild*, in Heisenberg’s terminology) is quite superfluous. However, we think that is is useful to emphasize the following points.
1. *“First” quantization*: the wave-function $\psi$ of Schrödinger and Dirac equations must be *complex* (see Pauli [@5]) – and this fact is sufficient to discard Majorana’s equation.
2. *Classical field theory*: the probability density and the probability current-density of the “first” quantization assumes a realistic meaning of matter density and matter current-density, which involves obviously a *non-real* classical field $\psi$: a Majorana’s equation does not make sense.
2.00cm
***APPENDIX C***
0.40cm We have emphasized in sect. **1** that a Dirac neutral particle coincides with its antiparticle. For a Weyl neutral particle things go otherwise [@6]. Indeed, the charge conjugate of Weyl equation describes an antiparticle, which is characterized by the opposite sign of particle *helicity* $\vec{\sigma} \cdot \vec{p} \, / |\vec{p}|$. This unique difference between particle and antiparticle has only a mathematical origin: Weyl equation is not invariant with respect to space reflections.
1.80cm
[9]{}
N.N. Bogoliubov and D.V. Shrikov, *Quantum Fields* (The Benjamin/Cummings Publ. Comp., Inc. – Reading, Mass.) 1983, Chapt. II, sect. 9.3.
E. Majorana, *Nuovo Cimento*, **14** (1937) 171.
F. Wilczek, *NATURE PHYSICS*, **5** (2009) 614; and references therein.
See *e.g.*: XENON Coll., *PRL*, **100**, 21303 (2008); K. Garrett and G. Duda, *arXiv:1006.2483* $[$hep-ph$]$ 12 Jun 2010 – and references therein; F. Arneodo, *arXiv:1301.0441* $[$astro-ph IM$]$ 3 Jan 2013.
W. Pauli, *Handb. der Physik*, Band V – Teil 1 (Springer-Verlag, Berlin, *etc.* 1958, p.p. 15-16 and 140.
See [@5]. p.150, and [@1].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'If the work per cycle of a quantum heat engine is averaged over an appropriate prior distribution for an external parameter $a$, the work becomes optimal at Curzon-Ahlborn efficiency. More general priors of the form $\Pi(a) \propto 1/a^{\gamma}$ yield optimal work at an efficiency which stays close to CA value, in particular near equilibrium the efficiency scales as one-half of the Carnot value. This feature is analogous to the one recently observed in literature for certain models of finite-time thermodynamics. Further, the use of Bayes’ theorem implies that the work estimated with posterior probabilities also bears close analogy with the classical formula. These findings suggest that the notion of prior information can be used to reveal thermodynamic features in quantum systems, thus pointing to a new connection between thermodynamic behavior and the concept of information.'
author:
- 'Ramandeep S. Johal[^1]'
title: Universal efficiency at optimal work with Bayesian statistics
---
The connection between thermodynamics and the concept of information is one of the most subtle analogies in our physical theories. It has played a central role in the exorcism of Maxwell’s demon [@Leffbook]. It is also crucial to how we may understand and exploit quantum information [@Vedral2008; @Ueda2009; @Jacobs2008]. To make nanodevices [@Joachim2000; @Serreli2007] that are functional and useful, we need to understand their performance with regard to heat dissipation and optimal information processing. To model such systems, standard thermodynamic processes and heat cycles have been generalised using quantum systems as the working media [@Hastapoulus76]-[@AJM2008]. It is well accepted that the maximal efficiency, $\eta_c = 1-T_2/T_1$, where $T_1 (T_2)$ is the hot (cold) bath temperature, is only obtained by a reversible heat engine which however involves infinitely slow processes. For heat cycles running in a finite time, the concept of power output becomes meaningful. Curzon and Ahlborn [@Curzon1975], first of all displayed an elegant formula in the so called endoreversible approximation for efficiency at maximum power, $\eta_{\rm CA} = 1-\sqrt{1-\eta_c}$. The appearance of an optimal efficiency in different models with a value close to Curzon-Ahlborn (CA) value, has raised the issue of its universality that has captured the imagination of workers in this area since many years [@Leff]. More recently [@AJM2008],[@Broeck2005]-[@Segal2010], a universal form for optimal efficiency, has been discussed within finite-time thermodynamics in the near-equilibrium regime (small value of $\eta_c$) as given by $\eta \approx {\eta_c}/{2} + O( {{\eta_c}^2}) $. On the other hand, in recent years, Bayesian methods of statistical inference have gained popularity in physics [@Dose]. In Bayesian probability theory, the central role is played by the concept of prior information. It represents state of our knowledge about a system before any experimental data is acquired. The assignment of a unique distribution to a given prior information is a non-trivial issue but may be argued on the basis of maximum entropy principle and certain requirements of invariance [@Jaynes1968; @Jaynes2003]. Using Bayes’ theorem [@Jeffreys1939] one can then update the prior probabilities based on the new information gathered from the data. Recently, Bayesian methods have been applied to modular structure of networks [@Hofman2008], inference of density of states [@Habeck2007], the interpretation of quantum probabilities [@Caves2002] and other inverse problems [@Lemm2000].
In this letter, a Bayesian approach is used to show that the efficiency at optimal work for a quantum heat engine is related to CA value, after the work per cycle is averaged over the prior distribution of an external parameter. In contrast to the finite-time models [@Esposito2009], the heat cycle considered here is performed infinitely slowly. The application of Bayes’ theorem gives the optimal efficiency exactly at CA value for a whole class of priors and for arbitrary bath temperatures. The present analysis thus provides a novel argument for the emergence of thermodynamic behavior in quantum heat engines from a Bayesian perspective.
As a model of a heat engine, consider a quantum system with Hamiltonian $H_1 =\sum_n \varepsilon_{n}^{(1)}
\vert n\rangle \langle n\vert$, where energy eigenvalues $\varepsilon_{n}^{(1)}= \varepsilon_n a_1$. The factor $\varepsilon_n$ depends on the energy level $n$ as well as other fixed parameters/constants of the system; $a_1$ is a controllable external parameter equivalent to say, the applied magnetic field for a spin-1/2 system. Other examples of this class are 1-d quantum harmonic oscillator ($a_1$ equivalent to frequency) and a particle in 1-d box ($a_1$ inversely proportional to the square of box-width). Initially, the quantum system is in thermal state $\rho(a_1) =\sum_n p_{n}^{(1)}\vert n\rangle \langle n\vert$ at temperature $T_1$ with its eigenvalues given by the canonical probabilities $p_{n}^{(1)}$. The quantum analogue of a classical Otto cycle between two heat baths at temperatures $T_1$ and $T_2$ involves the following steps [@Kieu2004]: (i) the system is detached from the hot bath and made to undergo the first quantum adiabatic process, during which the system hamiltonian changes to $H_2 =\sum_n \varepsilon_{n}^{(2)}
\vert n\rangle \langle n\vert$, where $\varepsilon_{n}^{(2)}= \varepsilon_n a_2$, without any transitions between the levels and so the system continues to occupy its initial state. For $a_2 < a_1$, this process is the analogue of an adiabatic expansion. The work done [*by*]{} the system in this stage is defined as the change in mean energy ${\cal W}_1= {\rm Tr}(\rho(a_1)[H_2-H_1])$; (ii) the system with modified energy spectrum $\varepsilon_{n}^{(2)}$ is brought in thermal contact with the cold bath and it achieves a thermal state $\rho(a_2) =\sum_n p_{n}^{(2)}\vert n\rangle \langle n\vert$. The modified canonical probabilities $p_{n}^{(2)}$ now correspond to temperature $T_2$. On average, heat rejected to the bath in this stage is defined as $Q_2 = {\rm Tr}([\rho(a_2)-\rho(a_1)] H_2)$; (iii) the system is now detached from the cold bath and made to undergo a second quantum adiabatic process (compression) during which the hamiltonian changes back to $H_1$. Work done [*on*]{} the system in this stage is ${\cal W}_2= {\rm Tr}(\rho(a_2)[H_1-H_2])$; (iv) finally, the system is brought in thermal contact with the hot bath again. Heat is absorbed by the system in this stage whence it recovers its initial state and its temperature attains back the value $T_1$. The total work done on average in a cycle is calculated to be $$\begin{aligned}
{\cal W} &=& \sum_{n} \left( \varepsilon_{n}^{(1)}- \varepsilon_{n}^{(2)} \right)
\left( p_{n}^{(1)}- p_{n}^{(2)} \right), \\
&=& (a_1-a_2)\sum_{n} \varepsilon_{n} \left( p_{n}^{(1)}- p_{n}^{(2)} \right) > 0.
\label{work}\end{aligned}$$ Similarly, heat exchanged with hot bath in stage (iv) is given by $Q_1=a_1 \sum_n \varepsilon_n \left( p_{n}^{(1)}- p_{n}^{(2)} \right) >0.$ Heat exchanged by the system with the cold bath is $Q_2 = {\cal W}-Q_1 <0$. The efficiency of the engine $\eta={\cal W }/Q_1$, is given by $$\eta = 1-\left(\frac{a_2}{a_1}\right).
\label{eta}$$ For convenience, we express ${\cal W} \equiv {\cal W}(a_1,\eta)$, using Eq. (\[eta\]). Consider an ensemble of such systems where now the value of parameter $a_1$ may vary from system to system. If the ensemble corresponds to an actual preparation according to a certain probability distribution $\Pi(a_1)$, then the state of the system can be expressed as $\hat{\rho} = \int \rho(a_1)\Pi(a_1) d a_1$. Each system in the ensemble is made to perform the quantum heat cycle described above, with a fixed efficiency $\eta$. We wish to study the optimal characteristics of the average work, in particular the efficiency at which the work becomes optimal. Clearly, choice of the probability distribution $\Pi(a_1)$ is expected to play a significant role in the conclusions. In the following, we analyse this problem by choosing a distribution $\Pi(a_1)$ according to the prior information available and show that the efficiency at optimal work is closely associated with CA value.
For simplicity, we now consider a two-level system as our working medium, with $\varepsilon_0 = 0$ and $\varepsilon_1 = 1$, so that the initial energy levels are $0$ and $a_1$. The work over a cycle in this case is $${\cal W}(a_1,\eta) = a_1 \eta \left[ \frac{1}{\left( 1+e^{a_1/T_1}\right) }
- \frac{1}{\left( 1+e^{a_1(1-\eta) /T_2}\right) } \right] > 0,
\label{worke}$$ where Boltzmann’s constant is put equal to unity. The average work with the initial state $\hat{\rho}$ for a given $\eta$, can be expressed as $$\overline{W} = \int_{a_{\rm min}}^{a_{\rm max}}
{\cal W}(a_1,\eta) \Pi (a_1) d a_1.
\label{avw}$$ A central issue in Bayesian probability is to assign a unique prior distribution corresponding to a given prior information. If the only prior information about the continuous parameter $a_1$ is that it takes positive real values but otherwise we have complete ignorance about it, then Jeffreys has suggested the prior distribution $\Pi(a_1) \propto {1}/{a_1}$ [@Jeffreys; @Jaynes1968], or in a finite range, $\Pi (a_1) = \left[ {\ln \left( a_{\rm max}/ a_{\rm min}\right)} \right]^{-1}
\left(1/a_1 \right)$, where $a_{\rm min}$ and $a_{\rm max}$ are the minimal and the maximal energy splitting achievable for the two-level system. For the above choice, we obtain $$\overline{W}= \left[ {\ln \left( \frac{a_{\rm max}}{ a_{\rm min}}\right)} \right]^{-1} \eta
\left[
\frac{T_2}{(1-\eta)} \ln \left( \frac{1+ e^{a_{\rm max}(1-\eta)/T_2}}
{1+ e^{a_{\rm min}(1-\eta)/T_2}} \right )
- T_1 \ln \left(\frac{1+ e^{a_{\rm max}/T_1}} {1+ e^{a_{\rm min}/T_1}} \right )\right ].
%- \ln \left( \right )
%
%\ln \left( \right ) \right ]
\label{av1w}$$ It can be seen that the average work $\overline{W}$ vanishes for $\eta = 0$ and $\eta = \eta_c$. In between these values of $\eta$, the average work exhibits a maximum. We look for the efficiency at which this work becomes maximal for the given range $[a_{\rm min},a_{\rm max}]$, by imposing the condition $\partial \overline{W} / \partial \eta = 0$. Here we consider the limit of $a_{\rm min} \to 0$ which gives $$\frac{T_2}{(1-\eta)^2} \ln\left[ \frac{1+ e^{a_{\rm max}(1-\eta)/T_2}}{2}\right]
-T_1 \ln\left[ \frac{1+ e^{a_{\rm max}/T_1}}{2} \right] -
\frac{\eta }{(1-\eta)}
\frac{a_{\rm max}}{\left( 1+ e^{-a_{\rm max}(1-\eta)/T_2}\right) } =0.
%
\label{etasol}$$ The solution $\eta$ of this equation has been plotted against $a_{\rm max}$ in Fig. 1. Interestingly, in the asymptotic limit of $a_{\rm max} >> T_1$, the above expression reduces to $$T_1 -\frac{T_2}{(1-\eta)^2} =0,
\label{eqca}$$ which yields the efficiency at optimal work as $\eta = 1- \sqrt{{T_2}/{T_1}}$, exactly the CA value. More significantly, the conclusion also holds in general i.e. for a working system with spectrum $\varepsilon_{n}^{(1)}= \varepsilon_n a_1$ and with Jeffreys’ prior. It is to be noted that in the asymptotic limits, the expression for average work (Eq. (\[av1w\])) diverges. However, the limits are taken after the derivative of work is set equal to zero in order to obtain well-defined expressions for the efficiency.
![ Efficiency versus $a_{\rm max}$ using Eq. (\[etasol\]). The curves correspond to $T_2 =1$ and $T_1$ taking values $2, 4, 6$ respectvely, from bottom to top. Apart from the approach to corresponding CA value at large $a_{\rm max}$, it is also seen that the limit is approached slowly for larger temperature differences.](fig1.eps){width="8cm"}
\[fig1\]
It is conceivable that other choices of the prior may yield similar results. To study consequences of deviations from the above choice, we consider a class of prior distributions, $\Pi (a_1) =
{N}{{a_1}^{-\gamma}}$, defined in the range $[0,a_{\rm max} ]$, where $N = (1-\gamma)/{a_{\rm max}}^{1-\gamma}$ and $\gamma <1$. Upon optimisation of the average work as defined in Eq. (\[avw\]) over $\eta$, we get $$\int_{0}^{a_{\rm max}} \left[ \frac{(a_1)^{1-\gamma}}{1+e^{a_1/T_1}} -
\frac{(a_1)^{1-\gamma}}{1+e^{a_1(1-\eta) /T_2}} \right] d a_1
- \frac{\eta}{T_2} \int_{0}^{a_{\rm max}}
\frac{(a_1)^{2-\gamma} e^{a_1(1-\eta)/T_2} } {( 1 + e^{a_1(1-\eta) /T_2} )^2 } d a_1 = 0.
\label{wgmodel}$$ In the limit $a_{\rm max}$ becoming very large, the above integrals can be evaluated using the standard results [@stint]. Then the above equation is simplified to $$(1-\eta^*)^{3-\gamma} -(1-\gamma)\theta^{2-\gamma}\eta^* -\theta^{2-\gamma} =0,
\label{cubeg}$$ where $\theta = T_2/T_1$. Now as $\gamma \to 1$, the above equation reduces to Eq. (\[eqca\]) and so CA value is also a limiting value for this model. Interestingly, even for other allowed values of $\gamma$, the solution $\eta^*$ of Eq. (\[cubeg\]) depends only on the ratio $\theta$, apart from the parameter $\gamma$. In particular, Laplace and Bayes have advocated a uniform prior to quantify the state of complete ignorance. For this case, we set $\gamma =0$. Then the above equation becomes $(1-\eta^*)^{3} - (1 + \eta^*) \theta^{2} = 0$, which has only one real solution given as $$\eta^{*} = 1+ \frac{\theta^{4/3}}{3\left(1 + \sqrt{1 + \frac{\theta^2}{27}} \right)^{1/3}} - \theta^{2/3}
\left(1 + \sqrt{1 + \frac{\theta^2}{27}} \right)^{1/3}.
\label{etacubic}$$ This solution along with other numerical solutions of (\[cubeg\]) for general $\gamma < 1$ are shown in Fig. 2. Remarkably, these curves stay very close to the CA value. However, at this point it is not possible to say in general what prior information may be quantified by the parameter $\gamma$. The curves in Fig. 2 are also closely similar to those observed in finite-time models at optimal power [@Esposito2009]. It is seen here that in the near-equilibrium regime, all the curves merge into each other and approach the CA value which is approximately $\eta_c/2$ in this limit. This can be shown as follows: taking $\theta$ to be close to unity in the near-equilibrium case, $\eta_c = (1-\theta)$ is close to zero. The efficiency $\eta^*$ being bounded from above by the Carnot value is thus small too. On using these facts in the expansion of Eq. (\[cubeg\]), we get $$\eta^* \approx \frac{\eta_c}{2} + \frac{(3-\gamma)}{16} {\eta_c}^2 + O({\eta_c}^3).
\label{cubegsolution}$$ Thus we recover the linear term ${\eta_c}/{2}$ mentioned earlier. For general values of $\theta$, the CA value is a lower bound for the efficiency at optimal work when $0<\gamma<1$ [@gammaamin].
{width="8cm"}
\[fig2\]
So far we have observed that the use of Jeffreys’ prior implies that efficiency at optimal work approaches CA value for arbitrary bath temperatures. Further, the efficiency also approaches a universal form for a class of priors, for nearly equal bath temperatures. In the following, we show that application of Bayes’ theorem can restore the efficiency back to the exact CA value even for the latter choice. Bayes’ theorem gives a prescription to convert the prior probabilities $\Pi (a_1) d a_1$ into posterior probabilities. Note that during the first quantum adiabatic process on the two-level system, the energy levels change from $(0,a_1)$ to $(0,a_2)$, but the system continues to occupy its initial state. The respective occupation probabilities are now interpreted as conditional probabilities, given by $p(\uparrow \vert a_1) = 1/(1+\exp(a_1/ T_1))$ and $p(\downarrow \vert a_1) = 1/(1+\exp(-a_1/ T_1))$. If the system is found in the up ($\uparrow$) state, the work done in this step is $(a_2-a_1) = -a_1 \eta$ and the posterior probabilities are given by $$p(a_1\vert \uparrow) d a_1 = \frac{p(\uparrow \vert a_1 ) \Pi (a_1) d a_1 }{\int
p(\uparrow \vert a_1 ) \Pi (a_1) d a_1}.
\label{Btheorem}$$ The average work for this process is now given by $W_1 = \int (-a_1 \eta) p(a_1\vert \uparrow) d a_1$. On the other hand, if the system is found in the down ($\downarrow$) state, the work is zero. Similarly, for the second quantum adiabatic process, the work performed can be either $(+a_1 \eta)$ or 0 and the avarage work $W_2$ for that process can be similarly calculated using the respective posterior probabilities. Now choosing the prior $\Pi (a_1) =
N {{a_1}^{-\gamma}}$ and with the system being in up state, the average work for the total cycle $(W_1 + W_2)$, in the limit of large $a_{\rm max}$ is given by $$W(\eta) = \frac{(2^{\gamma -1}-1) (1-\gamma) \zeta[2-\gamma] }
{(2^{\gamma}-1) \zeta[1-\gamma] } \eta \left(T_1 - \frac{T_2}{(1-\eta)} \right),
\label{finalw}$$ where $\gamma <1$. So using posterior probabilities, a well defined expression for average work is obtained even if the prior is non-normalisable in the asymptotic limit. More generally, given the value of external parameter $a_1$ and assuming canonical probabilities $p(n \vert a_1)$ to find the system in $n$th state, we infer the probability $p(a_1 \vert n) da_1$ about the value of $a_1$, if the system is actually found in the $n$th state. Remarkably, the work given by eq. (\[finalw\]) attains optimal value exactly at the CA efficiency, regardless of the value of $\gamma$ in the prior. Furthermore, the average work $W(\eta)$ shows the same dependence on efficiency $\eta$ as found for the classical Otto cycle in [@Leff].
In conclusion, we have argued the emergence of CA value as the efficiency at optimal work in quantum heat engines within a Bayesian framework. This effect of incorporating Bayesian probabilities leading to classical thermodynamic behavior in quantum systems has not been addressed before and may shed new light on the connection between information and thermodynamics. Due to current interest in small scale engines, the observation of similar curves (Fig. 2) as obtained in some recently proposed models of these engines, points to an interesting link between finite time models and our model based on the idea of prior information. Addressing these issues would hopefully lead to a broader perspective on the performance characteristics of small engines and also help to understand the limits of their performance based on principles of information.
ACKNOWLEDEMENTS {#acknowledements .unnumbered}
===============
The author expresses his sincere thanks to Arvind, Pranaw Rungta and Lingaraj Sahu for interest in the work and useful discussions.
H.S. Leff and A.F. Rex, *Maxwell’s Demon: Entropy, Information, Computing,* (Princeton: Princeton University Press, 1990); *Maxwell’s Demon 2: Entropy, Classical and Quantum Information, Computing* (Institute of Physics, Bristol, 2003). K. Muruyami, F. Nori and V. Vedral, Rev. Mod. Phys. [**81**]{}, 1 (2009). T. Sagawa and M. Ueda, Phys. Rev. Lett. [**102**]{}, 250602 (2009); Phys. Rev. Lett. [**100**]{}, 080403 (2008). K. Jacobs, Phys. Rev. A **80**, 012322 (2009). C. Joachim *et. al.*, Nature (London) [**408**]{}, 541 (2000). V. Serreli *et. al.*, Nature (London) [**445**]{}, 523 (2007). G. N. Hatsopoulos and E. P. Gyftopoulos, Found. Phys. **6**, 127 (1976). M. O. Scully, Phys. Rev. Lett. [**87**]{}, 220601 (2001) ; [*ibid*]{} [**88**]{}, 050602 (2002). A. E. Allahverdyan and Th. M. Nieuwenhuizen, Phys. Rev. Lett. **85**, 1799 (2000). J. Gemmer, M. Michel, and G. Mahler, [*Quantum thermodynamics*]{}, Springer, Berlin (2004) and references therein. H.E.D. Scovil and E.O. Schulz-Dubois, Phys. Rev. Lett. [**2**]{}, 262 (1959); J.E. Geusic, E.O. Schulz-Dubois, and H.E.D. Scovil, Phys. Rev. [**156**]{}, 343 (1967). R. Alicki, J. Phys. A, [**12**]{}, L103 (1979). E. Geva and R. Kosloff, J. Chem. Phys. [**96**]{}, 3054 (1992). T.D. Kieu, Phys. Rev. Lett. [**93**]{}, 140403 (2004); Eur. Phys. J. D [**39**]{}, 115 (2006). H.T. Quan, Yu-xi Liu, C.P. Sun, and F. Nori, Phys. Rev. E [**76**]{}, 031105 (2007). A. E. Allahverdyan, R. S. Johal, and G. Mahler, Phys. Rev. E [**77**]{}, 041118 (2008). F. Curzon and B. Ahlborn, Am. J. Phys. [**43**]{}, 22 (1975). H. S. Leff, Am. J. Phys. [**55**]{}, 8 (1987); P. T. Landsberg and H. Leff, J. Phys. A [**22**]{}, 4019 (1989). C. Van den Broeck, Phys. Rev. Lett. [**95**]{}, 190602 (2005). T. Schmiedl and U. Seifert, Europhys. Lett. **81**, 20003 (2008). Z.C. Tu, J. Phys. A: Math. Gen. **41**, 312003 (2008). M. Esposito, K. Lindenberg, and C. Van den Broeck, Phys. Rev. Lett. **102**, 130602 (2009). Y. Zhou and D. Segal, Phys. Rev. E **82**, 011120 (2010). V. Dose, Rep. Prog. Phys. **66**, 1421 (2003). H. Jeffreys, Theory of Probability (Clarendon Press, Oxford, 1939). E.T. Jaynes, IEEE Trans. Sys. Sc. and Cybernetics, **4**, 227 (1968). E.T. Jaynes, *Probability Theory: The Logic of Science* (Cambridge University Press, Cambridge, 2003). J.M. Hofman and C.H. Wiggens, Phys. Rev. Lett. **100**, 258701 (2008). M. Habeck, Phys. Rev. Lett. **98**, 200601 (2007). C. M. Caves, C. A. Fuchs, and R. Schack, Phys. Rev. A **65**, 022305 (2002). J.C. Lemm, J. Uhlig, and A. Weiguny, Phys. Rev. Lett. **84**, 2068 (2000). H. Jeffreys, *Scientific Inference* (Cambridge University Press, Cambridge, 1957). $\int_{0}^{\infty} {(x)^{1-\gamma}}/{\left(1+e^{x/T}\right)} d x =
(1-2^{\gamma -1}) T^{2-\gamma} \Gamma[2-\gamma] \zeta[2-\gamma]$, for $\gamma < 2$ and where $\Gamma[\cdot]$ and $\zeta[\cdot]$ are the Gamma function and Riemann zeta functions, respectively. If we start with the prior distribution $N/{a_1}^{\gamma}$ defined in the range $[a_{\rm min}, \infty]$, normalisable for $\gamma > 1$, and solve in the asymptotic limit $a_{\rm min} \to 0$, then CA value is an upper bound for optimal efficiency. This choice is also presented in Fig. 2 for $\gamma = 1.5$.
[^1]: electronic address: [email protected]
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We investigate whether explicit models of warped D-brane inflation are possible in string compactifications. To this end, we study the potential for D3-brane motion in a warped conifold that includes holomorphically-embedded D7-branes involved in moduli stabilization. The presence of the D7-branes significantly modifies the inflaton potential. We construct an example based on a very simple and symmetric embedding due to Kuperstein, $z_1 = $ constant, in which it is possible to fine-tune the potential so that slow roll inflation can occur. The resulting model is rather delicate: inflation occurs in the vicinity of an inflection point, and the cosmological predictions are extremely sensitive to the precise shape of the potential.'
author:
- Daniel Baumann
- Anatoly Dymarsky
- 'Igor R. Klebanov'
- Liam McAllister
- 'Paul J. Steinhardt'
title: A Delicate Universe
---
[**Introduction.**]{}String theory is a promising candidate for the theoretical underpinning of the inflationary paradigm [@Inflation], but explicit and controllable models of inflation in string theory have remained elusive. In this Letter we ask whether explicit working models are possible in the setting of slow roll warped D-brane inflation [@Dvali; @KKLMMT], in which the inflaton field is identified with the location of a mobile D3-brane in a warped throat region [@KS] of the compactification manifold. As explained in [@KKLMMT], moduli stabilization introduces potentially fatal corrections to the inflaton potential in this scenario. Some of these corrections arise from complicated properties of the compactification [@BHK] and have been computed only recently [@BDKMMM].
The attitude taken in most of the literature on the subject ([*[cf.]{}*]{} [@KKLMMT; @LindeReview]) is that because of the vast number and complexity of string vacua, in some nonzero fraction of them it should be the case that the different corrections to the inflaton potential cancel to high precision, leaving a suitable inflationary model. This expectation or hope has never been rigorously justified (but see [@Sandip] for a promising proposal), and there is no guarantee that the correction terms can ever cancel: for example, it may be the case that the correction terms invariably have the same sign, so that no cancellation can occur. In this Letter we report the results of a systematic investigation into whether or not this hope of fine-tuned cancellation can in fact be realized. Further details will appear in [@LongPaper]. The new ingredient that makes this work possible is the result of [@BDKMMM] for a correction to the volume-stabilizing nonperturbative superpotential. As explained in [@BHK; @GM; @BDKMMM], this effect corresponds to the interaction between the inflationary D3-brane and the moduli-stabilizing wrapped branes [@KKLT], [*[i.e.]{}*]{} D7-branes or Euclidean D3-branes wrapping a four-cycle of the compact space. The location of these wrapped branes therefore becomes a crucial parameter in the D3-brane potential.
In a recent paper [@Burgess], Burgess [*et al.*]{} showed that for a particular embedding of the D7-branes, the Ouyang embedding [@Ouyang], the correction to the inflaton potential from the term computed in [@BDKMMM] vanishes identically. In this case the potential is always too steep for inflation, independent of fine-tuning. We have found a similar problem [@LongPaper] in the large class of D7-brane embeddings described in [@Arean]. Here we consider a more promising case, a simple holomorphic embedding due to Kuperstein [@Kuperstein]. For fine-tuned values of the microphysical parameters, the potential for radial motion of a D3-brane in this background contains an approximate inflection point around which slow roll inflation can occur. This potential is not of the form anticipated by previous authors: the D7-branes have no effect on the quadratic term in the inflaton potential, but instead cause the potential to flatten in a small region far from the tip of the conifold. We emphasize that arranging for this inflection point to occur inside the throat region, where the metric is known and our construction is self-consistent, imposes a severe constraint on the compactification parameters. Moreover, inflation only occurs for a bounded range about the inflection point, which requires a high degree of control over the initial conditions of the inflaton field.
We employ natural units where $M_P^{-2} = 8\pi G \equiv 1$.
![Cartoon of an embedded stack of D7-branes wrapping a four-cycle, and a mobile D3-brane, in a warped throat region of a compact Calabi-Yau.[]{data-label="fig:throat"}](throatPRL.eps){width="40.00000%"}
[**The Compactification.**]{}Our setting is a flux compactification [@GKP; @FluxReview] of type IIB string theory on an orientifold of a Calabi-Yau threefold, or, more generally, an F-theory compactification. We suppose that the fluxes are chosen so that the internal space has a warped throat region, and that $n>1$ D7-branes supersymmetrically wrap a four-cycle that extends into this region (see Figure \[fig:throat\]). As a concrete example of this local geometry, we consider the warped version [@KS] of the deformed conifold $\sum z_i^2 = \varepsilon^2$, where $z_i$ are coordinates on $\mathbb{C}^4$. Assuming that the D3-brane is far from the tip of the conifold, we may neglect the deformation $\varepsilon$. We choose $z_\alpha=(z_1,z_2,z_3)$ as the three independent complex D3-brane coordinates, and use the conifold constraint to express $z_4$ in terms of them. We suppose that this throat is glued into a compact space, as in [@GKP], and for simplicity we take this space to have a single Kähler modulus $\rho$. Moduli stabilization [@KKLT] relies on the fact that strong gauge dynamics on suitable D7-branes generates a nonperturbative superpotential, $W_{\rm np} = A(z_\alpha) \exp[{-a \rho}]$, where $a=\frac{2\pi}{n}$. The D7-brane embedding is specified by a single holomorphic equation, $ f(z_\alpha) = 0 $, and the result of [@BDKMMM] is that \[equ:BDKMMMResult\] A(z\_) = A\_0 ( )\^[1/n]{} , where $A_0$ is independent of the D3-brane position $z_{\alpha}$. Including the flux superpotential [@GVW] $W_{\rm flux} = \int G\wedge \Omega \equiv W_0$, the total superpotential is $W = W_0 + A(z_\alpha) \exp[{-a\rho}] $. Next, the DeWolfe-Giddings Kähler potential [@DeWG] is \[equ:KKK\] [K]{}(,|,z\_, |z\_) = - 3 - 3 U , where $k(z_\alpha, \bar z_\alpha)$ is the Kähler potential of the Calabi-Yau space, and $\gamma$ is a constant [@LongPaper]. Well inside the throat but far from the tip, we may use the Kähler potential of the conifold [@Candelas], \[equ:randz\] k = ( \_[i=1]{}\^4 |z\_i|\^2 )\^[2/3]{} = r\^2 . Then the F-term potential is [@Burgess], [@LongPaper] \[equ:Fterm\] V\_F &=& , where k\^[|]{}= r . To this we add the contribution of an anti-D3-brane at the tip of the deformed conifold [@KKLMMT], V\_D = D(r)U\^[-2]{} , D(r) D (1- ) , where $D= 2T_3/h_0$, $T_3$ is the D3-brane tension, and $h_0$ is the warp factor [@KS] at the tip.\
[**Towards Fine-Tuned Inflation.**]{}To derive the effective single-field potential, we consider radial trajectories that are stable in the angular directions, so that the dynamics of the angular fields becomes trivial. We also integrate out the massive volume modulus, incorporating the crucial fact that the volume shifts as the D3-brane moves [@LongPaper]. Then the canonically-normalized inflaton $\phi \equiv r \sqrt{\frac{3}{2} T_3}$ parameterizes the motion along the radial direction of the throat. To investigate the possibility of sustained inflation, we consider the slow-roll parameter $\eta = V''/V$. We find $\eta = \frac{2}{3} + \Delta \eta(\phi)$, where $\Delta \eta$ arises from the dependence (\[equ:BDKMMMResult\]) of the superpotential on $\phi$. Slow-roll inflation is possible near $\phi=\phi_0$ if $\Delta \eta(\phi_0) \approx -\frac{2}{3}$. Here, using the explicit result of [@BDKMMM] for $A(\phi)$, we compute $\Delta \eta$ and determine whether the full potential can be flat enough for inflation.[^1]
A reasonable expectation implicit in prior work on the subject is that there exist fine-tuned values of the microphysical parameters for which $\Delta \eta(\phi) \approx -\frac{2}{3} $, [*[i.e.]{}*]{} the correction to the inflaton potential arising from $A(\phi)$ includes a term quadratic in $\phi$, which, for a fine-tuned value of its coefficient, causes $\eta$ to be small for a considerable range of $\phi$. However, we make the important observation that the functional form of (\[equ:BDKMMMResult\]) implies that there is actually [*no*]{} purely quadratic correction. To see this we note that $A$ is a holomorphic function of the $z_\alpha$ coordinates, which, by (\[equ:randz\]), scale with radius as $z_{\alpha} \propto \phi^{3/2}$. Thus, the presence of $A(\phi)$ in the form (\[equ:BDKMMMResult\]) does not lead to new quadratic terms in (\[equ:Fterm\]). This is concrete evidence against the hope of a fine-tuned cancellation of the inflaton mass over an extended range of $\phi$.
However, as we now explain, there exists a simple example in which a different sort of cancellation can occur. Kuperstein [@Kuperstein] studied the D7-brane embedding $z_1=\mu$, where we may assume that $\mu \in \mathbb{R}^{+}$. This embedding, and the potential in this background, preserve an $SO(3)$ subgroup of the $SO(4)$ global symmetry acting on the $z_i$ coordinates of the deformed conifold. To find a purely radial trajectory that is stable in the angular directions, we consider the variation $\delta z_1$ while keeping the radius $r$ fixed. We then require the first variation of the potential $\delta V = V(z_1+\delta z_1,r,\rho)- V(z_1,r, \rho)$ to vanish for all $r$, and the second variation $\delta^2V$ to be non-negative. The extremality constraint $\delta V=0$ specifies the radial trajectories $z_1=
\pm \frac{1}{\sqrt{2}} r^{3/2}$, $z_2 = \pm i z_1$. A detailed study [@LongPaper] of the angular mass matrix $\delta^2 V$ reveals that the trajectory along $z_1 = + \frac{1}{\sqrt{2}} r^{3/2}$ is unstable, while the trajectory along $z_1 = - \frac{1}{\sqrt{2}} r^{3/2}$ is stable in all angular directions. After integrating out the imaginary part of the Kähler modulus $\rho$, which amounts to the replacement $A \to |A|$ [@LongPaper], the potential along the latter trajectory is given in terms of the radius $r$ (or the canonical inflaton $\phi$) and the real-valued volume modulus $\sigma \equiv \frac{1}{2} (\rho + \bar \rho)$, as [@LongPaper] \[equ:VTwoField\] V(,) &=& g()\^[2/n]{} + . Here $g(\phi) \equiv \frac{f(\phi)}{f(0)} = 1 + \bigl(\frac{\phi}{\phi_\mu} \bigr)^{3/2}$, and $\phi_\mu^2 \equiv \frac{3}{2} T_3
(2\mu^{2})^{2/3}$ denotes the minimal radial location of the D7-branes. We have also introduced $ c^{-1} \equiv 4\pi\gamma(2\mu^2)^{2/3}$, used [@LongPaper] $\gamma = \sigma_0 T_3/3 $, and defined $U(\phi,\sigma) \equiv 2 \sigma - \frac{\sigma_0 }{3} \phi^2$. The parameter $\sigma_0$ is the stabilized value of the Kähler modulus in the absence of the D3-brane (or when the D3-brane is near the bottom of the throat). Now, for each value of $\phi$ we carry out a constrained minimization of the potential, [*i.e.*]{} we find $\sigma_\star(\phi)$ such that $\left.\frac{\partial V}{\partial \sigma}\right|_{\sigma_{\star}(\phi)} = 0$. The function $\sigma_\star(\phi)$ may either be computed numerically or fitted to high accuracy by the approximate expression [@LongPaper] \_() \_0 . Substituting $\sigma_\star(\phi)$ into (\[equ:VTwoField\]) gives our main result, the effective single-field potential $\mathbb{V}(\phi) \equiv V(\phi, \sigma_\star(\phi))$.
For generic values of the compactification parameters, $\mathbb{V}$ has a metastable minimum at some distance from the tip. In fact, one can show that the potential has negative curvature near the tip and positive curvature far away, so that by continuity, $\eta$ vanishes at some intermediate value $\phi_0$. Then, one can find fine-tuned values of the D7-brane position $\phi_{\mu}$ for which this minimum is lifted to become an inflection point (see Figure \[fig:pot\]). This transition from metastability to monotonicity guarantees that $\epsilon = \frac{1}{2}(V'/V)^2$ can be made extremely small, so that prolonged slow-roll inflation is possible. In our scenario, then, the potential contains an approximate inflection point at $\phi = \phi_0$, where $\mathbb{V}$ is very well approximated by the cubic \[equ:cubic\] = V\_0 + \_1 (- \_0) + \_3 (-\_0)\^3 , for some $V_0, \lambda_1, \lambda_3$.
![[**Inflaton potential $\mathbb{V}(\phi)$.**]{}\
Compactification data: $n=8$, $\phi_\mu = \frac{1}{4}$, $A_0 = 1$, $W_0= - 3.432 \times 10^{-4}$, $D = 1.2 \times 10^{-8} $, which imply $a \sigma_0 \approx 10.1$.[]{data-label="fig:pot"}](Potential.eps){width="40.00000%"}
The number of $e$–folds derived from the effective potential (\[equ:cubic\]) is \[equ:Ne\] N\_e() = \_[\_[end]{}]{}\^ = . ( ) |\_[\_[end]{}]{}\^ . Since $\eta$ is small only for a limited range of inflaton values, the number of $e$–folds is large only when $\epsilon$ is very small. This forces these models to be of the small field type. The scalar spectrum on scales accessible to cosmic microwave background (CMB) experiments can be red, scale-invariant, or blue, depending on how flat the potential is. That is, $n_s-1 = \left. ( 2\eta -6 \epsilon)\right|_{\phi_{\rm CMB}} \approx 2 \eta(\phi_{\rm CMB})$, where $\phi_{\rm CMB}$ corresponds to the field value when observable scales exit the horizon during inflation, say between $e$-folds 55 and 60. The sign of $\eta(\phi_{\rm CMB})$, and hence of $n_s - 1$, depends on where $\phi_{\rm CMB}$ is relative to the inflection point. If inflation only lasts for the minimal number of $e$–folds to solve the horizon and flatness problems then the scalar spectrum is blue. If the potential is made more flat, so that $\epsilon$ is smaller, inflation lasts longer, and $\phi_{\rm CMB}$ is reduced, the spectrum can be red. This sensitivity to the details of the potential reduces the predictivity of the scenario.\
[**Microscopic Constraints.**]{} A crucial consistency requirement is that the inflationary region around $\phi_0$, and the location $\phi_{\mu}$ of the tip of the wrapped D7-branes, should fit well inside the throat, where the metric is known. As shown in [@BM], the range of $\phi$ in Planck units is geometrically limited, \[equ:BM\] < , where $N \gg 1$ is the background D3-brane charge of the throat. When combined with the Lyth bound [@Lyth], this yields a sharp upper bound on the tensor signal in these models [@BM]. Here we find that this same bound actually poses an obstacle to inflation itself: for an explicit inflationary model with the Kuperstein embedding of D7-branes, $\phi_\mu$ and $\phi_0$ must obey (\[equ:BM\]). Although one can find examples [@LongPaper] in which this requirement is met, this imposes significant restrictions on the compactification. In particular, $N$ cannot be too large, implying that corrections to the supergravity approximation could be significant.\
[**Conclusions.**]{}We have assessed the prospects for an explicit model of warped D-brane inflation by including the known dangerous corrections to the inflaton potential. In particular, we have studied whether the hope of fine-tuning superpotential corrections to the inflaton potential to reduce the slow roll parameter $\eta$ can be justified. As shown in [@LongPaper], for a large class [@Arean] of holomorphic embeddings of wrapped D7-branes there are trajectories where the potential is too steep for inflation, with no possibility of fine-tuning to avoid this conclusion [@Burgess]. For the Kuperstein embedding [@Kuperstein], fine-tuning is possible in principle, and inflation can occur in a small region near an inflection point of the potential. The requirement (\[equ:BM\]) that this inflection point lies well inside the throat provides stringent constraints on the compactification. Detailed construction of compactifications where such constraints are satisfied remains an open problem.
This study illustrates the care that must be taken in assessing the prospects for inflationary cosmology in string theory. It appeared that warped D-brane inflation involved many adjustable parameters, including the D7-brane embedding and other compactification data, and so it was reasonable to expect that many working examples would exist. However, the compactification geometry constrains these microphysical parameters so that there is much less freedom to adjust the shape of the potential than simple parameter counting would suggest.
The problem of constructing a fully explicit model of inflation in string theory remains important and challenging. Diverse corrections to the potential that are negligible for many other purposes can be fatal for inflation, and one cannot reasonably claim success without understanding all these contributions. We have made considerable progress towards this goal, but have not yet succeeded: a truly exhaustive search for further corrections to the inflaton potential remains necessary.\
[**Acknowledgements.**]{}We thank C. Burgess, J. Conlon, O. DeWolfe, J. Distler, S. Kachru, J. Maldacena, A. Murugan, J. Polchinski, G. Shiu, E. Silverstein, H. Tye, and B. Underwood for helpful discussions. This research was supported in part by the National Science Foundation under Grant No. PHY-0243680 and by the Department of Energy under grant DE-FG02-90ER40542. The research of A.D. was supported in part by Grant RFBR 07-02-00878, and Grant for Support of Scientific Schools NSh-8004.2006.2. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of these funding agencies.
[10]{}
=14.5pt
A. H. Guth, Phys. Rev. D [**23**]{}, 347 (1981);
A. D. Linde, Phys. Lett. B [**108**]{}, 389 (1982);
A. Albrecht and P. J. Steinhardt, Phys. Rev. Lett. [**48**]{}, 1220 (1982).
G. Dvali and H. Tye, Phys. Lett. B [**450**]{}, 72 (1999). S. Kachru [*[et al.]{}*]{}, JCAP [**0310**]{}, 013 (2003). I. R. Klebanov and M. J. Strassler, JHEP [**0008**]{}, 052 (2000); I. R. Klebanov and A. A. Tseytlin, Nucl. Phys. B [**578**]{}, 123 (2000). M. Berg, M. Haack and B. Körs, Phys. Rev. D [**71**]{}, 026005 (2005). D. Baumann [*[et al.]{}*]{}, JHEP [**0611**]{}, 031 (2006). S. H. H. Tye, arXiv:hep-th/0610221;
J. M. Cline, arXiv:hep-th/0612129;
R. Kallosh, arXiv:hep-th/0702059.
N. Iizuka and S. P. Trivedi, Phys. Rev. D [**70**]{}, 043519 (2004).
D. Baumann [*[et al.]{}*]{}, to appear. S. B. Giddings and A. Maharana, Phys. Rev. D [**73**]{}, 126003 (2006).
S. Kachru [*[et al.]{}*]{}, Phys. Rev. D [**68**]{}, 046005 (2003). C. P. Burgess [*[et al.]{}*]{}, JHEP [**0703**]{}, 027 (2007). P. Ouyang, Nucl. Phys. B [**699**]{}, 207 (2004). S. Kuperstein, JHEP [**0503**]{}, 014 (2005). M. R. Douglas and S. Kachru, arXiv:hep-th/0610102.
S. B. Giddings, S. Kachru and J. Polchinski, Phys. Rev. D [**66**]{}, 106006 (2002).
S. Gukov, C. Vafa and E. Witten, Nucl. Phys. B [**584**]{}, 69 (2000) \[Erratum-ibid. B [**608**]{}, 477 (2001)\]. O. DeWolfe and S. B. Giddings, Phys. Rev. D [**67**]{}, 066008 (2003). P. Candelas and X. C. de la Ossa, Nucl. Phys. B [**342**]{}, 246 (1990).
D. Areán, D. E. Crooks and A. V. Ramallo, JHEP [**0411**]{}, 035 (2004). D. Baumann and L. McAllister, arXiv:hep-th/0610285.
D. H. Lyth, Phys. Rev. Lett. [**78**]{}, 1861 (1997).
[^1]: For the special case of the Ouyang embedding, $z_1 + i z_2 = \mu$, Burgess [*et al.*]{} proved a simple no-go result for fine-tuned brane inflation [@Burgess]. They found that for this particular example, $\Delta \eta$ vanishes along the angularly stable trajectory. We have found similar ‘delta-flat’ trajectories [@LongPaper] for all embeddings in the infinite class studied in [@Arean]. These trajectories cannot support slow roll inflation, no matter how the parameters of the potential are tuned. In this paper, we study an embedding for which there is no delta-flat direction.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this proceedings we present STAR measurements of two particle azimuthal correlations between trigger particles at mid-rapidity ($\abseta<$ 1) and associated particles at forward rapidities (2.7 $<\abseta<$ 3.9) in p+p, d+Au and Au+Au collisions at $\s $= 200 GeV. Two particle azimuthal correlations between a mid-rapidity trigger particle and forward-rapidity associated particles preferably probe large-x quarks scattered off small-x gluons in RHIC collisions. Comparison of the separate d- and Au-side measurements in d+Au collisions may potentially probe gluon saturation and the presence of Color Glass Condensate. In Au+Au collisions quark energy loss can be probed at large rapidities, which may be different from gluon energy loss measured at mid-rapidity.'
address: 'Purdue University, 525 Northwestern Avenue, West Lafayette, IN 47907-2036, USA; INFN Sezione di Bari, Via E. Orabona 4, 70126 Bari, Italy'
author:
- 'Levente Molnar (for the STAR Collaboration)'
title: 'Jet-like correlations between Forward- and Mid- rapidity in p+p, d+Au and Au+Au collisions from STAR at $\s $= 200 GeV'
---
Introduction
============
Jet-like azimuthal correlations at mid-rapidity in Au+Au collisions at RHIC energies have shown significant modifications, indicating the presence of a dense and strongly interacting medium [@StarWhite]. Such measurements at mid-rapidity mainly probe the energy loss of gluons, due to the dominance of gluon-gluon scattering at RHIC energies [@CTEQ]. STAR has the unique capability to extend two particle correlation measurements to forward rapidities (2.7 $<\abseta<$ 3.9) by utilizing the Forward Time Projection Chambers (FTPC) [@FtpcNim], with trigger particles at mid-rapidity ($\abseta<$ 1) in the STAR-TPC [@TpcNim]. This kinematics is sensitive to hard scattering of small-x gluons (from which the trigger particles are mostly produced) and large-x quarks (which fragment into associated particles at forward rapidities). The measurements may address the question of gluon saturation in d+Au collisions, energy loss of quark jets at large rapidities and the possible presence of long range $\deta$ correlations in Au+Au collisions [@Joern].
Analysis details
================
High-$\pt$ trigger particles (3 $< \pt^{trig} <$ 10 GeV/c, $\absetatrigger<$ 1) and associated particles (0.2 $< \pt^{assoc} <$ 2 GeV/c, 2.7 $< \absetaassoc <$ 3.9) are selected in the TPC and FTPCs, respectively. The two particle correlation functions are corrected for tracking efficiency and acceptance of the associated particles, and are normalized per trigger particle.
\[fig:ppdAu\]
The combinatorial background is constructed from the technique of event mixing. In p+p and d+Au collisions, combinatorial background is normalized to the signal region of $\absdphi<$ 1, because in this region no correlation structure is observed. In Au+Au collisions, elliptic flow ($v_2$) modulation is added pairwise to the mixed-events background. The background is normalized to the range of $0.8<\absdphi<1.2$ by the Zero Yield At 1 (ZYA1) method [@StarJet2; @Ulery:2006iw]. Trigger particle $v_2$ is taken to be the average of the results from the modified reaction plane and the 4-particle cumulant methods, and the range of the two results is taken to be the systematic uncertainty, as in [@StarJet2]. The associated particle $v_2$ used here is a parameterization of 2-particle cumulant measurements at forward rapidities by STAR [@StarFtpcV2], while the centrality dependence is parameterized from preliminary STAR results. The FTPC $v_{2}$ results obtained from different methods are in good agreement [@StarFtpcV2]. However, as a conservative estimate the same relative systematic uncertainty as at mid-rapidity is applied.
\[fig:AuAu\]
Results and discussion
======================
d+Au results - probing small-x gluons in nucleus
------------------------------------------------
Figure 1 shows the two particle azimuthal correlations for p+p and d+Au collisions at $\s$ = 200 GeV. The outgoing d- and Au-sides are shown separately. The p+p points are averaged over the positive and negative rapidities. As shown in Fig. 1, the away-side correlation shapes are similar. The d-side yield is suppressed by about a factor of two compared to the Au-side. The p+p result lies approximately at the average in between.
\[fig:TPCFTPC\]
Suppression of the d-side yield may be understood as the suppression of small-x gluons in the Au nucleus, as predicted by the Color Glass Condensate (CGC) picture [@CGC]. On the other hand, reduction of the d-side yield may arise from the energy degradation of the d-side quarks due to multiple scattering in the Au nucleus. Gluon anti-shadowing [@antishadowing] and the EMC effect [@EMC] would enhance the d-side yield relative to the Au-side.
Au+Au results - rapidity dependence of energy loss
--------------------------------------------------
Figure \[fig:AuAu\] shows the azimuthal correlations for 60-80% and 20-40% (left panel) and for 0-10% (right panel) Au+Au collisions at $\s$ = 200 GeV. Results for two associated $\pt$ ranges are shown for the 0-10% Au+Au data. The near-side correlations for 0.2 $< \pt^{assoc} <$ 2 GeV/c are consistent with zero within the systematic uncertainties. However, the central data at high associated $\pt$, with the reduced systematic uncertainty, are suggestive of non-zero correlation on the near-side. This result indicates that long range $\deta$ correlations, first observed in $\deta<$ 2 [@Joern], may extend out to $\deta \sim$ 4 in the FTPCs.
The away-side correlation shapes in Au+Au collisions are broadened with respect to p+p as shown in Fig. \[fig:AuAu\]. The broadening is present for each centrality and is similar to mid-rapidity measurements. Figure 3 shows the comparison of the azimuthal correlations measured at forward rapidity (red) and at mid-rapidity (blue). The away-side correlation shapes are identical within the systematic uncertainties. To understand the contributing physical processes of energy loss at mid- and forward-rapidities quantitative theory calculations are essential.
Summary
=======
In summary, we have presented two particle azimuthal correlations of charged hadrons at forward rapidities (2.7 $< \abseta <$ 3.9) with trigger particles selected at mid-rapidity ($\abseta <$ 1) in p+p, d+Au and Au+Au collisions at $\sqrt{s_{NN}}=$ 200 GeV.\
\
Near-side correlation is not observed in p+p and d+Au collisions. Near-side correlations in Au+Au collisions are consistent with zero with the systematic uncertainties for 0.2$<\pt^{assoc}<$2 GeV/c. However, the high associated $\pt$ data ($\pt>$1 GeV/c) in central Au+Au collisions suggest a finite near-side correlation, which may indicate the presence of long range $\deta$ correlations at forward rapidities.\
\
Significant away-side correlations are observed for all systems. In d+Au collisions, a factor of two suppression is observed for the d-side compared to the Au-side. The p+p measurement is approximately the average of the d-side and Au-side correlations. While the Color Glass Condensate gives a qualitative description of the observed relative d- and Au-side yields, due to suppression of small-x gluons in Au nucleus, other mechanisms may be also at work, such as $d$ energy degrading. Quantitative model calculations are needed for further understanding. The away-side correlation shape broadens from peripheral to central Au+Au collisions, similar to mid-rapidity observations. Moreover, the away-side correlation shapes are almost identical to those measured at mid-rapidity. To understand the contributing physical processes of energy loss at mid- and forward-rapidities quantitative theory calculations are essential.
References {#references .unnumbered}
==========
[50]{}
J. Adams [*et al.*]{} \[STAR Collaboration\], Nucl. Phys. A [**757**]{}, 102 (2005). H. L. Lai [*et al.*]{} \[CTEQ Collaboration\], Eur. Phys. J. C [**12**]{}, 375 (2000).
K. H. Ackermann [*et al.*]{}, Nucl. Instrum. Meth. A [**499**]{}, 713 (2003). M. Anderson [*et al.*]{}, Nucl. Instrum. Meth. A [**499**]{}, 659 (2003).
J. Putschke, these proceedings.
J. Adams [*et al.*]{} \[STAR Collaboration\], Phys. Rev. Lett. [**95**]{}, 152301 (2005). J. G. Ulery and F. Wang, arXiv:nucl-ex/0609016. J. Adams [*et al.*]{} \[STAR Collaboration\], Phys. Rev. C [**72**]{}, 014904 (2005).
L. V. Gribov, E. M. Levin and M. G. Ryskin, Phys. Rept. [**100**]{}, 1 (1983); D. Kharzeev, E. Levin and L. McLerran, Nucl. Phys. A [**748**]{}, 627 (2005). K. J. Eskola, V. J. Kolhinen and P. V. Ruuskanen, Nucl. Phys. B [**535**]{}, 351 (1998). J. J. Aubert [*et al.*]{} \[European Muon Collaboration\], Phys. Lett. B [**123**]{}, 275 (1983); J. Gomez [*et al.*]{}, Phys. Rev. D [**49**]{}, 4348 (1994).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The present paper deals with connected subtraction games in graphs, which are generalization of take-away games. In a connected subtraction game, two players alternate removing a connected subgraph from a given connected game-graph, provided the resulting graph is connected, and provided the number of vertices of the removed subgraph belongs to a prescribed set of integers. We derive general periodicity results on such games, as well as specific results when played on subdivided stars.'
author:
- 'Antoine Dailly [^1] $^,$[^2]'
- 'Julien Moncel [^3] $^,$[^4]'
- Aline Parreau
title: 'Connected Subtraction Games on Subdivided Stars[^5]'
---
**Keywords:** Combinatorial Games; Subtraction Games; Graphs
Connected subtraction games on graphs
=====================================
General description of the game
-------------------------------
In this paper, we study *connected subtraction games* on graphs, which are impartial combinatorial games where a player can remove a connected subgraph from a given connected game-graph, provided the moves lead to a new game-graph which remains connected, and provided the number of vertices that have been removed is legal, with respect to a given list of integers that characterizes the game. Such a game will be denoted a $CSG$ game.
More precisely, let $L$ be a set of positive integers and $G$ a connected graph. The game [$CSG(L)$]{} on $G$ is a 2-player game where, starting from $G$, a player can remove a connected subgraph $H$ from the current graph, whenever the number of vertices of $H$ belongs to $L$ and the remaining graph is still connected. The first player unable to play loses the game. See Figure \[fig:example\] for an example of a [$CSG(\{1,2,4\})$]{} game.
\(1) at (-0.5,0.5) ; (1) ++(0,0.25) node [1]{}; (2) at (0,0) ; (2) ++(-0.25,0) node [2]{}; (3) at (-0.5,-0.5) ; (3) ++(0,-0.25) node [3]{}; (4) at (-1.5,-0.5) ; (4) ++(0,-0.25) node [4]{}; (5) at (0.5,0.5) ; (5) ++(0,0.25) node [5]{}; (6) at (0.5,-0.5) ; (6) ++(0,-0.25) node [6]{}; (7) at (1,0) ; (7) ++(0.25,0) node [7]{}; (1)–(2)–(3)–(4); (2)–(5)–(7)–(6)–(2);
(2,0) to node\[above\][Player 1]{} (3,0) ;
\(5) at (0.5,0.5) ; (5) ++(0,0.25) node [5]{}; (6) at (0.5,-0.5) ; (6) ++(0,-0.25) node [6]{}; (7) at (1,0) ; (7) ++(0.25,0) node [7]{}; (5)–(7)–(6);
(5,0) to node\[above\][Player 2]{} (6,0) ;
\(6) at (0.5,-0.5) ; (6) ++(0,-0.25) node [6]{}; (7) at (1,0) ; (7) ++(0.25,0) node [7]{}; (7)–(6);
(8,0) to node\[above\][Player 1]{} (9,0) ;
at (0,0) [$\emptyset$]{};
In some sense, this game is a natural generalization of popular take-away games, that are usually played with heaps of counters, in which a player can remove a number of counters belonging to a given set $L$. In the theory of combinatorial games, this is called a (simple) subtraction game (see, for instance, Chapter 4 in the first volume of [@winningways]). From a graph theory point of view, such subtraction games can be seen as games [$CSG(L)$]{} played on paths.
Connected subtraction games on graphs have been addressed in [@033], where [$CSG(\{1,2\})$]{} is solved for subdivided stars and bistars. In the present paper we extend these results, by providing, on the first hand, insights for such games in general graphs, and, on the other hand, specific results for [$CSG(\{1,2,3\})$]{} and [$CSG(\{1,2,4\})$]{} in subdivided stars.
Notations for combinatorial games, outcomes and Grundy values
-------------------------------------------------------------
In this section, we recall basics of combinatorial games that we will need along the paper. For more about combinatorial games, the reader can refer to [@LIP; @winningways; @Siegel]. In the classical settings of combinatorial game theory, two players play in turn, without being able to pass. The game is finite (the number of positions is finite and we can not cycle through some positions), there are no random events, and the information is perfect (each player knows what is the situation and what are the possible moves). In addition, both players are identical, in the sense that they have the same available moves (such a game is called [*impartial*]{}). The first player having no more available moves loses the game (and the other player wins). Note that such games admit no draws.
In this setting, a given game is either a winning position or a losing position for the first player. We can recursively define losing and winning positions as follows. A given game is a losing position if it is either a game in which there are no available moves, or if it is a game in which each possible move leads to a winning position. A given game is a winning position if there exists a move leading to a losing position. A losing position is denoted as a ${\mathcal{P}}$-position, whereas a winning one is denoted as a ${\mathcal{N}}$-position.
We can also define recursively the Grundy value of a given game $G$, denoted ${\mathcal{G}}(G)$, as 0 if there are no possible moves, and as $\operatorname{mex}\{{\mathcal{G}}(G_1),\ldots,{\mathcal{G}}(G_k)\}$ otherwise, where $G_1,\ldots,G_k$ denote all the options of $G$ that are all the games we may obtain after playing one move on $G$. Recall that the $\operatorname{mex}$ of a finite sequence of nonnegative integers is defined as the smallest nonnegative integer not belonging to the sequence. Note that it follows from the definition that $G$ is a losing position if, and only if, ${\mathcal{G}}(G)=0$. Winning positions correspond to games $G$ for which we have ${\mathcal{G}}(G)>0$. Indeed, a game has Grundy value $g>0$ if, and only if, no move leads to a game having Grundy value $g$, and for any $0\leq g'<g$, there exists a move leading to a game having Grundy value $g'$. Therefore, the Grundy value is a refinement of the notion of outcome, i.e winning or losing positions. We generally consider that a game is completely solved if we know its Grundy value.
Grundy values are particularly useful when dealing with games that can be decomposed as sum of games. The sum of two games $G_1$ and $G_2$, denoted $G_1+G_2$, is the new game where a player can either play on $G_1$ (with any legal move on that game) or play on $G_2$ (with any legal move on that game).
It turns out that the Grundy value of a sum of games can be rather easily computed from the Grundy values of the games. It is easy to see that, for instance, ${\mathcal{G}}(G+G)=0$ for any game $G$. Indeed, we can show recursively that any move the first player makes on $G+G$ can be mimicked by the second player so as to leave a game $G'+G'$. Since there are no moves in $\emptyset+\emptyset$, this shows that ${\mathcal{G}}(G+G)=0$. More generally, the so-called Sprague-Grundy theorem [@Sprague] states that the binary value of ${\mathcal{G}}(G_1+G_2)$ can be computed as ${\mathcal{G}}(G_1)\oplus{\mathcal{G}}(G_2)$, where $\oplus$ denotes the bitwise XOR of the binary values of both games. For instance, the Grundy value of $G_1+G_2$ is 3 if ${\mathcal{G}}(G_1)=5$ and ${\mathcal{G}}(G_2)=6$, since $5_2=101$, $6_2=110$, and $101\oplus 110=011$, that is to say 3. A consequence of the Sprague-Grundy theorem is that two games $G_1$ and $G_2$ have the same Grundy value if and only if $G_1+G_2$ is a losing position. In this case we say that $G_1$ and $G_2$ are equivalent, denoted by $G_1 \equiv G_2$. We will often use this result to compute the Grundy value of a game.
In this paper, we will denote by ${\mathcal{G}}_L(G)$ the Grundy value of the game [$CSG(L)$]{} played on a graph $G$. If the context is clear, we will simply write ${\mathcal{G}}(G)$.
Literature review
-----------------
It is well-known (see for instance [@winningways]) that any simple subtraction game with a finite set $L$ is ultimately periodic, in the sense that there exists a period $T>0$ and an integer $k_0$ such that, for any $k\geq k_0$, we have ${\mathcal{G}}(G_{k+T})={\mathcal{G}}(G_{k})$, where $G_k$ denotes the game played with a heap of $k$ counters. Such ultimately periodic sequences of integers can be described using the $\overline{x_1,\ldots,x_k}$ notation, where $\overline{x_1,\ldots,x_k}$ denotes the infinite sequence $x_1,\ldots,x_k,x_1,\ldots,x_k,\ldots$. For instance, it is well-known and easy to check that $({\mathcal{G}}(G_k))_{k\in \mathbb N}=\overline{0123}$ for the subtraction game with $L=\{1,2,3\}$, meaning that, for instance, ${\mathcal{G}}(G_2)=2$ and ${\mathcal{G}}(G_4)=0$ (the sequence starts with the value corresponding to $k=0)$. Note that, in general, the sequence is ultimately periodic, meaning we may have a preperiod [@winningways Chapter 4]. For instance, for the subtraction game with $L=\{2,4,7\}$, the Grundy sequence has preperiod 8 and period $3$, since we have:
$$({\mathcal{G}}(G_k))_{k\in\mathbb N}=00112203\overline{102}.$$
Subtraction games are take-away games[^6], which is the family of games that includes more or less all combinatorial games plays on heaps of counters where players remove counters with some specific rules. The ultimate periodicity of such games has been conjectured by Guy [@Guy96] and is a key problem in Combinatorial Game Theory. See [@winningways] for more details on take-away games and their periodicity.
There exist several generalizations of take-away games in graphs. In [@Nim-graphs; @Nim-graphs2], the edges of the graphs are labeled with an integer and a token is moved along the edges. In Node-Kayles [@Node-Kayles], a move consists in removing a vertex and all its neighbours from the graph. If instead of removing vertices players remove edges, we obtain Arc-Kayles [@Node-Kayles]. In the game Grim defined and studied in [@Grim], a move consists in removing a vertex, deleting its adjacent edges, and deleting the resulting isolated vertices. Finally, a general definition of octal games on graphs and specifically of subtraction games has been recently given in [@033].
A natural question that arises in these generalizations in graphs is whether the (ultimate) periodicity of the Grundy sequences that appear in classical take-away games (that can be considered as a particular case of connected subtraction games played on paths) is still valid for more complicated graphs. This question has been studied for Node-Kayles [@fleischer] and Arc-Kayles [@Huggan] in the particular class of subdivided stars (which is the simplest generalization of paths). For both games, they are able to prove that the sequence is still (ultimately) periodic for specific subdivided stars with three rays. Huggan and Stevens [@Huggan] have conjectured that this is true for Arc-Kayles for all subdivided stars with three rays, one of which is of size 1, whereas Fleischer and Trippen [@fleischer] proved that this is not true for Node-Kayles. As for the game [$CSG(\{1,2\})$]{} on graphs, it is proved in [@033] that it is periodic in all subdivided stars and bistars. More precisely, if we denote by ${S_{\ell_1,\ldots,\ell_k}}$ the subdivided star obtained by appending to a single vertex $k$ finite paths of length $\ell_1,\ldots,\ell_k$, it is shown in [@033] that $${\mathcal{G}}_{1,2}({S_{\ell_1,\ldots,\ell_k}}) = {\mathcal{G}}_{1,2}({S_{\ell_1 \bmod 3,\ldots,\ell_k \bmod 3}})$$ for all $\ell_1,\ldots,\ell_k$. In other words, adding three vertices to any path of a subdivided star does not change its Grundy value. Note that this value 3 is also equal to the period of the subtraction game with $L=\{1,2\}$. Similar results for subdivided bistars are also derived in [@033], where a bistar is a graph obtained by connecting with a path the centers of two subdivided stars.
Outline of the paper
--------------------
In this paper we are interested in deriving general results on $CSG$ games, with a particular emphasis on the correlation of the structure of the graphs with the structure of the Grundy values.
Most of our results are about the behaviour of the Grundy value when one appends a path of varying size to a vertex of a graph. Let $G$ be a graph, $u$ one of its vertices and $k$ a positive integer. We denote by ${G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k}}$ the graph $G$ where a path $P_k$ of $k$ vertices is appended to $u$ ($u$ not being a part of this path). We extend this notation to the case where $G$ is empty, by setting in this case ${G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k}}=P_k$. We are interested by the periodicity of the function $$\begin{array}{ccccc}
f_{L,G,u} & : & \mathbb{N} & \to & \mathbb{N} \\
& & k & \mapsto & {\mathcal{G}}_L{({G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k}})}. \\
\end{array}$$ for various sets $L$.
In Section \[sec:generalperiod\], we give results for any subtraction game on a finite set. In particular, we extend the result of ultimate periodicity on paths to any graph. In other words, the function $f_{L,G,u}$ is ultimately periodic for any finite set $L$, graph $G$ and vertex $u$ of $G$. In the rest of the paper, we aim to prove periodicity without preperiod. In Section \[sec:123\], we derive general results for the game [$CSG(\{1,2,\ldots,N\})$]{}, that is the game where any connected subgraph of size up to $N$ can be removed. These results consider general graphs and subdivided stars to which paths are appended. They allow us to solve the particular case [$CSG(\{1,2,3\})$]{} in subdivided stars. In Section \[sec:124\], we derive periodicity results for [$CSG(\{1,2,4\})$]{} in subdivided stars.
General results on finite subtraction games {#sec:generalperiod}
===========================================
As said in the introduction, simple subtraction games with a finite set $L$ have ultimately periodic Grundy sequences (see Chapter 4 of [@Siegel] for a proof). Using our terminology, this means that if $L$ is finite, the function $f_{L,G,u}$ is ultimately periodic when $G$ is a path and $u$ is an endpoint of $G$. The main argument to prove this result is that when a path is long enough, all the moves in $L$ are possible and the number of moves is finite. Thus the Grundy value of a long path is the mex over a finite set. This set can only have a finite number of values which proves the periodicity. We extend this argument to any graph and any vertex.
\[thm:general\] Let $L$ be a finite set of positive integers, $G$ a graph and $u$ a vertex of $G$. Then the function $f_{G,L,u}$ is ultimately periodic.
We proceed by induction on $|G|$, where $|G|$ denotes the number of vertices of $G$. The case $|G|=1$ corresponds to the path and to simple subtraction games.
Let $G$ be a graph with at least two vertices and $u$ be a vertex of $G$. Let $G'\neq G$ be a connected subgraph of $G$ containing $u$. Assume that the function $f_{G',L,u}$ is ultimately periodic with period $T(G')$ and preperiod $k(G')$. Let $k_0$ and $T_0$ be the preperiod and the ultimate period of the Grundy sequence of the subtraction game with set $L$. Let $T$ be the lcm of $T_0$ and all the values $T(G')$, with $G'\neq G$ any connected subgraph of $G$ containing $u$, and let $k_{max}$ be the maximum of the preperiods among the preperiod $k(G')$ and $k_0$.
When $k\geq \max{L}$, then the set of available moves is fixed, and there are three kinds of moves from the position ${G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k}}$:
1. playing to ${G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k-i}}$ with $i\geq1$ (at most $|L|$ moves);
2. playing to ${G'\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k}}$, with $G'\neq G$ a connected subgraph of $G$ containing $u$ (at most $2^{|G|-1}$ moves);
3. playing to $P_{k-i}$ with $i\geq 1$ (at most $|L|$ moves).
Thus, the total number of moves is bounded by a constant $M$ only depending of $L$ and $G$ (but not depending of $k$). Thus the Grundy value of ${G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k}}$, $f_{G,L,u}(k)$, is bounded by $M$.
Let $k\in \mathbb N$ and let $A(k)\in \{0,\ldots,M\}^{|L|}$ be the $|L|$ consecutive values of the Grundy values of ${G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{i}}$ from $i=k+1$ to $i=k+|L|$: $$A(k)=(f_{G,L,u}(k+1),f_{G,L,u}(k+1),...,f_{G,L,u}(k+|L|)).$$
The vector $A(k)$ is an $|L|$-uplet with values in $\{0,...,M\}$. Since there are a finite number of such $|L|$-uplets, and since the number of such vectors $A(k)$ is infinite, then there exist $k_1$ and $k_2$ such that $A(k_1)=A(k_2)$. Moreover, we can assume that $k_{max}+|L|\leq k_1 \leq k_2$ and that $k_1\equiv k_2 \equiv 0 \bmod{T}$. Indeed, there are also infinitely many vectors $A(k)$ with $k\geq k_{max}+|T|$ and $k\equiv0\bmod{T}$, hence at least two of them must be equal.
Let $T_f=k_2-k_1$. We prove by induction that for any $k> k_1$, $f_{G,L,u}(k+T_f)=f_{G,L,u}(k)$ which will conclude the proof. By definition of $k_1$ and $k_2$, the result is true for $k_1< k \leq k_1+|L|$. Let $k>k_1+|L|$ and assume that the result is true for the $|L|$ values preceding $k$. We prove that this is still true for $k$. Remembering that, since $k \geq \max L$, then the set of available moves is fixed, and the value $f_{G,L,u}(k+T_f)$ is the mex of the following values:
1. $f_{G,L,u}(k+T_f-i)$ with $\leq i\leq |L|$;
2. $f_{G',L,u}(k+T_f)$, with $G'\neq G$ a connected subgraph of $G$ containing $u$;
3. ${\mathcal{G}}(P_{k+T_f-i})$ with $1\leq i\leq |L|$.
By induction hypothesis, $f_{G,L,u}(k+T_f-i)=f_{G,L,u}(k-i)$. Furthermore, $T_f$ is a multiple of $T$ and $k \geq k_{max}$, and thus $f_{G',L,u}(k+T_f)=f_{G',L,u}(k)$ and ${\mathcal{G}}(P_{k+T_f-i})={\mathcal{G}}(P_{k-i})$. Finally, since the set of available moves from ${G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k+T_f}}$ and from ${G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k}}$ are identical, then $f_{G,L,u}(k+T_f)$ and $f_{G,L,u}(k)$ are actually both a mex computed on the same set of values, and thus are equal.
The preperiod and the period obtained in Theorem \[thm:general\] can be arbitrary large, even compared to the period of the Grundy sequence of the subtraction game played on a path. However, in all the particular cases we have considered, the period is the same than for the path. We wonder if this is true for all subtraction games. Concerning the preperiod, for some simple subtraction games there is no preperiod and the Grundy sequence is purely periodic. This is the case for example when the set $L$ contains all the integers from $1$ to $N$. In the rest of the paper, we aim at proving “pure periodicity” theorems.
Some of our results are using the following lemma.
\[lem:general\] Let us consider the game $CSG(L)$, with $L$ a finite set, played on some family of graphs $\mathcal{F}$ which has the property that the Grundy value of the game when played on any graph $G$ of $\mathcal{F}$ depends only on the cardinality of $G$ modulo a given period. More precisely, let us assume that there exists $T>0$ and integers $\alpha_0,\ldots,\alpha_{T-1}$ satisfying $\{\alpha_0,\ldots,\alpha_{T-1}\} = \{0,\ldots,T-1\}$ such that ${\mathcal{G}}(G)=\alpha_{\Vert G\Vert}$ for any $G \in \mathcal{F}$, with $\Vert G\Vert$ defined as $|G| \bmod T$. Let $\mathcal{F}'$ be a family of graphs such that:
- any legal move played on a graph $G \in \mathcal{F}'$ leads to a graph of $\mathcal{F}$
- the set of legal moves played on a graph $G \in \mathcal{F}'$ satisfies $$\label{eq:legal_moves}
\{k \bmod T \mid \mbox{there exists a legal move removing } k \mbox{ vertices to } G\} = \{1,\ldots,T-1\}$$
Then, for any graph $G \in \mathcal{F}'$, we also have ${\mathcal{G}}(G)=\alpha_{\Vert G\Vert}$.
Let us denote $H^k$ the set of graphs of $\mathcal{F}$ that are obtained after playing any legal move consisting in removing $k$ vertices to a graph $G$ of $\mathcal{F}'$. By definition of $\mathcal{F}$, we have ${\mathcal{G}}(H) = \alpha_{\Vert H \Vert}$ for any $H \in H^k$. Since $|H| = |G|-k$, then we have ${\mathcal{G}}(H) = \alpha_{\Vert G \Vert-k \bmod T}$ for any such graph. Now, since (\[eq:legal\_moves\]) is assumed to be true, we have $$\begin{aligned}
{\mathcal{G}}(G) &=& \operatorname{mex}\{{\mathcal{G}}(H)\mid H\mbox{ is obtained by a legal move from }G\}\\
&=& \operatorname{mex}\{\alpha_{\Vert G \Vert-k\bmod T} \mid \mbox{there exists a legal move removing }k\mbox{ vertices to }G\}\\
&=&\operatorname{mex}\{\alpha_{\Vert G \Vert-1\bmod T},\alpha_{\Vert G \Vert-2\bmod T},\ldots,\alpha_{\Vert G \Vert-T+1\bmod T}\}
\end{aligned}$$ Since we assumed moreover that $\{\alpha_0,\ldots,\alpha_{T-1}\} = \{0,\ldots,T-1\}$, we conclude that ${\mathcal{G}}(G)=\alpha_{\Vert G \Vert}$, which is the desired result.
By inductively using this lemma, we will derive periodicity results for some specific families of subdivided stars, that will be useful for the study of $CSG(\{1,3\})$ and $CSG(\{1,2,4\})$ in subdivided stars.
In the sequel, this lemma will mainly be used in the special case where $\alpha_i=i$ for all $i$. In other words, $\mathcal{F}$ is such that ${\mathcal{G}}(G)=|G| \bmod T$ for all $G \in \mathcal{F}$.
We end this section with a simple observation.
\[obs:|V|\_bound\] For any graph $G$ and any set $L$ of integers, we have ${\mathcal{G}}_{L}(G) \leq |G|$.
We prove the result by induction on $G$. This is true if $G$ is empty since the Grundy value of the empty graph is 0. Let $G'$ be a graph obtained after a legal move on $G$. $G'$ has strictly less vertices than $G$. Thus by induction, ${\mathcal{G}}_L(G')\leq |G'|<|G|$. Finally, since ${\mathcal{G}}_L(G)=\operatorname{mex}\{{\mathcal{G}}_L(G') | G' \text{ option of } G\}$, we have ${\mathcal{G}}_L(G)\leq |G|$.
Connected subtraction games where we can remove up to $N$ vertices {#sec:123}
==================================================================
For a general finite set $L$, the period and the preperiod of the Grundy sequence of Theorem \[thm:general\] can be very large, as it is already the case for paths. Since the subtraction game on paths for the set $L=\{1,...,N\}=I_N$ is very simple, we focus on this game in this section and give periodicity results with small period, and without — or with a very short — preperiod. We first give general results, valid for any $N$, and any graph $G$. Then we focus on two families of subdivided stars and prove results that are still valid for any $N$. Finally, we apply these results to $N=3$ and completely solve it for subdivided stars. Note that the case $N=2$ is solved in [@033].
Results on general graphs
-------------------------
We first recall the Grundy sequence of the game [$CSG(I_N)$]{} on paths:
\[lem:In-paths\] Let $N$ and $k$ be two integers. We have ${\mathcal{G}}_{I_N}(P_k) = k \bmod (N+1)$, i.e. $({\mathcal{G}}_{I_N}(P_k))_{k\in\mathbb N} = \overline{012\ldots N}$, where $P_k$ denotes the path on $k$ vertices.
We now give some results on ${\mathcal{G}}_{I_N}(G)$ that are just depending on the number of vertices of $G$, when this last one is small enough. We will generalize this result when considering some specific stars (see Lemma \[lem:01s1kl\]).
\[lem:sizeOfGraph\] For any graph $G$ and any integer $N\geq 1$, we have ${\mathcal{G}}_{I_N}(G) = |G|$ if $|G|\in \{0,1\}$, ${\mathcal{G}}_{I_N}(G) \geq 2$ if $|G|\in \{2,\ldots,N\}$, ${\mathcal{G}}_{I_N}(G) = 0$ if $|G|=N+1$, and ${\mathcal{G}}_{I_N}(G) = 1$ if $|G|=N+2$.
There is no possible move in the empty graph and since $1\in I_N$, the only possible move from a single vertex is to take it to play to the empty graph. Thus it is clear that ${\mathcal{G}}_{I_N}(G)=|G|$ for $G$ with at most one vertex.
Assume now that $|G|\in \{2,\ldots,N\}$. There exists a vertex $u$ of $G$ such that $G-u$ is connected (take, for instance, $u$ as an endpoint of a diameter of $G$). Then removing the whole graph or removing $G-u$ are legal moves leading to a graph with at most one vertex. Thus ${\mathcal{G}}_{I_N}(G)\geq 2$.
Now, if $G$ has $N+1$ vertices, then every legal move leads to a graph $G'$ having between $1$ and $N$ vertices, and thus having a positive Grundy value. Hence ${\mathcal{G}}_{I_N}(G)=0$.
Similarly, if $G$ has $N+2$ vertices, one can play to a graph $G'$ with $N+1$ vertices since there exists a vertex not disconnecting the graph, thus to a graph with Grundy value $0$. But there is no possible move to a graph of Grundy value $1$. Finally ${\mathcal{G}}_{I_N}(G)=1$.
This lemma can be extended when one appends a path of size $N+1$ to $G$.
\[lem:general\_bounds\] Let $G$ be a graph, $u$ a vertex of $G$ and $G'={G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$. We have ${\mathcal{G}}_{I_N}(G') = |G|$ if $|G|\in \{0,1\}$, ${\mathcal{G}}_{I_N}(G') \geq 2$ if $|G|\in \{2,\ldots,N\}$, ${\mathcal{G}}_{I_N}(G') = 0$ if $|G|=N+1$, and ${\mathcal{G}}_{I_N}(G') = 1$ if $|G|=N+2$.
Let $n$ be the number of vertices of $G$. If $n\leq 1$, then $G'$ is a path and the result is true.
If $n \in \{2,\ldots,N\}$, then one can play from $G'$ to $P_{N+1}$ by removing $G$ and to ${G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+2-n}}$ which has $N+2$ vertices. By Lemma \[lem:sizeOfGraph\], these two graphs have Grundy values 0 and 1, thus ${\mathcal{G}}_{I_N}(G')\geq 2$.
Assume now that $n=N+1$. Let $H'$ be an option of $G'$. Either $H'={G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{i}}$ with $i\in \{1,...,N\}$ or $H'={H\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{i}}$ where $H$ is an option of $G$ not containing $u$ (taking $u$ but not the whole graph $G$ would disconnect $G'$). In the first case, one can play from $H'$ to $G$ which has Grundy value $O$. Thus ${\mathcal{G}}_{I_N}(H')\neq 0$. In the second case, as seen before, since $H$ has between $1$ and $N$ vertices, we also have ${\mathcal{G}}_{I_N}(H')\neq 0$. Finally, ${\mathcal{G}}_{I_N}(G')=0$.
Finally, assume that $n=N+2$. As before, the options $H'$ of $G'$ are either of the form $H'={G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{i}}$ with $i\in \{1,...,N\}$ or $H'={H\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{i}}$ where $H$ is a option of $G$ not containing $u$. In particular, one can take for $H$ in the second case the graph $G$ where a vertex not disconnecting $G$ and different from $u$ is removed. This option has Grundy value $0$. Thus ${\mathcal{G}}_{I_N}(G')>0$. In the first case, one can play to $G$ with Grundy value $1$ (by the previous lemma) and thus ${\mathcal{G}}_{I_N}(H')\neq 1$. In the second case, since $H$ has between $2$ and $N+1$ vertices, we also have ${\mathcal{G}}_{I_N}(H')\neq 1$. Finally, ${\mathcal{G}}_{I_N}(G')=1$.
In order to derive periodicity results, we will often prove that $G$ and ${G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$ have the same Grundy value. To prove such a result, a classical way is to prove the equivalent statement that $G+{G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$ is a ${\mathcal{P}}$-position. We will generally prove this by proving that any move of the first player can be answered by the second player to a move to a ${\mathcal{P}}$-position. The following proposition shows that, somehow, the only case to be addressed is the one where the first player removes vertices in $G$, and in particular removes the vertex $u$.
\[prop:winning\_move\] Let $N$ be an integer. Let $G$ be a graph and $u$ be a vertex of $G$. Assume $G$ is minimal in the sense that ${\mathcal{G}}_{I_N}(G)\neq{\mathcal{G}}_{I_N}({G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}})$ but ${\mathcal{G}}_{I_N}(G')={\mathcal{G}}_{I_N}({G'\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}})$ for any connected subgraph $G'$ of $G$ containing $u$. Since ${\mathcal{G}}_{I_N}(G+{G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}})\neq 0$, then there exists a winning move on $G+{G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$. Then every winning move of $G+{G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$ is in the component $G$, removes the vertex $u$ and leaves at least two vertices.
Any move in $P_{N+1}$ can be completed to leave $G + G$ to the first player, which is a ${\mathcal{P}}$-position. Any move in ${G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$ leading to ${G'\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$ with $G'$ a connected subgraph of $G$ containing $u$ can be replicated so as to leave $G'+{G'\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$, which is a ${\mathcal{P}}$-position since $G$ is assumed to be minimal. Now let us consider a move in ${G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$ leading to $P_{N+1-i}$. This implies that $|G|\leq N-i$. By Observation \[obs:|V|\_bound\], this implies $g={\mathcal{G}}_{I_N}(G) \leq N-i < N+1-i$, hence there exists a move from $P_{N+1-i}$ to $P_g$. In the whole game, the first player gets $G+P_g$ which has Grundy value $0$. Thus, if a winning move on $G+{G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$ exists, then this move is on $G$, and leads to a connected subgraph $G'$ of $G$. Moreover, if $u$ is still in $G'$ or if the move is taking the whole graph $G$, then this move can be replicated by the second player, so as to leave $G' +{G'\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$ or $P_{N+1}$, which is in both cases a ${\mathcal{P}}$-position. Finally, if $G'$ is a single vertex $v$, then the second player can take at the end of $P_{N+1}$ the same number of vertices, that is $|G|-1$, leaving a graph with $N+2$ vertices. Thus at the end, using Lemma \[lem:sizeOfGraph\], the two components of the sum have Grundy value 1 and thus the sum is a ${\mathcal{P}}$-position.
Simple stars
------------
In this subsection, we consider the case when $G$ is a simple star to which a path is appended. We call a simple star a graph with a central vertex connected to some vertices of degree 1, that we call leaves. We will denote by ${S_{1^t}}$ the simple star with $t$ leaves. We assume that $N\geq 3$ since the case $N=2$ is solved in [@033].
The periodicity of the Grundy values is an easy consequence of Proposition \[prop:winning\_move\].
\[prop:Stk\_general\] Let $t$ be a positive integer, $u$ the central vertex of ${S_{1^t}}$ and $N\geq 3$. For any $k\geq0$, we have ${\mathcal{G}}_{I_N}({{S_{1^t}}\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k}})={\mathcal{G}}_{I_N}({{S_{1^t}}\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k \bmod N+1}})$. In particular, the function $f_{I_N,{S_{1^t}},u}$ is purely periodic with period $N+1$.
Let $G={{S_{1^t}}\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k}}$ and $u'$ be the endpoint of $P_k$ in $G$ ($u'=u$ if $k=0$). Then ${{S_{1^t}}\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k+N+1}}$ can be written ${G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u'};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$. We need to prove that $G+{G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u'};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$ is a ${\mathcal{P}}$-position for any $k$ and $t$.
Assume that this is not true and let $(k,t)$ be the smallest integers (in lexicographic order) such that $G+{G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u'};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$ is ${\mathcal{N}}$-position.
Let $G'$ be a connected subgraph of $G$ containing $u'$. Either $G'$ is a path with endpoint $u'$, or $G'={{S_{1^{t'}}}\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k}}$ with $t'<t$. By minimality of $(k,t)$, and since the Grundy values of paths are periodic, then $G'+{G'\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u'};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$ is a ${\mathcal{P}}$-position.
Thus we can apply Proposition \[prop:winning\_move\] on $G$, and the only possible winning move in $G+{G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u'};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$ is a move in the component $G$ that is taking $u'$ and leaving at most two vertices. This corresponds to a move from $G$ to $G'={{S_{1^{t}}}\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k'}}$ with $k'<k$. Then the second player can play in the component ${G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u'};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$ to ${{S_{1^t}}\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k'+N+1}}$ by taking $k-k'$ vertices on the path. By minimality of $k$, the resulting position is a ${\mathcal{P}}$-position and thus the move from $G$ to $G'$ was not a winning move, a contradiction.
Since the function is purely periodic, if one wants to know the Grundy value of ${{S_{1^t}}\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k}}$ we just need to compute the values for $k\leq N$. The values for $k=0$, that are actually the Grundy values of ${S_{1^t}}$ are given in the following lemma.
\[lem:csg1nstarsareP\] Let $t\geq 1$ be a positive integer and $N\geq 3$. We have: $${\mathcal{G}}_{I_N}({S_{1^t}})=\left\{
\begin{array}{ll}
2 & \mbox{if } t \leq N-1 \mbox{ and } t \mbox{ odd} \\
3 & \mbox{if } t \leq N-1 \mbox{ and } t \mbox{ even} \\
0 & \mbox{if } t = N+i \mbox{ with } i \mbox{ even} \\
1 & \mbox{if } t = N+i \mbox{ with } i \mbox{ odd.}
\end{array}
\right.$$
We use induction on $t$. The base cases are $t=1$ and $t=2$ for which ${S_{1^t}}$ is simply $P_2$ and $P_3$ respectively and the result holds.
Assume now that $t\geq 3$. There are three cases:
**Case 1:** $t$ is odd, $3 \leq t \leq N-1$
We prove that ${S_{1^t}}+P_2$ is a ${\mathcal{P}}$-position. We examine every option of the first player:
- If the first player plays to either ${S_{1^t}} + P_1$ or $P_1 + P_2$, then the second player answers by playing to $P_1 + P_1$ which is a ${\mathcal{P}}$-position;
- If the first player plays to either $P_2$ or ${S_{1^t}}$, then the second player answers by playing by emptying the remaining graph;
- If the first player takes one leaf from ${S_{1^t}}$, then the second player answers by taking another leaf from ${S_{1^t}}$, leaving ${S_{1^{t-2}}} + P_2$. By induction hypothesis, this is a ${\mathcal{P}}$-position.
**Case 2:** $t$ is even, $4 \leq t \leq N-1$
We prove that ${S_{1^t}}+P_3$ is a ${\mathcal{P}}$-position. We examine every option of the first player:
- If the first player plays to ${S_{1^t}} + P_2$ (resp. to ${S_{1^{t-1}}}+P_3$), then the second player answers by taking one leaf from ${S_{1^t}}$ (resp. by reducing $P_3$ to $P_2$), leaving ${S_{1^{t-1}}}+P_2$, which is a ${\mathcal{P}}$-position by induction hypothesis;
- If the first player plays to either $P_3$ or ${S_{1^t}}$, then the second player answers by playing by emptying the remaining graph (this is always possible since $t\geq4$ and $t\leq N-1$ implies $N\geq3$);
- If the first player plays to either ${S_{1^t}} + P_1$ or $P_1 + P_3$, then the second player answers by playing to $P_1 + P_1$ which is a ${\mathcal{P}}$-position
**Case 3:** $t\geq N$
We write $t=N+i$ with $i\geq 0$. We use induction on $i$ to prove that ${\mathcal{G}}({S_{1^t}})=i \bmod 2$.
If $i=0$, then ${S_{1^t}}$ has $N+1$ vertices. By Lemma \[lem:sizeOfGraph\], this graph has Grundy value 0.
If $i>0$, then there is only one move from ${S_{1^t}}$, which is removing one leaf, leaving ${S_{1^{t-1}}}$. Thus, we have: $$\begin{array}{lll}
{\mathcal{G}}({S_{1^t}}) & = \operatorname{mex}({\mathcal{G}}({S_{1^{t-1}}})) & \mbox{ by definition of } {\mathcal{G}}\\
& = \operatorname{mex}(i-1 \bmod 2) & \mbox{ by induction hypothesis} \\
& = i \bmod 2&
\end{array}$$
One can get the values for $k>0$ from these first values by recursive computation. For instance, table \[tab:grundyS1tkI4\] gives the Grundy values for [$CSG(I_4)$]{} on ${S_{1^t,k}}$ for $t\leq 10$ and $k\leq 8$.
0 1 2 3 4 5 6 7 8 9 10
--- --- --- --- --- --- --- --- --- --- --- ----
0 1 2 3 2 0 1 0 1 0 1 0
1 2 3 2 0 1 0 1 0 1 0 1
2 3 4 0 1 2 3 2 3 2 3 2
3 4 0 1 4 3 2 3 2 3 2 3
4 0 1 5 3 4 5 4 5 4 5 4
5 1 2 3 2 0 1 0 1 0 1 0
6 2 3 2 0 1 0 1 0 1 0 1
7 3 4 0 1 2 3 2 3 2 3 2
8 4 0 1 4 3 2 3 2 3 2 3
: The Grundy values of ${S_{1^t,k}}$ for the game [$CSG(I_4)$]{}, for $t \leq 10$ and $k \leq 8$.[]{data-label="tab:grundyS1tkI4"}
Subdivided stars with three paths
---------------------------------
We now focus on subdivided stars with three paths, with one of the path of size $1$. We recall that we denote by $S_{1,k,\ell}$ the subdivided star with three paths of size respectively 1, $k$ and $\ell$. Technically speaking, all degree 1 vertices of $S_{1,k,\ell}$ are leaves, but, for the sake of simplicity, in the sequel, we will denote the degree 1 vertex corresponding to the path of size 1 as *the* leaf. Conversely, we will use the word *path* only to denote either $P_k$ or $P_{\ell}$, even if, technically speaking, $P_1$ is also a path. We prove that the Grundy values are purely periodic with period $N+1$ except for some values for which there is a preperiod $N+1$. More precisely, we will prove in this subsection the following theorem.
\[thm:S1lk\] Let $N\geq 3$, $k,\ell$ be two nonnegative integers. The Grundy value of the subdivided star ${S_{1,k,\ell}}$ for the game [$CSG(I_n)$]{} is $${\mathcal{G}}_{I_N}({S_{1,k,\ell}})=\left\{
\begin{array}{ll}
{\mathcal{G}}_{I_N}({S_{1,(k \bmod (N+1)) +N+1,N}}) & \mbox{if } k>N \mbox{ and } \ell \equiv N \bmod (N+1) \\
{\mathcal{G}}_{I_N}({S_{1,N,(\ell \bmod (N+1)) +N+1}}) & \mbox{if } \ell>N \mbox{ and } k \equiv N \bmod (N+1) \\
{\mathcal{G}}_{I_N}({S_{1,k \bmod (N+1),\ell \bmod (N+1)}}) & \mbox{otherwise}
\end{array}
\right.$$
Thanks to this theorem, for a fixed $N$ there are only $O(N^2)$ values to compute.
\[cor:S1lk-polAlgo\] Let $k,\ell$ be two nonnegative integers. Then there exists an algorithm computing the Grundy value of the subdivided star ${S_{1,k,\ell}}$ for the game [$CSG(I_N)$]{}, and this algorithm runs in time $\theta(N^2)$.
However, all of these values are needed to prove Theorem \[thm:S1lk\]. First, we give the Grundy values of ${S_{1,k,\ell}}$ when $\ell,k<N$.
\[lem:S1lkWithSmalllk\] Let $k,\ell$ be nonnegative integers. If $k,\ell<N$, then the Grundy value of the subdivided star ${S_{1,k,\ell}}$ for the game [$CSG(I_n)$]{} is $${\mathcal{G}}_{I_N}({S_{1,k,\ell}})=\left\{
\begin{array}{ll}
k+\ell & \mbox{if } k+\ell \leq N-2 \mbox{ and } k,\ell \mbox{ odd} \\
k+1 & \mbox{if } k+\ell \leq N-2 \mbox{ and } k=\ell\neq0 \mbox{ even} \\
k+\ell+2 \bmod (N+1) & \mbox{otherwise}
\end{array}
\right.$$
These values are compiled in Table \[tab:grundyValuesS1lk\].
In order to prove this result, we use induction on $(k,\ell)$ with lexicographic order. If $k$ or $\ell=0$, ${S_{1,k,\ell}}$ is simply the path $P_{\ell+k+2}$ on $\ell+k+2$ vertices, having Grundy value $k+\ell+2 \bmod (N+1)$.
Assume now that $k,\ell \neq 0$, and assume that the formula holds for all $(i,j) < (k,\ell)$. We prove that the Grundy value of ${S_{1,k,\ell}}$ satisfies the formula of Lemma \[lem:S1lkWithSmalllk\].
By symmetry, we can assume that $k\leq \ell$.
**Case 1:** $k+\ell \leq N-2$.
From ${S_{1,k,\ell}}$, then there are exactly three types of possible moves:
- Type 1 move: removing the leaf, leaving $P_{k+\ell+1}$ which has Grundy value $k+\ell+1$ (only one such move).
- Type 2 move: removing the leaf, the central vertex, one of the two paths and some vertices from the other path, leaving a path $P_{i}$, for $i \in \llbracket 0,\ell \rrbracket$ (hence Type 2 moves can have all the Grundy values between $0$ and $\ell$).
- Type 3 move: removing some vertices from one of the paths, leaving ${S_{1,i,\ell}}$ for $i \in \llbracket 0,k-1 \rrbracket$, or leaving ${S_{1,k,j}}$ for $j \in \llbracket 0,\ell-1 \rrbracket$, whose Grundy values may be computed by induction.
Depending on the parities of $k$ and $\ell$, the Grundy values of Type 3 moves may differ. The following subcases are dedicated to computing these Grundy values.
**Subcase 1.1:** $k,\ell$ are odd.
By induction hypothesis, the options ${S_{1,i,\ell}}$ with $i\in \llbracket 0,k-1\rrbracket$ have Grundy values:
- $\ell+i$ when $i$ is odd, giving all even values between $\ell+1$ and $\ell+k-2$;
- $\ell+i+2$ when $i$ is even, giving all odd values between $\ell+2$ and $\ell+k+1$.
That is to say, these options have all the values between $\ell+1$ and $\ell+k-1$, as well as value $\ell+k+1$. Similarly, the options ${S_{1,k,j}}$ with $j\in \llbracket 0,\ell-1 \rrbracket$ have all the Grundy values between $k+1$ and $\ell+k-1$, as well as value $\ell+k+1$.
Finally, among the options of ${S_{1,k,\ell}}$, we have all the values between $0$ and $\ell+k-1$, except the value $k+\ell$, hence ${\mathcal{G}}({S_{1,k,\ell}})=k+\ell$.
**Subcase 1.2:** $k=\ell$ and $k$ is even.
By induction hypothesis, the option ${S_{1,i,k}}$ have Grundy values $k+i+2>k+1$, for all $i\in \llbracket 0,k-1 \rrbracket$. Type 2 moves imply that ${\mathcal{G}}({S_{1,k,k}}) \geq k+1$. Since no option of ${S_{1,k,k}}$ has Grundy value $k+1$, this means that ${\mathcal{G}}({S_{1,k,k}})=k+1$.
**Subcase 1.3:** $k+\ell \leq N-2$, $k$ odd, $\ell$ even.
By induction hypothesis, the options ${S_{1,i,\ell}}$ with $i\in \llbracket 0,k-1\rrbracket$ have Grundy values $\ell+i+2$, giving all the values between $\ell+2$ and $\ell+k+1$ (since $i<k<\ell$ the case $\ell=i$ does not happen). As before, the options ${S_{1,k,j}}$ with $j\in \llbracket 0,j-1 \rrbracket$ give all the values between $k+1$ and $k+\ell$ and in particular the value $\ell+1$ since $k\geq 1$. Finally, all the values up to $\ell+k+1$ appear as options, except the value $\ell+k+2$, leading to ${\mathcal{G}}({S_{1,k,\ell}})=k+\ell+2$.
**Subcase 1.4:** $k+\ell \leq N-2$, $k$ even, $\ell$ odd.
By induction, the options ${S_{1,i,\ell}}$ give all the values between $\ell+1$ and $\ell+k$. Hence all the values between $0$ and $\ell+k$ are present in the options. Remember that value $k+\ell+1$ is done by the Type 1 move. Finally, the Grundy value of options ${S_{1,k,j}}$ are smaller than $k+\ell+1$, and we have again ${\mathcal{G}}({S_{1,k,\ell}})=k+\ell+2$.
**Subcase 1.5:** $k+\ell \leq N-2$, $k$ even, $\ell$ even, $k\neq \ell$.
The options ${S_{1,i,\ell}}$ give all the values between $\ell+2$ and $\ell+k+1$. The value $\ell+1$ is given by option ${S_{1,k,\ell-k-1}}$ (which exists since $k<\ell$). Thus all the values between $0$ and $\ell+k+1$ are present in the options. Finally, the Grundy value of options ${S_{1,k,j}}$ are smaller than $k+\ell+1$. Thus, we have again ${\mathcal{G}}({S_{1,k,\ell}})=k+\ell+2$.
**Case 2:** $k+\ell \geq N-1$ and $1 \leq k,\ell \leq N-1$.
Let $j=k+\ell+2 - N+1$. Then $0\leq j \leq N-1$. We prove that ${S_{1,k,\ell}}+P_j$ is a ${\mathcal{P}}$-position.
Assume that this is not true and take the smallest $j$ for which this result is not true.
If the first player takes $c$ vertices in ${S_{1,k,\ell}}$ then the second player can always take $N+1-c$ vertices in the same component and move to $P_j$. Then the sum is a ${\mathcal{P}}$-position. Indeed, if the first player takes $c$ vertices on the path of size $k$ and $c\leq k$, then the second player can take the $k-c\leq N-2$ vertices remaining on the path of size $k$, the central vertex, the leaf, and some vertices on the path of size $\ell$ to obtain $P_j$. If the first player plays from ${S_{1,k,\ell}}$ to a path, this one has $k+\ell+2-c>j$ vertices and the second player can take $N+1-c$ vertices on it to go to $P_j$.
Assume now that the first player plays $c$ vertices on $P_j$ to $P_{j'}$. By definition of $j$, we must have either $k\geq j$ or $\ell\geq j$. Assume that $k\geq j$. Then the first player takes $c$ vertices on the path of size $k$ to play to ${S_{1,k',j}}$. We still have $k'+\ell \geq N-1$ and by minimality, ${S_{1,k',\ell}}+P_{j'}$ is a losing position.
The following lemma gives the periodicity of values 0 and 1. It completes Lemma \[lem:sizeOfGraph\] for $S_{1,k,l}$.
\[lem:01s1kl\] Let $k,\ell$ be two integers.
If $|S_{1,k,\ell}|=0 \bmod (N+1)$ then ${\mathcal{G}}_{I_N}(S_{1,k,\ell})=0$.
If $|S_{1,k,\ell}|=1 \bmod (N+1)$ then ${\mathcal{G}}_{I_N}(S_{1,k,\ell})=1$.
Otherwise, ${\mathcal{G}}_{I_N}(S_{1,k,\ell})>1$.
We prove the whole lemma by induction on $(k,l)$ (in lexicographic order). Note that this result is trivially true for paths. By Lemma \[lem:sizeOfGraph\], the result is true if $|S_{1,k,\ell}|=k+\ell+2\leq N+2$.
Let $k,\ell$ with $k+\ell+2>N+2$. Assume first that $k+\ell+2=0\bmod (N+1)$. All the moves lead to a graph which is a path or a subdivided star with a number of vertices not congruent to 0 modulo $N+1$, thus, by induction the Grundy value is positive. Hence, ${\mathcal{G}}_{I_N}(S_{1,k,\ell})=0$.
Assume now that $k+\ell+2=1\bmod (N+1)$. As in the previous case, all the moves lead to a graph which is a path or a subdivided star with a number of vertices not congruent to 1 modulo $N+1$, and by induction the Grundy value is not equal to 1. Furthermore, it is always possible to remove one vertex, leading to a graph with 0 vertices modulo $N+1$ that has by induction Grundy value 0. Hence, ${\mathcal{G}}_{I_N}(S_{1,k,\ell})=1$.
Finally, assume that $k+\ell+2=i\bmod (N+1)$ with $i>1$. If one cannot remove $i$ vertices, it means that $k=\ell=i-1$. Hence the total number of vertices is $2i$ and must be congruent to $i$ modulo $N+1$. Since $i\neq 0$, this is not possible. Hence one can remove $i$ vertices and then the resulting graph has Grundy value 0. Similarly, if one cannot remove $i-1$ vertices, it means that $k=\ell=i-2$. Hence the total number of vertices is $2i-2$ and must be congruent to $i$ modulo $N+1$. It means that $i=2 \bmod (N+1)$ and so $i=2$. But then the graph is just $K_2$, which is not possible since there are at least $N+2$ vertices. Hence one can always remove $i-1$ vertices and the resulting graph has Grundy value 1. Finally, ${\mathcal{G}}_{I_N}(S_{1,k,\ell})>1$.
We now prove the “otherwise” part of Theorem \[thm:S1lk\] which is the simpliest case when both paths can be reduced modulo $N+1$.
\[lem:S1lkInternalPeriodicity\] Let $N\geq 3$, $k,\ell$ be two nonnegative integers satisfying the property:
$(P)$ $k\leq N$ or $\ell\leq N$ or $\ell$ and $k$ are not congruent to $N$ modulo $N+1.$
Then, we have ${\mathcal{G}}_{I_N}({S_{1,k,\ell}})=
{\mathcal{G}}_{I_N}({S_{1,k \bmod (N+1),\ell \bmod (N+1)}})$.
Let $k,\ell,N$ be integers satisfying $(P)$. Let $k_0=k \bmod{N+1}$ and $\ell_0=\ell \bmod{N+1}$.
We want to prove that ${S_{1,k,\ell}}+{S_{1,k_0,\ell_0}}$ is a ${\mathcal{P}}$-position. Consider a minimal counterexample, when minimizing $k+\ell$. We consider all the possible winning moves for the first player.
[**Case 1:**]{} The first player plays in the first component ${S_{1,k,\ell}}$.
[**Subcase 1.1:**]{} The first player takes $c$ vertices in the path of size $\ell$ leading to ${S_{1,k,\ell'}}$ with $\ell'<\ell$.
If $\ell>\ell_0$, then the second player can answer in the first component by ${S_{1,k,\ell-(N+1)}}$. Since $\ell-(N+1)=\ell_0 \bmod{N+1}$ and since $k,\ell-(N+1)$ satisfy $(P)$, by minimality the sum is a ${\mathcal{P}}$-position.
Assume now that $\ell=\ell_0$. Then the second player answers in the second component to ${S_{1,k_0,\ell'}}$. Since $k,\ell'$ satisfy $(P)$ ($\ell'\leq N$), then the sum is by minimality a ${\mathcal{P}}$-position.
By permuting $k$ and $\ell$, we can prove the symmetric case.
[**Subcase 1.2:**]{} The first player removes the leaf and thus move the first component to $P_{k+\ell+1}$. Then the second player can move the second component to $P_{k_0+\ell_0+1}$. Since $k=k_0 \bmod (N+1)$ and $\ell=\ell_0 \bmod (N+1)$, then these two paths have the same Grundy value and their sum is a ${\mathcal{P}}$-position.
[**Subcase 1.3:**]{} The first player removes the leaf, the path of size $k$, the central vertex and some vertices on the path of size $\ell$, leading the first component to $P_i$ with $i\leq \ell$. Then $k=k_0$ and the second player can remove the same number of vertices, leading to $P_{i-(\ell-\ell_0)}$, which has the same Grundy value as $P_i$.
By permuting $k$ and $\ell$ we get the symmetric case.
[**Case 2:**]{} The first player plays in the second component ${S_{1,k_0,\ell_0}}$.
[**Subcase 2.1:**]{} The first player takes $c$ vertices in the path of size $\ell_0$ leading to ${S_{1,k_0,\ell'_0}}$ with $\ell'_0<\ell_0$.
Let $\ell'=\ell-c$. Then $k,\ell'$ satisfy $(P)$. Indeed, $\ell'\neq N \bmod{N+1}$ since $\ell'=\ell'_0\bmod {N+1}$ and $0\leq \ell'_0<N$. Therefore, if the second player answer by playing in the first component to ${S_{1,k,\ell'}}$ we get a ${\mathcal{P}}$-position by minimality.
By permuting $k$ and $\ell$, we can prove the symmetric case.
[**Subcase 2.2:**]{} The first player removes the leaf and thus moves the second component to $P_{k_0+\ell_0+1}$. Then the second player can move the first component to $P_{k+\ell+1}$. Since $k=k_0 \bmod (N+1)$ and $\ell=\ell_0 \bmod (N+1)$, then these two paths have the same Grundy value and their sum is a ${\mathcal{P}}$-position.
[**Subcase 2.3:**]{} The first player removes the leaf, the path of size $k_0$, the central vertex and some vertices on the path of size $\ell_0$, leading the first component to $P_i$ with $i\leq \ell_0$. Let $c$ be the total number of vertices taken by the first player. If $k=k_0$, then the second player can play in the first component to $P_{i+\ell-\ell_0}$, which has the same Grundy value as $P_i$. Hence we can assume that $k>k_0$. Consider the position ${S_{1,k-c,\ell}}$.
Let $k'_0=k-c \bmod (N+1)$. We have $k'_0=k_0-c+N+1$ and $i+c=k_0+\ell_0+2$. Then $k'_0+\ell_0=k_0-c+N+1+\ell_0 =i+N+1$. In particular, $k'_0+\ell_0\geq N-1$. Hence, we are in the last case of Lemma \[lem:S1lkWithSmalllk\], and, by induction, $${\mathcal{G}}({S_{1,k-c,\ell}})={\mathcal{G}}({S_{1,k'_0,\ell_0}})=k'_0+\ell_0+2 \bmod (N+1)=i={\mathcal{G}}(P_i).$$
By permuting $k$ and $\ell$, we can prove the symmetric case.
We will now study the case where $(P)$ is not satisfied, which are the “if” parts of Theorem \[thm:S1lk\]. Since both parts are symmetrical, we only need to prove one of them.
\[lem:S1lkExternalPeriodicity\] Let $N\geq 3$, $k,\ell$ be two nonnegative integers such that $k>N$ and $\ell \equiv N \bmod (N+1)$. Then, we have ${\mathcal{G}}_{I_N}({S_{1,k,\ell}})={\mathcal{G}}_{I_N}({S_{1,(k \bmod (N+1))+N+1,N}})$.
We will first prove that the options ${S_{1,i,N}}$ with $i<N-2$ of ${S_{1,k,N}}$ (obtained by playing on the branch of size $N$) are not useful to compute ${\mathcal{G}}({S_{1,k,N}})$, when $N<k<2N$ . In order to prove this, we show how to compute the Grundy value of ${S_{1,k,N}}$ for $k \in \llbracket 1,2N \rrbracket$.
Let $r_N = \left \lfloor \frac{N}{2} \right \rfloor$ if $N \equiv 2,3 \bmod 4$, and $\left \lfloor \frac{N-2}{2} \right \rfloor$ otherwise.
This first technical claim will let us make explicit the Grundy values of the options of ${S_{1,k,N}}$ reached when playing on the path of length $N$:
\[clm:valuesOfColumn\] Let $k \in \llbracket 1,2N-1 \rrbracket$ be a positive integer with $k \not\in \{N-1,N,N+1\}$, and $k_0 = k \bmod (N+1)$.
1. If $k_0 \in \llbracket 1,r_N \rrbracket$ is even, then ${S_{1,k,N}}$ has options of all Grundy values in the interval $\llbracket 0,N \rrbracket$, except $2k_0+2$, by removing vertices from the path of length $N$.
2. If $k_0 \in \llbracket r_N+1,N-2 \rrbracket$ is even, then ${S_{1,k,N}}$ has options of all Grundy values in the interval $\llbracket 0,N \rrbracket$, except $k_0+1$, by removing vertices from the path of length $N$.
3. If $k_0$ is odd, then ${S_{1,k,N}}$ has options of all Grundy values in the interval $\llbracket 0,N \rrbracket$, except $N-1$ (if $N$ is odd), or $N$ (if $N$ is even), by removing vertices from the path of length $N$.
Let $k \in \llbracket 1,N-2 \rrbracket$ be a positive integer. We will make use of the Grundy values given in Lemma \[lem:S1lkWithSmalllk\].
If $k \in \llbracket 1,r_N \rrbracket$ is even, then every subdivided star ${S_{1,k,i}}$ (with $i \in \llbracket 0,N-1 \rrbracket$) has Grundy value $i+k+2 \bmod (N+1)$, except the subdivided star ${S_{1,k,k}}$ which has Grundy value $k+1$. Note that we have $2k+2 \leq N-2$ (by definition of $r_N$ and since $k$ is even). Thus, the subdivided star ${S_{1,k,N}}$ has options with all the Grundy values in the set $\{k+2 \bmod (N+1), \ldots , k+1+N \bmod (N+1)\}$, except $2k+2$, and an option of Grundy value $k+1$. This proves the first part of the claim.
Now, if $k \in \llbracket r_N+1,N-2 \rrbracket$ is even, then every subdivided star ${S_{1,k,i}}$ (with $i \in \llbracket 0,N-1 \rrbracket$) has Grundy value $i+k+2 \bmod (N+1)$. Thus, the subdivided star ${S_{1,k,N}}$ has options with all the Grundy values in the set $\{k+2 \bmod (N+1), \ldots , k+1+N \bmod (N+1)\}$, proving the second part of the claim.
Finally, if $k$ is odd, then the subdivided stars ${S_{1,k,i}}$ have the following Grundy values: $i+k$ if $i < N-k-1$ and $i$ is odd; and $i+k+2 \bmod (N+1)$ otherwise. Thus, the subdivided star ${S_{1,k,N}}$ has options with all the Grundy values in the set $\{0,\ldots,k\}$ (when $i \geq N-k-1$), and in the set $\{k+1, \ldots, N-2\}$. Furthermore, if $N$ is even then it has an option of Grundy value $N-1$, and if $N$ is odd then it has an option of Grundy value $N$. Indeed, the Grundy values of ${S_{1,k,0}},{S_{1,k,1}},{S_{1,k,2}},\ldots,{S_{1,k,N-k-4}},{S_{1,k,N-k-3}},{S_{1,k,N-k-2}}$ are $k+2,k+1,k+4,\ldots$ and either $\ldots,N-4,N-1,N-2$ if $N$ is even or $\ldots,N-2,N-3,N$ if $N$ is odd. This proves the third part of the claim.
If $k \in \llbracket N+2,2N-1 \rrbracket$, then the result holds by Lemma \[lem:S1lkInternalPeriodicity\]: the options of ${S_{1,k,N}}$ when removing vertices from the path of length $N$ have the same Grundy values as the equivalent options of ${S_{1,k-(N+1),N}}$.
We are now ready to compute the values of ${\mathcal{G}}({S_{1,k,N}})$ for $k\in \llbracket 1,2N \rrbracket$.
If $k=a(N+1)+b$ with $a \geq 0$ and $b \in \llbracket 1,N-1 \rrbracket$, then let $x_k$ be the number of subdivided stars ${S_{1,i,N}}$ with $i \in \llbracket a(N+1)+1,k-1 \rrbracket$ such that ${\mathcal{G}}({S_{1,i,N}})>N$.
\[claim:2N\] Let $N\geq 3$ be a nonnegative integer, and $k \in \llbracket 1,2N \rrbracket$, $k \notin \{N,N+1\}$. Let $m = \operatorname{mex}(\{{\mathcal{G}}({S_{1,k,i}}) | i<N\})$. We have the following: $${\mathcal{G}}({S_{1,k,N}}) = \left\{
\begin{array}{ll}
N & \mbox{if } k=N-1 \mbox{ or } k=2N \\
N-1 & \mbox{if } k=N-2 \mbox{ or } k=2N-1 \\
N+x_k+1 & \mbox{if } k \in \llbracket 1,r_N \rrbracket \mbox{ or } m \geq N-1 \\
m & \mbox{otherwise}
\end{array}
\right.$$
We prove the result by way of contradiction. Assume that $k \in \llbracket 1,2N \rrbracket$ is the smallest positive integer such that the result does not hold. Let $k_0 = k \bmod (N+1)$. From ${S_{1,k,N}}$, there are four kinds of moves:
- Type 1 move: removing the leaf, leaving $P_{N+1+k}$ which has Grundy value $k \bmod (N+1)$ (thus this option will never lead to a contradiction);
- Type 2 move: removing vertices from the path of length $N$, leaving stars ${S_{1,k,i}}$ for $i \in \llbracket 0,N-1 \rrbracket$;
- Type 3 move: removing vertices from the path of length $k$, leaving stars ${S_{1,i,N}}$ for $i \in \llbracket max(0,k-N),k-1 \rrbracket$;
- Type 4 move: if $k \leq N-2$, removing the path of length $k$, the leaf, the central vertex, and some vertices from the path of length $N$.
First, note that if $k \geq N+1$, then $x_k \leq x_{k-(N+1)}$. Indeed, by minimality of $k$, there are at most as many subdivided stars ${S_{1,i,N}}$ with Grundy values greater than $N$ when $i \in \llbracket N+2,k-1 \rrbracket$ than when $i \in \llbracket 1,k-(N+1)-1 \rrbracket$.
Observe that if $k=N$ (resp. $k=N+1$) then we have $|{S_{1,N,k}}|=2N+2$ (resp. $|{S_{1,N,k}}|=2N+3$) and by Lemma \[lem:01s1kl\] we have ${\mathcal{G}}({S_{1,N,k}})=0$ (resp. ${\mathcal{G}}({S_{1,N,k}})=1$). So $k \notin \{N,N+1\}$.
Assume that $k_0=N-1$, then by Lemmas \[lem:S1lkWithSmalllk\] and \[lem:S1lkInternalPeriodicity\], Type 2 moves leave subdivided stars with all Grundy values in $\llbracket 0,N-1 \rrbracket$, so ${\mathcal{G}}({S_{1,k,N}}) \geq N$. Furthermore, Type 3 moves will leave subdivided stars with Grundy values either smaller or greater than $N$ by minimality of $k$. Thus, ${\mathcal{G}}({S_{1,k,N}})=N$, a contradiction. So $k \notin \{N-1,2N\}$.
Now let $k \in \llbracket 1,2N-1 \rrbracket$, $k \notin \{N-1,N,N+1\}$. Assume first that $k_0$ is odd. Then, by Claim \[clm:valuesOfColumn\], the Type 2 moves leave subdivided stars with all Grundy values in $\llbracket 0,N \rrbracket$, except $N-1$ or $N$, depending on the parity of $N$.
Now if $k_0 < N-2$, then there are two cases:
1. If $k < N-2$ then ${S_{1,k,N}}$ has options of Grundy values $N-1$ and $N$ by Type 4 moves: leaving $P_{N-1}$ requires removing $k+2+1 \leq N$ vertices, and leaving $P_N$ requires removing $k+2 < N$ vertices.
2. If $k \in \llbracket N+2,2N-2 \rrbracket$, then Type 3 moves allow to reach ${S_{1,N-2,N}}$ and ${S_{1,N-1,N}}$, which have Grundy values $N-1$ and $N$.
Thus, the Grundy value of ${S_{1,k,N}}$ is greater than $N$. Furthermore, the Type 3 moves enable to reach $x_k$ options of Grundy values $N+1,\ldots,N+x_k$ by definition of $x_k$. And, if $k \in \llbracket N+2,2N-1 \rrbracket$, since $x_k \leq x_{k-(N+1)}$, the subdivided star ${S_{1,i,N}}$ such that ${\mathcal{G}}({S_{1,i,N}})=N+x_k+1$ verifies $i \leq k-(N+1)$, and as such is not an option of ${S_{1,k,N}}$. This implies that ${\mathcal{G}}({S_{1,k,N}})=N+x_k+1$, a contradiction. Thus if $k_0$ is odd then $k_0 = N-2$.
If $k_0 = N-2$, then $N$ is odd and ${S_{1,k,N}}$ has no option of Grundy value $N-1$ from Type 2 moves. Furthermore, Type 3 moves cannot lead to an option of Grundy value $N-1$. Indeed, those options have either $P_{N-1}$ (if $k=N-2$) or ${S_{1,N-2,N}}$ (if $k=2N-1$) as an option, and those have Grundy value $N-1$ by minimality of $k$. Those cannot be options of ${S_{1,k,N}}$ since it would require removing $N+1$ vertices. Finally, if $k=N-2$ then Type 4 moves can only leave $P_N$. Thus, ${\mathcal{G}}({S_{1,k,N}})=N-1$, a contradiction.
This implies that $k_0$ is even. Assume first that $k_0 \in \llbracket 1,r_N \rrbracket$. Then by Claim \[clm:valuesOfColumn\], Type 2 leave subdivided stars with all Grundy values in $\llbracket 0,N \rrbracket$, except $2k+2$, thus ${\mathcal{G}}({S_{1,k,N}}) \geq 2k+2$. We recall that $2k+2 \leq N-2$. There are two cases:
1. If $k \in \llbracket 1,r_N \rrbracket$, then the subdivided star ${S_{1,k,N}}$ has $P_{2k+2}$ as an option, by removing the path of length $k$, the leaf, the central vertex and $N-(2k+2)$ vertices from the path of length $N$. This move removes $N-k$ vertices, so this option is always available. Thus, the star ${S_{1,k,N}}$ has options of all Grundy values in $\llbracket 0,N \rrbracket$. Since Type 3 moves enable to reach $x_k$ options of Grundy values $N+1,\ldots,N+x_k$ by definition of $x_k$, we have ${\mathcal{G}}({S_{1,k,N}})=N+x_k+1$, a contradiction.
2. If $k \in \llbracket N-2,N+1+r_N \rrbracket$, then by minimality of $k$, no option of ${S_{1,k,N}}$ has Grundy value $2k_0+2$: the only options with a Grundy value smaller than $N$ have Grundy values either of the form $2i+2$ and are smaller than $2k_0+2$, or of the form $i+1$ with $i$ even, and thus are odd and different from $2k_0+2$. All this implies that ${\mathcal{G}}({S_{1,k,N}}) = 2k_0+2$, a contradiction.
Thus, we have $k_0 \in \llbracket r_N+1,N-2 \rrbracket$. By Claim \[clm:valuesOfColumn\] the second kind of move leaves stars with all Grundy values in $\llbracket 0,N \rrbracket$, except $k+1$, thus ${\mathcal{G}}({S_{1,k,N}}) \geq k+1$. The third kind of move can never leave a star with Grundy value $k+1$. Indeed, if $k \in \llbracket r_N+1,N-2 \rrbracket$ then by minimality of $k$, if an option by the third kind of move has a Grundy value smaller than $N$ then it is also smaller than $k+1$. Furthermore, if $k \in \llbracket N+2+r_N,2N-1 \rrbracket$ then any option by the third kind of move has ${S_{1,k_0,N}}$ as an option, which implies that all these options have a Grundy value different from $k_0+1$. Finally, if $k \in \llbracket r_N+1,N-2 \rrbracket$ then the fourth kind of move cannot leave $P_{k+1}$ since it would require removing $k+2+N-(k+1)=N+1$ vertices, which is impossible. Thus, the star ${S_{1,k,N}}$ has no option of Grundy value $k+1$, and as such we have ${\mathcal{G}}({S_{1,k,N}})=k+1$, a contradiction.
All of this implies that no such $k$ exists, so the claim is proven.
We now prove the lemma. Let $k,\ell$ be two nonnegative integers such that $k>N$ and $\ell \equiv N \bmod (N+1)$. Let $k_0=k\bmod (N+1) + N+1$.
A consequence of the previous proof is that the options ${S_{1,i,N}}$ with $i<N-2$ of ${S_{1,k_0,N}}$ (obtained by playing on the path of size $k_0$) are not useful to compute ${\mathcal{G}}({S_{1,k_0,N}})$.
We prove the result by induction on $k+\ell$. If $\ell=N$ and $k=N+1$ the result is clearly true.
Compare the options of ${S_{1,k,\ell}}$ and ${S_{1,k_0,N}}$. By Lemma \[lem:S1lkInternalPeriodicity\], options ${S_{1,k,\ell-i}}$ and ${S_{1,k_0, N-i}}$ have the same Grundy values. By induction and thanks to Claim \[claim:2N\], options ${S_{1,k-i,\ell}}$ and ${S_{1,k_0-i,N}}$, if $i\leq k_0+2$, have the same Grundy values. Options $P_{k+\ell+1}$ and $P_{k_0+N+1}$ have the same Grundy values by periodicity on the path (Lemma \[lem:In-paths\]).
Thus the only different options are ${S_{1,k-i,\ell}}$ and ${S_{1,k_0-i,N}}$ for $k_0+2<i\leq N$, if $k>k_0$ (if $k=k_0$, then they are the same by Lemma \[lem:S1lkInternalPeriodicity\] and thus we are done). By the previous remark, the second ones are not used to compute ${\mathcal{G}}({S_{1,k_0,N}})$. The first ones are different from ${\mathcal{G}}({S_{1,k-(N+1),\ell}})$ since one can play from ${S_{1,k-i,\ell}}$ to ${S_{1,k-(N+1),\ell}}$. By induction, ${\mathcal{G}}({S_{1,k-(N+1),\ell}})={\mathcal{G}}({S_{1,k_0,N}})$ and thus, we also have ${\mathcal{G}}({S_{1,k,\ell}})={\mathcal{G}}({S_{1,k_0,N}})$.
The proof of Theorem \[thm:S1lk\] is given by Lemma \[lem:S1lkInternalPeriodicity\] and Lemma \[lem:S1lkExternalPeriodicity\]. Note that the Grundy sequence is not purely periodic in general since for specific values of $N$ and $i$, we have ${\mathcal{G}}_{I_N}({S_{1,N,N+1+i}})\neq {\mathcal{G}}_{I_N}({S_{1,N,i}})$ (for example, ${\mathcal{G}}_{I_8}({S_{1,8,2}})=10$ whereas ${\mathcal{G}}_{I_8}({S_{1,8,11}})=6$). However, when $N$ is small enough, the “bad cases” do not happen and the sequence is purely periodic. This is the case for $N=3$. We will see in the next section, that actually the game [$CSG(I_3)$]{} is purely periodic on all the subdivided stars.
Application: study of [$CSG(1,2,3)$]{} on subdivided stars
----------------------------------------------------------
In this subsection, we use the previous general results to completely solve the game [$CSG(1,2,3)$]{} on subdivided stars.
\[thm:123\] For all $\ell_1,\ldots,\ell_t \geq0$, $t\geq0$, we have: $${\mathcal{G}}_{I_3}({S_{\ell_1,\ldots,\ell_t}}) = {\mathcal{G}}_{I_3}({S_{\ell_1 \bmod 4,\ldots,\ell_t \bmod 4}}).$$
We just need to prove that we can remove four vertices on any branch. Let $S={S_{\ell_1,\ldots,\ell_t,m}}$ and $S'={S_{\ell_1,\ldots,\ell_t,m+4}}$. We prove that $S+S'$ is a ${\mathcal{P}}$-position. If $S \in \{ \emptyset, P_1, P_2 \}$, then $S' \in \{ P_4, P_5, P_6 \}$ and the result holds by Lemma \[lem:In-paths\]. Hence we will suppose that $|S| \geq 3$.
We prove the result by way of contradiction. Let $u$ be the leaf of the path of length $m$ in $S$ (if $m=0$, then $u$ is the central vertex of $S$). Assume $S$ is minimal in the sense that $S \not\equiv S'$ but $S_1 \equiv {S_1\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{4}}$ for every sub-subdivided star $S_1$ of $S$. Note that, by sub-subdivided star, we mean a subdivided star $S_1 = {S_{\ell_1',\ldots,\ell_t'}}$ with $\ell_i' \leq \ell_i$ for all $i$, and with ant least one $i$ such that $\ell_i' < \ell_i$. By Proposition \[prop:winning\_move\] the only winning move in $S+S'$ is in $S$, removes the vertex $u$, and leaves at least two vertices in $S$.
If $u$ is the central vertex of $S$, then there are two cases:
1. $S$ is a path of length $k \geq 3$. In this case, $u$ can be a leaf, or at distance 1 or 2 from a leaf:
1. If $u$ is a leaf, then $S'$ is a path and $S \equiv S'$ by Lemma \[lem:In-paths\], a contradiction.
2. If $u$ is at distance 1 from a leaf, then $S'={S_{1,k-2,4}}$. If $k-2 \not\equiv 3 \bmod 4$, then by Theorem \[thm:S1lk\], we have ${\mathcal{G}}({S_{1,k-2,4}}) = {\mathcal{G}}(P_{k})$ and thus $S' \equiv S$. Otherwise, we have $k \equiv 1 \bmod 4$ and, as such, by Lemma \[lem:01s1kl\] we have ${\mathcal{G}}(S')=1={\mathcal{G}}(P_k)$. Both cases are contradictions.
3. If $u$ is at distance 2 from a leaf, then $S'={S_{2,k-3,4}}$ and the winning move consists in removing the path of length 2 along with $u$ in $S$, leaving $P_{k-3}$. We claim that removing three vertices from the path of length 4 in $S'$ is an equivalent move. Indeed, the move leaves ${S_{1,2,k-3}}$. If $k-3 \not\equiv 3 \bmod 4$, then by Theorem \[thm:S1lk\] and Lemma \[lem:S1lkWithSmalllk\] we have ${\mathcal{G}}({S_{1,2,k-3}}) ={\mathcal{G}}( {S_{1,2,k-3 \bmod 4}})= k+1 \bmod 4$ (the last equality comes from the fact that $2+(k-3 \bmod 4) \geq 2$). Finally, ${\mathcal{G}}({S_{1,2,k-3}})={\mathcal{G}}(P_{k-3})$. Otherwise, by Theorem \[thm:S1lk\], we have ${\mathcal{G}}({S_{1,2,k-3}})={\mathcal{G}}({S_{1,2,3}})$. By computation we have ${\mathcal{G}}({S_{1,2,3}})=3={\mathcal{G}}(P_{k-3})$. Both cases are contradictions.
2. $S={S_{1,1,k}}$. In this case, the winning move can only be removing the two leaves and $u$, leaving $P_k$. We claim that removing three vertices from the path of length 4 in $S'$ is an equivalent move. Indeed, this leaves ${S_{1^3,k}}$, which is equivalent to ${S_{1^3,k \bmod 4}}$ by Proposition \[prop:Stk\_general\]. There are four possibles values for $k \bmod 4$:
1. If $k \bmod 4 = 0$ then we have ${\mathcal{G}}({S_{1^3,k}})={\mathcal{G}}({S_{1^3}})=0$ by Lemma \[lem:csg1nstarsareP\].
2. If $k \bmod 4 = 1$ then we have ${\mathcal{G}}({S_{1^3,k}})={\mathcal{G}}({S_{1^4}})=1$ by Lemma \[lem:csg1nstarsareP\].
3. If $k \bmod 4 = 2$ then there are three options for ${S_{1^3,2}}$ which are ${S_{1^3}}$ (which has Grundy value 0 by Lemma \[lem:csg1nstarsareP\]), ${S_{1^4}}$ (which has Grundy value 1 by Lemma \[lem:csg1nstarsareP\]) and ${S_{1^2,2}}$ (which has Grundy value 1 by Lemma \[lem:S1lkWithSmalllk\]). Thus ${\mathcal{G}}({S_{1^3,2}})=2$.
4. If $k \bmod 4 = 3$ then there are four options for ${S_{1^3,3}}$ which are ${S_{1^3}}$ (which has Grundy value 0 by Lemma \[lem:csg1nstarsareP\]), ${S_{1^4}}$ (which has Grundy value 1 by Lemma \[lem:csg1nstarsareP\]), ${S_{1^3,2}}$ (which has Grundy value 2 by the above subcase) and ${S_{1^2,3}}$ (which has Grundy value 4 by computation, and using Lemma \[lem:S1lkWithSmalllk\]). Thus ${\mathcal{G}}({S_{1^3,2}})=3$.
In the four cases, we have ${\mathcal{G}}({S_{1^3,k \bmod 4}}) = k \bmod 4 = {\mathcal{G}}(P_k)$. Thus the only possible winning move in $S+S'$ actually leads to a an ${\mathcal{N}}$-position. This is a contradiction.
Assume now that $u$ is not the central vertex, which means that $u$ is a leaf. If the winning move of the first player does not take the central vertex, then the second player can take the same number of vertices in $S'$ in the branch of size $m+4$. So, necessarily, the winning move takes the central vertex, $u$, and leaves a path. This means that $S={S_{1,k,1}}$ and that the first player plays to $P_k$. We claim that removing $3$ vertices from the path of length $4$ is equivalent to $P_k$. Indeed, the second component becomes $S'_1={S_{1,k,2}}$ and by Theorem \[thm:S1lk\] and Lemma \[lem:S1lkWithSmalllk\], it has Grundy value $k\bmod {4}$.
All the cases lead to a contradiction, which proves that there is no $S$ such that $S+S'$ is an ${\mathcal{N}}$-position. Hence, we have ${\mathcal{G}}(S)={\mathcal{G}}(S')$.
The $CSG(1,2,4)$ game on subdivided stars {#sec:124}
=========================================
First, note that, if $M \not\equiv 0 \bmod (N+1)$, then the game [$CSG(I_N \cup \{ M \})$]{} is equivalent to the game [$CSG(I_N)$]{} when those games are played on paths. However, this is not the case when played on graphs, and even on subdivided stars.
\[obs:plus\_M\] For all $N \geq 2$, there exists $M \not\equiv 0 \bmod (N+1)$, such that there exists a graph $G$, and a distinguished vertex $u$ in $G$, such that ${\mathcal{G}}_L(G) \neq {\mathcal{G}}_L({G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}})$ where $L=I_N \cup \{ M \}$.
Let $S$ be a star with $N+2$ leaves and $u$ as a central vertex, and $M=2N+4$. $S + {S\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$ is an ${\mathcal{N}}$-position. The strategy for the first player is to empty ${S\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{N+1}}$ as a first move. The second player is left with a forced move on $S$. After two more forced moves, the first player will be able to empty $S$.
However, some particular games may retain this equivalence. In this section we study the game [$CSG(1,2,4)$]{} on subdivided stars and prove the following, which is analogous to Theorem \[thm:123\]:
\[thm:124\] Let $L=\{1,2,4\}$. For all $m\geq 0$ and $\ell_1,\ldots,\ell_t \geq0$, $t\geq0$, we have: $${\mathcal{G}}_L({S_{\ell_1,\ldots,\ell_t,m+3}}) = {\mathcal{G}}_L({S_{\ell_1,\ldots,\ell_t,m}}).$$
In order to prove this theorem, we first derive results for some specific subclasses of subdivided stars, and we then proceed by induction. For the rest of this section, all Grundy values will be computed for the game [$CSG(1,2,4)$]{}.
First, it is well-known and easy to see that this game is 3-periodic when played on paths.
\[lem:124-paths\] We have ${\mathcal{G}}(P_n) = |P_n| \bmod 3$, i.e. ${\mathcal{G}}(P_n) = \overline{012}$, where $P_n$ denotes the path on $n$ vertices.
Let us begin by some specific small subdivided stars that will be useful in the sequel.
\[lem:124-small-ones\] We have ${\mathcal{G}}(G)=|G| \bmod 3$ for all $G\in S$, with $$\begin{aligned}
S&=&\{{S_{1,1,1}}, {S_{1,1,2}}, {S_{1,1,3}}, {S_{1,1,1,3}},\\
&& {S_{1,2,2}}, {S_{1,2,3}}, {S_{1,1,2,2}}, {S_{1,1,2,3}}\}.
\end{aligned}$$ In addition, we have ${\mathcal{G}}({S_{1,1,1,1}})=0$ and ${\mathcal{G}}({S_{1,1,1,2}})=3$.
The proof is straightforward and a consequence of simple case analysis. It is easy to check that ${\mathcal{G}}({S_{1,1,1}})$ is indeed 1, since there are only 2 possible moves, that lead to $\emptyset$ and to the path $P_3$, that both have Grundy number 0. As for ${S_{1,1,2}}$, there are 4 possible moves, leading to paths $P_1$ and $P_3$, and to ${S_{1,1,1}}$, and the result follows from the application of the mex rule. Similarly, 4 moves are available from ${S_{1,1,3}}$, leading to paths and to ${S_{1,1,1}}$ and ${S_{1,1,2}}$. In the case ${S_{1,1,1,1}}$, we have only 2 possible moves, leading to ${S_{1,1,1}}$ and $P_1$, hence ${\mathcal{G}}({S_{1,1,1,1}})=0$. From ${S_{1,1,1,2}}$, 4 moves are available, leading to a path $P_2$ and to ${S_{1,1,1,1}}$, ${S_{1,1,1}}$, and ${S_{1,1,2}}$. We use this result to compute ${\mathcal{G}}({S_{1,1,1,3}})$, since one of the 4 possible moves leads to ${S_{1,1,1,2}}$, the other leading also to a path and to already known subdivided stars, namely ${S_{1,1,1,1}}$ and ${S_{1,1,3}}$. Similarly, there are 4 possible moves from ${S_{1,2,2}}$, all leading to already known graphs. There are 6 moves from ${S_{1,2,3}}$, again all leading to paths or previously studied subdivided stars. The cases ${S_{1,1,2,2}}$ and ${S_{1,1,2,3}}$ are similar: there are, respectively, 3 and 5 moves to study, that thankfully all lead to graphs whose Grundy value have been already computed above.
\[lem:124-S11k\] For all $k\geq0$, we have ${\mathcal{G}}({S_{1,1,k}})=|{S_{1,1,k}}| \bmod 3$, i.e. ${\mathcal{G}}({S_{1,1,k}}) = \overline{012}$.
We proceed by induction on $k$. The cases $k\leq3$ are all consequences of Lemma \[lem:124-paths\] and \[lem:124-small-ones\]. When $k\geq4$, we can apply Lemma \[lem:general\] since there are always exactly 5 moves available, leading to $P_{k+2}$, $P_{k-1}$, ${S_{1,1,k-1}}$, ${S_{1,1,k-2}}$, and ${S_{1,1,k-4}}$. Indeed, the lemma can be applied with $\mathcal{F}$ set to the set of all paths together with subdivided stars ${S_{1,1,k'}}$ such as $k'<k$, so as to get the desired result by induction (with $T$ in the statement of Lemma \[lem:general\] set to $3$ and $\alpha_i=i$ for $i\in\{0,1,2\}$).
\[lem:124-S111k\] For all $k\geq0$, we have ${\mathcal{G}}({S_{1,1,1,k}})=
\left\{\begin{array}{c}
1 \mbox{ if } k \equiv 0 \bmod 3\\
0 \mbox{ if } k \equiv 1 \bmod 3\\
3 \mbox{ if } k \equiv 2 \bmod 3
\end{array}\right.$, i.e. ${\mathcal{G}}({S_{1,1,1,k}}) = \overline{103}$.
The small cases $k\leq3$ are all consequences of Lemma \[lem:124-small-ones\]. When $k\geq4$, there are always exactly 5 possible moves, leading to ${S_{1,1,k}}$, $P_{k}$, ${S_{1,1,1,k-1}}$, ${S_{1,1,1,k-2}}$, and ${S_{1,1,1,k-4}}$. By induction, and thanks to Lemma \[lem:124-paths\] and \[lem:124-S11k\], we then have:
- If $k \equiv 0 \bmod 3$, then ${\mathcal{G}}({S_{1,1,1,k}})$ is computed as $\operatorname{mex}(0,0,3,0,3)=1$, and we are done.
- If $k \equiv 1 \bmod 3$, then ${\mathcal{G}}({S_{1,1,1,k}})$ is computed as $\operatorname{mex}(1,1,1,3,1)=0$, and we are done.
- If $k \equiv 2 \bmod 3$, then ${\mathcal{G}}({S_{1,1,1,k}})$ is computed as $\operatorname{mex}(2,2,0,1,0)=3$, and we are done.
\[lem:124-S12k\] For all $k\geq0$, we have ${\mathcal{G}}({S_{1,2,k}})=|{S_{1,2,k}}| \bmod 3$, i.e. ${\mathcal{G}}({S_{1,2,k}}) = \overline{120}$.
We proceed by induction on $k$. The cases $k\leq3$ are all consequences of Lemma \[lem:124-paths\] and \[lem:124-small-ones\]. When $k\geq4$, we can apply Lemma \[lem:general\] since there are always exactly 7 moves available, leading to $P_{k}$, $P_{k+2}$, $P_{k+3}$, ${S_{1,1,k}}$, ${S_{1,2,k-1}}$, ${S_{1,2,k-2}}$, and ${S_{1,2,k-4}}$. Indeed, the lemma can be applied with $\mathcal{F}$ set to the set of all paths together together with subdivided stars of Lemma \[lem:124-S11k\] and subdivided stars ${S_{1,2,k'}}$ such as $k'<k$, so as to get the desired result by induction (with $T$ in the statement of Lemma \[lem:general\] set to $3$ and $\alpha_i=i$ for $i\in\{0,1,2\}$).
\[lem:124-S112k\] For all $k\geq0$, we have ${\mathcal{G}}({S_{1,1,2,k}})=
\left\{\begin{array}{c}
2 \mbox{ if } k \equiv 0 \bmod 3\\
3 \mbox{ if } k \equiv 1 \bmod 3\\
1 \mbox{ if } k \equiv 2 \bmod 3
\end{array}\right.$, i.e. ${\mathcal{G}}({S_{1,1,2,k}}) = \overline{231}$.
The small cases $k\leq3$ are all consequences of Lemma \[lem:124-small-ones\]. When $k\geq4$, there are always exactly 6 possible moves, leading to ${S_{1,1,2,k-1}}$, ${S_{1,1,2,k-2}}$, ${S_{1,1,2,k-4}}$, ${S_{1,1,1,k}}$, ${S_{1,1,k}}$, and ${S_{1,2,k}}$. By induction, and thanks to Lemma \[lem:124-S11k\], \[lem:124-S111k\] and \[lem:124-S12k\], we then have:
- If $k \equiv 0 \bmod 3$, then ${\mathcal{G}}({S_{1,1,2,k}})$ is computed as $\operatorname{mex}(1,3,1,1,0,1)=2$, and we are done.
- If $k \equiv 1 \bmod 3$, then ${\mathcal{G}}({S_{1,1,2,k}})$ is computed as $\operatorname{mex}(2,1,2,0,1,2)=3$, and we are done.
- If $k \equiv 2 \bmod 3$, then ${\mathcal{G}}({S_{1,1,2,k}})$ is computed as $\operatorname{mex}(3,2,3,3,2,0)=1$, and we are done.
Note that none of the subdivided stars ${S_{1,1,2,k}}$ is a ${\mathcal{P}}$-position.
We also need the following detailed analysis of the subdivided star ${S_{3,3,3}}$.
\[lem:124-S333\] We have ${\mathcal{G}}({S_{3,3,3}})={\mathcal{G}}({S_{2,2,2}})=1$, ${\mathcal{G}}({S_{2,2,3}})=2$, with the Grundy numbers of subdivided stars that are accessible from ${S_{3,3,3}}$ as in Figure \[fig:124-S333\].
The proof is straightforward, and relies on the previous Lemmas \[lem:124-paths\] and \[lem:124-small-ones\]. Figure \[fig:124-S333\] speaks for itself.
(s333) at (0,0)
(dessin) at (0,1)
\(c) at (0,0) ; (1) at (1,1) ; (2) at (2,1) ; (3) at (3,1) ; (4) at (1,0) ; (5) at (2,0) ; (6) at (3,0) ; (7) at (1,-1) ; (8) at (2,-1) ; (9) at (3,-1) ; (c)–(1)–(3); (c)–(7)–(9); (c)–(6);
; (a) at (0,0) [${\mathcal{G}}({S_{3,3,3}})=1$]{}; (b) at (0,-0.5) [by $\operatorname{mex}$.]{};
; (s332) at (-3,-3)
(dessin) at (0,1)
\(c) at (0,0) ; (1) at (1,1) ; (2) at (2,1) ; (3) at (3,1) ; (4) at (1,0) ; (5) at (2,0) ; (7) at (1,-1) ; (8) at (2,-1) ; (9) at (3,-1) ; (c)–(1)–(3); (c)–(7)–(9); (c)–(5);
; (a) at (0,0) [${\mathcal{G}}({S_{3,3,2}})=0$]{}; (b) at (0,-0.5) [by $\operatorname{mex}$.]{};
; (s331) at (3,-3)
(dessin) at (0,1)
\(c) at (0,0) ; (1) at (1,1) ; (2) at (2,1) ; (3) at (3,1) ; (4) at (1,0) ; (7) at (1,-1) ; (8) at (2,-1) ; (9) at (3,-1) ; (c)–(1)–(3); (c)–(7)–(9); (c)–(4);
; (a) at (0,0) [${\mathcal{G}}({S_{3,3,1}})=2$]{}; (b) at (0,-0.5) [by $\operatorname{mex}$.]{};
; (s322) at (-4.5,-6)
(dessin) at (0,1)
\(c) at (0,0) ; (1) at (1,1) ; (2) at (2,1) ; (3) at (3,1) ; (4) at (1,0) ; (5) at (2,0) ; (7) at (1,-1) ; (8) at (2,-1) ; (c)–(1)–(3); (c)–(7)–(8); (c)–(5);
; (a) at (0,0) [${\mathcal{G}}({S_{3,2,2}})=2$]{}; (b) at (0,-0.5) [by $\operatorname{mex}$.]{};
; (s321) at (-1.5,-6)
(dessin) at (0,1)
\(c) at (0,0) ; (1) at (1,1) ; (2) at (2,1) ; (3) at (3,1) ; (4) at (1,0) ; (5) at (2,0) ; (7) at (1,-1) ; (c)–(1)–(3); (c)–(7); (c)–(5);
; (a) at (0,0) [${\mathcal{G}}({S_{3,2,1}})=1$]{}; (b) at (0,-0.5) [by Lemma \[lem:124-small-ones\].]{};
; (s33) at (1.5,-6)
(dessin) at (0,1)
\(c) at (0,0) ; (1) at (1,1) ; (2) at (2,1) ; (3) at (3,1) ; (7) at (1,-1) ; (8) at (2,-1) ; (9) at (3,-1) ; (c)–(1)–(3); (c)–(7)–(9);
; (a) at (0,0) [${\mathcal{G}}({S_{3,3}})=1$]{}; (b) at (0,-0.5) [by Lemma \[lem:124-paths\].]{};
; (s311) at (4.5,-6)
(dessin) at (0,1)
\(c) at (0,0) ; (1) at (1,1) ; (2) at (2,1) ; (3) at (3,1) ; (4) at (1,0) ; (7) at (1,-1) ; (c)–(1)–(3); (c)–(7); (c)–(4);
; (a) at (0,0) [${\mathcal{G}}({S_{3,1,1}})=0$]{}; (b) at (0,-0.5) [by Lemma \[lem:124-small-ones\].]{};
; (s222) at (0,-9)
(dessin) at (0,1)
\(c) at (0,0) ; (1) at (1,1) ; (2) at (2,1) ; (4) at (1,0) ; (5) at (2,0) ; (7) at (1,-1) ; (8) at (2,-1) ; (c)–(1)–(2); (c)–(7)–(8); (c)–(5);
; (a) at (0,0) [${\mathcal{G}}({S_{2,2,2}})=1$]{}; (b) at (0,-0.5) [by $\operatorname{mex}$.]{};
; (s221) at (-3,-9)
(dessin) at (0,1)
\(c) at (0,0) ; (1) at (1,1) ; (2) at (2,1) ; (4) at (1,0) ; (7) at (1,-1) ; (8) at (2,-1) ; (c)–(1)–(2); (c)–(7)–(8); (c)–(4);
; (a) at (0,0) [${\mathcal{G}}({S_{2,2,1}})=0$]{}; (b) at (0,-0.5) [by Lemma \[lem:124-small-ones\].]{};
; (s32) at (-6,-9)
(dessin) at (0,1)
\(c) at (0,0) ; (1) at (1,1) ; (2) at (2,1) ; (3) at (3,1) ; (7) at (1,-1) ; (8) at (2,-1) ; (c)–(1)–(3); (c)–(7)–(8);
; (a) at (0,0) [${\mathcal{G}}({S_{3,2}})=0$]{}; (b) at (0,-0.5) [by Lemma \[lem:124-paths\].]{};
; (s22) at (3,-9)
(dessin) at (0,1)
\(c) at (0,0) ; (1) at (1,1) ; (2) at (2,1) ; (7) at (1,-1) ; (8) at (2,-1) ; (c)–(1)–(2); (c)–(7)–(8);
; (a) at (0,0) [${\mathcal{G}}({S_{2,2}})=2$]{}; (b) at (0,-0.5) [by Lemma \[lem:124-paths\].]{};
;
(s333) to (s332); (s333) to (s331); (s332) to (s322); (s332) to (s321); (s332) to (s33); (s331) to (s321); (s331) to (s33); (s331) to (s311); (s322) to (s321); (s322) to (s222); (s322) to (s221); (s322) to (s32); (s222) to (s221); (s222) to (s22);
\[lem:124-S22k\] For all $k\geq0$, we have ${\mathcal{G}}({S_{2,2,k}})= |{S_{2,2,k}}|\bmod 3$, i.e. ${\mathcal{G}}({S_{2,2,k}}) = \overline{201}$.
We proceed by induction on $k$. The cases $k\leq3$ are consequences of Lemma \[lem:124-paths\], \[lem:124-small-ones\] and \[lem:124-S333\]. When $k\geq4$, there are always exactly 5 moves available, leading to ${S_{2,2,k-1}}$, ${S_{2,2,k-2}}$, ${S_{2,2,k-4}}$, ${S_{1,2,k}}$, and $P_{k+3}$. We then conclude by induction and thanks to Lemma \[lem:general\], since all these graphs (see Lemma \[lem:124-paths\] and \[lem:124-S12k\]) are such that ${\mathcal{G}}(G) = |G| \bmod 3$.
\[lem:124-S122k\] For all $k\geq0$, we have ${\mathcal{G}}({S_{1,2,2,k}})= |{S_{1,2,2,k}}| \bmod 3$, i.e. ${\mathcal{G}}({S_{1,2,2,k}}) = \overline{012}$.
We proceed by induction on $k$. The cases $k\leq1$ are consequences of Lemma \[lem:124-small-ones\]. For $k=2$, 3 moves are available from ${S_{1,2,2,2}}$, leading to ${S_{1,1,2,2}}$, ${S_{1,2,2}}$, and ${S_{2,2,2}}$, and we conclude thanks to Lemma \[lem:124-small-ones\] and \[lem:124-S22k\]. For $k=3$, 5 moves are available from ${S_{1,2,2,3}}$, leading to ${S_{1,2,2,2}}$, ${S_{1,1,2,2}}$, and ${S_{1,1,2,3}}$, ${S_{1,2,3}}$, and ${S_{2,2,3}}$, and we conclude thanks to Lemma \[lem:124-small-ones\] and \[lem:124-S22k\]. When $k\geq4$, there are always exactly 6 moves available, leading to ${S_{1,2,2,k-1}}$, ${S_{1,2,2,k-2}}$, ${S_{1,2,2,k-4}}$, ${S_{1,1,2,k}}$, ${S_{1,2,k}}$, and ${S_{2,2,k}}$. We conclude by induction, and thanks to Lemma \[lem:124-S12k\], \[lem:124-S112k\], and \[lem:124-S22k\], since we have:
- If $k \equiv 0 \bmod 3$, then ${\mathcal{G}}({S_{1,2,2,k}})$ is computed as $\operatorname{mex}(2,1,2,2,1,2)=0$, and we are done.
- If $k \equiv 1 \bmod 3$, then ${\mathcal{G}}({S_{1,2,2,k}})$ is computed as $\operatorname{mex}(0,2,0,3,2,0)=1$, and we are done.
- If $k \equiv 2 \bmod 3$, then ${\mathcal{G}}({S_{1,2,2,k}})$ is computed as $\operatorname{mex}(1,0,1,1,0,1)=2$, and we are done.
\[lem:124-S1111k\] For all $k\geq0$, we have ${\mathcal{G}}({S_{1,1,1,1,k}})= |{S_{1,1,1,1,k}}|+1 \bmod 3$, i.e. ${\mathcal{G}}({S_{1,1,1,1,k}}) = \overline{012}$.
We proceed by induction on $k$. The case $k=0$ is a consequence of Lemma \[lem:124-small-ones\]. For $k=1$, 1 move is available from ${S_{1,1,1,1,1}}$, leading to ${S_{1,1,1,1}}$, and we conclude again thanks to Lemma \[lem:124-small-ones\]. For $k=2$, 3 moves are available from ${S_{1,1,1,1,2}}$, leading to ${S_{1,1,1,1,1}}$, ${S_{1,1,1,1}}$, and ${S_{1,1,1,2}}$, and we conclude thanks to Lemma \[lem:124-small-ones\] (and thanks to the case $k=0$ above). The case $k=3$ is not harder: 3 moves lead to ${S_{1,1,1,1,2}}$, ${S_{1,1,1,1,1}}$, and ${S_{1,1,1,3}}$, and we can conclude thanks to Lemma \[lem:124-small-ones\] and thanks to the cases $k\leq2$ above. When $k\geq4$, there are always exactly 4 moves available, leading to ${S_{1,1,1,1,k-1}}$, ${S_{1,1,1,1,k-2}}$, ${S_{1,1,1,1,k-4}}$, and ${S_{1,1,1,k}}$. We conclude by induction, and thanks to Lemma \[lem:124-S111k\], since we have:
- If $k \equiv 0 \bmod 3$, then ${\mathcal{G}}({S_{1,1,1,1,k}})$ is computed as $\operatorname{mex}(2,1,2,1)=0$, and we are done.
- If $k \equiv 1 \bmod 3$, then ${\mathcal{G}}({S_{1,1,1,1,k}})$ is computed as $\operatorname{mex}(0,2,0,0)=1$, and we are done.
- If $k \equiv 2 \bmod 3$, then ${\mathcal{G}}({S_{1,1,1,1,k}})$ is computed as $\operatorname{mex}(1,0,1,3)=2$, and we are done.
Note that Lemma \[lem:general\] is not invoked in the two previous proofs. This is due to the cases ${S_{1,1,2,k}}$ and ${S_{1,1,1,k}}$, which appear in the general case $k\geq4$, but do not satisfy the properties ${\mathcal{G}}(G) = |G| \bmod 3$ and ${\mathcal{G}}(G) = |G|+1 \bmod 3$, respectively.
We are now ready to prove Theorem \[thm:124\].
The proof is similar to that of Theorem \[thm:123\]. We denote A (for Alice) the first player, and B (for Bob) the second player, and we will show by induction that for any subdivided star $S={S_{\ell_1,\ldots,\ell_t,n}}$ with $\ell_1,\ldots,\ell_t,n \geq 0$, then $S + S'$ is always a ${\mathcal{P}}$-position, with $S'$ defined as ${S_{\ell_1,\ldots,\ell_t,n+3}}$.
Let $b(S)$ be the number of nontrivial branches of $S$, that is to say $b(S) = |\{ i \mid \ell_i >0\}|$ if $n=0$ and $b(S) = |\{ i \mid \ell_i >0\}| + 1$ otherwise.
If $b(S)=0$ then we are trivially done thanks to Lemma \[lem:124-paths\] since $S=P_1$ and $S'=P_4$.
If $b(S)=1$, then we conclude again by Lemma \[lem:124-paths\] since $S$ and $S'$ are both paths.
If $b(S)\geq 2$, then several cases follow.
If A plays in $S'$ so as to leave $S + {S_{\ell_1,\ldots,\ell_t,n+3-k}}$, with $k\in\{1,2,4\}$, then B replies by taking $3-k$ vertices in $S'$ if $k\in\{1,2\}$, so as to leave $S + {S_{\ell_1,\ldots,\ell_t,n}} = S + S$, which is a ${\mathcal{P}}$-position, and if $k=4$ then B replies by taking 1 vertex in $S$, so as to leave ${S_{\ell_1,\ldots,\ell_t,n-1}} + {S_{\ell_1,\ldots,\ell_t,n-1}}$, which is a ${\mathcal{P}}$-position.
If A plays elsewhere in $S'$, then B can always copy A’s move on $S$. For instance, if A removes $k\geq \ell_1$ vertices in $S'$, so as to leave $S \cup {S_{\ell_1-k,\ldots,\ell_t,n+3}}$, then B removes $k$ vertices in $S$ so as to leave ${S_{\ell_1-k,\ldots,\ell_t,n}} \cup {S_{\ell_1-k,\ldots,\ell_t,n+3}}$. When B replies by mimicking A’s move on $S$, B manages to leave $T + T'$ to A, with $|T| < |S|$, and we thus can conclude by induction that this is a ${\mathcal{P}}$-position.
If A plays in $S$, then we have to be careful. If A’s move does not involve taking the center of $S$, then we are done, since B is again always able to copy A’s move in $S'$, so as to leave $T + T'$ to A, with $|T| < |S|$, and we again conclude by induction that this is a ${\mathcal{P}}$-position.
The tricky cases are the ones where A takes the center of $S$. Indeed, in this case, B can not always simply copy A’s move in $S'$, and we have to carry a more detailed analysis. For instance, if $S=\star{1,1,1,2}$ and $S'={S_{1,1,1,2,3}}$, and if A removes 4 vertices so as to leave $P_2 + {S_{1,1,1,2,3}}$, then B can not mimic A’s move in $S'$ since it would disconnect ${S_{1,1,1,2,3}}$, so as to leave $P_2 + P_2 + P_3$. In other words, A’s move copied in $S'$ is not always a legal move.
Let us assume now that A’s move in $S$ can not be copied by player B in $S'$. In this case, A was able to remove the centre of $S$, and this implies that $S$ has at most 4 nontrivial branches, and that $S'$ has at most one more nontrivial branch than $S$. Note also that A removed strictly more than 1 vertex to $S$, and exactly $b(S)-1$ branches of $S$ (if A empties $S$ then B can trivially copy A’s move on $S'$ so as to leave $P_3$). Three cases and several subcases follow. We will denote by $k$ the number of vertices removed by A.
**Case (i): $b(S)=2$.** Without loss of generality let us assume $S={S_{\ell_1,\ell_2}}$ with $1\leq\ell_1 \leq \ell_2$. As in the case of Theorem \[thm:123\], if there is a long enough branch in $S$ (w.r.t the number of vertices removed by A), then we will proceed by a mirror argument, since $S$ is actually a path $P_{\ell_1 + \ell_2 +1}$. More precisely, if $k\leq\ell_2$, then we may consider A actually played so as to leave ${S_{\ell_1,\ell_2-k}} + S'$, and in this case player B can copy A’s move on $S'$, and we are done.
Let us assume now that $k>\ell_2$. If $k=2$, then $S = {S_{1,1}}$, and we may have either $S'={S_{1,4}}=P_6$ or $S'={S_{1,1,3}}$. If $S'={S_{1,4}}=P_6$, then A may remove 2 vertices from $P_6$ so as to leave $P_1 + P_4$, which is a ${\mathcal{P}}$-position. If $S'={S_{1,1,3}}$, then A may remove 2 vertices from ${S_{1,1,3}}$ so as to leave $P_1 + {S_{1,1,1}}$, which is a ${\mathcal{P}}$-position since we know from Lemma \[lem:124-small-ones\] that ${\mathcal{G}}({S_{1,1,1}}) = 1 = {\mathcal{G}}(P_1)$.
If $k=4$, then there are several subcases.
- If $S={S_{1,2}}$, then A leaves $\emptyset +S'=S'$ and we may have either $S'=P_7$ or $S'={S_{1,2,3}}$. If $S'=P_7$, then A may remove 1 vertex from $P_7$ so as to leave $P_6$, which is a ${\mathcal{P}}$-position. If $S'={S_{1,2,3}}$, then A may remove 1 vertex from ${S_{1,2,3}}$ so as to leave $P_6$, which is a ${\mathcal{P}}$-position.
- If $S={S_{1,3}}$, then A leaves $P_1 +S'$ and we may have either $S'=P_8$ or $S'={S_{1,3,3}}$. If $S'=P_8$, then A may remove 1 vertex from $P_8$ so as to leave $P_1 + P_7$, which is a ${\mathcal{P}}$-position. If $S'={S_{1,3,3}}$, then A may remove 1 vertex from ${S_{1,3,3}}$ so as to leave $P_1 + {S_{1,2,3}}$, which is a ${\mathcal{P}}$-position thanks to Lemma \[lem:124-small-ones\].
- If $S={S_{2,2}}$, then A leaves $P_1 +S'$ and we may have either $S'=P_8$ or $S'={S_{2,2,3}}$. If $S'=P_8$, then A may remove 1 vertex from $P_8$ so as to leave $P_1 + P_7$, which is a ${\mathcal{P}}$-position. If $S'={S_{2,2,3}}$, then A may remove 1 vertex from ${S_{2,2,3}}$ so as to leave $P_1 + {S_{2,2,2}}$, which is a ${\mathcal{P}}$-position thanks to Lemma \[lem:124-S333\].
- If $S={S_{2,3}}$, then A leaves $P_2 +S'$ and we may have either $S'=P_9$ or $S'={S_{2,3,3}}$. If $S'=P_9$, then A may remove 1 vertex from $P_9$ so as to leave $P_2 + P_8$, which is a ${\mathcal{P}}$-position. If $S'={S_{2,3,3}}$, then A may remove 1 vertex from ${S_{2,3,3}}$ so as to leave $P_2 + {S_{2,2,3}}$, which is a ${\mathcal{P}}$-position thanks to Lemma \[lem:124-S333\].
- If $S={S_{3,3}}$, then A leaves $P_3 +S'$ and we may have either $S'=P_{10}$ or $S'={S_{3,3,3}}$. If $S'=P_{10}$, then A may remove 1 vertex from $P_{10}$ so as to leave $P_3 + P_9$, which is a ${\mathcal{P}}$-position. If $S'={S_{3,3,3}}$, then A may remove 1 vertex from ${S_{3,3,3}}$ so as to leave $P_3 + {S_{2,3,3}}$, which is a ${\mathcal{P}}$-position thanks to Lemma \[lem:124-S333\].
**Case (ii): $b(S)=3$.** Let us denote $S={S_{\ell_1,\ell_2,\ell_3}}$. In this case, we must have $k=4$. Two subcases follow.
- If $S = {S_{1,1,\ell_3}}$, then A leaves $P_{\ell_3-1} + S'$, and we may have $S'={S_{1,1,3,\ell_3}}$, $S'={S_{1,1,\ell_3+3}}$, or $S'={S_{1,4,\ell_3}}$.
- If $S'={S_{1,1,3,\ell_3}}$, then B’s answer depends on $\ell_3 \bmod 3$. If $\ell_3 \equiv 0\bmod 3$, B may reply by taking 1 vertex in $S'$, so as to leave $P_{\ell_3-1} + {S_{1,1,2,\ell_3}}$. Indeed, from Lemma \[lem:124-S112k\], we know that ${\mathcal{G}}({S_{1,1,2,\ell_3}})=2={\mathcal{G}}(P_{\ell_3-1})$, thus $P_{\ell_3-1} + {S_{1,1,2,\ell_3}}$ is a ${\mathcal{P}}$-position. If $\ell_3 \equiv 1\bmod 3$, B may reply by taking 2 vertices in $S'$, so as to leave $P_{\ell_3-1} + {S_{1,1,1,\ell_3}}$. Indeed, from Lemma \[lem:124-S111k\], we know that ${\mathcal{G}}({S_{1,1,1,\ell_3}})=0={\mathcal{G}}(P_{\ell_3-1})$, thus $P_{\ell_3-1} + {S_{1,1,1,\ell_3}}$ is a ${\mathcal{P}}$-position. If $\ell_3 \equiv 2\bmod 3$, B may reply by taking 1 vertex in $S'$, so as to leave $P_{\ell_3-1} + {S_{1,1,2,\ell_3}}$. Indeed, from Lemma \[lem:124-S112k\], we know that ${\mathcal{G}}({S_{1,1,2,\ell_3}})=1={\mathcal{G}}(P_{\ell_3-1})$, thus $P_{\ell_3-1} + {S_{1,1,2,\ell_3}}$ is a ${\mathcal{P}}$-position.
- If $S'={S_{1,1,\ell_3+3}}$, then B removes 4 vertices from $S'$, so aas to leave $P_{\ell_3-1}+P_{\ell_3-1}$, which is a ${\mathcal{P}}$-position.
- If $S'={S_{1,4,\ell_3}}$, then B removes 4 vertices from $S'$, so aas to leave $P_{\ell_3-1}+P_{\ell_3+2}$, which is a ${\mathcal{P}}$-position.
- If $S = {S_{1,2,\ell_3}}$, then A leaves $P_{\ell_3} + S'$, and we may have $S'={S_{1,2,3,\ell_3}}$, $S'={S_{1,2,\ell_3+3}}$, $S'={S_{1,5,\ell_3}}$, or $S'={S_{2,4,\ell_3}}$. If $S'={S_{1,2,3,\ell_3}}$, B can reply by taking 1 vertex in $S'$, so as to leave $P_{\ell_3} + {S_{1,2,2,\ell_3}}$, which is a ${\mathcal{P}}$-position. Indeed, from Lemma \[lem:124-S122k\], we know that ${\mathcal{G}}({S_{1,1,2,\ell_3}})={\mathcal{G}}(P_{\ell_3})= \ell_3 \bmod 3$. If $S'={S_{1,2,\ell_3+3}}$, then B may reply by removing 1 vertex from $S'$, so as to leave $P_{\ell_3} + P_{\ell_3+3}$, which is a ${\mathcal{P}}$-position. If $S'={S_{1,5,\ell_3}}$, then B may reply by removing 1 vertex from $S'$, so as to leave $P_{\ell_3} + P_{\ell_3+6}$, which is a ${\mathcal{P}}$-position. Finally, if $S'={S_{2,4,\ell_3}}$, then B may reply by removing 4 vertices from $S'$, so as to leave $P_{\ell_3} + P_{\ell_3+3}$, which is a ${\mathcal{P}}$-position.
**Case (iii): $b(S)=4$.** In this case, we must have $k=4$ and $S={S_{1,1,1,\ell_4}}$, and A leaves $P_{\ell_4} + S'$, with $S'={S_{1,1,1,3,\ell_4}}$, $S'={S_{1,1,1,\ell_4+3}}$, or $S'={S_{1,1,4,\ell_4}}$. If $S={S_{1,1,1,3,\ell_4}}$, then a good answer for B is then to remove 2 vertices from $S'$, so as to leave $P_{\ell_4} + {S_{1,1,1,1,\ell_4}}$, which is a ${\mathcal{P}}$-position, since we know from Lemma \[lem:124-S1111k\] that ${\mathcal{G}}({S_{1,1,1,1,\ell_4}})={\mathcal{G}}(P_{\ell_4})= \ell_4 \bmod 3$. If $S={S_{1,1,1,\ell_4+3}}$, then B may remove 4 vertices from $S'$, so as to leave $P_{\ell_4} + P_{\ell_4+3}$, which is a ${\mathcal{P}}$-position. If $S={S_{1,1,4,\ell_4}}$, then B may remove 4 vertices from $S'$, so as to leave $P_{\ell_4} + {S_{1,1,\ell_4}}$, which is a ${\mathcal{P}}$-position, since we know from Lemma \[lem:124-S11k\] that ${\mathcal{G}}({S_{1,1,\ell_4}})=P_{\ell_4}=\ell_4 \bmod 3$.
Conclusion
==========
In this paper, we derived results for connected subtraction games in graphs. Our main results deal with graphs to which we append paths, for which we were able to derive a very general result showing that there is always ultimate periodicity, in the sense that, for any finite set $L$, there exists two integers $k_0$ and $T$ such that $${\mathcal{G}}_L({G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k+T}}) = {\mathcal{G}}_L({G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k}}),$$ whenever $k\geq k_0$ (Theorem \[thm:general\]). Is it true that the period $T$ is the same as the period of the game played on paths?
We then managed to prove pure periodicity results in some particular cases. We showed that, for some classes of graphs and sets $L$, equation ${\mathcal{G}}_L({G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k+T}}) = {\mathcal{G}}_L({G\tikz[baseline=-4]{\draw (0,0) node[below] {\footnotesize u};\draw (0,0) node[circle,fill=black,minimum size=0,inner sep=1]{} -- (0.3,0) node[circle,fill=black, minimum size=0, inner sep=1] {};} P_{k}})$ holds starting from $n_0=0$. This is the case, for instance, for simple stars when playing [$CSG(\{1,\ldots,N\})$]{} (Proposition \[prop:Stk\_general\]), for subdivided stars when playing [$CSG(\{1,2,3\})$]{} or [$CSG(\{1,2,4\})$]{} (Theorems \[thm:123\] and \[thm:124\]).
However, even in graphs as structurally simple as subdivided stars, pure periodicity is not the general setting, even if the game is purely periodic when played in paths. We proved, for instance, in Theorem \[thm:S1lk\] that we only have ultimate periodicity in subdivided stars of type ${S_{1,k,l}}$ for the game [$CSG(\{1,\ldots,N\})$]{}. For instance, as a corollary of Theorem \[thm:S1lk\], we have ${\mathcal{G}}({S_{1,2,8}})=10$ whereas ${\mathcal{G}}({S_{1,2+9k,8}})=6$ for all $k\geq1$, for the game [$CSG(\{1,\ldots,8\})$]{}. Note that this game is purely periodic when played in paths, and that subdivided stars of the form ${S_{1,k,l}}$ can be seen as paths to which we attached a vertex. Hence there is a complexity gap in the computation of the Grundy values when considering subdivided stars instead of paths. This complexity gap between paths and stars is also illustrated by Observation \[obs:plus\_M\], which states that, for subdivided stars – and even for stars – the Grundy sequence of [$CSG(I_N \cup \{M\})$]{} may be different from the one of [$CSG(I_N)$]{}, for $M \not\equiv 0 \bmod{N+1}$.
[apalike]{}
R. Adams, J. Dixon, J. Elder, J.Peabody, O. Vega and K. Will, [*Combinatorial Analysis of a Subtraction Game on Graphs*]{}, Retrieved from arXiv:1507.05673 (2015).
M. Albert, R. Nowakowski and D. Wolfe, [*Lessons in play: an introduction to combinatorial game theory*]{}, CRC Press (2007).
L. Beaudou, P. Coupechoux, A. Dailly, S. Gravier, J. Moncel, A. Parreau and É. Sopena, [*Octal Games on Graphs: The game 0.33 on subdivided stars and bistars*]{}, Theoret. Comput. Sci. (to appear).
E. R. Berkelamp, J. H. Conway and R. K. Guy, [*Winning ways for your mathematical plays*]{}, 2nd ed., 4 volumes, A K Peters (2001–2004).
R. Fleischer and G. Trippen, [*Kayles on the way to the stars*]{}, International Conference on Computers and Games, Springer Berlin Heidelberg (2004).
M. Fukuyama, [*A Nim game played on graphs*]{}, Theoret. Comput. Sci. **304(1-3)** (2003), 387–399.
M. Fukuyama, [*A Nim game played on graphs II*]{}, Theoret. Comput. Sci. **304(1-3)** (2003), 401–419.
R. K. Guy, [*Unsolved problems in Combinatorial Games*]{}, Games of No Chance [**29**]{}, 1996.
M. Huggan and B. Stevens, [*Polynomial Time Graph Families for Arc-Kayles*]{}, Integers, **16** (2016).
T. J. Schaeffer, [*On the complexity of some two-person perfect-information games*]{}, J. Comput. System Sci. **16(2)** (1978), 185–225.
A. N. Siegel, [*Combinatorial game theory*]{}, American Mathematical Society **146** (2013).
R. Sprague, [*Uber mathematische Kampfspiele*]{}, Tohoku Mathematical Journal, First Series **41** (1935), 438–444.
[^1]: Corresponding author
[^2]: Univ Lyon, Université Lyon 1, LIRIS UMR CNRS 5205, F-69621, Lyon, France.
[^3]: LAAS-CNRS, Université de Toulouse, CNRS, Université Toulouse 1 Capitole - IUT Rodez, Toulouse, France.
[^4]: Fédération de Recherche Maths à Modeler, Institut Fourier, 100 rue des Maths, BP 74, 38402 Saint-Martin d’Hères Cedex, France.
[^5]: This work has been supported by the ANR-14-CE25-0006 project of the French National Research Agency.
[^6]: Also called Nim-like games, or octal games.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'H. Ozawa'
- 'N. Grosso'
- 'T. Montmerle'
title: |
The X-ray emission from Young Stellar Objects\
in the $\rho$ Ophiuchi cloud core as seen by XMM-Newton
---
Introduction
============
Young low-mass stars are known to be ubiquitous X-ray emitters (Montmerle et al. 1993; Feigelson and Montmerle 1999). Their X-ray luminosities are $\sim 10^4-10^5$ times higher than that of the present-day Sun. YSOs are classified from IR and sub-millimeter spectral energy distributions (André & Montmerle 1994) into four classes: Class 0 and Class I protostars, Class II (classical T Tauri stars) and Class III (weak-lined T Tauri stars) sources, interpreted as a time evolution sequence from Class 0 (age $\sim 10^4$ yrs) to Class III (age $\ga 10^6$ yrs). It is generally accepted that enhanced magnetic activity of YSOs produces X-rays emitted by high temperature plasmas as a result of heating by magnetic reconnection events, which is called the solar paradigm because of the analogy with the X-ray emission of the Sun (e.g. as seen with [*Yohkoh*]{}).
The $\rho$ Ophiuchi cloud is a well-known nearby active star-forming region at $d \sim 140$ pc, which has been observed over twenty years by practically all X-ray observatories: [*Einstein*]{} (Montmerle et al. 1983), [*ROSAT*]{} (Casanova 1994; Casanova et al. 1995; Grosso et al. 1997; Grosso et al. 2000; Grosso 2001), [*ASCA*]{} (Koyama et al. 1994; Kamata et al. 1997; Tsuboi et al. 2000), and [*Chandra*]{} (Imanishi et al. 2001a; Imanishi et al. 2001b; Imanishi et al. 2003). To investigate the YSO X-ray emission in the dense CO core F of the $\rho$ Ophiuchi cloud (Loren et al. 1990), we took with [*XMM-Newton*]{} a 33 ks exposure of this region, as part of the EPIC Guarantee Time program, and we report here the results of this observation. In Sect. \[sec\_obana\], we present observation and data analysis, including source detection, extraction of X-ray light curves and X-ray spectra from detected sources. We compare our [*XMM-Newton*]{} results with previously obtained [*Chandra*]{} results in this region in Sect. \[sec\_chan\]. We discuss the IR properties of detected sources in Sect. \[sec\_ir\]. We present the X-ray detections of young bona fide brown dwarfs in Sect. \[sec\_bd\], and the X-ray properties of Class I, II and III sources in Sect. \[sec\_x\]. Finally, we summarize all results in Sect. \[sec\_sum\].
Observation and data analysis {#sec_obana}
=============================
The main (densest) core F of the $\rho$ Ophiuchi cloud was observed with [*XMM-Newton*]{} for 33 ks on February 19, 2001 pointing at $\alpha_{\rm{J2000}}$ $\rm{= 16^{h}27^{m}26.0^{s}}$ and $\delta_{\rm{J2000}}$ $\rm{=-24\degr40\arcmin48.0\arcsec}$. [*XMM-Newton*]{} is the X-ray astronomical observatory of the European Space Agency which was launched on December 10, 1999. [*XMM-Newton*]{} has three co-aligned X-ray telescopes with 6$\arcsec$-FWHM angular resolution at 1.5 keV and with large total effective area up to 5000 cm$^{-2}$ at 1 keV (Jansen et al. 2001). The [*European Photon Imaging Camera*]{} (EPIC), i.e. two MOS (namely, MOS1 and MOS2) and one PN X-ray CCD arrays, are placed on each focal plane of the three X-ray telescopes. EPIC provides imaging capability over a wide field-of-view of 30$\arcmin$ diameter, and spectroscopic capability with moderate spectral resolution ($\sim$ 60 eV at 1 keV) in the 0.2$-$10.0 keV band. In this observation, EPIC was operated with the medium optical blocking filter and in the Full-Frame mode.
We reduced the data using the [*XMM-Newton*]{} [*Science Analysis Software*]{} (SAS version 5.4). The photon event lists were obtained by running [emchain]{} and [epchain]{}. We rejected bad pixels, and selected single, double, triple, and quadruple pixel events for MOS, and single and double pixel events for PN using [evselect]{}. We excluded the time intervals where the X-ray counts from whole field of view are extremely high, above 4.4 counts s$^{-1}$ for MOS1 and MOS2, and 20 counts s$^{-1}$ for PN, indicating a high level of irradiation by solar protons, and hence a high background level on EPIC. The sum of the final good-time-intervals are 31 ks for MOS1, 32 ks for MOS2, and 29 ks for PN.
Source detection
----------------
Fig. \[fig\_tri\] shows the resulting EPIC image where the MOS1, MOS2, and PN images were corrected for vignetting and co-added. Red, green, and blue colors code the X-ray events in the 0.5$-$1.0 keV, 1.0$-$2.4 keV, 2.4$-$8.0 keV energy bands, respectively. The image is smoothed to enhance the contrast.
We perform the source detection on the individual detector X-ray images, using the standard source detection method of SAS. This method includes 4 steps, heritage of the [*Einstein*]{} and [*ROSAT*]{} source detection algorithms.
1\) The source candidates are obtained by the sliding cell method in [eboxdetect]{}. A small size cell surrounded by a larger size cell is moved over the X-ray image. The background counts are estimated from the area surrounding the small size cell. If the counts in the small size cell are significantly larger than in the background, the cell is considered to have excess counts. After sliding the cell over the whole region, we obtain “islands” of excess counts, which are considered as source candidates.
2\) The local background image is created by [esplinemap]{}. The source candidates obtained in step 1 are subtracted from the X-ray image, and a spline fit is applied to this “swiss-cheese" image to obtain the background image.
3\) The sliding cell method is again performed with [eboxdetect]{} to obtain more reliable source candidates. This time, only the small size cell is moved over the X-ray image. The local background counts are estimated locally using the background image obtained in the previous step.
4\) A two-dimensional fit is performed on the X-ray source candidates in [emldetect]{} using a model where the X-ray counts come from a point source, contamination by nearby point sources, and the background. The point spread function is used as the model function of the point source. The local background created at step 2 is used as the background model. Through this task, the significance of the source candidates is checked precisely and source counts are obtained accurately.
This method can deal with images in different energy bands and different detectors simultaneously. We use images in the 0.3$-$2.0 keV and 2.0$-$8.0 keV energy bands, and we deal with the MOS1, MOS2, and PN X-ray images simultaneously to obtain good statistics. For the areas covered by only two detectors or one detector, we can analyze only the images of the corresponding detector set.
As a result of this protocol, we detect in total 87 X-ray sources with a likelihood threshold above 12, corresponding to the significance of $\sim$ 4.4 $\sigma$ for Gaussian statistics[^1], within the 30$^\prime$ diameter field-of-view of the EPIC detectors. The positions of the detected 87 X-ray sources are indicated in the lower panel of Fig. \[fig\_tri\]. Table 1 gives the coordinates, count rates, and the hardness ratios of the detected sources. For convenience, these sources are designated here ROXN-$n$ (for “Rho Oph X-ray sources, Newton, number $n$")[^2]. The hardness ratios are calculated after the formula $HR = (H - S)/(H + S)$, where $S$ and $H$ are counts in the 0.3$-$2.0 keV, and in the 2.0$-$8.0 keV energy band, respectively. [^3]
The detected sources are cross-correlated with: X-ray source catalogs from [*Einstein*]{} (Montmerle et al. 1983), [*ROSAT*]{} (Casanova et al. 1995; Grosso et al. 2000), [*ASCA*]{} (Kamata el al. 1997), and [*Chandra*]{} (Imanishi et al. 2001a); near-IR source catalogs (e.g. Barsony et al. 1997; 2MASS = Two Micron All-Sky Survey); ISOCAM mid-IR catalog (Bontemps et al. 2001); and an optical catalog (Monet 1996). Table 2 shows the resulting cross-correlation: 62 ROXN sources are identified with previously known X-ray sources, hence we find 25 new X-ray sources from our observation, and 68 ROXN sources have 2MASS near-IR counterparts. Table 2 also lists the IR classifications from Bontemps et al. (2001), or André & Montmerle (1994).
As a result of these cross-identifications, we find 7 Class I sources, 26 Class II sources, and 17 Class III sources. We present 15 previously unknown Class III candidates in section \[sec\_ir\].
\
------ ----------------------- ----------------------- ------------- ------------- --------------- ------------- --------------- ------------- ---------------
ROXN $\alpha_{\rm{J2000}}$ $\delta_{\rm{J2000}}$ err
($\arcsec$) (cts/ks) $H.R.$ (cts/ks) $H.R.$ (cts/ks) $H.R.$
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
1 16 26 23.7 -24 43 16.2 0.1 716.6(16.6) 0.22(0.02)
2 16 26 27.5 -24 41 51.3 1.0 6.7( 1.1) 0.86(0.12) 6.8( 1.1) 1.00 11.7( 1.6) 0.77(0.11)
3 16 26 35.1 -24 42 32.6 1.6 0.8( 0.5) 0.72(0.47) 1.3( 0.6) 0.73(0.38) 4.8( 1.2) 0.51(0.21)
4 16 26 44.2 -24 34 49.4 1.2 2.1( 0.5) 0.92(0.16) 3.2( 0.7) 0.66(0.20) 3.3( 0.8) 0.86(0.20)
5 16 26 44.3 -24 43 13.9 0.3 18.4( 1.3) $-$0.62(0.06) 16.2( 1.2) $-$0.71(0.06) 45.7( 2.6) $-$0.69(0.04)
6 16 26 44.5 -24 47 13.4 1.2 2.1( 0.5) $-$0.87(0.22) 0.9( 0.4) $-$1.00 3.5( 0.7) $-$1.00
7 16 26 45.3 -24 52 06.8 1.6 7.8( 1.5) $-$0.04(0.19)
8 16 26 47.1 -24 44 28.9 1.5 1.9( 0.5) $-$0.66(0.27) 1.6( 1.1) $-$1.00 4.7( 0.9) $-$0.75(0.19)
9 16 26 48.1 -24 42 06.6 1.4 1.1( 0.4) $-$0.15(0.37) 0.7( 0.3) $-$0.16(0.49) 2.2( 0.6) $-$1.00
10 16 26 51.3 -24 32 42.7 1.0 1.2( 0.3) $-$1.00 1.1( 0.4) $-$1.00 5.4( 0.8) $-$1.00
11 16 26 58.6 -24 45 35.8 0.2 40.2( 1.7) 0.03(0.04) 31.7( 1.6) 0.01(0.05) 111.4( 2.7) 0.02(0.02)
12 16 26 59.2 -24 34 58.0 0.9 2.1( 0.4) 1.00 3.3( 0.6) 1.00 4.9( 0.8) 1.00
13 16 27 03.8 -24 45 56.7 1.8 0.5( 0.3) $-$0.11(0.59) 0.7( 0.3) 0.87(0.30) 1.8( 0.6) 0.49(0.32)
14 16 27 04.6 -24 42 59.9 0.3 13.3( 0.9) $-$0.32(0.06) 13.0( 0.8) $-$0.40(0.06) 35.3( 1.5) $-$0.41(0.04)
15 16 27 04.6 -24 42 14.2 0.4 4.2( 0.5) $-$0.47(0.11) 6.3( 0.6) $-$0.26(0.09) 14.3( 1.0) $-$0.41(0.07)
16 16 27 05.2 -24 41 13.2 1.8 0.7( 0.2) 1.00 0.7( 0.2) 1.00 1.9( 0.5) 1.00
17 16 27 06.5 -24 41 49.7 1.3 0.4( 0.2) $-$0.44(0.55) 1.1( 0.3) $-$0.32(0.30) 0.9( 0.4) $-$0.51(0.49)
18 16 27 06.8 -24 38 15.5 0.3 11.0( 0.8) 0.79(0.04) 10.1( 0.7) 0.72(0.05) 30.4( 1.4) 0.80(0.03)
19 16 27 07.6 -24 31 52.4 1.0 0.6( 0.3) 0.20(0.57) 0.4( 0.3) 1.00 2.9( 0.7) $-$0.79(0.24)
20 16 27 09.0 -24 34 09.3 1.3 2.7( 0.6) $-$0.27(0.20) 2.2( 0.9) $-$0.64(0.38) 2.2( 0.6) 0.33(0.24)
21 16 27 09.4 -24 43 21.2 1.4 1.3( 0.3) $-$0.70(0.24) 0.6( 0.3) $-$0.18(0.46) 2.1( 0.5) 0.22(0.25)
22 16 27 09.4 -24 37 19.4 0.2 23.9( 1.1) 0.90(0.02) 23.6( 1.1) 0.83(0.03) 54.4( 1.8) 0.87(0.02)
23 16 27 10.9 -24 51 37.8 1.6 0.9( 0.4) 0.87(0.32) 1.6( 0.5) 0.19(0.31) 1.9( 0.8) $-$0.45(0.47)
24 16 27 11.0 -24 52 08.0 1.4 1.6( 0.5) 0.55(0.31) 1.0( 0.4) 0.22(0.41) 1.0( 0.8) 0.77(0.72)
25 16 27 11.7 -24 38 32.8 0.5 3.0( 0.4) 1.00 3.3( 0.5) 0.91(0.09) 8.2( 0.8) 0.92(0.07)
26 16 27 12.1 -24 34 48.8 1.2 1.2( 0.3) 1.00 1.5( 0.3) 1.00 2.1( 0.6) 0.85(0.20)
27 16 27 14.4 -24 51 32.5 0.6 4.5( 1.2) $-$0.84(0.18) 6.0( 1.1) $-$0.82(0.12) 22.0( 2.3) $-$0.83(0.08)
28 16 27 15.1 -24 51 38.7 0.1 21.9( 1.5) $-$0.78(0.04) 18.5( 1.3) $-$0.82(0.05) 63.8( 2.8) $-$0.87(0.03)
29 16 27 15.5 -24 26 40.0 0.9 10.4( 1.3) 0.40(0.11) 12.4( 1.4) 0.46(0.10)
30 16 27 15.9 -24 38 43.6 0.2 17.4( 0.9) 0.40(0.05) 19.5( 1.0) 0.39(0.04) 44.7( 1.6) 0.31(0.03)
31 16 27 16.4 -24 31 14.2 0.7 7.2( 0.8) 1.00 6.0( 0.8) 0.94(0.07) 10.5( 1.1) 0.96(0.05)
32 16 27 18.2 -24 28 52.7 0.2 29.7( 2.2) 0.91(0.03) 34.3( 2.2) 0.83(0.04) 67.9( 3.2) 0.92(0.03)
33 16 27 18.4 -24 39 15.4 0.5 3.2( 0.4) 0.84(0.08) 4.4( 0.5) 0.89(0.06) 13.7( 1.7) 1.00
34 16 27 18.6 -24 29 05.9 0.6 8.8( 1.1) 0.20(0.12) 11.2( 1.3) 0.40(0.10) 19.2( 1.9) 0.48(0.09)
35 16 27 19.4 -24 28 45.4 0.8 7.3( 1.2) 0.88(0.09) 7.0( 1.2) 0.88(0.11) 18.5( 2.1) 0.53(0.09)
36 16 27 19.5 -24 41 40.9 0.1 120.7( 2.2) $-$0.75(0.01) 112.6( 2.1) $-$0.77(0.01) 406.6( 4.4) $-$0.81(0.01)
37 16 27 21.4 -24 41 42.3 0.5 3.4( 0.4) $-$0.56(0.11) 2.4( 0.4) $-$0.84(0.14) 11.6( 0.9) $-$0.59(0.07)
38 16 27 21.7 -24 29 53.2 0.9 4.2( 0.7) 0.83(0.12) 3.0( 0.6) 1.00 8.8( 1.2) 1.00
39 16 27 21.8 -24 43 35.8 0.3 10.6( 0.7) 0.57(0.05) 10.5( 0.7) 0.63(0.05) 27.5( 1.3) 0.67(0.04)
40 16 27 22.8 -24 48 09.0 1.3 0.9( 0.3) 0.15(0.35) 1.2( 0.3) 0.03(0.27) 2.5( 0.6) $-$0.89(0.20)
41 16 27 24.6 -24 29 33.9 1.1 1.2( 0.5) $-$0.12(0.40) 3.4( 0.7) 0.59(0.15) 5.7( 1.1) 0.56(0.17)
42 16 27 26.4 -24 39 23.7 1.0 1.3( 0.4) 0.67(0.21) 1.9( 0.4) 1.00 6.0( 0.8) 0.65(0.10)
43 16 27 26.9 -24 40 50.4 0.2 20.5( 0.9) 0.78(0.03) 22.0( 0.9) 0.73(0.03) 53.8( 1.6) 0.73(0.02)
44 16 27 27.0 -24 32 18.0 0.9 1.6( 0.4) 0.34(0.24) 3.0( 0.5) 0.25(0.16) 6.3( 0.9) 0.62(0.12)
45 16 27 27.4 -24 31 17.2 0.4 10.0( 0.9) 0.19(0.09) 12.5( 1.0) $-$0.03(0.08) 29.8( 1.7) 0.01(0.06)
46 16 27 28.0 -24 39 33.8 0.2 28.9( 1.1) 0.85(0.02) 30.4( 1.1) 0.81(0.02) 75.5( 1.9) 0.80(0.02)
47 16 27 28.7 -24 54 33.2 1.1 2.8( 0.7) $-$0.54(0.23) 3.5( 0.7) $-$1.00 8.4( 1.5) $-$0.76(0.19)
48 16 27 28.7 -24 27 21.4 2.4 2.0( 0.6) 0.64(0.30) 2.5( 0.8) $-$0.30(0.33)
49 16 27 29.7 -24 33 35.0 1.3 0.5( 0.2) 1.00 1.2( 0.3) 0.38(0.26) 1.9( 0.5) 0.71(0.24)
50 16 27 30.2 -24 27 43.4 0.9 7.3( 0.9) 0.44(0.11) 10.9( 1.2) 0.51(0.09)
51 16 27 30.6 -24 52 08.4 0.5 7.4( 0.9) 0.18(0.12) 7.5( 0.8) 0.22(0.11) 23.6( 1.8) 0.30(0.07)
52 16 27 30.9 -24 47 27.7 0.3 13.4( 0.9) $-$0.10(0.07) 13.1( 0.8) $-$0.11(0.07) 35.8( 1.9) $-$0.10(0.05)
53 16 27 31.0 -24 34 03.9 0.9 1.7( 0.3) $-$0.10(0.19) 1.2( 0.4) $-$0.34(0.33) 3.9( 0.7) $-$0.45(0.17)
54 16 27 31.3 -24 31 10.1 1.2 1.0( 0.4) 1.00 1.4( 0.4) 0.88(0.21) 3.2( 0.7) 1.00
55 16 27 32.7 -24 33 24.9 1.1 1.4( 0.4) 0.67(0.24) 2.2( 0.5) 0.28(0.20) 3.7( 0.7) 0.02(0.19)
56 16 27 32.9 -24 32 34.6 1.1 2.3( 0.5) 0.59(0.17) 1.8( 0.7) 0.75(0.32) 4.3( 0.8) 0.10(0.17)
57 16 27 33.1 -24 41 14.3 0.1 44.8( 1.3) 0.20(0.03) 44.7( 1.3) 0.18(0.03) 120.0( 2.4) 0.20(0.02)
58 16 27 35.3 -24 38 33.0 0.7 2.1( 0.3) $-$0.59(0.16) 1.7( 0.3) $-$0.89(0.14) 4.0( 0.6) $-$0.80(0.17)
59 16 27 35.7 -24 45 31.7 0.8 1.8( 0.3) $-$0.76(0.18) 1.5( 0.3) $-$0.74(0.17) 4.7( 0.7) $-$0.82(0.13)
60 16 27 38.2 -24 30 42.2 1.4 0.4( 0.3) 0.18(0.78) 1.0( 0.3) $-$1.00 3.6( 0.8) 0.04(0.23)
61 16 27 38.3 -24 36 58.6 0.1 37.4( 1.3) 0.12(0.04) 38.0( 1.4) 0.08(0.04) 97.1( 2.4) 0.09(0.02)
62 16 27 38.6 -24 38 39.6 0.5 4.8( 0.5) $-$0.13(0.10) 5.4( 0.5) 0.00(0.09) 14.1( 1.3) $-$0.10(0.09)
63 16 27 38.9 -24 40 20.9 0.6 5.6( 0.5) 0.98(0.03) 6.0( 0.6) 0.95(0.05) 15.6( 3.2) 1.00
64 16 27 39.1 -24 47 23.1 0.5 5.5( 0.6) 0.78(0.07) 4.8( 0.5) 0.59(0.09) 12.8( 1.1) 0.71(0.06)
65 16 27 39.4 -24 39 15.4 0.1 67.4( 1.7) 0.13(0.03) 68.9( 1.7) 0.10(0.03) 225.3(11.7) 0.06(0.05)
66 16 27 39.8 -24 43 15.1 0.3 9.6( 0.7) 0.66(0.05) 9.9( 0.7) 0.68(0.05) 25.6( 1.3) 0.64(0.04)
67 16 27 43.6 -24 31 27.4 0.7 3.1( 0.5) 0.77(0.13) 2.2( 0.5) 0.79(0.16) 10.3( 1.1) 1.00
68 16 27 47.1 -24 45 33.5 1.3 1.7( 0.5) $-$0.30(0.32) 1.0( 0.6) $-$0.18(0.55) 3.9( 0.7) 0.17(0.17)
69 16 27 49.1 -24 39 06.7 1.7 0.8( 0.2) 0.70(0.28) 0.8( 0.2) 0.38(0.31) 1.0( 0.4) 0.82(0.33)
70 16 27 50.3 -24 31 50.0 1.2 2.3( 0.5) 0.46(0.18) 2.4( 0.5) 0.64(0.18) 6.5( 1.1) 0.71(0.14)
71 16 27 51.9 -24 31 46.9 1.2 1.0( 0.3) 1.00 0.7( 0.3) 0.47(0.47) 5.5( 1.1) 0.33(0.20)
72 16 27 51.9 -24 46 29.7 0.6 4.0( 0.5) $-$0.24(0.13) 3.2( 0.5) $-$0.20(0.14) 8.1( 1.0) 0.01(0.12)
73 16 27 52.1 -24 40 50.1 0.1 50.5( 1.8) $-$0.27(0.03) 54.9( 1.6) $-$0.25(0.03) 171.2( 3.3) $-$0.37(0.02)
74 16 27 55.6 -24 44 51.3 1.1 1.1( 0.3) $-$0.44(0.29) 0.8( 0.3) 0.13(0.38) 2.5( 0.5) $-$1.00
75 16 27 55.8 -24 41 16.5 0.8 2.9( 0.5) 0.74(0.13) 1.6( 0.3) 0.70(0.17) 4.5( 0.7) 0.68(0.14)
76 16 27 57.8 -24 40 02.3 0.6 7.0( 0.8) 0.22(0.11) 7.7( 0.7) $-$0.03(0.09) 19.6( 2.4) 0.06(0.12)
77 16 28 00.0 -24 48 19.8 0.5 6.3( 0.7) $-$0.71(0.09) 5.1( 0.7) $-$0.67(0.11) 18.4( 1.5) $-$0.54(0.08)
78 16 28 04.6 -24 34 55.7 0.4 8.8( 0.8) 0.33(0.09) 5.7( 0.7) 0.45(0.10) 19.8( 1.6) 0.22(0.08)
79 16 28 04.7 -24 37 11.9 0.9 2.4( 0.5) 0.09(0.22) 1.8( 0.4) $-$0.27(0.23) 6.9( 0.9) 0.20(0.13)
80 16 28 04.8 -24 50 17.6 1.2 1.0( 0.4) $-$0.29(0.42) 1.8( 0.5) 0.14(0.29) 6.0( 1.2) 0.20(0.19)
81 16 28 06.8 -24 48 30.3 0.8 2.1( 0.5) 1.00 1.5( 0.4) 0.88(0.19) 7.7( 1.2) 0.76(0.13)
82 16 28 08.2 -24 45 13.8 1.1 0.7( 0.3) $-$0.93(0.27) 1.8( 0.4) $-$0.21(0.23) 3.5( 0.8) $-$0.19(0.24)
83 16 28 08.8 -24 46 23.0 1.7 0.7( 0.3) 1.00 0.8( 0.4) 0.31(0.43) 3.3( 0.8) 1.00
84 16 28 11.3 -24 46 14.8 1.6 0.9( 0.3) 1.00 0.5( 0.3) 1.00 2.9( 0.8) 0.43(0.27)
85 16 28 12.7 -24 31 22.9 1.2 4.6( 0.9) 1.00 4.3( 0.8) 0.56(0.15)
86 16 28 23.7 -24 41 41.5 0.9 8.0( 0.9) 0.16(0.11) 9.9( 1.0) 0.10(0.10)
87 16 28 25.2 -24 45 00.6 0.7 7.1( 0.9) $-$1.00 8.7( 1.0) $-$1.00
------ ----------------------- ----------------------- ------------- ------------- --------------- ------------- --------------- ------------- ---------------
\(1) [*XMM-Newton*]{} source ID number, “Rho Oph X-ray sources, Newton”.; (5-10) The hyphens indicate that the X-ray sources are outside the field of view of the instrument.; (6),(8),(10) The hardness ratio is calculated by the equation, $H.R. = (H - S) / (H + S)$, where $S$ and $H$ are counts in the 0.3$-$2.0 keV band and in the 2.0$-$8.0 keV band, respectively. Note that due to pile-up, $H.R.$ is maybe not reliable for source 1. \[xraytable\]
---------- ----------- ----- ----- ----- ----- ----- ----- ------------------ ---------- ---------- -------- --------- --------- ------------------------- --------
E R1 R2 A RH C ISO 2MASS J NAME Class
(2) (3) (4) (5) (6) (7) (8) (9) (15) (16)
1 - F1 - 16262367-2443138 9.39 8.40 7.85 0.99 0.55 DoAr25 II
2$\ast$ - - 16262753-2441535 14.04 11.42 9.98 2.62 1.44 GY33 II
3$\ast$ - -
4 - 1 65 16264419-2434483 16.76 13.76 11.60 3.00 2.16 WL12 I
5 31 - F2 - 66 16264429-2443141 10.99 10.02 9.57 0.97 0.45 GY112 III
6$\ast$ - - 16264441-2447138 11.83 11.00 10.63 0.83 0.37 BKLT162644-244711 nIII
7$\ast$ - - - -
8 C10 - F3 - 69 16264705-2444298 12.33 11.12 10.56 1.21 0.56 GY122 III
9$\ast$ - - 16264810-2442033 12.45 11.28 10.65 1.17 0.63 GY125 nIII
10$\ast$ - 16265128-2432419 15.30 14.47 13.89 0.83 0.58 GY141 bd
11 C13 35 - F7 - 88 16265850-2445368 9.75 8.16 7.06 1.59 1.10 SR24S II
12 - 10 90 16265916-2434588 $>$17.26 $>$17.82 15.13 $>$2.69 WL22 II
13$\ast$ - -
14 - F8 - 96 16270451-2442596 12.02 10.55 9.84 1.47 0.71 GY193 III
15 - F9 97 16270456-2442140 12.21 10.54 9.81 1.67 0.73 GY194 III
16 - 15
17 - 17 102 16270659-2441488 12.43 11.40 10.77 1.03 0.63 GY204 II
18 - C5 18 103 16270677-2438149 $>$17.28 14.30 10.97 $>$2.98 3.33 WL17 II
19$\ast$ -
20 37 - 7 21 105 16270910-2434081 12.55 10.19 8.91 2.36 1.28 WL10 II
21$\ast$ - - 16270931-2443196 14.30 12.21 11.06 2.09 1.15 GY212 nIII
22 - 8 23 108 16270943-2437187 16.79 11.05 7.14 5.74 3.91 EL29 I
23$\ast$ - - - -
24$\ast$ - - - -
25 - 25 114 16271171-2438320 $>$18.60 15.06 11.06 $>$3.54 4.00 WL19 III
26$\ast$ - 115 16271213-2434491 15.62 13.11 11.49 2.51 1.62 WL11 II
27 20 - 22 - - - 16271448-2451334 11.45 10.69 10.38 0.76 0.31 ROXs20A III
28 20 - - F14 - - 16271513-2451388 10.62 9.77 9.39 0.85 0.38 ROXs20B III
29 C6 - 26 16271545-2426398 17.42 13.46 10.79 3.96 2.67 IRS34 II
30 C16 39 - F16 29 121 16271587-2438433 13.89 11.26 9.59 2.63 1.67 WL20 II
31 - 30 16271643-2431145 $>$18.70 $>$17.69 15.09 $>$2.60 16271643-24311 nIII
32 40 - 34 125 16271817-2428526 $>$17.87 14.69 10.56 $>$3.18 4.13 WL5 III
33 - 35 127 16271838-2439146 $>$18.67 15.54 12.23 $>$3.13 3.31 GY245 II
34 - C6 36 128 16271848-2429059 14.61 11.50 9.68 3.11 1.82 WL4 II
35 - 38 129 16271921-2428438 $>$17.61 14.66 11.49 $>$2.95 3.17 WL3 II
36 21 41 - 9A F17 40 130 16271951-2441403 9.42 8.63 8.41 0.79 0.22 SR12A-B III
37 - 43 132 16272146-2441430 15.22 11.25 8.48 3.97 2.77 IRS42 II
38 C7? - 10 44 134 16272180-2429533 $>$18.65 15.38 10.83 $>$3.27 4.55 WL6 I
39 C18 - C7 - 133 16272183-2443356 17.27 13.20 10.78 4.07 2.42 GY253 III
40 - F19 - - 16272297-2448071 10.92 9.86 9.39 1.06 0.47 WSB49 II
41 - 50 16272463-2429353 $>$18.71 14.78 12.43 $>$3.93 2.35 GY259 nIII
42 - 53 16272648-2439230 15.69 12.07 9.95 3.62 2.12 GY262 II
43 43 - F21 54 141 16272693-2440508 $>$18.53 13.52 9.74 $>$5.01 3.78 IRS43/YLW15 I
44 - 55 16272706-2432175 $>$17.81 14.51 12.42 $>$3.30 2.09 GY266 nIII
45 44 - F22 56 142 16272738-2431165 12.35 10.38 9.32 1.97 1.06 VSSG25 II
46 $\dagger$ - 9B 57 143 16272802-2439335 $>$16.56 13.68 10.38 $>$2.88 3.30 IRS44/YLW16A I
47 - - F24 - - 16272873-2454317 12.60 12.02 11.69 0.58 0.33 0600-20531406$\ddagger$
48 - 59 144 16272844-2427210 15.74 12.31 10.10 3.43 2.21 IRS45 II
49 - 61 146 16272996-2433365 $>$18.70 15.58 12.43 $>$3.12 3.15 BKLT162730-243336 nIII
50 - 63 147 16273018-2427433 15.32 11.52 9.02 3.80 2.50 IRS47 II
51$\ast$ - - - -
52 46 - F26 - 149 16273084-2447268 12.24 10.42 9.50 1.82 0.92 BKLT162730-244726 III
53 47 - 64 148 16273105-2434032 13.43 11.36 10.39 2.07 0.97 GY283 nIII
54 - 65
55 - 66 152 16273267-2433239 16.15 12.74 10.90 3.41 1.84 GY289 III
56 48 - 67 154 16273285-2432348 16.19 12.74 10.96 3.45 1.78 GY291 II
57 - F27 69 155 16273311-2441152 11.32 9.13 7.81 2.19 1.32 GY292 II
58 - 70 156 16273526-2438334 11.28 10.23 9.67 1.05 0.56 GY295 nIII
59 - F28 - 157 16273566-2445325 12.71 11.47 10.88 1.24 0.59 GY296 III
60 C10 - 74 16273812-2430429 12.54 10.54 9.66 2.00 0.88 IRS50 III
61 C20 - F29 75 163 16273832-2436585 11.38 9.43 8.27 1.95 1.16 IRS49 II
62 - 76 164 16273863-2438391 13.27 11.93 11.08 1.34 0.85 GY310 II, bd
63 - 77 165 16273894-2440206 16.54 13.91 12.29 2.63 1.62 GY312 II
64$\ast$ - - -
65 51 - F30 78 166 16273942-2439155 10.75 9.21 8.46 1.54 0.75 GY314 II
66 - 79 167 16273982-2443150 17.05 12.13 8.99 4.92 3.14 IRS51 I
67 - 82
68$\ast$ - - 177 16274709-2445350 15.75 12.81 11.13 2.94 1.68 GY352 II
69 - 83
70 - 84
71 C11 - 182 16275180-2431455 14.68 11.19 8.71 3.49 2.48 IRS54 I
72 - - F32 - 183 16275191-2446296 14.05 11.61 10.37 2.44 1.24 GY377 III
73 54 - F33 85 184 16275209-2440503 10.00 8.72 8.13 1.28 0.59 IRS55 III
74$\ast$ - - - 186 16275565-2444509 12.34 11.15 10.47 1.19 0.68 GY398 nIII
75$\ast$ - -
76 C13 - - 188 16275782-2440017 12.65 10.73 9.87 1.92 0.86 GY410 III
77 55 - - F34 - - 16275996-2448193 10.81 9.76 9.27 1.05 0.49 BKLT162800-244819 nIII
78$\ast$ - - 16280464-2434560 16.92 13.18 10.97 3.74 2.21 GY463 nIII
79$\ast$ - - 16280478-2437100 13.62 11.45 10.34 2.17 1.11 GY465 nIII
80$\ast$ - - - -
81$\ast$ - - - -
82$\ast$ - - - 16280810-2445121 14.12 12.86 12.15 1.26 0.71 GY478 nIII
83$\ast$ - - -
84$\ast$ - - -
85$\ast$ - -
86 - 33 - - 16282373-2441412 11.73 9.83 9.01 1.90 0.82 BKLT162823-244140 nIII
87 - 34 - F37 - 16282516-2445009 6.72 6.60 6.51 0.12 0.09 HD148352$\dagger$
---------- ----------- ----- ----- ----- ----- ----- ----- ------------------ ---------- ---------- -------- --------- --------- ------------------------- --------
(1)[*XMM-Newton*]{} source ID number, “Rho Oph X-ray sources, Newton”. The asterisks indicate new X-ray sources found in this observation; (2$-$9) The identification with source catalogs of (2) [*Einstein*]{} (Montmerle et al. 1983), ROX$-$; [*$\dagger$ Grosso (2001) shows that one of these [*Einstein*]{} observation has detected an X-ray source associated with the YLW16 IR source group*]{}; (3)[*ROSAT PSPC*]{} (Casanova et al. 1995), ROXR1$-$; (4)[*ROSAT PSPC*]{} (Casanova 1994), ROXR2$-$; (5)[*ASCA*]{} (Katama et al. 1997), ROXA$-$; (6)[*ROSAT HRI*]{} (Grosso et al. 2000), ROXR$-$; (7)[*Chandra*]{} (Imanishi et al. 2001a); (8)[*ISO ISOCAM*]{} (Bontemps et al. 2001), (9) 2MASS. The dotted lines indicate that the [*XMM-Newton*]{} sources have no known counterpart. The hyphens indicate that the [*XMM-Newton*]{} sources are outside the field of view of the instruments. (10$-$14)If only the lower limits of the values are available, “$>$” is put before the values. (15) For the source naming convention, see André & Montmerle (1994) and references therein. The source names beginning with BKLT are from Barsony et al. (1997). 0600-2053 is from the catalog USNO-A1.0 (Monet et al. 1996). $\ddagger$ indicate foreground stars. (16) The classifications of Class I, II, and III are from Bontemps et al. (2001) or André & Montmerle (1994). The Class III candidates proposed in this paper are indicated by “nIII”. The bona fide brown dwarfs are also indicated in this column by “bd”. \[irtable\]
X-ray light curves
------------------
We extract background-subtracted X-ray light curves for the ROXN sources. X-ray counts are selected from circular regions around the sources where 90% of the X-ray counts from the sources are included at 1.5 keV.[^4] In crowded regions, we adjust the circle radius to avoid contamination by neighbouring sources. Background X-ray counts for each source are extracted from annular regions with a 60$\arcsec$ inner radius and a 150$\arcsec$ outer radius around the sources. If the annuli contain other X-ray sources, regions within a 40$\arcsec$ radius around those sources are excluded from the corresponding annuli. To avoid decreasing too much the area of the annuli by excluding nearby X-ray sources, the outer radius was adjusted to keep a constant area for the background extraction region.
Since YSOs are known in general to exhibit X-ray flares, we systematically look for flares in our sources. We find that nearly half of the sources (40) have enough X-ray counts ($N > 150$, where $N$ is the sum of MOS1, MOS2 and PN counts) to search for X-ray flares, using the following method. We extract X-ray light curves with a uniform reference time binsize. A bin-to-bin variation was then defined to be a flare if the X-ray counts in a given time bin are 1.5 times higher than at least one of the three preceding bins with a statistical significance above 4 $\sigma$. We apply this method varying the reference time binsizes from 1000 s to 4000 s by steps of 1000 s, which covers the typical rise times of X-ray flares from YSOs. To plot the light curves, we use an adaptive binning where the time binsize is tuned to keep a constant number of counts inside each time bin. This enables us to catch any rapid time variability such as rising phase of X-ray flares. Depending on the total number of counts, the time bins are tuned so as to have enough counts in each bin to see the overall time profile properly (see tuning given in Table 3).
Fig. \[fig\_lc\] shows a sample of background-subtracted X-ray light curves. The X-ray flares seen in Fig. \[fig\_lc\] show a variety of profiles, broadly characterized by a fast increase and a roughly exponential decay, although some profiles are more symmetrical, with comparatively slow rise and decay phases of similar durations. Contrary to previous X-ray observations which had limited statistics, a “typical" YSO X-ray flare profile cannot, in general, be reduced to a simple fast rise followed by an exponential decay. Such a complex behavior is also seen in the large X-ray flare database obtained in 40 ks [*Chandra*]{} observations of the Orion Nebula Cluster (Feigelson et al. 2002a).
In our $\rho$ Oph observations, X-ray flares are detected from 1/4 Class I, 6/13 Class II, and 3/12 Class III sources. These ratios are 2–3 times lower than those obtained by [*Chandra*]{} (Imanishi et al. 2001a; 14/18 for Class I + Class I candidates, 12/20 for Class II+III), which is to be expected since our [*XMM-Newton*]{} exposure (33 ks) is three times shorter than that of [*Chandra*]{} (100 ks). The unclassified sources did not show any strong flares.
Source total counts Counts inside a time bin
--------------------- -------------------------- --
$~~~0$ – $~500$ 15
$~500$ – $1000$ 25
$1000$ – $2000$ 35
$2000$ – $4000$ 50
$4000$ – $6000$ 80
$6000 < ~~~~~$ 100
: Tuning of the adaptive binning. Counts inside a time bin for a given source total counts.
X-ray spectra {#subsec_spec}
-------------
We now compute background-subtracted X-ray spectra for sources with $N > 150$ cts (4 Class I, 13 Class II, 12 Class III sources, and 11 unclassified sources). The way to extract X-ray counts and background counts is the same as for extracting the light curves. [^5] Since the difference between the responses of MOS1 and MOS2 is not large because they use the same type of CCD array, we add the MOS1 and MOS2 spectra to improve the statistics, i.e., we add the ancillary response files, and average the photon redistribution matrixes of MOS1 and MOS2 weighted by the exposure time of each detector. We treat the PN spectra separately. Fig. \[spec\_fig\] shows examples of the X-ray spectra obtained in our observation.
We first fit the separate PN and MOS1+MOS2 spectra with a single temperature thin thermal emission model (MEKAL) combined with an absorption model (WABS)[^6]. For sources which do not have enough counts to yield the plasma metal abundance, we fix it to 0.3 solar, a typical value for YSOs obtained in $\rho$ Oph with [*ASCA*]{} (Kamata et al. 1997) and [*Chandra*]{} (Imanishi et al. 2001a), and also in other star-forming regions. For calculations of emission measures and luminosities, we adopt $d = 140$ pc, based on [*Hipparcos*]{} measurements (see discussion in Bontemps et al. 2001), and for consistency with previous [*ROSAT*]{} study (Grosso et al. 2000).
As a result, we obtain acceptable one-temperature spectral fits for all the sources except for ROXs20A + ROXs20B, SR12A$-$B and IRS55 (see below). The values of the fitting parameters are listed in Table 4. The individual metal abundances obtained from the fits are found to be distributed around 0.3 solar, which is consistent with most existing YSO X-ray observations. If we alternatively choose to fix the metal abundance to 0.3 solar for all the sources, we also obtain acceptable one-temperature fits with better constraints on other spectral parameters except for the Class I source EL29 (ROXN-22). EL29 did not show X-ray flares during our observation, whereas it showed X-ray flares during the [*ASCA*]{} and [*Chandra*]{} observations (Kamata et al. 1997, Imanishi et al. 2001a). We find 1.0 (0.80$-$1.3) for the metal abundance of EL29 relative to solar, which is far from the value, $\sim$0.3, observed for the quiescent level by [*ASCA*]{} and [*Chandra*]{} (Kamata et al. 1997, Imanishi et al. 2001a). Actually, the K$\alpha$ line from He-like iron ions at 6.7 keV is clearly seen in the spectra of EL29 (see Fig. \[spec\_fig\]). We note a residual in the spectral fitting around 6.4 keV in both PN and MOS1+MOS2, however adding a gaussian line at this energy does not change the metallicity at all. This suggests that the metal abundances of the X-ray emitting plasma in YSOs could be variable. For example, such metalicity enhancement could be explained by photospheric evaporation produced by flare (Güdel et al. 2001). Then we decide to list the values of the metal abundances obtained from X-ray spectral fitting in our observation in Table 4.
Since it is often found that X-ray spectra with low statistics can be modeled by either a single temperature or a two-temperature MEKAL $+$ absorption model, we also test a two-temperature MEKAL $+$ absorption model for all the ROXN sources. The metal abundance is then fixed to 0.3 solar except for EL29. We thus obtain acceptable two-temperature fits for all the sources, but the emission measures of the soft components are either extremely large or very low for most sources. This means that the large absorption of the sources does not allow us to constrain the parameters of the soft component, so that one-temperature fits are sufficient in practice. However, for the bright Class III sources ROXs20A and ROXs20B (i.e., ROXN-27 and ROXN-28, which are spatially resolved but with too few counts, hence we use a circular area encompassing both sources to obtain a spectrum), SR12A$-$B (ROXN-36) and IRS55 (ROXN-73), two-, three-, and two-temperature MEKAL $+$ absorption models, respectively, are needed to obtain acceptable fits (see Fig. \[spec\_fig\]).
[cccccccccc]{} &&&&&&&\
ROXN & NAME & & $kT$ & $N_{\rm{H}}$ & abundance & $E.M.$ & $L_{\rm{X}}$ & $\chi^2$/d.o.f\
&&& & ($10^{22}$ cm$^{-2}$) & &($10^{52}$ cm$^{-3}$) & ($10^{29}$ erg s$^{-1}$) &\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9)\
\
22 & EL29 & & 4.3(3.6$-$5.1) & 4.9(4.4$-$5.3) & 1.0(0.80$-$1.3) & 16.0(13.6$-$18.9) & 27.5& 90.0/ 63.0\
43 & IRS43/YLW15 & & 3.9(3.0$-$5.3) & 3.3(2.9$-$3.8) & \[0.30\] & 13.1(10.2$-$17.3) & 17.3& 36.5/ 35.0\
& &f& 3.2(2.7$-$3.9) & 4.0(3.5$-$4.5) & \[0.30\] & 29.1(22.8$-$37.0) & 34.8& 56.3/ 45.0\
46 & IRS44/YLW16A & & 2.7(2.4$-$3.1) & 5.3(4.9$-$5.7) & 0.28(0.19$-$0.38) & 45.6(38.3$-$53.9) & 49.7& 77.0/ 72.0\
66 & IRS51 & & 2.1(1.8$-$2.8) & 4.2(3.5$-$4.8) & 0.62(0.27$-$1.18) & 13.8(10.2$-$18.5) & 16.2& 51.7/ 48.0\
\
\
&&&&&&&\
ROXN & NAME && $kT$ & $N_{\rm{H}}$ & abundance & $E.M.$ & $L_{\rm{X}}$ & $\chi^2$/d.o.f\
ID &&& & ($10^{22}$ cm$^{-2}$) & &($10^{52}$ cm$^{-3}$) & ($10^{29}$ erg s$^{-1}$) &\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9)\
\
1 & DoAr25& & 6.4(4.9$-$8.8) & 1.1(0.96$-$1.2) & 0.26(0.02$-$0.51) & 157(141$-$177) & 247 & 80.1/ 72.0\
11 & SR24S & & 2.3(2.0$-$2.7) & 1.2(1.0$-$1.3) & 0.24(0.07$-$0.43) & 16.3(13.4$-$19.4) & 15.9& 107.7/ 75.0\
18 & WL17 & & 2.9(2.0$-$4.3) & 4.1(3.1$-$5.2) & \[0.30\] & 8.4(5.0$-$14.6) & 9.6& 27.5/ 19.0\
& &f& 3.2(2.5$-$4.2) & 4.5(3.8$-$5.4) & 0.66(0.31$-$1.2) & 18.1(13.1$-$25.7) & 24.8& 35.4/ 25.0\
29 & IRS34 & & 1.8(1.1$-$3.7) & 2.7(1.6$-$4.4) & \[0.30\] & 12.8($<$41.5) & 11.7 & 4.7/ 5.0\
30 & WL20 & & 2.3(1.8$-$3.0) & 2.0(1.6$-$2.4) & \[0.30\] & 8.1(5.9$-$11.4) & 8.1& 16.2/ 17.0\
& &f& 3.4(2.7$-$4.5) & 2.0(1.7$-$2.3) & 0.34(0.04$-$0.75) & 16.4(13.3$-$20.2) & 20.4& 34.7/ 29.0\
33 & GY245 & & 2.1(1.3$-$5.0) & 7.7(5.2$-$11.0) & \[0.30\] & 4.5(1.5$-$16.4) & 4.4& 10.9/ 9.0\
45 & VSSG25&f& 2.0(1.6$-$2.5) & 1.1(0.9$-$1.4) & \[0.30\] & 6.3(4.7$-$8.2) & 5.9& 29.1/ 24.0\
50 & IRS47 &f& 2.9(1.7$-$7.4) & 2.3(1.5$-$3.3) & \[0.30\] & 7.5($<$15.5) & 8.5 & 1.2/ 2.0\
57 & GY292 & & 2.3(2.1$-$2.5) & 1.7(1.5$-$1.8) & 0.23(0.14$-$0.36) & 31.2(28.6$-$35.8) & 30.8& 159.3/153.0\
61$^a$ & IRS49 & & 2.0(1.9$-$2.2) & 1.6(1.4$-$1.7) & 0.18(0.08$-$0.29) & 27.1(23.9$-$30.7) & 24.1& 131.1/114.0\
62 & GY310 &f& 2.5(1.7$-$3.6) & 0.82(0.59$-$1.2) & \[0.30\] & 1.8(1.4$-$2.4) & 1.9& 22.4/ 13.0\
63 & GY312 & & 2.3(1.7$-$3.3) & 7.6(6.0$-$8.7) & \[0.30\] & 11.4(6.4$-$22.5) & 11.6& 25.0/ 17.0\
65 & GY314 & & 2.1(2.0$-$2.3) & 1.5(1.4$-$1.6) & 0.16(0.07$-$0.26) & 31.6(28.4$-$35.0) & 28.4& 113.2/113.0\
\
\
&&&&&&&\
ROXN & NAME && $kT$ & $N_{\rm{H}}$ & abundance & $E.M.$ & $L_{\rm{X}}$ & $\chi^2$/d.o.f\
&&& & ($10^{22}$ cm$^{-2}$) & &($10^{52}$ cm$^{-3}$) & ($10^{29}$ erg s$^{-1}$) &\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9)\
\
5 & GY112 & & 0.98(0.79$-$1.1) & 0.36(0.31$-$0.50) & 0.10(0.06$-$0.16) & 7.5(6.1$-$11.8) & 4.3& 40.9/ 37.0\
14$^a$ & GY193 & & 1.27(1.19$-$1.74) & 0.87(0.56$-$1.07) & \[0.30\] & 7.0(4.6$-$8.2) & 6.0& 29.4/ 38.0\
15 & GY194 & & 1.5(1.2$-$1.9) & 0.79(0.6$-$1.1) & \[0.30\] & 3.2(2.6$-$3.9) & 2.8& 20.7/ 13.0\
25 & WL19 & & 2.5(1.5$-$4.6) & 8.62(6.0$-$12.8) & \[0.30\] & 9.2(3.7$-$30.7) & 9.8& 8.4/ 10.0\
27,28$^{a,b}$ & ROXs20A, ROXs20B & & & 0.13(0.06$-$0.18) & 0.42(0.23$-$1.21) & & 4.6& 59.2/ 40.0\
& & & 0.66(0.56$-$0.77) & & & 1.7(0.8$-$2.2) & 1.9&\
& & & 1.66(1.30$-$2.29) & & & 2.8(1.9$-$3.9) & 2.8&\
32,34,35$^{b,c}$ & WL5,WL4,WL3 & &2.5(2.2$-$3.0) & 4.7(4.2$-$5.2) & \[0.30\] & 60.4(47.4$-$76.9) & 64.2& 104.6/ 74.0\
36 & SR12A$-$B & & & 0.11(0.08$-$0.13) & 0.12(0.08$-$0.17) & & 21.9& 248.6/205.0\
& & & 0.25(0.23$-$0.27) & & & 18.2(11.1$-$30.2) & 4.4&\
& & & 0.97(0.89$-$1.03) & & & 18.6(17.0$-$25.2) & 11.1&\
& & & 3.5(2.5$-$6.5) & & & 5.6(3.1$-$8.3) & 6.5&\
& &f& & 0.06(0.03$-$0.10) & 0.24(0.13$-$0.55) & & 24.0& 74.0/ 60.0\
& & & 0.32(0.28$-$0.38) & & & 7.8(4.3$-$17.1) & 4.2&\
& & & 1.22(1.09$-$1.35) & & & 15.3(5.5$-$22.5) & 12.0&\
& & & 12.0($>$3.1) & & & 4.9(3.0$-$10.4) & 7.7&\
39 & GY253 & & 1.6(1.4$-$2.0) & 4.5(3.9$-$5.2) & \[0.30\] & 24.9(17.5$-$34.9) & 22.0& 57.7/ 52.0\
52 & BKLT162730-244726 & & 1.9(1.3$-$2.3) & 1.0(0.8$-$1.5) & \[0.30\] & 6.6(5.3$-$11.2) & 6.1& 44.1/ 43.0\
72 & GY377 & & 1.8(1.1$-$2.7) & 1.1(0.7$-$1.8) & \[0.30\] & 1.7(0.9$-$3.6) & 1.5& 7.1/ 11.0\
73 & IRS55 & & & 1.1(1.0$-$1.2) & 0.31(0.20$-$0.50) & & 62.3& 139.3/141.0\
& & & 0.64(0.60$-$0.68) & & & 46.9(28.0$-$67.2) & 45.7&\
& & & 2.5(2.2$-$3.0) & & & 15.7(12.7$-$19.2) & 16.6&\
76 & GY410 & & 1.8(1.3$-$2.4) & 1.5(1.1$-$1.9) & \[0.30\] & 3.8(2.5$-$5.9) & 3.4& 8.9/ 23.0\
\
\
&&&&&&&\
ROXN & NAME && $kT$ & $N_{\rm{H}}$ & abundance & $E.M.$ & $L_{\rm{X}}$ & $\chi^2$/d.o.f\
ID &&& & ($10^{22}$ cm$^{-2}$) & &($10^{52}$ cm$^{-3}$) & ($10^{29}$ erg s$^{-1}$) &\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9)\
31 & 16271643-24311 && 4.2(2.2$-$15.1) & 5.2(3.5$-$7.4) & \[0.30\] & 6.1(3.2$-$14.3) & 8.3& 25.8/ 26.0\
41 & GY259 && 3.4($>$1.3) & 2.2(1.1$-$4.5) & \[0.30\] & 1.9(0.8$-$6.3) & 2.3& 0.5/ 2.0\
44 & GY266 && 2.3(1.1$-$6.1) & 2.8(1.6$-$5.0) & \[0.30\] & 2.4(1.0$-$8.2) & 2.5& 7.5/ 6.0\
53 & GY283 && 1.4(0.5$-$2.6) & 0.47(0.2$-$1.5) & \[0.30\] & 0.5(0.3$-$3.0) & 0.5& 2.1/ 4.0\
58 & GY295 && 1.9(1.2$-$2.8) & 0.34(0.2$-$0.6) & \[0.30\] & 0.7(0.5$-$0.8) & 0.6& 12.3/ 11.0\
77 & BKLT162800-244819 && 0.5(0.4$-$0.7) & 0.97(0.8$-$1.1) & \[0.30\] & 8.7(4.4$-$17.2) & 7.8& 24.2/ 25.0\
78 & GY463 && 1.6(1.2$-$2.1) & 2.6(2.0$-$3.4) & \[0.30\] & 11.0(6.8$-$20.2) & 9.6& 16.0/ 20.0\
86 & BKLT162823-244140 && 4.3(2.2$-$13.2) & 0.87(0.5$-$1.3) & \[0.30\] & 4.0(2.8$-$6.7) & 5.5 & 9.7/ 9.0\
\
\
&&&&&\
ROXN & NAME && $kT$ & $N_{\rm{H}}$ & abundance & $E.M.$ & $L_{\rm{X}}$ & $\chi^2$/d.o.f\
ID &&& & ($10^{22}$ cm$^{-2}$) & &($10^{52}$ cm$^{-3}$) & ($10^{29}$ erg s$^{-1}$) &\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9)\
51 & && 10.5($>$3.9)& 0.92(0.6$-$1.4) & \[0.30\] & 2.5(2.0$-$3.8) & 4.4& 40.1/ 28.0\
64 & && 82.4($>$20.5) & 2.2(1.7$-$2.9)& \[0.30\] & 3.5(2.9$-$4.1) & 5.5& 64.5/ 43.0\
67 & && 3.6($>$1.4) & 9.5(4.8$-$18.0) & \[0.30\] & 6.7(2.4$-$46.9) & 8.6& 25.4/ 29.0\
\
\
\
(1)[*XMM-Newton*]{} source ID number, “Rho Oph X-ray sources, Newton”. (2) The source name from column (15) in Table 2. (3) “f” means that the X-ray spectra are extracted during the flare phase. (4$-$7) The 90% confidence regions are given between parenthesis. (5) We use WABS for absorption model in XSPEC where standard solar metal abundances are used. The use of reviese solar abundances (Holweger 2001, Allend Prieto et al. 2001, 2001) would increase N$_\mathrm{H}$ by 20 $\sim$% (Vuong et al. 2003). (7),(8) [The emission measures and X-ray luminosities were calculated from the best-fit PN parameters assuming a distance of 140 pc. Those from MOS1+MOS2 parameters are consistent within errors.]{}\
$^a$ These sources showed small flare during the observation, but the X-ray counts in the flares are not enough to make spectra.\
$^b$ It is difficult to obtain individual spectrum without contamination from neighbouring sources. Then we extracted X-ray spectrum from a circular region which contain all the sources.\
$^c$ Since the X-ray counts from the source 32 is dominant, we put the spectral parameters in the table for Class III.
\[spec\_tab\]
Comparison with [*Chandra*]{} source detection {#sec_chan}
==============================================
[*Chandra*]{} observed the $\rho$ Ophiuchi cloud core F region with an exposure of 100 ks (Imanishi et al. 2001a), although, contrary to our [*XMM-Newton*]{} observations, the pointing was somewhat off-centered with respect to the peak of the CO emission of core F. Fig. \[fig\_im\] shows the [*XMM-Newton*]{} and [*Chandra*]{} X-ray source positions superimposed on the 0.3$-$8 keV band image obtained by [*XMM-Newton*]{}. [*Chandra*]{} detected in 100 ks 87 X-ray sources in its field of view (17$\arcmin$x17$\arcmin$). This the same number of sources as in our three times shorter [*XMM-Newton*]{} observations, in a field of view which is nearly four times as large. We shall return to this coincidence below. In the region covered by the field of view of both observatories, 47 X-ray sources are detected with [*XMM-Newton*]{} and 81 X-ray sources are detected with [*Chandra*]{}, while 43 X-ray sources are detected with both observatories.
Fig. \[fig\_chanc\] shows [*XMM-Newton*]{} vs. [*Chandra*]{} count rates for all these X-ray sources. Upper limits correspond to the 99.9989% confidence level threshold (i.e., 4.4 $\sigma$ level for Gaussian statistics) for the [*Chandra*]{} sources undetected by [*XMM-Newton*]{}, and for the [*XMM-Newton*]{} sources undetected by [*Chandra*]{}, using a method explained in Appendix A. The median of the count rate upper limits of sources detected with only [*Chandra*]{} is the dotted horizontal line in Fig. \[fig\_chanc\], which indicates also the 4.4 $\sigma$ detection threshold for [*XMM-Newton*]{}. There is a good correlation between [*XMM-Newton*]{} and [*Chandra*]{} count rates. The median of the ratios between [*XMM-Newton*]{} and [*Chandra*]{} count rates for X-ray sources detected by both observatories is 6.8, which is indicated by a continuous line in Fig. \[fig\_chanc\]: “the median ratio line". Most of the data points are scattered close to this median ratio line. A few sources display large discrepancies between [*XMM-Newton*]{} and [*Chandra*]{} observations, which suggests that their X-ray luminosities could have changed due to time variability. We check the two most extreme cases, sources ROXN-33 (GY245) and ROXN-46 (YLW16A), which are far from the median ratio line. GY245 did not show any flare during the [*Chandra*]{} observation, but showed a flare during the [*XMM-Newton*]{} observation, while YLW16A showed an X-ray flare during the [*Chandra*]{} observation, but did not show any flare during the [*XMM-Newton*]{} observation. Hence, flaring activity explains well the large apparent discrepancies observed for these two sources.
There is a large difference between the number of X-ray sources detected with [*XMM-Newton*]{} and detected with [*Chandra*]{} in the region common to the field of view of both observatories, which we now seek to explain. The observation time of [*Chandra*]{}, 100 ks, is three times longer than that of [*XMM-Newton*]{}, 33 ks. If the observation time of [*Chandra*]{} had been as short as that of [*XMM-Newton*]{}, the number of sources detected by [*Chandra*]{} would have been smaller. More precisely, if we assume that $\sim$ 6 counts are needed to make a [*Chandra*]{} detection and that the count rates of the X-ray sources are constant, in a 33 ks [*Chandra*]{} observation, only those sources with count rates above 0.18 counts ks$^{-1}$ would have been detected (see the dotted line in Fig. \[fig\_chanc\]). Fig. \[fig\_chanc\] shows that most of the 23 X-ray sources detected by [*Chandra*]{} only with count rates above 0.18 counts ks$^{-1}$ are distributed well below the median ratio line, which indicates that these sources probably showed X-ray flares only during the 100 ks [*Chandra*]{} observation, but could disappear in a shorter, 33 ks observation. All in all, when the exposure time is set to about 30 ks for both observatories, and taking into account the differences in fields of view and detection thresholds, [*Chandra*]{} and [*XMM-Newton*]{} detect roughly the same number of sources.
In summary, we find that the respective merits of [*XMM-Newton*]{} and [*Chandra*]{}, for the same type of observations (here a typical low-mass star-forming region) are as follows. [*XMM-Newton*]{} has an effective area about three times larger than [*Chandra*]{}, so that [*XMM-Newton*]{} detects more source counts than [*Chandra*]{} for the same observation time: as shown in Fig. \[fig\_chanc\], the [*XMM-Newton*]{} count rate is 6.8 times larger than the [*Chandra*]{} count rate. On the other hand, [*Chandra*]{} has a lower background, and a sharper point spread function than [*XMM-Newton*]{}, which means [*Chandra*]{} suffers less noise than [*XMM-Newton*]{} for point source detection. For relatively short exposures (but typical in a proposal request) of 30 ks, though these values depend on position with respect to the pointing axis, the signal-to-noise ratios are similar for both observatories. For longer exposure times, [*Chandra*]{} provides a better detection ability than [*XMM-Newton*]{}, while for shorter exposure times, [*XMM-Newton*]{} provides a slightly better detection ability than [*Chandra*]{}, because the statistics with [*Chandra*]{} is photon-dominated, while for [*XMM-Newton*]{} it is background-dominated. Because it detects more photons in a given exposure, [*XMM-Newton*]{} allows a more detailed spectral and time variability analysis for moderately bright sources (above $\sim 150$ counts), typical of YSOs, than does [*Chandra*]{}.
IR properties of the [*XMM-Newton*]{} sources {#sec_ir}
=============================================
ISOCAM provides a reliable classification of YSOs in the $\rho$ Ophiuchi cloud in terms of Class I or II sources, from the detection of IR excess produced by circumstellar material (Bontemps et al. 2001). We detect X-rays from 7/11 Class I sources, 25/61 Class II sources, and 14/15 Class III sources in the overlapping region between the field of view of [*XMM-Newton*]{} and the ISOCAM survey. [^7] We find a high detection rate for Class I sources (64 %), which is higher than that of Class II sources (48 %). The [*XMM-Newton*]{} detection rate of Class I sources is comparable to that obtained from the [*Chandra*]{} observation, but the detection rate for Class II sources is lower by more than 70 % than that obtained by [*Chandra*]{} (Imanishi et al. 2001a). The Class III sample is biased because these sources were mostly identified by X-rays, which explains why their high detection rate (93 %).
Fig. \[fig\_av\] is a scatter plot of stellar luminosities ($L_\star$) vs. extinction (A$_\mathrm{V}$), for Class II (filled circles) and Class III sources (open circles) from Bontemps et al. (2001) results. Sources having [*XMM-Newton*]{} X-ray counterparts are indicated by crosses. Most of the X-ray non-detected Class II sources are distributed in the lower-right region (low $L_\star$ and high A$_\mathrm{V}$) of Fig. \[fig\_av\]. Assuming that there is an $L_X/L_\star$ correlation (see below, Sect. 6), combined with the high extinction, the X-ray non-detected Class II sources in this region could simply lie below the detection threshold of [*XMM-Newton*]{} (see Grosso et al. 2000) which, as discussed above, is higher than for [*Chandra*]{}.
However, some X-ray non-detected Class II sources, like SR24N , WSB37, and GY3, are not located in the lower-right region of Fig. \[fig\_av\]. This could be explained in two ways. First, their individual $L_X/L_\star$ ratios may be lower than typical ones. Second, their extinction may be higher than given by the statistical extinctions estimated from the NIR data (Bontemps et al. 2001). This could happen if the circumstellar disk of these Class II sources is seen close to edge-on: in such a case, the central star is not visible, and we just observe its scattered light so that a statistical value of A$_\mathrm{V}$ underestimates the real value. Conversely, two luminous Class III sources with very high extinction, WL5 (ROXN-32) and WL19 (ROXN-25), are located in the upper-right corner of Fig. \[fig\_av\]. Their X-ray spectral fits provide a column density consistent with high absorption (see Table 4). These two Class III sources are probably located [*behind*]{} the dense core F of cloud. A similar conclusion, in the much wider field of view of [*ROSAT*]{}, was reached for many Class III sources by Grosso et al. (2000). The correlation between X-ray absorption column density and optical extinction was studied in detail by Vuong et al. (2003), who compare the gas and dust properties of the dense interstellar matter in nearby star-forming regions using X-rays from Class III sources.
Considering that ISOCAM observations of Bontemps et al. (2001) were sensitive enough to detect IR excess, if any, from the new X-ray sources found in our observation, they could be Class III candidates. There are 37 [*XMM-Newton*]{} X-ray sources without IR classification in Table 2, 18 of which have 2MASS counterparts. With the exception of ROXN-10 (identified with the bona fide brown dwarf GY 141, see below, §5), we check now whether the near-IR photometry of these X-ray sources is consistent with that of Class III sources.
Fig. \[fig\_cc\] shows the color-color diagram of the [*XMM-Newton*]{} X-ray sources identified with the 2MASS near-IR sources. Class I, II, and III sources and brown dwarfs are indicated by filled diamonds, filled circles, open circles, and open diamonds, respectively; and sources without IR classification are indicated by asterisks labeled with the ROXN numbers of Table 2. As ROXN-31 is too faint to have measurable $J$ and $H$ magnitudes, it is removed from our sample, and 16 unclassified sources are plotted in Fig. \[fig\_cc\]. For ROXN-41, -44, and -49, only lower limits are available for the $J-H$ color. ROXN-87, corresponding to [*ROSAT*]{} source ROXR-F37, was identified as a foreground F2 V star, HD148352 (Grosso et al. 2000), which is consistent with its position on the locus of the intrinsic colors (Bessel & Brett 1988). This is also consistent with the fact that its X-ray hardness ratio is -1.0 for MOS1 and -1.0 for MOS2 from Table 1, which means there is no hard X-ray emission above 2 keV. ROXN-47 is also located at the position of a late-type M dwarf without extinction in Fig. \[fig\_cc\]; its hardness ratio is relatively low, $-$0.54, $-$1.00, and $-$0.76 for MOS1, MOS2, and PN, respectively. We thus consider this source also as a foreground star
Fig. \[fig\_cm\] shows the color-magnitude diagrams of $H$ vs. $J-H$ in the left panel, and $K_\mathrm{s}$ vs. $H-K_\mathrm{s}$ in the right panel for [*XMM-Newton*]{} sources having 2MASS counterparts. (Note that ROXN-31 appears only in the $K_\mathrm{s}$ vs. $H-K_\mathrm{s}$ diagram since this source is detected only in the $K_\mathrm{s}$ band.) We plot for comparison the 1 Myr isochrone (Baraffe et al. 1998) and reddening vectors (Cohen et al. 1981). The positions of the remaining 15 sources without IR classification are well mixed with those of well-known Class II and III sources. Therefore, we propose these 15 X-ray sources as new Class III candidates (labeled ‘nIII’ in Table 2), therefore doubling the present number of Class III sources in this area. Spectroscopic follow-up is now needed to determine the effective temperature of these objects, and to put them in an H-R diagram to confirm their pre-main sequence status. We estimate stellar luminosities and extinctions of these new Class III candidates using $J$ and $H$ band 2MASS magnitudes (see formulas in Bontemps et al. 2001), and plot them in Fig. \[fig\_av\]. The new Class III candidates are also mixed with the Class II and III sources in Fig. \[fig\_av\].
In summary, we find a high [*XMM-Newton*]{} detection rate for Class I and Class III sources, consistent with the [*Chandra*]{} results, and a lower detection rate for Class II sources. For those sources, the difference is probably due to a combination of high extinction (interstellar + circumstellar) and of lower detection efficiency of [*XMM-Newton*]{} compared with [*Chandra*]{}. We however detect 15 previously unknown X-ray sources, which we propose as new Class III candidates, pending spectroscopic follow-ups to confirm their nature. Our [*XMM-Newton*]{} observations thus allow for a significant improvement of the YSO census in the $\rho$ Oph cloud core F region (15 new YSOs in addition to a total of 87 previously known from X-ray/IR observations, and a potential doubling of the number of Class III sources).
X-ray detection of young bona fide brown dwarfs {#sec_bd}
===============================================
Since [*ROSAT*]{} observations, it is known that young brown dwarfs also emit X-rays (Neuh[ä]{}user & Comer[ó]{}n 1998; Neuh[ä]{}user et al. 1999; Imanishi et al. 2001a; Imanishi et al. 2001b; Preibisch & Zinnecker 2001; Mokler & Stelzer 2002; Feigelson et al. 2002b; Tsuboi el al. 2003). We thus look for X-rays from young [*bona fide*]{} brown dwarfs in our observation, i.e., from young objects with substellar status confirmed by spectroscopy. There are in our field-of-view only three [*bona fide*]{} brown dwarfs[^8]: GY202, GY141, and GY310 (Mart[í]{}n et al. 1999; Cushing et al. 2000; Wilking et al. 1999). Bontemps et al. (2001) classified GY310 as Class II, and recently Mohanty et al. (2004) detected it from the ground with [*Subaru*]{} at 8.6 and 11.7$\mu$m, confirming the presence of significant mid-infrared excess arising from an optically thick, flared dusty disk.
A low S/N ratio X-ray detection of GY202 was reported by Neuh[ä]{}user et al. (1999) from the [*ROSAT/PSPC*]{} pointing observation of Casanova et al. (1995). We find that the identification of this [*ROSAT/PSPC*]{} source with GY202, located 19 away, is dubious because the closest counterpart is in fact WL1, an embedded (A$_\mathrm{V} \sim 21$mag) Class II source (Bontemps et al. 2001), located only 9 away. Moreover, in spite of the fact that we do not detect WL1, its X-ray emission is confirmed by the [*Chandra*]{} observation (source 13 of Imanishi et al. 2001a) with a luminosity roughly consistent with the [*ROSAT/PSPC*]{} estimate, whereas GY202 is not detected either by [*Chandra*]{} (Imanishi et al. 2001b; Imanishi et al. 2003) and [*XMM-Newton*]{} (this work).
A very weak X-ray emission has been reported from GY141 with [*Chandra*]{} (Imanishi et al. 2001b; source BF-S2 in Imanishi et al. 2003): only $\sim$8 X-ray photons were collected during the 100ks exposure. According to Fig. \[fig\_chanc\] this low count rate is well below the sensitivity of our [*XMM-Newton*]{} observation. However we detect GY141 (ROXN-10) with a count rate of 7.7ctsks$^{-1}$, i.e., at a level $\sim$ 90 times higher than during the [*Chandra*]{} observation. Although we detect this brown dwarf during a phase of intense X-ray activity, we do not have enough statistics to investigate further this high X-ray state (see its background-subtracted light curve in Fig. \[fig\_lc\_bd\]).
X-ray emission was also reported from GY310 with [*Chandra*]{} (Imanishi et al. 2001a; Imanishi et al. 2001b; Imanishi et al. 2003). The [*Chandra*]{} X-ray light curve shows no clear flare, but exhibits aperiodic variability by a factor 2 within the 100ks exposure, around $\sim2\times10^{-3}$ctss$^{-1}$ (Imanishi et al. 2001b). During our observation, GY310 (ROXN-62) displayed an X-ray flare (see Fig. \[fig\_lc\_bd\]). To our knowledge this is the first X-ray flare from a young [*bona fide*]{} brown dwarf with enough counts to derive its X-ray spectrum $--$ due to [*XMM-Newton*]{}’s large effective area. The observed count rate with MOS1+MOS2 increased from 0.007ctss$^{-1}$ to 0.04ctss$^{-1}$. The quiescent level gives an equivalent [*Chandra*]{} count rate of $2\times0.007/6.8 \sim 2\times 10^{-3}$ctss$^{-1}$, consistent with the low value previously observed by [*Chandra*]{}. We obtain enough X-ray counts from GY310 during the flare to make spectral analysis presented in Sect. \[subsec\_spec\] (see Fig. \[fig\_spec\_bd\] and Table 4). The derived column density is identical to the one found by Imanishi et al. (2001b). We find a somewhat higher plasma temperature in our flare observation (2.5keV, 1.7–3.6keV), compared to the quiescent [*Chandra*]{} value (1.7keV, 0.9–2.2keV), a fairly general characteristic of stellar X-ray flares.
Similar changes in the level of X-ray activity, X-ray flares, and high plasma temperatures, are ubiquitous in low-mass protostars and T Tauri stars. This suggests that X-ray emission from brown dwarfs in their early phase of evolution are produced basically by the same solar-like, magnetic activity mechanism at work in low-mass protostars and T Tauri stars, and more generally in late-type stars, as also noticed in previous studies (Imanishi et al. 2001b, Feigelson et al. 2002b).
X-ray properties of Young Stellar Objects in the $\rho$ Ophiuchi cloud {#sec_x}
======================================================================
To investigate a possible correlation between the X-ray luminosity $L_X$ and the stellar luminosity $L_{*}$ of YSOs, we plot $L_X$ vs. $L_{*}$ for Class I, II and III sources, new Class III candidates, and brown dwarfs in Fig. \[lx\_lbol\]. We calculate absorption corrected $L_X$ in the 0.5$-$10 keV band using the best fit model for the X-ray spectra, which are listed in Table 4. In Fig. \[lx\_lbol\], we plot only the ROXN sources which are bright enough to determine their spectral parameters, except for ROXs20A (ROXN-27), ROXs20B (ROXN-28), WL5 (ROXN-32), WL4 (ROXN-34), and WL3 (ROXN-35), for which the individual X-ray spectra cannot be resolved. We use $L_{*}$ from Bontemps et al. (2001) for Class I, II, and III sources and brown dwarfs, and use our estimation of $L_{*}$ for new Class III candidates presented in section \[sec\_ir\]. For Class I sources, as the stellar luminosity is unknown because the central star is invisible due to the remnant dust envelopes, we use the bolometric luminosity as an upper limit to the stellar luminosity. The correlation index between $L_X$ and $L_{*}$ of Class II and III sources in the quiescent state is then $-$0.08, which indicates a weak correlation. Indeed the $L_X / L_{*}$ ratios show a large spread, from $4.5 \times 10^{-6}$ to $1.6 \times 10^{-2}$. New Class III candidates have similar $L_X$ and $L_{*}$ properties as other YSOs, which is consistent with the assumption of their YSO nature.
To compare the characteristics of the X-ray spectra of Class I, II, and III sources, we show in Fig. \[kt\_nh\] the scatter plot of their X-ray determined absorption column density, $N_{\rm{H}}$, vs. plasma temperature, $kT$. For ROXN-36 (SR12A-B) and ROXN-73 (IRS55), we calculated the average temperature from the multiple components weighted by their emission measures (see above, §2.3). An interesting property emerges from the $N_\mathrm{H}$ vs. $kT$ diagram: only a few data points are seen in the upper-left region and the lower-right region. The upper-left region is where the X-ray sources have a low plasma temperature and large absorption: since the soft X-rays from low temperature plasmas are easily absorbed, it is natural that X-rays should be hard to detect from these sources. On the other hand, the lower-right region is where the X-ray sources have a high plasma temperature and small absorption, and are thus easy to detect, but this region contains surprisingly few data points. This could be explained by an intrinsic effect if both the maximum temperature of YSOs and their circumstellar material causing absorption of X-rays decrease in the course of their evolution. Class I sources, which are in the early stage of the evolution, are indeed located in the upper-right region of this diagram. Class II and III sources, which are in a later stage of the evolution, are located in the region of lower temperature and smaller absorption than Class I sources, which confirms the results obtained by [*Chandra*]{} (Imanishi et al. 2001a). There is also a tendency for the temperature and absorption of Class III sources to be lower than those of Class II. We conclude that, although our statistics are still limited, an evolutionary effect seems to be present from the Class I stage (high X-ray temperature, high extinction) to the Class III stage (lower X-ray temperature, lower extinction).
Summary and conclusions {#sec_sum}
=======================
The main results of our [*XMM-Newton*]{} 33 ks, 30’ diameter observation of the $\rho$ Ophiuchi cloud core F (i.e., the main molecular core) are as follows :
1. [ We detect 87 X-ray sources; 62 sources are already known from previous X-ray observations, hence we find 25 new X-ray sources. ]{}
2. [ In the region covered by the field-of-view of both [*XMM-Newton*]{} and [*Chandra*]{} observations, 4 X-ray sources are detected only with [*XMM-Newton*]{} and 38 X-ray sources are detected only with [*Chandra*]{}, while 43 X-ray sources are detected with both observatories. We have shown that these differences are explained by the difference in exposure time and sensitivity to point source detection, along with intrinsic time variability of X-ray sources (flares). We find that [*Chandra*]{} and [*XMM-Newton*]{} are equally sensitive for point source detection in star forming-regions for exposures of [$\sim$ 30 ks]{}. ]{}
3. [ We detect X-ray emission from 7 Class I sources, 26 Class II sources, and 17 Class III sources. For Class I sources, our detection rate is high (64 %), consistent with that already observed by [*Chandra*]{} (Imanishi et al. 2001a). ]{}
4. [ For Class II sources, our detection rate is much lower (48 % vs. 70 %), which can be explained by a combination of instrumental reasons (better sensitivity for point source detection by [*Chandra*]{}) and intrinsic reasons (better probability to detect a source via X-ray flares in a longer exposure). ]{}
5. [ We find a high detection rate for Class III sources, which is not surprising in view of the fact that most Class III sources were discovered in X-rays. But we propose 15 previously unknown X-ray sources as new class III candidates, in addition to the already known 87 YSOs in the $\rho$ Oph core region (Bontemps et al. 2001). Pending spectroscopic confirmation, this would double the number of Class III sources known in this area. ]{}
6. [ We detect X-ray emission from two young bona fide brown dwarfs, GY310 and GY141. GY310 showed an X-ray flare during our observation, and GY141 appeared brighter by nearly two orders of magnitude than in the [*Chandra*]{} observation (Imanishi et al. 2001a). Their X-ray properties suggest that the X-ray emission from young brown dwarfs is the same as in low-mass protostars and T Tauri stars, i.e., resulting from solar-like magnetic activity, as proposed by previous studies (Imanishi et al. 2001b, Feigelson et al. 2002b). ]{}
7. [ We are able to extract X-ray light curves and spectra from a dozen sources, and find some of them showed X-ray flares. Altogether, from spectral fitting we confirm that there is an evolutionary trend, from Class I sources showing higher temperature and larger absorption, to Class II and III sources showing lower temperature and smaller extinction, as previously reported by the [*Chandra*]{} observation (Imanishi et al. 2001a). ]{}
H.O. acknowledges the Postdoctoral Fellowship for Research Abroad of the Japan Society for the Promotion of Science, and Conseil national des astronomes et physiciens.
[99]{} Allende Prieto, C., Lambert, D., & Asplund, M. 2001, ApJ, 556, L63 Allende Prieto, C., Lambert, D., & Asplund, M. 2002, ApJ, 573, L137 André, P., Ward-Thompson, D., Barsony, M. 1993, ApJ, 406, 122 André, P. & Montmerle, T. 1994, ApJ, 420, 837 Baraffe, I., Chabrier, G., Allard, F., Hauschildt, P. H. 1998, A&A, 337, 403 Barsony, M., Kenyon, S. J., Lada, E. A., Teuben, P. J. 1997, ApJS, 112, 109 Bessell, M. S., Brett, J. M. 1988, 100, 1134 Bontemps, S., André, P., Kaas, A. A., et al. 2001, A&A, 372, 173 Carpenter, J. M. 2001, AJ, 121, 2851 Casanova, S., 1994, Ph.D. thesis, Paris VI University Casanova, S., Montmerle, T., Feigelson, E.D., & André, P. 1995, ApJ, 439, 752 Cohen, J. G., Persson, S. E., Elias, J. H., & Frogel, J. A. 1981, ApJ, 249, 481 Cushing, M. C., Tokunaga, A. T., & Kobayashi, N. 2000, AJ, 119, 3019 Cutri, R. M., Skrutskie, M. F., van Dyk, S. et al. 2003, 2MASS All-Sky Catalog of Point Sources (Pasadena: IPAC) Feigelson, E. D. and Montmerle, T. 1999, ARA&A, 37, 363 Feigelson, E. D., Garmire, G. P., Pravdo, S. H. 2002a, ApJ, 572, 335 Feigelson, E. D., Broos, P., Gaffney, J. et al. 2002b, ApJ, 574, 258 Grosso, N., Montmerle, T., Feigelson, E. D., André, P., Casanova, S., & Gregorio-Hetem, J. 1997, Nature, 387, 56 Grosso, N., Montmerle, T., Bontemps, S., André, P., & Feigelson, E. D. 2000, A&A, 359, 113 Grosso, N. 2001, A&A, 370, L22
G[" u]{}del, M., Audard, M., Magee, H., Franciosini, E., Grosso, N., Cordova, F. A., Pallavicini, R., & Mewe, R. 2001, , 365, L344 Holweger, H. 2001, Joint SOHO/ACE workshop Solar and Galactic Composition, AIP Conf. Proc., 598, 23 Imanishi, K., Koyama, K., & Tsuboi, Y. 2001a, ApJ, 557, 747 Imanishi, K., Tsujimoto, M., & Koyama, K. 2001b, ApJ, 563, 361 Imanishi, K., Nakajima, H., Tsujimoto, M., Koyama, K., & Tsuboi, Y. 2003, PASJ, 55, 653 Jansen, F., Lumb, D., Altieri, B., et al. 2001, A&A, 365, L1 Kamata, Y., Koyama, K., Tsuboi, Y., & Yamauchi, S. 1997, PASJ, 49, 461 Koyama, K., Maeda, Y., Ozaki, M., Ueno, S., Kamata, Y., Tawara, Y., Skinner, S., & Yamauchi, S. 1994, PASJ, 46, L125 Loren, R. B.,Wootten, A., & Wilking, B. A. 1990, ApJ, 365, 269 Martín, E. L., Delfosse, X., Basri, G., Goldman, B., Forveille, T., Zapatero O., & Maria R. 1999, AJ, 118, 1005 Meyer, M. R., Calvet, N., Hillenbrand, L. A. 1997, AJ, 114, 288 Mokler, F., & Stelzer, B. 2002, A&A, 391, 1025 Mohanty, S., Jayawardhana, R., Natta, A., Fujiyoshi, T., et al. 2004, ApJL, 609, L33 Monet,D. 1996, USNO A$-$1.0 Montmerle, T., Koch-Miramond, L., Falgarone, E., & Grindlay, J. E. 1983, ApJ, 269, 182 Montmerle, T., Feigelson, E. D., Bouvier, J., André, P. 1993, in Protostars and Planets III Neuhäuser, R., Briceno, C., Comerón, F. et al. 1999, A&A, 343, 883 Neuhäuser, R. & Comerón, F. 1998, Science, 282, 83 Preibisch, T. & Zinnecker, H. 2001, AJ, 122, 866 Tsuboi, Y., Imanishi, K., Koyama, K., Grosso, N., & Montmerle, T. 2000, ApJ, 532, 1089 Tsuboi, Y., Maeda, Y., Feigelson, E. D., et al. 2003, ApJ 587, L51 Vuong, M. H., Montmerle, T., Grosso, N., Feigelson, E.D., Verstraete, L., and Ozawa, H. 2003, A&A, 408, 581 Wilking B.A., Greene T.P., Meyer M.R., 1999, AJ 117, 469
Method to compute the X-ray count rate upper limit
===================================================
XMM-Newton data
---------------
For each instrument, MOS1, MOS2, and PN, we compute with the SAS command [eexpmap]{} the corresponding exposure map with a spatial resolution of 2$^{\prime\prime}$ for X-ray photons of 1.5keV energy, and normalize it by its maximum value located on-axis. The resulting bidimensionnal array, $\cal E$, represents the spatial effective area variations, including CCD gaps, bad pixels, and vignetting, relative to the on-axis value. From $\cal E$ we define an exposure mask, $\cal M$, having $\cal M$(i)=0 if $\cal E$(i)=0, and $\cal M$(i)=1 if $\cal E$(i)$\neq$0. This exposure mask shows only CCD gaps.
We estimate for each position of undetected Chandra source the local average background. We extract the event number, $N_\mathrm{bgd}$, inside the 15-radius disk, $\cal D_\mathrm{bgd}$, of geometrical area $A_\mathrm{bgd}^\mathrm{geo}$ centered on the Chandra source position. This extraction region contains $n^{\cal E}_\mathrm{bgd}$ sky pixels of $\cal E$, and $n^{*{\cal M}}_\mathrm{bgd}$ sky pixels of $\cal M$ where $\cal M$ is equal to 1, i.e. which are not CCD gaps. The net pixel area of $\cal D_\mathrm{bgd}$ is : $$A_\mathrm{bgd}^\mathrm{net} = A_\mathrm{bgd}^\mathrm{geo} \, n^{*{\cal M}}_\mathrm{bgd}\,/\,n^{\cal E}_\mathrm{bgd}.$$ The average local background per sky pixel is obtained by : $$<\!B\!>\,= N_\mathrm{bgd}/A_\mathrm{bgd}^\mathrm{net}.$$
This 15-radius size is a compromise for crowded region. With this local method, local background variations as PSF wings of neighbouring X-ray source, or trails due to out-of-time events of bright sources, are automatically included in our background estimate. This would not be the case if we had used the spline-smoothed background built by [esplinemap]{}, because it uses a ‘Swiss-cheese’ method which subtracts the source PSF from the X-ray image.
Then we compute the PSF radius encircling $f_\mathrm{ee}=50\%$ of the PSF energy, $R(f_\mathrm{ee})$, using the radial average of the calibration images in the CCF. We extract the event number, $N_\mathrm{src+bgd}(f_\mathrm{ee})$, within the $R(f_\mathrm{ee})$-radius disk, $\cal D_\mathrm{src}$, of geometrical area $A_\mathrm{src}^\mathrm{geo}$ centered on the Chandra source position. These counts are the sum of background and source counts, $N_\mathrm{src}(f_\mathrm{ee})$ and $N_\mathrm{bgd}(f_\mathrm{ee})$, respectively. The extraction region $A_\mathrm{src}^\mathrm{geo}$ contains $n^{\cal E}_\mathrm{src}$ sky pixels of $\cal E$, and $n^{*{\cal M}}_\mathrm{src}$ sky pixels of $\cal M$ where $\cal M$ is equal to 1. The net pixel area of $\cal D_\mathrm{src}$ is : $$A_\mathrm{src}^\mathrm{net} = A_\mathrm{src}^\mathrm{geo} \, n^{*{\cal M}}_\mathrm{src} \,/\, n^{\cal E}_\mathrm{src}.$$ We compute for $N_\mathrm{src+bgd}(f_\mathrm{ee})$ its 99.9989% confidence level (corresponding to 4.4$\sigma$ in Gaussian statistics, i.e. the threshold of our detection algorithm) upper limit, $N^\mathrm{ul}_\mathrm{src+bgd}(f_\mathrm{ee})$, using the approximate formula (10) of Gehrels (1986). Assuming that the background is constant on $\cal D_\mathrm{src}$, the upper limit on the source count in $\cal D_\mathrm{src}$ is : $$\begin{aligned}
N^\mathrm{ul}_\mathrm{src}(f_\mathrm{ee}) & = & \left\{ N_\mathrm{src+bgd}^\mathrm{ul}(f_\mathrm{ee})
- <\!B\!> A_\mathrm{src}^\mathrm{net} \right\} \nonumber \\
& & \times \frac{n^{\cal E}_\mathrm{src}}{\sum_{\mathrm{i} \in {\cal D_\mathrm{src}}} {\cal E}(i)}\,.\end{aligned}$$ The right hand side multiplicative factor corrects in average from loss of sensitivity due to bad pixels, CCD gaps, and vignetting. For Chandra sources which are over CCD gaps or bad pixels of XMM-Newton (a minority in our sample), a more accurate correction would involve a convolution of the PSF shape with the exposure mask.
The upper limit on the total source count is hence : $$N^\mathrm{ul}_\mathrm{src}=N^\mathrm{ul}_\mathrm{src}(f_\mathrm{ee})\,/\,f_\mathrm{ee} \,.$$ Finally, the count rate upper limit is obtained by dividing $N^\mathrm{ul}_\mathrm{src}$ by the livetime of the corresponding CCD.
We compute the count rate upper limit for the sum of the MOS1, MOS2, and PN data, the so-called EPIC instrument, $CR^\mathrm{ul}_\mathrm{src,EPIC}$, using this set of straightforward formulae using the previous notation, where $i \in [\mathrm{MOS1,MOS2,PN}]$ :
$$N_\mathrm{src+bgd,EPIC}(f_\mathrm{ee})=\sum_\mathrm{i}\,N_\mathrm{src+bgd,i}(f_\mathrm{ee})\, ,$$
$$\begin{aligned}
N^\mathrm{ul}_\mathrm{src,EPIC}(f_\mathrm{ee}) & = & \left\{ N_\mathrm{src+bgd,i}^\mathrm{ul}(f_\mathrm{ee}) - \sum_\mathrm{i} <\!B\!>_\mathrm{i} A_\mathrm{src,i}^\mathrm{net} \right\} \nonumber \\
& & \times \frac{\sum_\mathrm{i} n^{{\cal E}_\mathrm{i}}_\mathrm{src} }{\sum_\mathrm{i} \sum_{\mathrm{j} \in {\cal D}_\mathrm{src,i}} {\cal E}_\mathrm{i}(j)}
\, ,\end{aligned}$$
$$N^\mathrm{ul}_\mathrm{src,EPIC} = N^\mathrm{ul}_\mathrm{src,EPIC}(f_\mathrm{ee})\,/\,f_\mathrm{ee} \, ,$$
$$CR^\mathrm{ul}_\mathrm{src,EPIC}=N^\mathrm{ul}_\mathrm{src,EPIC}\,/\/\sum_\mathrm{i} LIVETIME_\mathrm{\,i}\, .$$
Chandra data
------------
We get the level 2 data of the Core F observation of the $\rho$ Ophiuchi dark cloud (sequence number 200060, observation ID 635) from the Chandra data archive[^9]. We compute the exposure map with a resolution of 4$^{\prime\prime}$ for X-ray photons of 1.5keV energy, and normalize it by its maximum value (on-axis), from which we derive an exposure mask. We proceed as for XMM-Newton data, using for the PSF radius encircling 50% of the PSF energy the formula given by Feigelson et al. (2002b) : $R(50\%)=0.43-0.1\,\theta +0.05\,\theta\,^2$, where $\theta$ is the off-axis angle in arcmin.
We note that for Chandra data, thanks to the large satellite wobbling, bad pixels and CCD gaps are smooth in the exposure map, they do not produce holes with sharp edges, hence with the above notation we have $n^{*{\cal M}} / n^{\cal E} =1$. By combining (1)–(3), we find a formula identical to the formula (6) of Feigelson et al. (2002b) when replacing the 90% upper limit count by the observed count.
Comparison between [*Chandra*]{} and [*XMM-Newton*]{} determination of $N_\mathrm{H}$
=====================================================================================
To check the reliability of the X-ray derived values of $N_\mathrm{H}$, we make a plot of the $N_\mathrm{H}$ values determined by [*Chandra*]{} vs. those determined by [*XMM-Newton*]{} for the X-ray sources with spectral fitting data (see Table. 4) in the overlapping field-of-views of the two satellites (see Fig. \[fig\_nh\]). A best fit value of the ratio of the $N_\mathrm{H}$ values from [*Chandra*]{} and [*XMM-Newton*]{} obtained from a linear fit is 0.96 ( 0.93$-$0.99, 90 % confidence region), indicating that the both determinations of $N_\mathrm{H}$ values are consistent better than 10%.
Uncertainties of the X-ray derived $N_\mathrm{H}$ values were discussed for both [*Chandra*]{} and [*XMM-Newton*]{} in Vuong et al. (2003). They showed that the use of the recently revised solar abundances (Holweger 2001, Allend Prieto et al. 2001, 2002) in X-ray spectral fitting increases $N_\mathrm{H}$ values by $\sim$20 %. Both in the X-ray spectral fitting by [*Chandra*]{} (Imanishi et al. 2001a) and by [*XMM-Newton*]{} (this work, see Sect. \[subsec\_spec\]), the $N_\mathrm{H}$ values were determined using WABS absorption model in XSPEC where solar standard metal abundances are used. If the abundance effect shown by Vuong et al. (2003) is taken into account, all the $N_\mathrm{H}$ values derived by [*Chandra*]{} and [*XMM-Newton*]{} would increase and all the data points in Fig. \[fig\_nh\] would move towards upper-right direction by $\sim$20 % along the diagonal.
[^1]: There is a coding error in [*emldetect*]{} until SAS version 5.4 ([*XMM-Newton*]{}-News \#29, 11-Mar-2003), producing an overestimate of the source likelihood. We corrected the source likelihood using the table provided in the [*XMM-Newton*]{}-Newsletter.
[^2]: The official naming convention for [*XMM-Newton*]{} sources is XMMU JHHMMSS.s$+/-$DDMMSS.
[^3]: ROXN-1 was so bright that the CCD pixels on the source are strongly affected by pile-up. Hence $HR$ computed for this source is not reliable.
[^4]: See calibration files XRT1 XPSF 0005.CCF, XRT2 XPSF 0005.CCF, and XRT3 XPSF 0004.CCF.
[^5]: For ROXN-1, we exclude a circular region of 17.5$^{\prime\prime}$ radius centered on the source to keep the spectrum from being affected by pile-up.
[^6]: We check consistency of $N_\mathrm{H}$ determination between [*Chandra*]{} and [*XMM-Newton*]{} in Appendix B. We mention also uncertainties of X-ray derived $N_\mathrm{H}$ values especially for an effect of metal abundances for the absorption model.
[^7]: Note that the prototype of Class 0 sources, VLA1623 (André et al. 1993) is located in the dense core A, which is not covered by our observation.
[^8]: We note in Fig. \[fig\_cm\] that ROXN-82 is located well below the substellar limit of $\sim$ 0.08M$_{\sun}$ and could be a brown dwarf candidate, but it was found to be a very low-mass M star by a near-IR spectroscopical follow-up (Grosso et al. 2004, in preparation).
[^9]: [http://asc.harvard.edu/cgi-gen/cda/retrieve5.pl.]{}
| {
"pile_set_name": "ArXiv"
} | ArXiv |
=24.8cm=-1.2cm=15.3cm -0.1cm-0.1cm
------------------
HEPHY-PUB 714/99
UWThPh-1999-30
hep-ph/9905556
May 1999
------------------
\
[**THE ONE-DIMENSIONAL SPINLESS\
RELATIVISTIC COULOMB PROBLEM**]{}\
\
Institut für Hochenergiephysik,\
Österreichische Akademie der Wissenschaften,\
Nikolsdorfergasse 18, A-1050 Wien, Austria\
[**Franz F. SCHÖBERL**]{}\
Institut für Theoretische Physik,\
Universität Wien,\
Boltzmanngasse 5, A-1090 Wien, Austria
Motivated by a recent analysis which presents explicitly the general solution, we consider the eigenvalue problem of the spinless Salpeter equation with a (“hard-core amended”) Coulomb interaction potential in one dimension. We prove the existence of a critical coupling constant (which contradicts the assertions of the previous analysis) and give analytic upper bounds on the energy eigenvalues. These upper bounds seem to disprove the previous explicit solution.
[*PACS*]{}: 03.65.Ge, 03.65.Pm
Introduction: The Spinless Salpeter Equation in One Dimension
=============================================================
The spinless Salpeter equation arises either as a standard reduction of the well-known Bethe–Salpeter formalism [@SSE] for the description of bound states within the framework of relativistic quantum field theory or as a straightforward relativistic generalization of the nonrelativistic Schrödinger equation. This semirelativistic equation of motion with a static interaction described by the Coulomb potential (originating, for instance, from the exchange of a massless particle between the bound-state constituents) defines what we call, for short, the “spinless relativistic Coulomb problem.[^1]”
Recently, confining the configuration space to the positive half-line (and mimicking thereby the effect of a “hard-core”), the relativistic Coulomb problem has been studied in one dimension [@Brau99]. This one-dimensional case may serve as a toy model which might prove to be instructive for the analysis of the still unsolved three-dimensional problem. In view of its potential importance, we re-analyze this nontrivial and delicate problem.
The spinless Salpeter equation may be regarded as the eigenvalue equation $$H|\chi_k\rangle=E_k|\chi_k\rangle\ ,\quad k=1,2,3,\dots\ ,$$for the complete set of Hilbert-space eigenvectors $|\chi_k\rangle$ and corresponding eigenvalues $$E_k\equiv\frac{\langle\chi_k|H|\chi_k\rangle}{\langle\chi_k|\chi_k\rangle}$$ of a self-adjoint operator $H$ of Hamiltonian form, consisting of a momentum-dependent kinetic-energy operator and a coordinate-dependent interaction-potential operator:$$H=T+V\
,\label{Eq:SRH}$$where $T$ is the “square-root” operator of the relativistic kinetic energy of some particle of mass $m$ and momentum $p$,$$T=T(p)\equiv\sqrt{p^2+m^2}\
,\label{eq:RKE}$$and $V=V(x)$ is an arbitrary, coordinate-dependent, static interaction potential. The action of the kinetic-energy operator $T$ on an element $\psi$ of $L_2(R)$, the Hilbert space of square-integrable functions on the real line, $R$, is defined by (cf. also Eq. (3) of Ref. [@Brau99]) $$(T\psi)(x)=\frac{1}{2\pi}\int\limits_{-\infty}^{+\infty} {\rm
d}p\int\limits_{-\infty}^{+\infty}{\rm d}y\,\sqrt{p^2+m^2}\,\exp[{\rm
i}\,p\,(x-y)]\,\psi(y)\ .\label{Eq:AOT}$$
In Ref. [@Brau99], the domain of $H$ is restricted to square-integrable functions $\Psi(x)$ with support on the positive real line, $R^+$, only, vanishing at $x=0$ (cf. Eq. (28) of Ref. [@Brau99]): $$\Psi(x)=0\quad\mbox{for}\ x\le0\ .$$This restriction may be interpreted as due to the presence of a “hard-core” interaction potential effective for $x\le0$. For $x>0$, the interaction potential $V$ is chosen to be of Coulomb type, its strength parametrized by a positive coupling constant $\alpha$, i.e., $\alpha>0$:$$V(x)=V_{\rm C}(x)=-\frac{\alpha}{x}\quad\mbox{for}\ x>0\
.$$Let the Coulomb-type semirelativistic Hamiltonian $H_{\rm C}$ be the operator defined in this way.
Concerns — Dark Clouds Appear at the Horizon
============================================
Now, according to the analysis of Ref. [@Brau99], the point spectrum of the Hamiltonian $H_{\rm
C}$ consists of the set of eigenvalues (cf. Eq. (33) of Ref. [@Brau99])$$\widetilde
E_n=\frac{m}{\displaystyle\sqrt{1+\frac{\alpha^2}{n^2}}}\ ,\quad
n=1,2,3,\dots\ .\label{Eq:EEVs}$$The corresponding eigenfunctions $\Psi_n(x)$ must be of the form (cf. Eq. (28) of Ref. [@Brau99])$$\Psi_n(x)=\psi_n(x)\,\Theta(x)\ ,\quad
n=1,2,3,\dots\ ,\label{Eq:HEF}$$where $\Theta(x)$ denotes the Heaviside step function, defined here by$$\begin{aligned}
\Theta(x)&=&1\quad\mbox{for}\ x>0\
,\\[1ex]\Theta(x)&=&0\quad\mbox{for}\ x\le0\ .\end{aligned}$$In particular, the (not normalized) eigenfunctions $\psi_n(x)$, $n=1,2,3,$ corresponding to the lowest energy eigenvalues $\widetilde E_n$ are explicitly given by (cf. Eqs. (37)–(40) of Ref. [@Brau99])$$\begin{aligned}
\psi_1(x)&=&x\,\exp(-\beta_1\,x)\ ,\nonumber\\[1ex]
\psi_2(x)&=&x\left(x-\frac{m^2}{S_2^2\,\beta_2}\right)\exp(-\beta_2\,x)\
,\nonumber\\[1ex]
\psi_3(x)&=&x\left(x^2-\frac{3\,m^2}{S_3^2\,\beta_3}\,x+
\frac{3\,m^2\,(\beta_3^2+m^2)}{2\,S_3^4\,\beta_3^2}\right)\exp(-\beta_3\,x)\
,\label{Eq:LEF}\end{aligned}$$with (cf. Eq. (32) of Ref. [@Brau99])$$\beta_n\equiv
\frac{m\,\alpha}{n\,\displaystyle\sqrt{1+\frac{\alpha^2}{n^2}}}=
\frac{\alpha}{n}\,\widetilde E_n\ ,\quad n=1,2,3,\dots\ ,$$and the abbreviation (cf. Eq. (26) of Ref. [@Brau99])$$S_n\equiv\sqrt{m^2-\beta_n^2}\ ,\quad n=1,2,3,\dots\ .$$
However, there are some facts which cause severe doubts about the validity of this solution:
Boundedness from below:
: For coupling constants $\alpha$ larger than some critical value $\alpha_{\rm c}$ (which has yet to be determined), the operator $H_{\rm C}$ is not bounded from below. This may be seen, for instance, already from the expectation value of $H_{\rm C}$ with respect to the (normalized) trial state $|\Phi\rangle$ defined by the configuration-space trial function $$\Phi(x)=\varphi(x)\,\Theta(x)$$with $$\varphi(x)=2\,\mu^{3/2}\,x\,\exp(-\mu\,x)\ ,\quad\mu>0\ ,$$and satisfying the normalization condition$$\||\Phi\rangle\|^2\equiv\langle\Phi|\Phi\rangle=
\int\limits_0^\infty{\rm d}x\,|\varphi(x)|^2=1\ .$$Apart from the arbitrariness of the variational parameter $\mu$, this trial function $\Phi$ coincides, in fact, with the ground-state solution $\Psi_1$ as given in Eqs. (\[Eq:HEF\]), (\[Eq:LEF\]). The expectation value of the Coulomb interaction-potential operator $V_{\rm C}$ with respect to the trial state $|\Phi\rangle$ reads$$\langle\Phi|V_{\rm
C}|\Phi\rangle=-\alpha\int\limits_0^\infty{\rm
d}x\,\frac{1}{x}|\varphi(x)|^2=-\mu\,\alpha\ .$$There is a trivial (but nevertheless fundamental) inequality for the expectation values of a self-adjoint (but otherwise arbitrary) operator ${\cal O}={\cal O}^\dagger$ and its square, taken with respect to an arbitrary Hilbert-space state $|\psi\rangle$ in the domain ${\cal D}({\cal O})$ of this operator ${\cal
O}$:$$\frac{|\langle\psi|{\cal
O}|\psi\rangle|}{\langle\psi|\psi\rangle}\le\sqrt{\frac{\langle\psi|{\cal
O}^2|\psi\rangle}{\langle\psi|\psi\rangle}}\quad\mbox{for all}\
|\psi\rangle\in{\cal D}({\cal O})\ .$$Application of this inequality to the kinetic-energy operator $T$ of Eq. (\[eq:RKE\]) allows to get rid of the troublesome square-root operator:$$\langle\Phi|T|\Phi\rangle\le
\sqrt{\langle\Phi|T^2|\Phi\rangle}\equiv\sqrt{\langle\Phi|p^2|\Phi\rangle+m^2}\
.$$The expectation value of $p^2$ required here reads$$\langle\Phi|p^2|\Phi\rangle=\mu^2\ .$$Thus the expectation value of the Coulomb-like semirelativistic Hamiltonian $H_{\rm C}$ with respect to the trial state $|\Phi\rangle$ is bounded from above by $$\langle\Phi|H_{\rm C}|\Phi\rangle=
\langle\Phi|T+V_{\rm C}|\Phi\rangle\le\sqrt{\mu^2+m^2}-\mu\,\alpha\
.\label{Eq:EVOCH}$$When inspecting this inequality in the limit of large $\mu$, that is, for $\mu\to\infty$, one realizes that, for $\alpha$ large enough, the operator $H_{\rm C}$ is not bounded from below. In fact, the expectation value of the kinetic-energy operator $T$ with respect to the trial state $|\Phi\rangle$,$$\langle\Phi|T|\Phi\rangle
=\int\limits_{-\infty}^{+\infty}{\rm d}x\,\Phi^\ast(x)\,(T\Phi)(x)
=\frac{4\,\mu^3}{\pi}\int\limits_0^\infty{\rm d}p\,
\frac{\displaystyle\sqrt{p^2+m^2}}{(p^2+\mu^2)^2}\
,\label{Eq:EVKE}$$is simple enough to be investigated explicitly. For $\mu\gg m$, this expectation value simplifies to$$\langle\Phi|T|\Phi\rangle=\frac{2\,\mu}{\pi}\quad\mbox{for}\ \mu\gg m\
.$$Consequently, in the (ultrarelativistic) limit $\mu\to\infty$, the expectation value of $H_{\rm C}$ behaves like$$\lim_{\mu\to\infty}\frac{\langle\Phi|H_{\rm
C}|\Phi\rangle}{\mu}=\frac{2}{\pi}-\alpha\ .$$This clearly indicates that for the Hamiltonian $H_{\rm C}$ to be bounded from below the Coulomb coupling constant $\alpha$ has to be bounded from above by the critical value$$\alpha_{\rm c}\le\frac{2}{\pi}\ .$$(This upper bound on $\alpha_{\rm
c}$ is, in fact, identical to the critical coupling constant $\alpha_{\rm c}$ found in the case of the three-dimensional spinless relativistic Coulomb problem [@Herbst].)
Upper bound on lowest eigenvalue:
: As rather trivial consequence of the famous minimum–maximum principle [@MMP], the expectation value$$\frac{\langle\psi|H|\psi\rangle}{\langle\psi|\psi\rangle}$$of a self-adjoint operator $H$ bounded from below, with respect to some arbitrary state $|\psi\rangle$ in the domain of $H$, ${\cal D}(H)$, is always larger than or equal to the lowest eigenvalue $E_1$ of $H$:[^2]$$E_1\le
\frac{\langle\psi|H|\psi\rangle}{\langle\psi|\psi\rangle}\quad\mbox{for all}\
|\psi\rangle\in{\cal D}(H)\ .$$Accordingly, minimizing the expression on the right-hand side of inequality (\[Eq:EVOCH\]) with respect to the variational parameter $\mu$ yields a simple analytic upper bound $\widehat
E_1$ on the ground-state energy eigenvalue $E_1$ of the Coulomb-like semirelativistic Hamiltonian $H_{\rm C}$:$$E_1\le\widehat
E_1$$with$$\widehat E_1=m\,\sqrt{1-\alpha^2}\
.\label{Eq:AUBGSE}$$The same analytic upper bound on the ground-state energy $E_1$ has been found in the case of the three-dimensional spinless relativistic Coulomb problem [@Lucha94varbound; @Lucha96:AUB; @Lucha98:T]. Reality of this latter expression requires again the existence of a critical coupling constant $\alpha_{\rm c}$ and indicates that this critical value of $\alpha$ is less than or equal to 1:$$\alpha_{\rm c}\le1\ .$$Moreover, at least for the energy eigenvalue $E_1$ corresponding to the ground state of the Hamiltonian $H_{\rm
C}$, the supposedly [*exact*]{} value of Eq. (\[Eq:EEVs\]), $$\widetilde E_1=\frac{m}{\displaystyle\sqrt{1+\alpha^2}}\
,\label{Eq:GSE}$$is in clear conflict with the naive [*upper bound*]{} $\widehat E_1$ of Eq. (\[Eq:AUBGSE\]):$$\frac{\widehat
E_1}{\widetilde E_1}=\sqrt{1-\alpha^4}$$and therefore$$\widehat
E_1<\widetilde E_1\quad\mbox{for}\ \alpha>0\ .$$For larger values of the Coulomb coupling constant $\alpha$, the upper bound (\[Eq:AUBGSE\]) on the ground-state energy can be easily improved by fixing in the expectation value (\[Eq:EVKE\]) of the kinetic-energy operator $T$ the variational parameter $\mu$ to the value $\mu=m$. In this case, this expectation value reads$$\langle\Phi|T|\Phi\rangle=\frac{4\,m}{\pi}\
.$$Accordingly, the ground-state energy eigenvalue $E_1$ is bounded from above by$$E_1\le\left(\frac{4}{\pi}-\alpha\right)m\
.\label{Eq:AUBGSE-mu=m}$$For the Coulomb coupling constant $\alpha$ in the range $$\frac{2}{\pi}-\sqrt{\frac{1}{2}-\frac{4}{\pi^2}}<\alpha\le\frac{2}{\pi}\
,$$the above expression represents a genuine improvement of the upper bound (\[Eq:AUBGSE\]).
Eigenstate expectation values vs. eigenvalues:
: The expectation value (\[Eq:EVKE\]) of the kinetic-energy operator $T$ with respect to the trial state $|\Phi\rangle$ may be written down explicitly:$$\langle\Phi|T|\Phi\rangle=\frac{2}{\pi}\,m\left[\frac{\mu}{m}
+\displaystyle\frac{\arccos\displaystyle\frac{\mu}{m}}
{\displaystyle\sqrt{1-\frac{\mu^2}{m^2}}}\right].$$Now, for$$\mu=\beta_1=\frac{m\,\alpha}{\displaystyle\sqrt{1+\alpha^2}}\ ,$$the trial function $\Phi$ coincides with the normalized ground-state eigenfunction $\Psi_1$. In this case, the corresponding expectation value of the Hamiltonian $H_{\rm C}$ becomes$$\langle\Psi_1|H_{\rm
C}|\Psi_1\rangle=
\frac{m}{\displaystyle\sqrt{1+\alpha^2}}\left[\frac{2}{\pi}\left(\alpha+
(1+\alpha^2)\,\mbox{arccot}\,\alpha\right)-\alpha^2\right].
\label{Eq:GSE-expl}$$Unfortunately, the above expectation value does not agree with the ground-state energy (\[Eq:GSE\]) deduced from Eq. (\[Eq:EEVs\]):$$\langle\Psi_1|H_{\rm C}|\Psi_1\rangle\ne\widetilde E_1\
.$$
Orthogonality of eigenstates:
: Eigenstates $|\chi_i\rangle$, $i=1,2,3,\dots,$ of some self-adjoint operator $H$ corresponding to distinct eigenvalues of $H$ are mutually orthogonal:$$\langle\chi_i|\chi_k\rangle\propto\delta_{ik}\ ,\quad
i,k=1,2,3,\dots\ .$$This feature is definitely not exhibited by the overlaps$$\langle\Psi_i|\Psi_k\rangle=\int\limits_{-\infty}^{+\infty}{\rm
d}x\,\Psi_i^\ast(x)\,\Psi_k(x)=\int\limits_0^\infty{\rm
d}x\,\psi_i^\ast(x)\,\psi_k(x)\ ,\quad i,k=1,2,3,\dots\ ,$$of the lowest eigenfunctions $\Psi_i(x)$, $i=1,2,3,$ given in Eqs. (\[Eq:HEF\]), (\[Eq:LEF\]). For instance, the overlap $\langle\Psi_1|\Psi_2\rangle$ of the ground state $|\Psi_1\rangle$ and the first excitation $|\Psi_2\rangle$ is given by$$\langle\Psi_1|\Psi_2\rangle=
\frac{2\,[3\,S_2^2\,\beta_2-m^2\,(\beta_1+\beta_2)]}
{(\beta_1+\beta_2)^4\,S_2^2\,\beta_2}\ ,$$revealing thus, beyond doubt, the non-orthogonality of the vectors $|\Psi_1\rangle$ and $|\Psi_2\rangle$.
Exact Analytic Upper Bounds on Energy Levels
============================================
In view of the above, let us try to collect unambiguous results for the one-dimensional spinless relativistic Coulomb problem. With the help of the definition (\[Eq:AOT\]) of the action of a momentum-dependent operator in coordinate space, it is easy to convince oneself of the validity of the operator inequality$$T\le
T_{\rm NR}\equiv m+\frac{p^2}{2\,m}\ ;$$the relativistic kinetic-energy operator $T$ is bounded from above by its nonrelativistic counterpart $T_{\rm
NR}$: when introducing the Fourier transform $\widetilde\psi(p)$ of the coordinate-space representation $\psi(x)$ of the Hilbert-space vector $|\psi\rangle$,$$\widetilde\psi(p)\equiv\frac{1}{\sqrt{2\pi}}
\int\limits_{-\infty}^{+\infty} {\rm d}x\,\exp[-{\rm i}\,p\,x]\,\psi(x)\
,$$one finds$$\begin{aligned}
\langle\psi|T_{\rm NR}-T|\psi\rangle&=&
\int\limits_{-\infty}^{+\infty}{\rm d}x\,\psi^\ast(x)\left[(T_{\rm
NR}\psi)(x)-(T\psi)(x)\right]\\[1ex]&=&
\int\limits_{-\infty}^{+\infty}{\rm d}p\,|\widetilde\psi(p)|^2\left(
m+\frac{p^2}{2\,m}-\sqrt{p^2+m^2}\right)\\[1ex]&\ge&0\ .\end{aligned}$$ Hence, adding the Coulomb interaction potential $V_{\rm C}$, the semirelativistic Hamiltonian $H_{\rm C}$ is, of course, bounded from above by the corresponding nonrelativistic Hamiltonian $H_{\rm C,NR}$:$$H_{\rm C}\le
H_{\rm C,NR}\equiv T_{\rm NR}+V_{\rm C}\ .$$Now, upon invoking the minimum–maximum principle [@MMP] (which requires the operator $H_{\rm
C}$ to be both self-adjoint and bounded from below) and combining this principle with the above operator inequality, we infer that every eigenvalue $E_n$, $n=1,2,3,\dots,$ of $H_{\rm C}$ is bounded from above by a corresponding eigenvalue $E_{n,\rm NR}$, $n=1,2,3,\dots,$ of $H_{\rm
C,NR}$:[^3]$$E_n\le E_{n,\rm NR}\quad\mbox{for}\
n=1,2,3,\dots\ .$$It is a simple and straightforward exercise to calculate the latter set of eigenvalues:$$E_{n,\rm
NR}=m\left(1-\frac{\alpha^2}{2\,n^2}\right),\quad n=1,2,3,\dots\ .$$These upper bounds on the energy eigenvalues $E_n$ may be easily improved by the same reasoning as before. Introducing an arbitrary real parameter $\eta$ (with the dimension of mass), we find a set of operator inequalities for the kinetic energy $T$ [@Lucha96:AUB; @Lucha98:T; @Martin], namely, $$T\le\frac{p^2+m^2+\eta^2}{2\,\eta}\quad\mbox{for all}\ \eta>0$$and, consequently, a set of operator inequalities for the Coulomb-type semirelativistic Hamiltonian $H_{\rm C}$ [@Lucha96:AUB; @Lucha98:T]: $$H_{\rm C}\le\widehat H_{\rm
C}(\eta)\equiv\frac{p^2+m^2+\eta^2}{2\,\eta}+V_{\rm C}\quad\mbox{for all}\
\eta>0\ .$$Accordingly, every eigenvalue $E_n$, $n=1,2,3,\dots,$ of $H_{\rm
C}$ is bounded from above by the minimum, with respect to the mass parameter $\eta$, of the corresponding eigenvalue [@Lucha96:AUB; @Lucha98:T] $$\widehat E_{n,\rm C}(\eta)=\frac{1}{2\,\eta}\left[m^2+\eta^2
\left(1-\frac{\alpha^2}{n^2}\right)\right],\quad n=1,2,3,\dots\ ,$$of $\widehat H_{\rm C}(\eta)$:$$E_n\le\displaystyle\min_{\eta>0}\widehat
E_{n,\rm C}(\eta)=m\,\sqrt{1-\frac{\alpha^2}{n^2}}\quad\mbox{for all}\
\alpha\le\alpha_{\rm c}\ .$$For $n=1$, this (variational) upper bound coincides with the previous upper bound (\[Eq:AUBGSE\]). It goes without saying that these upper bounds are violated by the energy eigenvalues $\widetilde E_n$ given in Eq. (\[Eq:EEVs\]):$$\frac{1}{\widetilde
E_n}\,\displaystyle\min_{\eta>0}\widehat E_{n,\rm
C}(\eta)=\sqrt{1-\frac{\alpha^4}{n^4}}<1\quad\mbox{for}\ \alpha\ne 0\
,\quad\mbox{for all}\ n=1,2,3,\dots\
,$$means$$\displaystyle\min_{\eta>0}\widehat E_{n,\rm C}(\eta)<\widetilde
E_n\quad\mbox{for}\ \alpha\ne 0\ ,\quad\mbox{for all}\ n=1,2,3,\dots\ !$$
Moreover, for $\mu=m\,\alpha$, our generic trial state $|\Phi\rangle$ becomes the lowest eigenstate of the nonrelativistic Hamiltonian $H_{\rm C,NR}$, corresponding to the ground-state eigenvalue[^4]$$E_{1,\rm
NR}=m\left(1-\frac{\alpha^2}{2}\right),$$which may be easily seen:$$(T_{\rm
NR}\varphi)(x)=\left(m-\frac{1}{2\,m}\,\frac{{\rm d}^2}{{\rm
d}x^2}\right)\varphi(x)=
\left(m-\frac{\mu^2}{2\,m}+\frac{\mu}{m}\,\frac{1}{x}\right)\varphi(x)\quad
\mbox{for}\ x>0$$implies (with $\mu=m\,\alpha$)$$H_{\rm
C,NR}|\Phi\rangle=E_{1,\rm NR}|\Phi\rangle\ .$$It appears rather unlikely that the same functional form represents also the eigenstate of the semirelativistic Hamiltonian $H_{\rm C}$.
Summary, Further Considerations, Conclusions
============================================
This work is devoted to the study of the one-dimensional spinless relativistic Coulomb problem on the positive half-line. Assuming a (dense) domain in $L_2(R^+)$ such that the semirelativistic Coulombic Hamiltonian $H_{\rm C}$ defined in the Introduction is self-adjoint, analytic upper bounds on the energy eigenvalues $E_k$, $k=1,2,3,\dots,$ have been derived:$$E_k\le
m\,\sqrt{1-\frac{\alpha^2}{k^2}}\quad\mbox{for all}\ k=1,2,3,\dots\
.\label{Eq:AUBOELFSRCP}$$Surprisingly, the explicit solution presented in Ref. [@Brau99] does not fit into these bounds.
In order to cast some light into this confusing situation, let us inspect the action (\[Eq:AOT\]) of the kinetic-energy operator $T$ in more detail. Consider not normalized Hilbert-space vectors $|\Phi_n\rangle$, $n=0,1,2,\dots,$ defined, as usual, by the coordinate-space representation $$\Phi_n(x)=x^n\,\exp(-\mu\,x)\,\Theta(x)\ ,\quad\mu>0\ ,\quad n=0,1,2,\dots\
.$$These vectors certainly belong to the Hilbert space $L_2(R)$ for all $n=0,1,2,\dots,$ since$$\||\Phi_n\rangle\|^2\equiv\langle\Phi_n|\Phi_n\rangle=
\int\limits_{-\infty}^{+\infty}{\rm d}x\,|\Phi_n(x)|^2=
\int\limits_0^\infty{\rm d}x\,x^{2\,n}\,\exp(-2\,\mu\,x)
=\frac{\Gamma(2\,n+1)}{(2\,\mu)^{2\,n+1}}<\infty\ .$$[*However*]{}: The norm $\|T|\Phi_n\rangle\|$ of the vectors $T|\Phi_n\rangle$, $n=0,1,2,\dots,$ may be found from$$\|T|\Phi_n\rangle\|^2=\int\limits_{-\infty}^{+\infty}{\rm
d}x\,|(T\Phi_n)(x)|^2=
\frac{[\Gamma(n+1)]^2}{2\pi}\int\limits_{-\infty}^{+\infty}{\rm
d}p\,\frac{p^2+m^2}{(p^2+\mu^2)^{n+1}}\ .$$This observation might be a hint that the vector $|\Phi_0\rangle$, that is, $\Phi_0(x)=\exp(-\mu\,x)\,\Theta(x)$, does [*not*]{} belong to the domain of the kinetic-energy operator $T$. If this is indeed true, it is by no means obvious how to make sense of Eq. (16) of Ref. [@Brau99] for the case $n=0$.
Trivially, [*if*]{} Eq. (16) of Ref. [@Brau99] is correct for $n=0$, all these relations for arbitrary $n=1,2,\dots$ may be obtained by a simple differentiation of the relation for $n=0$ with respect to the (generic) parameter $\mu$, taking advantage of$$T\,x^n\,\exp(-\mu\,x)=\left(-\frac{{\rm
d}}{{\rm d}\mu}\right)^n\,T\,\exp(-\mu\,x)\ .$$
Similarly, it is somewhat hard to believe that Eq. (16) of Ref. [@Brau99] holds for $n=1$. In our notation, Eq. (16) of Ref. [@Brau99] [*would*]{} read for $n=1$$$(T\Phi_1)(x)=\left(S+\frac{\mu}{S\,x}\right)\Phi_1(x)$$with $$S\equiv\sqrt{m^2-\mu^2}\ .$$Considering merely the norms of the vectors on both sides of this equation, we find, for the norm of the vector on the left-hand side,$$\|T|\Phi_1\rangle\|^2=\frac{m^2+\mu^2}{4\,\mu^3}$$but, for the norm of the vector on the right-hand side,$$\left\|\left(S+\frac{\mu}{S\,x}\right)|\Phi_1\rangle\right\|^2=
\frac{m^4+\mu^4}{4\,\mu^3\,S^2}\ .$$These two expressions for the norms become equal only for the—excluded—case $\mu=0$. Unfortunately, precisely the above relation forms the basis for the assertion in Ref. [@Brau99] that $\Phi_1(x)$ with $\mu=\beta_1$ is the ground-state eigenfunction of the (“hard-core amended”) one-dimensional spinless relativistic Coulomb problem as defined in the Introduction.
In conclusion, let us summarize our point of view as follows: The energy eigenvalues $E_k$, $k=1,2,3,\dots,$ of the one-dimensional spinless relativistic Coulomb problem (with hard-core interaction on the nonpositive real line) are bounded from above by Eq. (\[Eq:AUBOELFSRCP\]). For the ground-state energy eigenvalue $E_1$, this upper bound may be improved to some extent, by considering appropriately the minimum of the bounds of Eq. (\[Eq:AUBGSE-mu=m\]), Eq. (\[Eq:GSE-expl\]), or Eq. (\[Eq:AUBOELFSRCP\]) for $k=1$, that is, Eq. (\[Eq:AUBGSE\]). To our knowledge, these upper bounds represent the only information available at present about the exact location of the energy levels of the (“hard-core amended”) one-dimensional spinless relativistic Coulomb problem.
Acknowledgements {#acknowledgements .unnumbered}
================
We would like to thank H. Narnhofer for stimulating discussions and a critical reading of the manuscript.
Combining Minimum–Maximum Principle with Operator Inequalities [@Lucha96:AUB; @Lucha98:T] {#App:MMPOI}
=========================================================================================
There exist several equivalent formulations of the well-known “min–max principle” [@MMP]. For practical purposes, the most convenient one is perhaps the following:
- Let $H$ be a self-adjoint operator bounded from below.
- Let $E_k$, $k=1,2,3,\dots,$ denote the eigenvalues of $H$, defined by$$H|\chi_k\rangle=E_k|\chi_k\rangle\ ,\quad k=1,2,3,\dots\
,$$and ordered according to$$E_1\le E_2\le E_3\le\dots\ .$$
- Consider only the eigenvalues $E_k$ below the onset of the essential spectrum of $H$.
- Let $D_d$ be some $d$-dimensional subspace of the domain ${\cal
D}(H)$ of $H$: $D_d\subset{\cal D}(H)$.
Then the $k$th eigenvalue $E_k$ (when counting multiplicity) of $H$ satisfies the inequality $$E_k\le\displaystyle\sup_{|\psi\rangle\in D_k}
\frac{\langle\psi|H|\psi\rangle}{\langle\psi|\psi\rangle}\quad\mbox{for}\
k=1,2,3,\dots\ .$$The min–max principle may be employed in order to compare eigenvalues of operators:
- Assume the validity of a generic operator inequality of the form$$H\le{\cal O}\ .$$Then$$E_k\equiv
\frac{\langle\chi_k|H|\chi_k\rangle}{\langle\chi_k|\chi_k\rangle}
\le\sup_{|\psi\rangle\in D_k}
\frac{\langle\psi|H|\psi\rangle}{\langle\psi|\psi\rangle}
\le\sup_{|\psi\rangle\in D_k}\frac{\langle\psi|{\cal
O}|\psi\rangle}{\langle\psi|\psi\rangle}\ .$$
- Assume that the $k$-dimensional subspace $D_k$ in this inequality is spanned by the first $k$ eigenvectors of the operator ${\cal O}$, that is, by precisely those eigenvectors of ${\cal O}$ that correspond to the first $k$ eigenvalues $\widehat E_1,\widehat E_2,\dots,\widehat E_k$ of ${\cal O}$ if the eigenvalues of ${\cal O}$ are ordered according to$$\widehat E_1\le\widehat
E_2\le\widehat E_3\le\dots\ .$$Then$$\sup_{|\psi\rangle\in
D_k}\frac{\langle\psi|{\cal O}|\psi\rangle}{\langle\psi|\psi\rangle}=\widehat
E_k\ .$$
Consequently, every eigenvalue $E_k$, $k=1,2,3,\dots,$ of $H$ is bounded from above by the corresponding eigenvalue $\widehat E_k$, $k=1,2,3,\dots,$ of ${\cal O}$:$$E_k\le\widehat E_k\quad\mbox{for}\
k=1,2,3,\dots\ .$$
[30]{} E. E. Salpeter and H. A. Bethe, Phys. Rev. [**84**]{} (1951) 1232;\
E. E. Salpeter, Phys. Rev. [**87**]{} (1952) 328. W. Lucha and F. F. Schöberl, in: [*Proceedings of the International Conference on Quark Confinement and the Hadron Spectrum*]{} (Como, Italy, June 1994), eds. N. Brambilla and G. M. Prosperi (World Scientific, River Edge, N. J., 1995) p. 100, hep-ph/9410221. W. Lucha and F. F. Schöberl, HEPHY-PUB 692/98 (1998), hep-ph/9807342. W. Lucha and F. F. Schöberl, HEPHY-PUB 701/98 (1998), hep-ph/9812368, Int. J. Mod. Phys. A (in print);\
W. Lucha and F. F. Schöberl, HEPHY-PUB 706/98 (1998), hep-ph/9812526, to appear in the Proceedings of the International Conference on “Nuclear & Particle Physics with CEBAF at Jefferson Lab,” Nov. 3 – 10, 1998, Dubrovnik, Croatia. F. Brau, J. Math. Phys. [**40**]{} (1999) 1119. I. W. Herbst, Commun. Math. Phys. [**53**]{} (1977) 285; [**55**]{} (1977) 316 (addendum). M. Reed and B. Simon, [*Methods of Modern Mathematical Physics IV: Analysis of Operators*]{} (Academic Press, New York, 1978) Section XIII.1;\
A. Weinstein and W. Stenger, [*Methods of Intermediate Problems for Eigenvalues – Theory and Ramifications*]{} (Academic Press, New York/London, 1972) Chapters 1 and 2;\
W. Thirring, [*A Course in Mathematical Physics 3: Quantum Mechanics of Atoms and Molecules*]{} (Springer, New York/Wien, 1990) Section 3.5. W. Lucha and F. F. Schöberl, Phys. Rev. D [**50**]{} (1994) 5443, hep-ph/9406312. W. Lucha and F. F. Schöberl, Phys. Rev. A [**54**]{} (1996) 3790, hep-ph/9603429. A. Martin and S. M. Roy, Phys. Lett. B [**233**]{} (1989) 407.
[^1]: The present state-of-the-art of the three-dimensional relativistic Coulomb problem has been reviewed, for instance, in Refs. [@Lucha94:C; @Lucha98:R; @Lucha98:T].
[^2]: This statement constitutes what is sometimes simply called “Rayleigh’s principle.”
[^3]: The line of arguments leading to the general form of this statement may be found, for instance, in Refs. [@Lucha96:AUB; @Lucha98:T]. It is summarized in Appendix \[App:MMPOI\]. For a rather brief account of the application of these ideas to the three-dimensional spinless relativistic Coulomb problem, see, e.g., Ref. [@Lucha98:R].
[^4]: The Coulomb problem involves no dimensional parameter other than the particle mass $m$. Therefore, both the energy eigenvalues $E_n$ and the parameter(s) $\mu$ have to be proportional to $m$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
In this paper, we will introduce an special kind of graph homomorphisms namely semi-locally-surjective graph homomorphisms and show some relations between semi-locally-surjective graph homomorphisms and colorful colorings of graphs and then we prove that for each natural number $k$, the Kneser graph $KG(2k+1,k)$ is $b$-continuous. Finally, we introduce some special conditions for graphs to be $b$-continuous.
\
[**Keywords:**]{} [ graph colorings, colorful colorings, Kneser graphs, semi-locally-surjective graph homomorphisms.]{}\
[**Subject classification: 05C**]{}
---
[**On $b$-continuity of Kneser Graphs of Type $KG(2k+1,k)$**]{}\
\
[*Department of Mathematical Sciences*]{}\
[*Institute for Advanced Studies in Basic Sciences [(]{}IASBS[)]{}*]{}\
[*P.O. Box [45195-1159]{}, Zanjan, Iran*]{}\
[s\[email protected]]{}\
\
Introduction
============
All graphs considered in this paper are finite and simple (undirected, loopless and without multiple edges). Let $G=(V,E)$ be a graph and $k\in \mathbb{N}$ and let $[k]:=\{i|\ i\in
\mathbb{N},\ 1\leq i\leq k \}$. A $k$-coloring (proper k-coloring) of $G$ is a function $f:V\rightarrow [k]$ such that for each $1\leq i\leq k$, $f^{-1}(i)$ is an independent set. We say that $G$ is $k$-colorable whenever $G$ has a $k$-coloring $f$, in this case, we denote $f^{-1}(i)$ by $V_{i}$ and call each $1\leq i\leq
k$, a color (of $f$) and each $V_{i}$, a color class (of $f$). The minimum integer $k$ for which $G$ has a $k$-coloring, is called the chromatic number of G and is denoted by $\chi(G)$.
Let $G$ be a graph and $f$ be a $k$-coloring of $G$ and $v$ be a vertex of $G$. The vertex $v$ is called $b$-dominating ( or colorful or color-dominating) ( with respect to $f$) if each color $1\leq i\leq k$ appears on the closed neighborhood of $v$ ( $f(N[v])=[k]$ ). The coloring $f$ is said to be a colorful $k$-coloring of $G$ if each color class $V_{i}\ (1\leq i\leq k) $ contains a $b$-dominating vertex $x_{i}$. Obviously, every $\chi(G)$-coloring of $G$ is a colorful $\chi(G)$-coloring of $G$. We denote ${\rm B}(G)$ the set of all positive integers $k$ for which $G$ has a colorful $k$-coloring. The maximum of ${\rm
B}(G)$, is called the $b$-chromatic number of $G$ and is denoted by $b(G)$ (or $\phi(G)$ or $\chi_{b}(G)$). The graph $G$ is said to be $b$-continuous if each integer $k$ between $\chi(G)$ and $b(G)$ is an element of ${\rm B}(G)$. There are graphs that are not $b$-continuous, for example, the 3-dimensional cube $Q_{3}$ is not $b$-continuous, because $2\in {\rm B}(G)$ and $4\in {\rm
B}(G)$ but $3\notin {\rm B}(G)$ ( [@irv] ). We have to note that the problem of deciding whether graph $G$ is $b$-continuous is NP-complete ( [@bar] ). The colorful coloring of graphs was introduced in 1999 in [@irv] with the terminology $b$-coloring.
Let $m,n\in \mathbb{N}$ and $m\leq n$. $KG(n,m)$ is the graph whose vertex set is the set of all subsets of size $m$ of $[n]$ in which two vertices $X$ and $Y$ are adjacent iff $X\bigcap
Y=\emptyset$. Note that $KG(5,2)$ is the famous Petersen graph. It was conjectured by Kneser in 1955 ( [@KG] ), and proved by Lov[á]{}sz in 1978 ( [@lov] ), that if $n\geq 2m$, then $\chi
(KG(n,m))=n-2m+2$. Lov[á]{}sz’s proof was the beginning of using algebraic topology in combinatorics. Colorful colorings of Kneser graphs have been investigated in [@haj] and [@jav]. Javadi and Omoomi in [@jav] showed that for $n\geq 17$, $KG(n,2)$ is $b$-continuous. Only a few classes of graphs are known to be $b$-continuous (see [@bar], [@fai] and [@jav]). We want to prove that for each natural number $k$, $KG(2k+1,k)$ is $b$-continuous. In this regard, first we introduce an special kind of graph homomorphisms which is related to colorful colorings of graphs.
Let $G$ and $H$ be graphs. A function $f:V(G)\rightarrow V(H)$ is called a semi-locally-surjective graph homomorphism from $G$ to $H$ if $f$ is a surjective graph homomorphism from $G$ to $H$ and satisfies the following condition :\
\
$\forall u \in V(H):\ \exists a\in f^{-1}(u)\ s.t\ \forall v\in N_{H}(u):\ \exists b\in f^{-1}(v)\ s.t\ \{a,b\}\in E(G) $.
We know that a graph $G$ is $k$-colorable iff there exists a graph homomorphism from $G$ to the complete graph $K_{k}$ and the chromatic number of $G$ is the least natural number $k$ for which there exists a graph homomorphism from $G$ to $K_{k}$. Indeed, we can think of graph homomorphisms from graphs to complete graphs instead of graph colorings. The following obvious theorem shows such a similar relation between colorful colorings of graphs and semi-locally-surjective graph homomorphisms. Indeed, we can think of semi-locally-surjective graph homomorphisms from graphs to complete graphs instead of colorful colorings of graphs.
[ Let $G$ be a graph and $k\in \mathbb{N}$. Then $k\in {\rm B}(G)$ iff there exists a semi-locally-surjective graph homomorphism from $G$ to $K_{k}$. Also, the chromatic number of $G$ ( $\chi(G)$ ) and the $b$-chromatic number of $G$ ( $b(G)$ ) are respectively the least and the greatest natural numbers $k$ for which there exists a semi-locally-surjective graph homomorphism from $G$ to $K_{k}$ . ]{}
We know that the composition of two graph homomorphisms is again a graph homomorphism. A similar theorem holds for composition of semi-locally-surjective graph homomorphisms.
[ Let $G_{1}$, $G_{2}$ and $G_{3}$ be graphs. If $g$ is a semi-locally-surjective graph homomorphism from $G_{2}$ to $G_{1}$ and $f$ is a semi-locally-surjective graph homomorphism from $G_{3}$ to $G_{2}$, then $gof$ is a semi-locally-surjective graph homomorphism from $G_{3}$ to $G_{1}$. ]{}
The following theorem shows another relation between semi-locally-surjective graph homomorphisms and colorful colorings of graphs.
[\[thm1\] Let $G_{1}$ and $G_{2}$ be graphs. If there exists a semi-locally-surjective graph homomorphism from $G_{1}$ to $G_{2}$, then ${\rm B}(G_{2}) \subseteq {\rm B}(G_{1})$. ]{}
[ Let $f$ be a semi-locally-surjective graph homomorphism from $G_{1}$ to $G_{2}$, $k\in {\rm B}(G_{2})$, and $V_{1},\ldots ,V_{k}$ be color classes of a colorful $k$-coloring of $G_{2}$ and $x_{1},\ldots ,x_{k}$ be some $b$-dominating vertices of $G_{2}$ with respect to this $k$-coloring and $x_{i}\in V_{i}\ (1\leq i\leq k)$. Obviously, $f^{-1}(V_{1}),\ldots
,f^{-1}(V_{k})$ are nonempty color classes of a $k$-coloring of $G_{1}$ and $f^{-1}(x_{1}),\ldots ,f^{-1}(x_{k})$ are some $b$-dominating vertices of $G_{1}$ with respect to this $k$-coloring and $f^{-1}(x_{i})\in f^{-1}(V_{i})\ (1\leq i\leq k)$. Therefore, $G_{1}$ has a colorful $k$-coloring and $k\in {\rm B}(G_{1})$. Hence, ${\rm B}(G_{2})
\subseteq {\rm B}(G_{1})$. ]{}
Now we prove that for each natural number $k$, $KG(2k+1,k)$ is $b$-continuous.
[ For each $k\in \mathbb{N}$, $KG(2k+1,k)$ is $b$-continuous. ]{}
For each $k\in \mathbb{N}$, $\chi (KG(2k+1,k))=3$. Note that ${\rm B}(KG(3,1))={\rm B}(K_{3})=\{3\}$ and therefore, for $k=1$ the assertion follows. Blidia, et al. in [@bli] proved that the $b$-chromatic number of the Petersen graph is 3 and therefore, ${\rm B}(KG(5,2))={\rm B}(Petersen\ graph)=\{3\}$. Hence, $KG(2k+1,k)$ is $b$-continuous for $k=2$. For $k\geq3$, the function $f:V(KG(2k+3,k+1))\rightarrow V(KG(2k+1,k))$ which assigns to each $A\subseteq [2k+3]$ with $|A\bigcap
\{2k+2,2k+3\}|\leq1$, $f(A)=A\setminus\{\max A\}$ and to each $A\subseteq [2k+3]$ with $\{2k+2,2k+3\}\subseteq A$, $f(A)=(A\setminus\{2k+2,2k+3\})\bigcup \{\max ([2k+1]\setminus
A)\}$, is a surjective graph homomorphism from $KG(2k+3,k+1)$ to $KG(2k+1,k)$. Now for each $X\in V(KG(2k+1,k))$, $(X \bigcup \{2k+2\})\in f^{-1}(X)$ and for each $Y\in N_{KG(2k+1,k)}(X)$, $(Y \bigcup \{2k+3\})\in f^{-1}(Y)$ and $\{X \bigcup \{2k+2\},Y \bigcup \{2k+3\}\}\in E(KG(2k+3,k+1))$. Hence, $f$ is a semi-locally-surjective graph homomorphism from $KG(2k+3,k+1)$ to $KG(2k+1,k)$. Consequently, Theorem \[thm1\] implies that ${\rm B}(KG(2k+1,k)) \subseteq {\rm B}(KG(2k+3,k+1))$, besides, ${\rm B}(KG(7,3))\subseteq {\rm B}(KG(9,4))\subseteq ...\subseteq
{\rm B}(KG(2n+1,n)) \subseteq ...$ . (I)
On the other hand, Javadi and Omoomi in [@jav] showed that for $k\geq 3$, $b(KG(2k+1,k))=k+2$ and $k+2\in {\rm
B}(KG(2k+1,k))$. Therefore, for each $k\geq 3$, $\{i+2|\
i\in\mathbb{N},\ 3\leq i\leq k\}\subseteq {\rm B}(KG(2k+1,k))$. Also, since $\chi (KG(2k+1,k))=3$, $3\in {\rm B}(KG(2k+1,k))$. So, constructing a colorful 4-coloring of $KG(2k+1,k)\ (k\geq3)$ completes the proof. (I) implies that it is enough to construct a colorful 4-coloring of $KG(7,3)$. Set $$\begin{array}{cc}
V_{1}:=\{\ \{1 ,2 ,3 \},\{1 ,4 ,5 \},\{2 , 5,6 \},\{ 1,2 ,6
\},\{ 1,2 ,7 \},\{ 1,3 ,6 \},\{ 1,6 ,7 \}, & \\
\{ 1, 4, 6\}\ \}, & \\
V_{2}:=\{\ \{5 , x, y\}\ |\ x,y\in \{1,2,3,4,6,7\},\ x\neq y\
\}\setminus\{\ \{1 , 4,5 \},\{2 ,5 ,6 \}, & \\
\{4 ,5 , 7\}\ \}, & \\
V_{3}:=\{\ \{1 ,2 ,4 \},\{1 ,3 , 7\},\{4 ,5 ,7 \},\{1 ,4 ,7
\},\{2 ,6 ,7 \}\ \}, & \\
V_{4}:=(\{\ \{4 ,x ,y \}\ |\ x,y\in \{1,2,3,6,7\},\ x\neq y
\}\setminus\{\ \{1 ,2 ,4 \},\{ 1,4 ,6 \}, & \\
\{1 ,4 ,7 \}\ \})\ \bigcup\ \{\ \{2 ,3 ,6 \},\{2 ,3 ,7 \},\{3 ,6
,7 \}\ \}. &
\end{array}$$
Now, one can check that $V_{1},\ V_{2},\ V_{3},\ V_{4}$ are color classes of a colorful 4-coloring of $KG(7,3)$ that $ \{1 ,2 ,3
\}\in V_{1},\ \{5 ,6 ,7 \}\in V_{2},\ \{2 ,6 ,7 \}\in V_{3}$ and $\{1 ,3 ,4 \}\in V_{4}$ are some $b$-dominating vertices with respect to this $4$-coloring.
The semi-locally-surjective graph homomorphism $f$ in above Theorem can be generalized as follows.
[ Let $n,m\in\mathbb{N}$ with $n>2m$. Then ${\rm B}(KG(n,m))\subseteq {\rm B}(KG(n+2,m+1))$.]{}
[The function $f:V(KG(n+2,m+1))\rightarrow V(KG(n,m))$ which assigns to each $A\subseteq [n+2]$ with $|A\bigcap
\{n+1,n+2\}|\leq1$, $f(A)=A\setminus\{\max A\}$ and to each $A\subseteq [n+2]$ with $\{n+1,n+2\}\subseteq A$, $f(A)=(A\setminus\{n+1,n+2\})\bigcup \{\max ([n]\setminus A)\}$, is a surjective graph homomorphism from $KG(n+2,m+1)$ to $KG(n,m)$. Now for each $X\in V(KG(n,m))$, $(X \bigcup \{n+1\})\in f^{-1}(X)$ and for each $Y\in N_{KG(n,m)}(X)$, $(Y \bigcup \{n+2\})\in f^{-1}(Y)$ and $\{X \bigcup \{n+1\},Y \bigcup \{n+2\}\}\in E(KG(n+2,m+1))$. Hence, $f$ is a semi-locally-surjective graph homomorphism from $KG(n+2,m+1)$ to $KG(n,m)$ and therefore, Theorem \[thm1\] implies that ${\rm B}(KG(n,m)) \subseteq {\rm B}(KG(n+2,m+1))$.]{}
[ Let $a,b\in\mathbb{N}\bigcup\{0\}$ and $a>2b$. Also, for each $i\in \mathbb{N}\setminus\{1\}$, let ${\rm B}_{i}:={\rm B}(KG(2i+a,i+b))$ and ${\rm b_{i}}:=b(KG(2i+a,i+b))$. Then ${\rm B}_{2}\subseteq {\rm B}_{3}\subseteq {\rm
B}_{4}\subseteq ...\subseteq {\rm B}_{n}\subseteq {\rm
B}_{n+1}\subseteq ...$ , and ${\rm b_{2}}\leq {\rm b_{3}}\leq {\rm
b_{4}}\leq ...\leq {\rm b_{n}}\leq {\rm b_{n+1}}\leq ...$]{} .
Now we introduce some special conditions for graphs to be $b$-continuous. But first we note that in a graph $G$ with at least one cycle, the girth of $G$ ($g(G)$), is the minimum of all cycle lengths of $G$ and if $G$ has not any cycles, the girth of $G$ is defined $g(G)=+\infty$ .\
Blidia, et al. proved the following theorem.
[( [[@bli]]{} )\[thm5\] If $d\leq6$, then for every $d$-regular graph $G$ with girth $g(G)\geq5$ which is different from the Petersen graph, $b(G)=d+1$. ]{}
By using this theorem, we prove the following theorem.
[ Let $3\leq d\leq6$ and for each $2\leq i\leq d$, $G_{i}$ be an $i$-regular graph with girth $g(G_{i})\geq5$ which is different from the Petersen graph. Also, suppose that for each $3\leq i\leq d$, there exists a semi-locally-surjective graph homomorphism $f_{i}$ from $G_{i}$ to $G_{i-1}$. Then for each $2\leq i\leq d$, $G_{i}$ is $b$-continuous. ]{}
Theorem \[thm5\] implies that for each $2\leq i\leq d$, $b(G_{i})=i+1$ and therefore, $i+1 \in {\rm B}(G_{i}) $. Also, since for each $3\leq i\leq d$, there exists a semi-locally-surjective graph homomorphism $f_{i}$ from $G_{i}$ to $G_{i-1}$, theorem \[thm1\] implies that ${\rm
B}(G_{i-1})\subseteq {\rm B}(G_{i})$ and consequently, ${\rm
B}(G_{2})\subseteq {\rm B}(G_{3})\subseteq ... \subseteq {\rm
B}(G_{d}).$ Hence, for each $2\leq i\leq d$, $\{j+1|2\leq j\leq
i\}\subseteq {\rm B}(G_{i})$ and therefore, $\{3,4,...,i+1\}\subseteq {\rm B}(G_{i})$. Now, there are 2 cases:
Case 1) The case that $G_{i}$ is bipartite. In this case, $\chi(G_{i})=2$ and therefore, $2\in {\rm B}(G_{i})$ and $\{2,3,...,i+1\}\subseteq {\rm B}(G_{i})$, so ${\rm
B}(G_{i})=\{2,3,...,i+1\}$ and $G_{i}$ is $b$-continuous.
Case 2) The case that $G_{i}$ is not bipartite. In this case, $\chi(G_{i})\geq 3$ and since $\{3,...,i+1\}\subseteq B(G_{i})$, so ${\rm B}(G_{i})=\{3,...,i+1\}$ and $G_{i}$ is $b$-continuous.\
\
Therefore, for each $2\leq i\leq d$, $G_{i}$ is $b$-continuous.
[8]{}
D. Barth, J. Cohen, and T. Faik, On the $b$-continuity property of graphs, [*Discrete Applied Mathematics*]{}, [**155**]{} (2007), 1761-1768.
M. Blidia, F. Maffray, and Z. Zemir, On $b$-colorings in regular graphs, [*Discrete Applied Mathematics*]{}, [**157**]{} (2009), 1787-1793.
T. Faik, About the $b$-continuity of graphs, [*Electronic Notes in Discrete Mathematics*]{}, [**17**]{} (2004), 151-156.
H. Hajiabolhassan, On the $b$-chromatic number of Kneser graphs, [*Discrete Applied Mathematics*]{}, [**158**]{} (2010), 232-234.
R. W. Irving and D. F. Manlove, The $b$-chromatic number of a graph, [*Discrete Applied Mathematics*]{}, [**91**]{} (1999), 127-141.
R. Javadi and B. Omoomi, On $b$-coloring of the Kneser graphs, [*Discrete Mathematics*]{}, [**309**]{} (2009), 4399-4408.
M. Kneser, Aufgabe 300, [*Jber. Deutsch. Math.-Verein.*]{}, [**58**]{} (1955), 27.
L. Lov[á]{}sz, Kneser’s conjecture, chromatic number, and homotopy, [*Journal of Combinatorial Theory Ser. A*]{}, [**25**]{} (1978), 319-324.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We broaden the domain of application of the recently proposed thermal boundary condition of the wave function of the Universe, which has been suggested as the basis of a dynamical selection principle on the landscape of string solutions.'
address: |
$^{1}$Centro Multidisciplinar de Astrofísica - CENTRA, Departamento de Física,\
Instituto Superior Técnico, Av. Rovisco Pais 1, 1096 Lisboa, Portugal\
$^{2}$Departamento de Física, Universidade da Beira Interior,\
Rua Marquês d’Avila e Bolama, 6200 Covilhã, Portugal
author:
- 'Mariam Bouhmadi-López$^{1,2}$[^1] and Paulo Vargas Moniz$^{2,1}$[^2]'
title: 'On the thermal boundary condition of the wave function of the Universe[^3]'
---
Introduction {#intro}
============
The existence of a multiverse of vast solutions [@Bousso:2000xa; @Douglas:2003um; @Susskind:2003kw] to string theory constitutes currently an important challenge: How to select a Universe or a class from the multiverse that will bear significant similarities to ours?
The framework of quantum cosmology [@Vilenkin:1983xq; @Hartle:1983ai; @Linde:1983cm] provides a methodology to establish a probability distribution for the dynamical parameters of the Universe. In this context, Brustein and de Alwis proposed in [@Brustein:2005yn], using FRW quantum cosmology, a dynamical selection principle[^4] on the landscape of string solutions [@Susskind:2003kw].
We prove that the thermal boundary condition applied in [@Brustein:2005yn] corresponds to the particular physical situation where the amount of radiation is very large. We then provide a broader and improved analysis of the [*generalised thermal boundary condition*]{} that is independent of such restrictive limit [@BLVM2006]; i.e. we consider an arbitrary amount of radiation consistent with the tunnelling of a closed radiation-filled Universe with a positive cosmological constant.
The generalised thermal boundary condition
==========================================
The thermal boundary condition for the wave function of the Universe states that the Universe emerges from the string era in a thermally excited state above the Hartle-Hawking (HH) vacuum [@Brustein:2005yn]. Furthermore, the primordial thermal bath is effectively described by a radiation fluid whose energy density $\rho$ “depends” on the cosmological constant $\lambda$ $$\begin{aligned}
\rho=\frac{3\tilde{K}}{8\pi\textrm{G}}\frac{1}{a^4}, \quad
\tilde{K}\simeq \frac{\nu}{\lambda^2}. \label{K1}\end{aligned}$$ In the previous expressions $\tilde{K}$ and $\nu$ are parameters that quantify the amount of radiation, $\textrm{G}$ is the gravitational constant and $a$ the scale factor. Therefore, the transition amplitude of a closed radiation-filled FRW Universe[@Vilenkin:1998rp; @Bouhmadi-Lopez:2002qz] to tunnel from the first Lorentzian region ($a<a_-$, see Fig. 1-a) to the larger Lorentzian region ($a_+<a$, see also Fig. 1-a) within a WKB approximation reads [@Brustein:2005yn; @BLVM2006]
$$\mathcal{A}=\exp(\epsilon 2I), \quad
I=\frac{\pi}{2^{\frac32}\textrm{G}}\frac{1}{\nu}g,
\label{transition2}$$
where $\epsilon=\pm 1$ and $$\begin{aligned}
g=\frac{\nu}{\lambda}{\sqrt{1+m}}\left[\textrm{E}(\alpha_{II})-(1-m)\textrm{K}(\alpha_{II})\right],\,\,
\alpha_{II}=\sqrt{\frac{2m}{1+m}}, \,\, m=\sqrt{1-4\tilde{K}\lambda}, \label{defg}\end{aligned}$$ with $K(m)$ and $ E(m)$ as complete elliptic integrals of the first and second kind, respectively. Consequently, the thermal boundary condition implies a switch in the usual features of the HH [@Hartle:1983ai] ($\epsilon=1$ choice in Eq. (\[transition2\])) and the tunnelling ($\epsilon=-1$ choice in Eq. (\[transition2\])) [@Vilenkin:1983xq; @Linde:1983cm] wave functions: The HH wave function, once the thermal boundary condition of [@Brustein:2005yn] is assumed, favours a non-vanishing cosmological constant; $\lambda\simeq 8.33\nu$, larger than the one preferred by the tunnelling wave function; $\lambda\simeq 4\nu$, (see Fig. 1-b).
![Figure 1-a corresponds to the potential barrier ($V_0-\tilde{K}$) separating the two Lorentzian regions of a closed radiation-filled FRW Universe. $\tilde{K}_{\rm{max}}$ corresponds to the maximum amount of radiation consistent with the tunnelling of the Universe. We will refer to this situation as a large amount of radiation. Figures 1-b and 1-c corresponds to $g$ defined in Eq. (\[defg\]) for the thermal boundary condition and the [*generalised thermal boundary condition*]{}, respectively.[]{data-label="amplitudeplot2"}](figura9.eps "fig:"){width="4cm"}![Figure 1-a corresponds to the potential barrier ($V_0-\tilde{K}$) separating the two Lorentzian regions of a closed radiation-filled FRW Universe. $\tilde{K}_{\rm{max}}$ corresponds to the maximum amount of radiation consistent with the tunnelling of the Universe. We will refer to this situation as a large amount of radiation. Figures 1-b and 1-c corresponds to $g$ defined in Eq. (\[defg\]) for the thermal boundary condition and the [*generalised thermal boundary condition*]{}, respectively.[]{data-label="amplitudeplot2"}](figura4.eps "fig:"){width="4cm"}![Figure 1-a corresponds to the potential barrier ($V_0-\tilde{K}$) separating the two Lorentzian regions of a closed radiation-filled FRW Universe. $\tilde{K}_{\rm{max}}$ corresponds to the maximum amount of radiation consistent with the tunnelling of the Universe. We will refer to this situation as a large amount of radiation. Figures 1-b and 1-c corresponds to $g$ defined in Eq. (\[defg\]) for the thermal boundary condition and the [*generalised thermal boundary condition*]{}, respectively.[]{data-label="amplitudeplot2"}](figura8-1.eps "fig:"){width="4cm"}\
1-a 1-b 1-c
It turns out that the thermal effect considered in [@Brustein:2005yn] corresponds to a large amount of radiation where $\tilde{K}$ is close to $\tilde{K}_{\rm{max}}$ (see Fig. 1-a); i.e. the turning points $a_-$ and $a_+$ are very close or equivalently the height of the barrier separating both Lorentzian regions is very small.
Next we consider a [*generalised thermal boundary condition*]{} for the wave function [@BLVM2006] where we will assume instead an arbitrary amount of radiation, consistent with a tunnelling of the Universe; i.e. $\tilde{K}<\tilde{K}_{max}$ (see Fig. 1-a). Consequently, Eqs. (\[K1\])-(\[defg\]) are replaced by consistent and more general relations where the amount of radiation as measured by $\tilde{K}$ depends also on $\lambda$ and reads $$\tilde{K}= \frac{4\nu\lambda^{-2}}{(1+4\nu\lambda^{-1})^2}.
\label{eq20}$$ Within this broader range, the relevant features is that the transition amplitude as a function of $\nu/\lambda$ will be unlike the one deduced in [@Brustein:2005yn]. Indeed, this is the case as is shown in Figs. 1-b and 1-c.
Regarding the HH wave function, it now favours a vanishing cosmological constant ($\nu /\lambda
\rightarrow \infty$) and $\tilde{K} \sim 1
/(4\nu)$. This physical case is represented schematically by a star in Fig. 1-c. In this manner, the role of the HH wave function and subsequent transition amplitude is returned to its “original” implication, with the thermal boundary condition being implemented in a fully consistent manner and not restricted to a narrow (perhaps not fully valid) limit.
Concerning the tunnelling wave function, it favours two possible physical situations depicted by a circle and a square in Fig. 1-c. On the one hand, the “circle” option corresponds to a large cosmological constant ($\nu / \lambda \rightarrow 0$) and a small amount of radiation as measured by $\tilde{K}$ ($\tilde{K}
\lambda \rightarrow 0$). On the other hand, the “square” option implies no tunnelling, that is, $4 \nu / \lambda
\rightarrow 1$ or equivalently $4\tilde{K} \lambda \rightarrow 1$; i.e. both turning points coincide. In order to select one of these two possibilities for the tunnelling wave function, we employed the DeWitt’s argument [@DeWitt1; @Bouhmadi-Lopez:2004mp], since there is a curvature singularity at small scale factors. It turns out that the preferred value of the cosmological constant in this case is a large one. Moreover, this condition implies a small amount of radiation (as measured by the parameter $\tilde K$) allowing consequently the tunnelling of the Universe.
Conclusions
===========
We prove that the thermal boundary condition applied in [@Brustein:2005yn] corresponds to the particular physical situation where the amount of radiation is very large. We then provide a broader and improved analysis of the [*generalised thermal boundary condition*]{} that is independent of such restrictive limit [@BLVM2006].
Acknowledgments {#acknowledgments .unnumbered}
===============
MBL acknowledges the support of a Marcel Grossmann fellowship to attend the meeting. MBL also acknowledges the support of CENTRA-IST BPD (Portugal) as well as the fellowship FCT/BPD/26542/2006.
[00]{}
R. Bousso and J. Polchinski, JHEP [**0006**]{}, 006 (2000).
M. R. Douglas, JHEP [**0305**]{}, 046 (2003). L. Susskind, arXiv:hep-th/0302219.
J. B. Hartle and S. W. Hawking, Phys. Rev. D [**28**]{}, 2960 (1983).
A. Vilenkin, Phys. Rev. D [**27**]{}, 2848 (1983).
A. D. Linde, Sov. Phys. JETP [**60**]{}, 211 (1984) \[Zh. Eksp. Teor. Fiz. [**87**]{}, 369 (1984)\].
R. Brustein and S. P. de Alwis, Phys. Rev. D [**73**]{}, 046009 (2006). S. Sarangi and S. H. Tye, arXiv:hep-th/0505104.
R. Holman and L. Mersini-Houghton, Phys. Rev. D [**74**]{}, 043511 (2006).
For more details see: M. Bouhmadi-López and P. Vargas Moniz, arXiv:hep-th/0612149.
A. Vilenkin, arXiv:gr-qc/9812027.
M. Bouhmadi-López, L. J. Garay and P. F. González-Díaz, Phys. Rev. D [**66**]{}, 083504 (2002). B. DeWitt, Phys. Rev. [**160**]{}, 1113 (1967).
M. Bouhmadi-López and P. Vargas Moniz, Phys. Rev. D [**71**]{}, 063521 (2005); M. Bouhmadi-Lopez and P. Vargas Moniz, AIP Conf. Proc. [**736**]{}, 188 (2005).
[^1]: [email protected]
[^2]: [email protected]
[^3]: This research work was supported by the grants POCI/FP/63916/2005,FEDER-POCI/P/FIS/57547/2004 and Acções Integradas (CRUP-CSIC) Luso-Espanholas E-138/04.
[^4]: See also Refs. [@Sarangi:2005cs; @Holman:2005eu].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this paper, we propose a model which simulates odds distributions of pari-mutuel betting system under two hypotheses on the behavior of bettors: 1. The amount of bets increases very rapidly as the deadline for betting comes near. 2. Each bettor bets on a horse which gives the largest expectation value of the benefit. The results can be interpreted as such efficient behaviors do not serve to extinguish the FL bias but even produce stronger FL bias.'
address:
- '2-1-1, Tsuda-machi, Kodaira-shi, Tokyo, Japan'
- ' 1030, Shimo-Ogino, Atsugi, Kanagawa, Japan'
author:
- Kazutaka Kurihara
- Yohei Tutiya
title: 'Efficiency in Micro-Behaviors and FL Bias'
---
pari-mutuel system ,time series of odds ,FL bias ,market efficiency ,horse-racing
History and Motivation
======================
The market efficiency hypothesis has been one of the most argued subjects in Economics ([@Bach; @Fam] for standard references). It claims that if an arbitrage occurs in a market, news of this will spread so rapidly that on average people are unable to benefit. This notion has also been applied to the study of Pari-mutuel betting systems in the late 1900’s. For example, if a horse wins too often in relation to its odds, it quickly becomes well known and people begin to bet more money on it. Hence, we may well expect its odds to decrease rapidly, to the point where there is no more benefit in betting on this horse compared to any other. However, this is not always the case in real horse racing. It has been reported that expectation values of the benefit of betting on the favorites tend to be higher than the perceived benefit of betting on longshots ([@Gr] for the first report). This is called favorite-longshot (FL) bias. It should be noted here that the weak form of market efficiency is still observed under the bias. In other word, although the rate of return for favorites is large, it does not exceed one.\
After the report [@Gr], FL bias has been observed in many races all over the world. It is also notable that inverse FL bias, i.e. a phenomenon that rate of return for longshots becomes higher than that for favorites, was observed in several parimutuel systems other than horse racing [@SobRyan]. In both FL and inverse FL bias, what is theoretically important is that the rates of return are different in favorites and longshots. This has been one of the central subjects in the study of the pari-mutuel betting system. Many preceding studies attempted to explain such bias by assuming the existence of non-efficient bettors. In [@Col; @BruJoh], for example, the bias is explained as an equilibrium of powers between two groups, one of which is constituted of efficient bettors while the other puts an importance on pursuing the leisure or thrill of betting. There also many studies which explain the bias in terms of misperception of the true underlying probabilities for winning [@ChadQua; @ThaZie]. If the existence of non-efficient bettors or misperception of probabilities account for the bias, the next question will be how and why the ratio of such bettors or misperceptions are determined so coincidentally as to produce the real distribution of odds with weak efficiency being a possibility. Before going to this stage, this study simulates what will happen if all the betters behave efficiently. The result suggests that the FL bias occurs even if all the bettors behave to maximize their expectation values of their bets. The bettors in the model bet their money on horses whose odds are large in relation to their probabilities of winning. And the odds fluctuate over time with such bets until the race begins. Conditions for the simulations will vary in several ways. We will change how efficiently bettors behave. It will also be changed how strong favorites are in a race. We will then change how rapidly bets accumulate over time.\
It should be noted here that there are already some studies which explain FL bias by considering how odds tables are driven by wagering over time. Some of the most successful examples will be Ali [@Ali], Blough [@Blo], Brown and Lin [@BroLin]. In their studies, bettors hold heterogeneous subjective probabilities about the true winning probability of each horse. Contrary to that, in our model, the beliefs of bettors are homogeneous. But the attitudes for efficiency fluctuate randomly. According to our results, FL bias is even amplified when bettors behave too efficiently.
Modeling
========
Let us consider the win betting system in a race consists of $N$ horses. We denote the ratio of the amount of bets for the $i$-th horse to that of all horses at time $t$ as $V_i(t)$. In pari-mutuel betting system, the odds for this horse is $(1-\alpha)/V_i(t)$, where the constant $\alpha$ is the rate of track take fixed in each racetrack. Since $V_i(t)$ varies over time, the odds table also varies over time. In most racetracks, odds tables changes at one or two minute intervals. And, the change of the odds table becomes one of major reasons to promotes changes in $V_i(t)$ because people make their decisions based on the odds table. According to this mechanism, $V_i(t)$ changes over time until the race starts. These are, at this time, the general setting of the pari-mutuel system. Now we introduce a positive constant $p_i$, which represents a rational value of $V_i(t)$. And we regard $p_i/V_i(t)$ to represents how much it is efficient to bet on $i$-th horse. That is to say, the larger $p_i/V_i$ , the more the $i$-th horse appealing to efficient bettors. If the track take is zero, $p_i$ can be regarded as the true probability that $i$-th horse wins, then, $p_i/V_i$ is nothing but the rate of return of betting on $i$-th horse. One might wonder that, if it were roulette, bettors are in advance informed of the probability that each number will come up. But, in the case of horse racing, people do not know any such exact values to judge the efficiency beforehand. More precise explanations on this issue will be given in section 2.3. Another remarkable tendency of efficient bettors is that they like to put off their bets for as long as possible. To reflect this pattern, our model only deals with the time evolution of bets in the last $20$ minutes before races start. Thus, we arrive at the following model.
- Let $t\ (t=-20,-19,\cdots,0[{\rm min}])$ denotes time after a race starts. The total amount of minimum unit of bets in each race is fixed to be 200,000.And the rate of bets finished before time $t$ will be denoted as $F(t)$. Hence, $F(0)=1$ is assumed. We will also use $f(t)=F(t)-F(t-1)$ to represent the rate of bets which are accumulated between $t-1$ and $t$.
- The number of horses in each race is fixed to be $N$ . And each horse equips a constant which will be referred to as consensus probability $p_i\ (i=1,2,\cdots,N)$ [@Blo]. We assume $p_1+ p_2+ \cdots+p_N=1$ and $p_1\geq p_2\geq \cdots\geq p_N$ for convenience.
- At time $t$, the all $N$ horses are sorted according to the magnitude of $p_i/V_i(t)$. Then we introduce a probability $q_n$ with which a bet is made on a horse having the $n$-th largest magnitude. $\{q_n\}_{n=1,2,\cdots,N}$ will be denoted as the preference probability distribution, where $q_1+q_2+ \cdots+q_N=1$ and $q_1\geq q_2\geq \cdots\geq q_N$ are assumed.
Now we proceed to choose the parameters of this model.
The shape of $F(t)$
-------------------
If the bettors are genuinely efficient, they will not make any bet until the last minute before each race starts. For example, [@GraMac] reports that 40% of the total bets rash to the last few minutes before a race starts. According to this idea, the most suitable distribution is a delta function, i.e. $f(t)$ takes positive value only when $t=1$ and $f(t)=0$ otherwise. But, in reality, “the last minute" can be different from bettor to bettor. So there must be a fluctuation from the delta function. Here we exhibit an example of $F(t)$ in a real race track in FIG. \[fig:invest\_final\].
![\[fig:invest\_final\] an average of $F(t)$ in HK racetrack and exponential fittings.](invest_final.png){width=".9\linewidth"}
The real line represents the average $F(t)$ in the HongKong racetrack [@Hong] in the period from October 2012 to May 2014, where the average is taken over entire 2000 races during the period. The horizontal axis represents the time $t$. It should be noted here that numbers of bettors in real data change until $t=2$ or 3 probably because of a lag of publication. The regression curve of dotted line is of the form $\exp(c_0+c_1t)$ and the dashed one $\exp(c_0+c_1t+c_2t^2+c_3t^3)$.It would be an interesting future study to investigate the mechanism that generates the shape of $F(t)$. Instead of sticking to the exact shape, in this paper, we will change $f(t)$ in several ways. As for the speed of growth, we will try exponential growth, constant growth and exponential decay. Another important factor is the initial value $F(-20)$ . We will consider $F(-20)=0.35$ and 0.01. Up to Section 5, we only deal with exponentially growing $f(t)$ with $F(-20)=0.35$, which gives the nearest shape for HK data in Fig.\[fig:invest\_final\]. Henceforth in this paper the HongKong data will be regarded as standard for setting parameters.\
It should be noted that very few pari-mutuel racetracks disclose data on $F(t)$. And the HongKong racetrack is one of the most notable examples of world famous racetracks which openly disclose their data on $F(t)$.
Choice of $N$
-------------
We set as $N=14$. This is not only because the number of horses in a HongKong race is 14, but also because most races in the world are held with 10 to 16 horses.
Choice of consensus probability
--------------------------------
As stated in the rule, we associated the consensus probability $p_i$ with each horse. This might seem strange because people do not know the probability that each horse will win. However, each bettor must possesses his/her own valuation on each horse and the average of these valuations must exist as $p_i$. Since it is just an average, it can differ from valuations of individual betters. And the belief in $p_i$ can fluctuate from time to time. The value of $p_i$ itself might also fluctuate over time. In this paper, we take a viewpoint that such variance and fluctuation in $p_i$ are rounded to choice of preference distribution $\{q_n\}$ . Thus, we set, $$\begin{aligned}
\displaystyle p_i=C_a\exp\left(-ai\right)\label{ref_prob_eq}\end{aligned}$$ where $C_a$ is the normalization constant to make $\sum_{j=1}^{14}p_j=1$ hold. We carried out the simulations with $a=0.05,\, 0.1,\, 0.3,\, 0.5,\, 1.3$. Judging from the frequency of winning of favorites in the data, $a$ is estimated to be $0.3\sim0.5$ in real HongKong races.
![\[fig:ref\_prob\] Shapes of $\{p_i\}$ for various values of $a$.](ref_prob.png){width=".7\linewidth"}
Choice of preference probability
--------------------------------
As for $q_n$, the only assumption one can naturally make is that $q_n$ decreases with $n$. To make variety of ${q_n}$ wide enough, we use the following function. $$\begin{aligned}
q_n=D_b\left(\frac{15-n}{14}\right)^b\end{aligned}$$ where $D_b$ is a normalization constant which make $\sum_{j=1}^{14}q_j=1$ hold. We set $b=1/400, 1/54, 1/20, 1/7, 1/4, 1/2, 1, 2, 4, 7, 20, 54, 400$. The shape of $\{q_n\}$ for these values of $b$ are listed in FIG. \[fig:pref\_prob\].
![\[fig:pref\_prob\] Shapes of $\{q_n\}$ for various values of $b$.](pref_prob.png){width=".7\linewidth"}
Results of simulations
======================
We carried out 2000 simulations for each set of parameters. The source cords and HongKong data of $F(t)$ are kept at\
https://github.com/qurihara/EconomicSimulation.
Preference distribution and FL bias
-----------------------------------
We first present a result of the case $a=0.3$ in FIG.\[fig:standard\]. The accumulation of bets is exponentially growing type and initial rate of bets $F(-20)$ is set to be 0.35.
![\[fig:standard\] popularity rank and rate of bets compared to consensus probabilities](standard.png){width=".8\linewidth"}
In FIG.\[fig:standard\], $b=400$ implies that all the bettors bet their money on the most beneficial horse. While, the almost uniform distribution, $b=1/400$, implies all the bettors bet their money on a randomly chosen horse regardless whether the horse is beneficial or not. It is not surprising that a FL bias appeared in the latter case, because people bet their money on longshots in equal probability as on favorites. It is truly astonishing that a strong FL bias is formed in $b=400$ case. The bias seems weakest around $b=4$.
![\[fig:1to3\] $V_i(0)/p_i$ for top 3 favorites.](1to3.png){width=".9\linewidth"}
Seeing FIG.\[fig:standard\], it is sufficient to compare $V_1$ with $p_1$ to measure the bias. Hence, in FIG.\[fig:1to3\], $V_i(0)/p_i$ for $i=1,2,3$ are exhibited, i.e., for top 3 favorites. In this figure, $V/p<1$ means FL bias. The result suggests that there exists a value of $b$ which maximize $V_i/p_i$. The simulation claims that FL bias occurs even if all the betters behave efficiently. Moreover, it claims that a bias can even be weakened when there is proper amount of inefficient betters.
Uniformness of consensus distribution and FL bias
-------------------------------------------------
Now, it is natural to change consensus distribution because bias can not happen when neither favorite nor longshot exists, i.e., consensus probability distribution is uniform. The bias must depend on how much stronger favorites are than longshots. Seeing the equation (\[ref\_prob\_eq\]), the distribution tends to uniform distribution when $a$ goes to $0$. While, as $a$ becomes greater than $0$, the favorite becomes stronger. FIG.\[fig:p\_change\] represents results of simulations for $a=0.05,\ 0.1,\ 0.3,\ 0.5,\ 1.3$.
![\[fig:p\_change\] How FL bias depends on consensus distributions.](p_change.png){width=".9\linewidth"}
As the result shows, basically, the more outstaning the favorite, the stronger the bias becomes. Another notable feature is how the bias depends on preference distribution. Every time we fix $a$, there exists a value of $b$ which best weaken the FL bias. And such $b$ becomes greater as $a$ becomes greater. That is to say, the stronger the favorite becomes, the more we should bet efficiently to eliminate the FL bias. But bias does not disappear regardless. And yet another mystery is that the bias is again enforced when $b$ exceed the best value.
Rapidity of accumulation of betting and FL bias
-----------------------------------------------
Since FL bias is a kind of over/under-estimation, it must depend on the amount of bettors who are making estimations at time $t$. Hence, in this section, we will change the time dependence of the number of bettors. We introduced two viewpoints in section 2.1. First, we change the percentage of what we call pre-bettors, i.e., the bettors already finished their bets more than 20 minutes before a race start. In real Hong-Kong races, a typical percentage of pre-bettors is 35%. But it is apparent that a percentage of pre-bettors has a large effect on the final distribution of bettors. Hence we not only examined 35%, but also a condition that the percentage of pre-bettors is very small, i.e., 1%. In both cases, we made the odds distributions of pre-bettors correspond to consensus probability distribution with a certain amount of random fluctuation. Second of all, we changed the rapidity of the growth of the numbers of bettors during the last 20 minutes. We examined exponential growth, linear growth and exponential decay. In all, we examined $2\times3=6$ models for the growth of numbers of bettors. The mathematical expression for each model becomes unique due to the conditions $F(0)=1$ and F(-20)=0.35 (or 0.01).\
abbreviation growth of $f(t)$ pre-bettors expression of $F(t)$
-------------- -------------------- ------------- ----------------------
0.35-growth exponential growth 35% $0.35^{-t/20}$
0.35-linear linear growth 35% 0.0325t+1
0.35-decay exponential decay 35% $1.35-0.35^{1+t/20}$
0.01-growth exponential growth 1% $0.01^{-t/20}$
0.01-linear linear growth 1% 0.0495t+1
0.01-decay exponential decay 1% $1.01-0.01^{1+t/20}$
: models for $f(t)$.[]{data-label="expfit"}
FIG.\[fig:invest\_change\] represents how FL bias depends on the rapidity of growth of bettors. The vertical axis again represents $V_1/P_1$. The parameter $a$ is set to be $0.3$.
![\[fig:invest\_change\] How $V/p$ changes with $f(t)$](invest_change.png){width=".9\linewidth"}
As is shown, FL bias appears more radically when the percentage of pre-bettors is 0.01. Now we have to be mindful that the pre-bettors bet in accordance with the consensus probability distribution. This might partially explain why $V_1/p_1$ becomes higher for 1% case than for 35% case. Another outstanding feature is that the bias is strongest in 0.01-deay model. We do not yet know any quantitative explanation for these result.
Discussion and future works
===========================
The model in this paper assumes that each bettor behaves to make $V_i$ near to $p_i$. However, $V_i$ becomes greater than $p_i$ for longshots and smaller for favorites. This can be interpreted as a FL bias. Since it is a purely statistical model with homogeneous bettors, the reason for this bias should exist in macro-micro interaction, i.e., sorting of horses. We believe it is interesting that such simple and uniform rules in micro behavior result in unexpected macro statistics. It is also observed that the strength of bias depends on consensus probability distributions and how bets accumulate over time. The more favorites outstand, the more bias get stronger. The bias becomes stronger if $f(t)$ has points where it increases rapidly. Qualitatively speaking, a minimum unit of bet for longshots makes a larger disproportionate effect on odds distribution than for favorites because the number of bettors in longshots is smaller than those in favorites. For example, assume that the number of bettors for longshots is 1 and for favorites 10. Then, $V$ for longshots is $1/11$ and favorites 10/11. Now assume that one bet is added to longshots. Then, $V$ for longshots becomes $2/12$. If it is added to favorites, $V$ becomes $11/12$. In this example, it is apparent that the change in $V$ is much greater in longshots. But we yet do not know whether this explains our results enough. Due to the sorting process, it is difficult to give mathematical analysis. Probably, analysing $N=2$ case will be a first future work. A master equation for $N=2$ case is exhibited in Appendix.\
Another branch for this study is that it can serve to analyze real systems similar to pari-mutuel system. For example, it simulates games such as roulette and baccarat where consensus distributions are disclosed to bettors. Another type of application is for growth of queues at counters in shops or gates on highways. In these settings bettors are interpreted as customers. each customer choose a queue to join, judging from the consensus length of the queue. Favorites now implies counters or gates which have shorter consensus length. In the field of stochastic process, our model is a kind of anomalous walk where the drifts can vary depending on the position of walkers.\
In terms of pari-mutuel betting systems, the most necessary modification for the model will be to introduce a track take. In real horse races, efficient bettors will bet their money only when the benefit is positive even if the track take is considered. That probably means we should set a threshold for betting in terms of $V_i/p_i$.
Master equation for $N=2$ case
==============================
Let $P(\cdot ,t)$ denote the probability density function for $V_1(t)$. Then, the state $k<V_1(t)<k+dk$ occurs with the probability $P(k,t)dk$.The aim of this appendix is to show that the below is the time evolution equation of $P(k,t)$ for $N=2$ case. $$\begin{aligned}
&&P(k,t)=\int_{-\infty}^{p}P(s,t-1)\, \phi\left(k-s;\, \frac{(s-q)f(t)}{F(t)},\, \frac{q\overline{q}f(t)}{mF(t)^2}\right)ds\nonumber\\
&&\hspace{2em}+
\int_{p}^{\infty}P(s,t-1)\, \phi\left(k-s;\, \frac{(s-\overline{q})f(t)}{F(t)},\, \frac{q\overline{q}f(t)}{mF(t)^2}\right)ds.
\label{master}\\
&&\displaystyle\phi(x;\, \mu,\sigma^2)=
\frac{1}{\sqrt{2\pi}\sigma}\exp\left(-\frac{(x-\mu)^2}{2\sigma^2}\right)
\nonumber\end{aligned}$$ Here, we defined $q=q_1,\ \overline{q}=q_2(=1-q)$ and $p=p_1$ for simplicity. $m$ is the total number of bets in a race fixed as 200,000 in this paper. $\phi(\cdot\,;\mu,\sigma^2)$ is the probability density function for ${\rm N}(\mu,\sigma^2)$, the normal distribution with the average $\mu$ and variance $\sigma^2$.\
First notice for deriving (\[master\]) is that, if $V_1(t-1)<p$, the amount of bets accumulating to the favorite at time $t$ yields ${\rm B}(mf(t), q)$. Here, ${\rm B}(n, r)$ denotes the binomial distribution with $n$ being total number of experiments and $r$ the probability of a successful result. Then, since $mf(t)$ is large, ${\rm B}(mf(t), q)$ is approximated by ${\rm N}(mqf(t),\ mq\overline{q}f(t))$. Then, we interpret this observation in terms of $V_1(t)$. Suppose $V_1(t-1)$ increases from $y$ to $y+z$ at time $t$, i.e. $V_1(t)=y+z$. This implies that $m(zF(t)-yf(t))$ bets are made at time $t$. Hence, the increment $z$ yields ${\rm N}((y-q)f(t)/F(t),\ f(t)q\overline{q}/mF(t)^2)$. We can analyse the $V_1(t-1)\geq p$ case in the same manner. In this case, $z$ yields ${\rm N}((y-\overline{q})f(t)/F(t),\ f(t)q\overline{q}/mF(t)^2)$.\
Now, we proceed to consider a recurrence equation for $P(k,t)$. One way to reach the state $V_1(t)=k$ is that $V_1$ increases by $k-s$ after $V_1(t-1)=s$ occurred. The probability density of this is a multiple $$P(s,t-1)\cdot \phi\left(k-s;\, \frac{(s-q)f(t)}{F(t)},\, \frac{q\overline{q}f(t)}{mF(t)^2}\right),$$ if $s$ is less than $p$. The r.h.s. of the equation (\[master\]) is nothing but an integration of such multiples for various values of $s$.
References
==========
[9]{}
L. Bachelier, ‘Théorie de la speculation’, Paris: Gautier-Villars, reprinted in English in P. Cootner (ed.) ‘The Random Character of Stock Market Prices’, Cambridge, MA: MIT Press. (1964) E. F. Fama, ‘Efficient Capital Markets II’ Journal of Finance 46 (5) (1991) pp 1575-1617. R. M. Griffith, ‘Odds Adjustments by American Horse-Race Bettors’, American Journal of Economic Psychology, 62 (1949) pp 290-294. R. S. Sobel and M. E. Ryan, ‘Unifying the Favorite-Longshot Bias with Other Market Anomalies’ in D B Hausch and W T Ziemba (ed) Handbook of Sports and Lottery Markets (Handbooks in Finance) (Elsevier, Amsterdam , NL, 2008) pp 137-160. L. Coleman, ‘New light on the longshot bias’ , Applied Economics 36 (2004) pp 315-326. A. C. Bruce and J. E. V Johnson, ‘Toward an Explanation of Betting as a Leisure Pursuit’ , Leisure Studies 11 (1992) pp 201-218. S. Chad and R. E. Quandt ‘Betting Bias and Market Equilibrium in Racetrack Betting’ , Applied Financial Economics 6 (3) (1996) pp 287-292. R. Thaler and W. Ziemba, ‘Parimutuel Betting Markets: Racetracks and Lotteries’ , Journal of Economic Perspectives 2 (1988) pp 161-174. M. M. Ali, ‘Probability and utility estimates for racetrack bettors’ , Journal of Political Economics 84 (1977) pp.803-815. S. Blough, ‘Differences of opinions at the racetrack’ . In: Hausch, D., Lo, V., Ziemba, W. (Eds.), Efficiency of Racetrack Betting Markets, (1994) pp. 323-341. R. D. Brown and Y. Lin ‘Racetrack betting and consensus of subjective probabilities’ , Statistics & Probability Letters 62 (2003) pp.175-187. M. Gramm and C. N. Mckinney ‘The effect of late money on betting market efficiency’ , Applied Economics Letters 16 (4) (2009) pp. 369-72 The URL of HongKong racetracks,\
http://www.hkjc.com/home/english/index.aspx
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The inclusive production of $\omega$ and $\phi$ mesons is studied in the backward region of the interaction of 12 GeV protons with polyethylene, carbon, and copper targets. The mesons are measured in $e^+ e^-$ decay channels. The production cross sections of the mesons are presented as functions of rapidity $y$ and transverse momentum $p_T$. The nuclear mass number dependences ($A$ dependences) are found to be $A^{0.710 \pm 0.021(\text{stat}) \pm 0.037(\text{syst})}$ for $\omega$ mesons and $A^{0.937 \pm 0.049(\text{stat}) \pm 0.018(\text{syst})}$ for $\phi$ mesons in the region of $0.9 < y < 1.7$ and $p_T < 0.75$ GeV/$c$.'
author:
- 'T. Tabaru'
- 'H. En’yo'
- 'R. Muto'
- 'M. Naruki'
- 'S. Yokkaichi'
- 'J. Chiba'
- 'M. Ieiri'
- 'O. Sasaki'
- 'M. Sekimoto'
- 'K. H. Tanaka'
- 'H. Funahashi'
- 'Y. Fukao'
- 'M. Kitaguchi'
- 'M. Ishino'
- 'H. Kanda'
- 'S. Mihara'
- 'T. Miyashita'
- 'K. Miwa'
- 'T. Murakami'
- 'T. Nakura'
- 'F. Sakuma'
- 'M. Togawa'
- 'S. Yamada'
- 'Y. Yoshimura'
- 'H. Hamagaki'
- 'K. Ozawa'
bibliography:
- 'draft09.bib'
title: 'Nuclear mass number dependence of inclusive production of $\omega$ and $\phi$ mesons in 12 GeV $p$ + $A$ collisions'
---
\[sec:introduction\]Introduction
================================
The modification of the vector meson spectral function in hot and/or dense matter is currently a hot subject in terms of spontaneous chiral symmetry breaking and partial restoration of the symmetry in nuclear matter. Currently, this is the focus of several experiments [@ceres; @tagx; @pionic; @taps]. The experiment KEK-PS E325 was performed to measure the vector meson spectral functions in dense matter, i.e., nucleus. Thus far, we have reported the signature of the mass modification of vector mesons [@ozawa; @naruki; @muto]. These observations are fairly remarkable; hence, we also performed analyses to determine the absolute cross sections and nuclear mass number dependences of the production of these mesons in order to understand the underlying production mechanism.
The nuclear mass number dependence of the cross sections for the particle production is usually parameterized as $$\sigma(A)=\sigma_0 A^\alpha
\label{eq:alpha}$$ for a target nucleus with mass number $A$. When the collision energy $\sqrt{s_{NN}}$ is sufficiently large, the parameter $\alpha$ in the production of light mesons like pions or $\rho$ mesons is about 2/3 [@Ono; @Binkley]. This can be interpreted by considering such productions to be dominated by primary collisions at the front surface of a target nucleus. Note that the mean free path of incident protons in a nucleus is as small as 1.4 fm. On the other hand, $\alpha$ tends to unity in the case of hard reactions like the production of $J/\psi$ at high energies, $\sqrt{s_{NN}} \gtrsim 20$ GeV [@Binkley; @NA50; @HERA-B].
The present experiment was performed at $\sqrt{s_{NN}} = 5.1$ GeV. At higher energy, $\alpha$ for $\phi$ meson production was reported to be $0.81 \pm 0.06$, $0.96 \pm 0.04$, and $0.86 \pm 0.02$ at $\sqrt{s_{NN}} = 11.6$ [@Aleev], 14.2 [@Daum], and 15.1 GeV [@Bailey], respectively. However, there is no reason to believe that these values are applicable at our energy. A few heavy ion induced experiments at $\sqrt{s_{NN}}$ from 4.9 to 5.4 GeV [@E917; @E802] reported $\phi$ meson production data, to which the present experiment can provide complementary data. Note that there have been no measurements of $\omega$ mesons and $\phi$ mesons with $p$ + $A$ reactions at $\sqrt{s_{NN}}$ around 5 GeV.
A $\phi$ meson is almost a pure $s \bar s$ state. Therefore, the production of a $\phi$ meson without other accompanying strange particles is suppressed by the Okubo-Zweig-Iizuka rule [@OZI]. This results in the effective threshold energy being as high as $\sqrt{s_{NN}} = 3.9$ GeV, which corresponds to $2m_p + 2m_{s \bar s}$, since two $s \bar s$ pairs are effectively needed to realize the $\phi$ production. Our collision energy is fairly close to this effective threshold. In addition, the importance of additional mechanisms for $\phi$ meson production, such as an intrinsic $s \bar s$ component in nucleons [@Ellis] and $\phi \rho \pi$ coupling [@vonRotz; @Nakayama2], have been suggested by theorists; however, they have thus far been insufficiently studied at our energy from an experimental viewpoint. Thus, to understand the production mechanism at our energy, basic measurements such as production cross sections and $\alpha$ parameters are indispensable.
In this article, we present inclusive production cross sections and $\alpha$ parameters of $\omega$ and $\phi$ mesons measured via the $e^+ e^-$ decay channels in 12 GeV $p + p$, $p$ + C, and $p$ + Cu collisions. The results are compared to those from nuclear cascade simulation [<span style="font-variant:small-caps;">jam</span>]{} [@jam], and the implications to production mechanisms are discussed.
\[sec:experiment\]Experiment
============================
The spectrometer was built in the EP1-B beam line at the 12 GeV proton synchrotron in KEK. The beam line was designed to deliver a 12 GeV primary proton beam with an intensity of up to $4 \times 10^9$ protons per spill. The beam was extracted for 1.8 s with a repetition rate of 1/4 Hz. The beam intensity was monitored with about 10% accuracy using ionization chambers [@IonChamber] located downstream of the spectrometer.
Figures \[fig:setup\] and \[fig:setup2\] show a top view and a side view of the experimental setup, respectively. The spectrometer was designed to simultaneously measure $e^+ e^-$ and $K^+ K^-$ pairs. A dipole magnet and tracking devices were commonly used together with electron and kaon identification counters.
![ \[fig:setup\] Schematic view of the experimental setup from the top, which is designed symmetrically with respect to the beam. The components of this setup are referred as left and right arms in this article. ](fig1)
![ \[fig:setup2\] Cross section of the experimental setup along the plane of the center of the kaon identification counters. ](fig2)
The magnet was operated at 0.71 T at the center of the dipole gap and provided 0.81 Tm of field integral for the tracking. The magnetic field map was calculated by the finite-element-analysis software [<span style="font-variant:small-caps;">tosca</span>]{} [@TOSCA]. The calculated map agreed well with the measured map, and the difference was negligible when compared to the momentum resolution of this spectrometer. During the data collection periods, the field strength was monitored every 4 s with a nuclear-magnetic-resonance (NMR) probe located at the center of the surface of the lower pole piece. The magnetic field map was scaled run by run according to the NMR data. The fluctuation of the magnetic field was found to be less than $10^{-5}$ within a typical run of two hours long.
The targets were aligned inline along the beam axis at the center of the magnet. The target materials and configurations are shown in Table \[table:config\]. =6.8mm
[ c c d r@[.]{}l r@[.]{}l d@[ ]{}l c ]{} year & target & & & & & trigger\
& CH$_2$ & $-48$ & 0 & 111 & 0&195 & & & electron\
1999 & C & 38 & 0 & 106 & 0&213 & 3.0 & $\times 10^{13}$ & / kaon\
& Cu & $-7$ & 0 & 0391 & 0&412 & & &\
2002 & C & 0 & 0 & 213 & 0&431 & 3.2 & $\times 10^{14}$ & electron\
& Cu & & $4 \times 0$ & 0539 & $4 \times 0$&565 & &\
The beam profile in the horizontal direction was measured by counting the interaction rate by changing the beam position at the target. The beam position was moved by the bending magnet located about 10 m upstream of the target. In this measurement, the center target, whose thickness was 1 mm, was rotated by $90^\circ$ around the vertical axis, and was used as a 1 mm-wide probe. The typical beam size in the horizontal direction was found to be 2.0 mm in full-width half-maximum. The beam size in the vertical direction was known to be almost the same as in the horizontal direction as seen in a view of a luminescence plate which was temporarily inserted during the beam tuning.
Three tracking devices—the vertex tracking chamber (VTC), cylindrical drift chamber (CDC), and barrel drift chamber (BDC)—were used to determine the trajectories of the charged particles. In the present analysis, the momentum was determined using CDC and BDC. The momentum resolution $\sigma_p$ was $\sigma_p = \sqrt{(1.37\% \cdot p)^2 + 0.41\%^2}\cdot p$ (GeV/$c$), where $p$ is the momentum of a particle.
For the electron identification, two types of gas Čerenkov counters (FGC and RGC) and three types of lead glass calorimeters (SLG, RLG, and FLG) were employed. The gas Čerenkov counters were horizontally segmented into 6$^\circ$. The radiator of the gas Čerenkov counters was isobutane at room temperature and atmospheric pressure. The refractive index is 1.00127 at the standard temperature and pressure, which corresponds to a momentum threshold of 2.26 GeV/$c$ for charged pions. All lead glass detectors were built with SF6W [@SF6W] and were typically segmented to 3.5$^\circ$ horizontally. The typical energy resolution for 1 GeV electrons is 0.15 GeV. The acceptance for electrons ranged from $\pm$12$^\circ$ to $\pm$90$^\circ$ in the horizontal direction and from $-22^\circ$ to 22$^\circ$ in the vertical direction.
The counters for the kaon identification were start timing counters (STC), hodoscopes (HC), aerogel Čerenkov counters (AC), and time of flight counters (FTOF). The acceptance for kaons ranged from $\pm$12$^\circ$ to $\pm$54$^\circ$ in the horizontal direction, and from $-6^\circ$ to 6$^\circ$ in the vertical directions. The three counters (STC, AC, and FTOF) were horizontally segmented to 6$^\circ$, and HC was typically segmented horizontally to 3$^\circ$. The time of flight (TOF) of charged particles was measured using STC and FTOF with a resolution of 0.36 ns and flight length of about 3.7 m. Kaons and pions were separated in a momentum range from 0.53 to 1.88 GeV/$c$ using an aerogel with a refractive index of 1.034.
The electron trigger signal required a hit in FGC accompanied by a geometrical coincidence with RGC, SLG, or RLG. To select events containing $e^+ e^-$ pairs with large opening angles, both the left and the right arms were required to contain at least one $e^+$ or $e^-$ candidate. The typical efficiency of the trigger for the electron pairs in the acceptance was 92.4%.
The kaon trigger signal was obtained from the coincidence of STC, HC, and FTOF. The charged pion contamination was reduced by using AC as a veto trigger. Proton contamination was reduced by setting a TOF window for the kaons by using a rough momentum value calculated by combining the hits in STC, HC, and FTOF in the trigger.
The number of recorded events for electron data were $7.41 \times 10^7$ and $5.08 \times 10^8$ in 1999 and 2002, respectively. A detailed description of the spectrometer can be found in Ref. [@nim].
\[sec:analysis\]Data Analysis
=============================
\[ssec:e analysis 2002\] Analysis of $\omega$ and $\phi \to e^+ e^-$ in $p$ + C and $p$ + Cu collisions
--------------------------------------------------------------------------------------------------------
### \[sssec: track reconstruction\] Event reconstruction
The charged tracks were reconstructed from hit positions in the drift chambers by using the Runge-Kutta fitting method. After the tracks were reconstructed, tracks corresponding to momenta between 0.4 and 2.0 GeV/$c$ were selected for further analyses. The lower limit in the momentum range was set based on the threshold of the trigger, whereas the upper was set in order to avoid pion contamination above the gas Čerenkov threshold.
Pairs of positive and negative tracks were required to satisfy the trigger condition. All the $e^+$ and $e^-$ candidates were reexamined so that an FGC hit association could be obtained with an RGC, SLG, or RLG association, depending on the location of the track. For candidates associated with lead glass calorimeters, the momentum ratio $E/p$ should be larger than 0.5 for energy in the calorimeters to be obtained. We chose the value of 0.5 balancing the purity of $e^+$ and $e^-$ with the statistics of the present data. Figure \[fig:E/p\] shows the distribution of the energy and momentum of the present data with FGC associations. In this figure, it is clearly seen that electrons range along the line of $E=p$. After an $e^+ e^-$ pair was detected, we simultaneously refit the $e^+$ and $e^-$ tracks by constraining them to have the same vertex point on the interaction target. Finally, we identified $5.69 \times 10^5$ $e^+ e^-$ pairs.[^1]
![ \[fig:E/p\] A contour plot of the distribution of energy and momentum of the present data with FGC associations. The solid line shows the criterion of electron identification described in this article. ](fig3)
The reconstruction efficiency of the tracks from the targets was evaluated by both an eye-scan and a detector simulation using [<span style="font-variant:small-caps;">geant4</span>]{} [@g4].
In the former method, first we used only one of the two arms and determined the target in which an interaction occurred. Then we visually scanned about 200 event displays and found track candidates in the other arm using the drift chamber hit information with help from the interaction point. The tracking efficiency was evaluated by seeing if the tracking program could find those eye-scanned track candidates. The efficiency was found to be 67% on average.
In the detector simulation method, the reconstruction efficiency was evaluated using simulation tracks embedded onto the real data. According to this method, the efficiency was found to be 78%.
In order to combine the results of the two evaluations, we simply assumed the average 73% to be the efficiency in the present analysis and considered half the difference as a systematic error. The reason of this discrepancy is unknown, and this is one of the major sources of the systematic uncertainty in the present data analysis.
The inefficiency of the vertex reconstruction process was evaluated as follows. We took all the combinations of $e^+$ and $e^-$ tracks regardless of the position of the closest point of each pair, and obtained the yield of the $\omega$ meson peak. Then we used the vertex reconstruction program and the required the event vertex belong to any of the target disks. In addition, we fit the $e^+$ and $e^-$ tracks together by constraining them to have the same vertex point on the interaction target, and required that the $\chi^2$ over the number of degrees of freedom (NDF) should be less than 5. As a result, we lost 6.0% of the $\omega$ meson yield, so that the vertex reconstruction efficiency was 94.0%.
Contaminations due to the misidentification of pions and other particles in the present data were evaluated at the mass region of the $\omega$ meson, i.e., from 0.75 to 0.80 GeV/$c^2$, by tightening the electron identifications with gas Čerenkov counters and lead glass calorimeters until the misidentification becomes negligibly small. In this mass region, in the $p$ + C data, we found that 18% of the events result from misidentification and 18% are from uncorrelated $e^+ e^-$ pairs.
### \[sssec:correction\]Corrections
Besides the tracking efficiency, several detector effects were evaluated and corrected as described below. The efficiencies of the electron identification counters were evaluated as a function of momentum by using pure electron samples from $\gamma$ conversions and Dalitz decays. These electron samples were identified as a zero-mass peak in the $e^+ e^-$ spectra, and they were not required to participate in the trigger to avoid trigger bias. The obtained efficiencies were typically 85% for FGC, 86% for RGC, and 97% for the lead glass counters.
The energy losses of the tracks through the detectors were estimated by using [<span style="font-variant:small-caps;">geant4</span>]{} simulation. Typically, the reconstructed momentum gave a value lower by 3 MeV/$c$ for a 1 GeV/$c$ electron due to the energy loss. The momentum difference was corrected track by track by using a correction table obtained by the simulation. It should be noted that this correction compensated only mean energy loss. For effects that cause an eventual large energy loss like bremsstrahlung, the corrections were carried out in a different manner, as described later.
The geometrical acceptances for vector mesons $V$ were obtained as functions of the invariant mass, rapidity $y$, and transverse momentum $p_T$ by the simulation. The acceptances were averaged over azimuth $\varphi$, and isotropic decays of $V \to e^+ e^-$ were assumed. The effects of the trigger, i.e., requirements of a geometrical correlation between the electron identification counters, were also considered. In order to obtain the yields of vector mesons, mass spectra were corrected for the acceptance in the mass range above 0.55 GeV/$c^2$. Below 0.55 GeV/$c^2$, the correction was too large to evaluate reliable values. The obtained acceptances at the $\omega$ and the $\phi$ meson masses are tabulated in Fig. \[table:Acc Sample\].
 
### \[sssec:spectrum decomposition\]Spectrum decomposition
In Fig. \[fig:ee uncorr mass\], the left and the right panels show the spectra of the invariant mass of $e^+ e^-$ pairs in the range of $0.9 < y < 1.7$ and $p_T < 0.75$ GeV/$c$ without and with the acceptance correction, respectively. The peaks of $\omega$ and $\phi$ are distinctly observed.
The $e^+ e^-$ mass spectra were fit and decomposed into the dielectron decays $\omega \to e^+ e^-$, $\phi \to e^+ e^-$, and $\rho^0 \to e^+ e^-$; the Dalitz decays $\eta \to \gamma e^+ e^-$, $\omega \to \pi^0 e^+ e^-$, $\phi \to \pi^0 e^+ e^-$, and $\phi \to \eta e^+ e^-$; and the combinatorial background. The origins of the combinatorial background were pairs which were picked up from two independent Dalitz decays or $\gamma$ conversions, and pairs like $e^\pm \pi^\mp$ due to the misidentification. The Dalitz decay $\pi^0 \to \gamma e^+ e^-$ contribution is negligible in the acceptance of the present data.
For the invariant mass distributions of the $\omega \to e^+ e^-$ and $\phi \to e^+ e^-$ decays, we used in the fit a Breit-Wigner function $$\frac{d\sigma}{dm} = \frac{N}
{\left( m - m_0 \right)^2 + \Gamma^2_\text{tot}/4}
\label{eq:BW}$$ convoluted with a Gaussian function for the experimental resolution. If the relativistic Breit-Wigner shape is used instead, the results do not change significantly. Here, $\sigma$ is the cross section; $m$, the invariant mass of the $e^+ e^-$ pair; $N$, a normalization factor; $m_0$, the meson mass; and $\Gamma_\text{tot}$, the natural decay width. For the shape of the $\rho^0 \to e^+ e^-$ decay, we used the relativistic Breit-Wigner shape $$\frac{d\sigma}{dm} = \frac{N}
{\left( m^2 - m^2_{\rho^0} \right)^2 + m^2 \Gamma^2_\text{tot}},
\label{eq:Naruki}$$ instead of the Breit-Wigner function. To obtain the mass distribution in the spectrometer acceptance, we used the [<span style="font-variant:small-caps;">geant4</span>]{} simulation with an input momentum distribution of $\rho^0$ mesons that was obtained using [<span style="font-variant:small-caps;">jam</span>]{}. The $e^+ e^-$ invariant mass spectra from the Dalitz decays of $\eta$, $\omega$, and $\phi$ mesons were also obtained by the simulation. Their $e^+ e^-$ distributions were determined by following the vector meson dominance model given in Ref. [@VDMFormula] by using the mother meson distributions obtained by [<span style="font-variant:small-caps;">jam</span>]{}. The combinatorial background shape was evaluated using an event mixing method, combining $e^+$ and $e^-$ tracks picked from different events.
The free parameters of the fit were the yields, the peak positions, and the mass resolutions of $\omega$ and $\phi$ mesons; the yields of $\eta$ and $\rho^0$ mesons; and the number of the background events. As mentioned earlier, the spectra were corrected in a mass range only above 0.55 GeV/$c^2$. Therefore, below 0.55 GeV/$c^2$, the uncorrected spectra were used in the fit mainly to accurately obtain the amount of the background. Although the fit region was from 0.20 GeV/$c^2$ to 1.2 GeV/$c^2$, the mass regions from 0.600 to 0.765 GeV/$c^2$ and from 0.955 to 0.985 GeV/$c^2$ were excluded in order to avoid the effect of the excesses below the $\omega$ and $\phi$ peaks, which in other publications [@ozawa; @naruki; @muto] we have claimed as the signal of the mass modification. In the present analysis, however, we did not assume any underlying physics for the excess, and we aimed only to obtain the yields of $\omega$ and $\phi$ mesons correctly, as discussed later.
The fit results are shown by thin solid lines in Fig. \[fig:ee uncorr mass\]. The $\chi^2/\text{NDF}$ were obtained as 157/152 and 192/152 for the $p$ + C and $p$ + Cu interactions, respectively.
 
After obtaining the raw yields of the $\omega$ and $\phi$ mesons by the fit, the correction for hard energy losses such as bremsstrahlung or a radiative tail was applied. The hard energy loss causes a low mass tail in the invariant mass distribution. The loss of the yield due to this tail could not be evaluated using the procedure described above; hence, we performed studies using the [<span style="font-variant:small-caps;">geant4</span>]{} simulation. The yields of the tails with respect to the integrals of the Breit-Wigner peaks were found to be $11.9 \pm 1.0$% for the $\omega$ mesons and $10.4 \pm 0.8$% for the $\phi$ mesons. These values were simply added to the peak yields. The uncertainty of these corrections includes an ambiguity with respect to the cross sections of such hard energy losses in [<span style="font-variant:small-caps;">geant4</span>]{}.
By the fit procedure described above, the peak position and mass resolution of the $\phi$ meson were obtained as $1019 \pm 1$ MeV/$c^2$ and $11.8 \pm 1.0$ MeV/$c^2$, respectively. The peak position of the $\phi$ meson is consistent with the values of the Particle Data Group [@PDG], and the mass resolution is consistent with the simulation value of 10.7 MeV/$c^2$; however, the peak position of the $\omega$ meson was found to be lower by 2.2 MeV/$c^2$. The width of the omega mesons was also broader than that expected by the simulation.[^2] Here, we adopted a conservative approach in order to obtain the yield of unmodified $\omega$ mesons by estimating the additional systematic error due to this peak shift. The error for the $\omega$ meson yield was evaluated as 0.48% for a $p$ + C interaction and 1.8% for a $p$ + Cu interaction, which were obtained by forcing the peak position in the fit to the higher value of 2.2 MeV/$c^2$. The momentum scale and resolution were also verified using the $K_s \to \pi^+ \pi^-$ data [@muto]. The measured peak position of $K_s$ was $496.8 \pm 0.3$ MeV/$c^2$, which was consistent with that of the simulation result, 496.9 MeV/$c^2$. The measured mass resolution of $K_s$ was $3.9 \pm 0.4$ MeV/$c^2$, which was expected to be 3.5 MeV/$c^2$ by the simulation.
### \[sssec:systematic errors\]Systematic errors
In addition to the uncertainties described previously, the following systematic uncertainties were studied to obtain the cross sections.
The uncertainty in the background shapes could be a source of the systematic error. The background was a result of uncorrelated pairs that were obtained from two independent Dalitz decays, $\gamma$ conversions, or other meson decays. The background shape was obtained by an event mixing technique. In the mixing process, it is possible that the correlated $e^+ e^-$ pairs from the decays of $\rho^0$, $\omega$, and $\phi$ mesons deform the estimated background shape; this could result in a systematic error. In order to estimate the systematic error, two methods were used in the event mixing. One was to use all electrons except for those belonging to the $\omega$ and $\phi$ meson mass regions, which are 0.765 to 0.800 GeV/$c^2$ and 0.995 to 1.035 GeV/$c^2$, respectively. In the other method, we used all the pairs in the event mixing but with weights in order to obtain a self-consistent shape of the background. The weights were obtained as a function of the $e^+ e^-$ mass as the ratio of the background shape to the real mass spectrum, and they were self-consistently determined by repeating the fit several times. We adopted the latter method to determine the background shape. The difference between the two methods was assigned as a systematic error. The difference in the yield of $\omega$ mesons was 0.23% and that in $\phi$ mesons was 0.60%.
The normalization and spectral shape of the combinatorial background were affected by all the other correlated $e^+ e^-$ components since they were obtained by the fit. The systematic errors in the peak-yield determination due to the ambiguity of the Dalitz decays of $\eta$, $\omega$, and $\phi$ mesons were evaluated by doubling or eliminating those yields and refitting the mass spectra. These systematic errors were found to be 0.62% and 0.85% for $\omega$ and $\phi$ mesons, respectively.
It was difficult to consider the uncertainty of the $\rho^0$ shape, since the mass region from 0.600 to 0.765 GeV/$c^2$ could not be represented by known sources, and the shape of the mass modification was not well understood. In this analysis, we simply considered the relativistic Breit-Wigner shape for the $\rho^0$ distribution and performed a fit by excluding the excess region. In order to evaluate the systematic error, we fixed the $\rho^0$ yield at zero as an extreme case and reperformed the fit. In this fit, the $\omega$ yield increased by 3.19% for the carbon target and 7.65% for the copper target; further, the $\phi$ meson yield increased by 3.45% and 5.48% for each of the above targets, respectively. We considered these values as systematic errors due to the unknown $\rho^0$ distribution. It should be noted that the $\omega$ mesons can also be modified such that the $\omega$ cross sections obtained in the present analysis are only for that component whose shape is consistent with the un-modified shape.
A possible cause of another error might lie in the efficiency estimations for electron identification. These uncertainties were evaluated by using the error bands of efficiency curves of electron identification counters and were obtained as 2.92% for $\omega$ mesons and 2.60% for $\phi$ mesons.
All the systematic errors are summarized in Table \[table:sys error 2002\]. In the evaluation of the absolute cross section, the errors of the beam intensity, target thicknesses, and efficiency of the trigger electronics were also considered. In summary, the systematic errors for the cross sections for $\omega$ mesons were 19.5% for the carbon target, and 20.8% for the copper target, and those of $\phi$ mesons were 19.5% and 20.0%, respectively. =6.6mm
[ c d@[%]{} c d@[%]{} ]{} & &&\
&\
acceptance & 1.23 && 1.25\
hard energy loss & 1.18 && 0.94\
electron identification & 2.92 && 2.60\
background shape & 0.23 && 0.60\
mass scale for $p$ + C & 0.48 &&\
mass scale for $p$ + Cu & 1.83 &&\
Dalitz yield & 0.62 && 0.85\
$\rho^0$ from $p$ + C & 3.2 && 3.5\
$\rho^0$ from $p$ + Cu & 7.6 && 5.5\
total for $p$ + C & 19.5 && 19.5\
total for $p$ + Cu & 20.8 && 20.0\
The systematic errors for the $\alpha$ parameters are only due to the items that differ between carbon and copper targets. These are listed in Table \[table:alpha sys error 2002\]. In summary, the uncertainties for the $\alpha$ parameters of $\omega$ and $\phi$ mesons were 5.2% and 1.9%, respectively. =7.5mm
[ c d@[%]{} c d@[%]{} ]{} & &&\
acceptance & 1.47 && 1.13\
target thickness & 0.52 && 0.39\
background shape & 0.27 && 0.54\
mass scale & 1.13 &&\
Dalitz yield & 0.37 && 0.48\
$\rho^0$ yield & 4.89 && 1.25\
total & 5.2 && 1.9\
\[ssec:e analysis 1999\] Extraction of the cross section in $p + p$ interaction
--------------------------------------------------------------------------------
The target subtraction method was employed to obtain the cross section in $p + p$ collisions; this was achieved by using the data taken in 1999. The production cross sections $\sigma(p)$ in $p + p$ collisions were calculated by the formula $\sigma(p)=(\sigma($CH$_2)-\sigma($C$))/2$ in the region of $0.9 < y < 1.7$ and $p_T < 0.6$ GeV/$c$. The analysis was performed in almost the same manner as that for the data in 2002.
To minimize the uncertainty due to experimental differences between the 1999 and 2002 data, we normalized the yield of $\omega$ mesons with carbon targets in the 1999 data to that in 2002 data and extracted the absolute cross sections in $p + p$ collisions. Another uncertainty arose from the difference in the target thicknesses that were measured with an accuracy of 0.8%. We obtained 20.1% and 20.0% as the systematic uncertainty in $p + p$ collisions for the $\omega$ meson production and $\phi$ meson production, respectively.
\[sec:result\]Result and discussion
===================================
\[ssec:raw CS\]Global feature of the obtained results
-----------------------------------------------------
From the decay branching fractions listed in Table \[table:branching ratio\], the differential cross sections of the inclusive $\omega$ meson production are obtained as $14.30 \pm 0.34(\text{stat}) \pm 2.79(\text{syst})$ mb$\cdot c$/GeV and $46.63 \pm 1.25(\text{stat}) \pm 9.70(\text{syst})$ mb$\cdot c$/GeV in $p$ + C and $p$ + Cu collisions in the region of $0.9 < y < 1.7$ and $p_T < 0.75$ GeV/$c$, respectively; further, those of $\phi$ meson production are $0.270 \pm 0.017(\text{stat}) \pm 0.053(\text{syst})$ mb$\cdot c$/GeV and $1.290 \pm 0.070(\text{stat}) \pm 0.258(\text{syst})$ mb$\cdot c$/GeV, respectively.[^3] The $\alpha$ parameters of $\omega$ and $\phi$ mesons are obtained as $0.710 \pm 0.021(\text{stat}) \pm 0.037(\text{syst})$ and $0.937 \pm 0.049(\text{stat}) \pm 0.018(\text{syst})$, respectively. The difference is $0.227 \pm 0.054(\text{stat}) \pm 0.041(\text{syst})$, and it is statistically significant. =10.0mm
branching fraction
---------- -------------------- ----------------------------------
$\omega$ $e^+ e^-$ $(6.96 \pm 0.15) \times 10^{-5}$
$\phi$ $e^+ e^-$ $(2.96 \pm 0.04) \times 10^{-4}$
$\phi$ $K^+ K^-$ $49.2^{+0.6}_{-0.7}$%
: \[table:branching ratio\] Branching fractions in the tables of the Particle Data Group [@PDG].
In order to compare the production cross sections in the $e^+ e^-$ decay channels with those in the $K^+ K^-$ decay channel and the previous measurement of 12 GeV/$c$ $p + p \to \rho^0 X$ by Blobel [*et al.*]{} [@Blobel], the present data were extrapolated to the backward hemisphere—region of $x_F < 0$ or $y < 1.66$—where $x_F$ is Feynman’s $x$. The production cross sections in the backward hemisphere are listed in Table \[table:CS\]. The correction factors for this extrapolation from the measured regions—$0.9 < y < 1.7$ and $p_T < 0.75$ GeV/$c$—to the backward hemisphere were calculated by using [<span style="font-variant:small-caps;">jam</span>]{}, since the shapes of $p_T$ and $y$ spectra are consistent with the result of [<span style="font-variant:small-caps;">jam</span>]{} calculation, as described later. =11.4mm
---- ---- --- --- --- --- --- ---- --- --- --- ----- --- ----- --- ----- --- ---- --- ---- --- ----
H 2 . 3 1 . 1 0 . 5 0 034 0 045 0 007 0 03 0 19 0 01
C 13 . 4 0 . 3 2 . 6 0 240 0 015 0 047 0 39 0 05 0 19
Cu 49 . 0 1 . 3 10 . 2 1 21 0 07 0 24 1 84 0 27 0 89
Pb 6 0 1 8 2 9
---- ---- --- --- --- --- --- ---- --- --- --- ----- --- ----- --- ----- --- ---- --- ---- --- ----
The cross sections of the inclusive $\phi$ meson production in the $K^+ K^-$ decay channel are also listed in Table \[table:CS\]. These were obtained from the previous analysis in this experiment [@yokkaichi; @ishino]. The $\alpha$ parameter of the $\phi$ meson production measured in the $K^+K^-$ decay channel is obtained as $1.01 \pm 0.09$ using the $p + p$, $p$ + C, $p$ + Cu, and $p$ + Pb data in the spectrometer acceptance; this value is statistically consistent with the present $e^+ e^-$ analysis.
Figure \[fig:backwardCS\] shows the cross sections in the backward hemisphere as a function of the nuclear mass number. The dotted lines represent the result of the [<span style="font-variant:small-caps;">jam</span>]{} calculation and the solid and dashed lines represent the measured $\alpha$ parameterization. The $\alpha$ parameters shown in the figure were 0.710, 0.937, and 1.01 for the data of $\omega \to e^+ e^-$, $\phi \to e^+ e^-$, and $\phi \to K^+ K^-$, respectively. It should be noted that these values were obtained from the data within the spectrometer acceptance before they were extrapolated to the backward hemisphere.
The previous measurement [@Blobel] yielded the total cross section of $1.8 \pm 0.25$ mb for $\rho^0$ mesons. By assuming a $\rho^0/\omega$ ratio of $1.0 \pm 0.2$,[^4] we obtain the $\omega$ meson production in the backward hemisphere as $0.90 \pm 0.22$ mb for a comparison with the present measurement. The triangle in Fig. \[fig:backwardCS\] represents the obtained value. The measured $\omega$ meson production cross section in the present $p + p$ collision data is consistent with their $\rho^0$ cross section within the error.
![ \[fig:backwardCS\] Production cross section in the backward hemisphere as a function of mass number. The circles, open squares, and filled squares show the cross section of $\omega$ mesons, $\phi$ mesons measured in the $e^+ e^-$ decay channel, and those in the $K^+ K^-$ decay channel, respectively. The vertical lines represent the statistical errors, and the brackets represent the systematic errors. The previous $p + p \to \rho^0 X$ measurement [@Blobel] is indicated by a triangle. The dotted lines represent the prediction by [<span style="font-variant:small-caps;">jam</span>]{} simulation results. The solid and dashed lines represent the $\alpha$ parameterization (see text). ](fig6)
\[ssec:diff CS\] Differential cross section of $\omega$ and $\phi$ production measured in $e^+ e^-$ decays
-----------------------------------------------------------------------------------------------------------
In a region $0.9 < y < 1.7$ and $p_T < 0.75$ GeV/$c$, the differential cross sections of $\omega$ and $\phi$ mesons were obtained for each $y$ or $p_T$ bin, as shown in Fig. \[fig:DCS\]. These are also listed in Table \[table:DiffCS\].
=2.8mm
[ c r@[.]{}l@[–]{}r@[.]{}l d@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l d@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l d@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{} & & &\
& 0&9 & 1&0 & 14.7 & 1&5 & 2&9 & 47.8 & 4&6 & 10&0 & 0.708 & 0&083 & 0&042\
& 1&0 & 1&1 & 14.2 & 1&2 & 2&8 & 51.2 & 4&2 & 10&7 & 0.770 & 0&070 & 0&041\
& 1&1 & 1&2 & 15.0 & 1&1 & 3&0 & 48.8 & 4&6 & 10&2 & 0.706 & 0&071 & 0&040\
$y$ & 1&2 & 1&3 & 14.5 & 0&8 & 2&8 & 48.3 & 2&9 & 10&1 & 0.724 & 0&049 & 0&039\
& 1&3 & 1&4 & 14.4 & 0&8 & 2&8 & 45.8 & 2&7 & 9&6 & 0.693 & 0&048 & 0&039\
& 1&4 & 1&5 & 14.4 & 0&7 & 2&8 & 45.3 & 3&2 & 9&5 & 0.689 & 0&052 & 0&039\
& 1&5 & 1&6 & 15.3 & 0&7 & 3&0 & 46.9 & 2&9 & 9&8 & 0.671 & 0&045 & 0&038\
& 1&6 & 1&7 & 14.2 & 0&7 & 2&8 & 46.7 & 3&8 & 9&7 & 0.716 & 0&056 & 0&039\
& 0&00 & 0&12 & 5.7 & 0&4 & 1&1 & 18.8 & 1&2 & 3&9 & 0.714 & 0&057 & 0&038\
& 0&12 & 0&24 & 15.1 & 0&7 & 3&0 & 49.2 & 3&2 & 10&2 & 0.709 & 0&047 & 0&038\
$p_T$ & 0&24 & 0&36 & 18.1 & 0&8 & 3&5 & 63.1 & 3&4 & 13&2 & 0.750 & 0&042 & 0&039\
(GeV/$c$) & 0&36 & 0&48 & 19.5 & 0&9 & 3&8 & 64.2 & 3&6 & 13&4 & 0.714 & 0&044 & 0&038\
& 0&48 & 0&60 & 16.8 & 1&3 & 3&3 & 52.6 & 3&6 & 11&0 & 0.687 & 0&063 & 0&040\
& 0&60 & 0&75 & 13.9 & 1&1 & 2&7 & 41.3 & 3&9 & 8&6 & 0.655 & 0&075 & 0&041\
& 14.30 & 0&34 & 2&79 & 46.63 & 1&25 & 9&70 & 0.710 & 0&021 & 0&037\
& & &\
& 0&9 & 1&1 & 0.348 & 0&046 & 0&068 & 1.60 & 0&17 & 0&32 & 0.916 & 0&101 & 0&022\
$y$ & 1&1 & 1&3 & 0.232 & 0&032 & 0&045 & 1.33 & 0&13 & 0&27 & 1.050 & 0&101 & 0&020\
& 1&3 & 1&5 & 0.277 & 0&029 & 0&054 & 1.20 & 0&11 & 0&24 & 0.881 & 0&084 & 0&020\
& 1&5 & 1&7 & 0.255 & 0&037 & 0&050 & 0.93 & 0&13 & 0&19 & 0.780 & 0&119 & 0&019\
$p_T$ & 0&00 & 0&25 & 0.185 & 0&024 & 0&036 & 0.93 & 0&10 & 0&19 & 0.971 & 0&101 & 0&019\
(GeV/$c$) & 0&25 & 0&50 & 0.405 & 0&032 & 0&079 & 1.78 & 0&13 & 0&36 & 0.890 & 0&066 & 0&019\
& 0&50 & 0&75 & 0.255 & 0&032 & 0&050 & 1.19 & 0&16 & 0&24 & 0.924 & 0&111 & 0&021\
& 0.270 & 0&017 & 0&053 & 1.290 & 0&070 & 0&258 & 0.937 & 0&049 & 0&018\
The previous $p + p \to \rho^0 X$ measurement [@Blobel] was plotted using triangles, as shown in Fig. \[fig:DCS\]. It should be noted that the data points were scaled by factors of 5.81 (= $12^{0.71}$) for clarity.
The distributions obtained by the [<span style="font-variant:small-caps;">jam</span>]{} calculation are compared with the measurements in Fig. \[fig:DCS\]. The total cross sections obtained from the [<span style="font-variant:small-caps;">jam</span>]{} calculation are larger than that obtained from the data. Hence, they are scaled by the factors 0.489 for $p$ + C $\to \omega X$, 0.421 for $p$ + Cu $\to \omega X$, 1.006 for $p$ + C $\to \phi X$, and 0.686 for $p$ + Cu $\to \phi X$. These scale factors were determined as the ratio of the total cross sections in the acceptance between the data and [<span style="font-variant:small-caps;">jam</span>]{}. The shapes of the differential cross sections of [<span style="font-variant:small-caps;">jam</span>]{} are consistent with the present data, although the absolute cross sections are systematically larger than the data.
Figure \[fig:alpha\] shows the $\alpha$ parameters of $\omega$ and $\phi$ mesons obtained in the region of $0.9 < y <1.7$ and $p_T < 0.75$ GeV/$c$. These are also listed in Table \[table:DiffCS\]. The flat lines represent the averaged $\alpha$ parameters with the errors shown in the left column. The dotted and dashed curves represent the results of [<span style="font-variant:small-caps;">jam</span>]{} for $p$ + C and $p$ + Cu collisions. Although the $\alpha$ parameters of the [<span style="font-variant:small-caps;">jam</span>]{} calculation are significantly larger than those of the data, the difference between $\omega$ and $\phi$ mesons in the $\alpha$ parameters is similar to what is seen in the [<span style="font-variant:small-caps;">jam</span>]{} calculation.
\[ssec:discussion\] Discussion
------------------------------
An interesting characteristic is the difference in the $y$ dependence between $\omega$ and $\phi$ meson production cross sections. While the $\omega$ production is essentially independent of $y$, the $\phi$ production increases towards the smaller $y$ values. In addition, the $\alpha$ parameters of $\omega$ and $\phi$ meson production are different by $0.227 \pm 0.054(\text{stat}) \pm 0.041(\text{syst})$. This significant difference in the $\alpha$ parameters confirms that the production mechanisms of $\omega$ and $\phi$ mesons are different.
One possible mechanism of $\phi$ meson production is a hard reaction between incident and projectile nucleons. If the hard reaction is dominant like $J/\psi$ production at higher energies, $\alpha$ is expected to be around unity and independent of $y$. Further, the $y$ distribution of the cross section is expected to be symmetric with respect to $y_{\text{c.m.}}$, which is not the case in the present $\phi$ production data.
Another possible effect to explain the observed characteristics of the $\alpha$ parameter and cross sections is the effect of secondary collisions in a target nucleus. In this case, no hard reaction is necessary. These effects are expected to increase in a smaller rapidity region for $p$ + $A$ interaction, and the cross sections and $\alpha$ are expected to be larger in the backward region.
Although the [<span style="font-variant:small-caps;">jam</span>]{} does not reproduce the data quantitatively, it yields fairly similar shapes in the inclusive cross sections of $\omega$ and $\phi$ meson productions; further, it also predicts the difference between the $\alpha$ parameters of these mesons. In the [<span style="font-variant:small-caps;">jam</span>]{} calculation, more than 90% of the $\phi$ mesons are produced by secondary collisions mostly between non-strange mesons and nucleons in a target nucleus, while $\omega$ mesons are produced both in primary and secondary reactions. For example, 30% and 50% of the $\omega$ mesons are produced in secondary reactions in the cases of 12 GeV $p$ + C and $p$ + Cu reactions in [<span style="font-variant:small-caps;">jam</span>]{}, respectively. These differences in the production mechanism in [<span style="font-variant:small-caps;">jam</span>]{} will surely be an important step to understand the measured differential cross sections.
The difference between the scaling factors applied to $\phi$ meson production in $p$ + C and $p$ + Cu collisions, in [<span style="font-variant:small-caps;">jam</span>]{}, described in Sec. \[ssec:diff CS\], can be understood qualitatively by the overestimated contribution of the secondary collisions in [<span style="font-variant:small-caps;">jam</span>]{}. Although in those calculations almost all the $\phi$ mesons are produced in secondary collisions, the present data suggest that the contribution of the primary collisions can be larger. For the case of the $\omega$ meson, [<span style="font-variant:small-caps;">jam</span>]{} just predicts larger absolute production yields than seen in the present data.
In summary, we measured the inclusive differential cross sections of the $\omega$ and $\phi$ meson production in $p$ + $A$ collisions in the backward region. The difference in the $\alpha$ parameters between $\omega$ and $\phi$ mesons confirms that the production mechanisms of $\omega$ and $\phi$ mesons are different. The results are compared to the nuclear cascade calculations.
We have greatly appreciated all of the staff members of KEK-PS, especially the beam channel group, for their helpful support. We would like to thank C. Lourenco for proof reading of the manuscript. This work was partly funded by the Japan Society for the Promotion of Science, RIKEN Special Post doctoral Researchers Program, and a Grant-in-Aid for Scientific Research from the Japan Ministry of Education, Culture, Sports, Science and Technology (MEXT). Finally, we also thank the staff members of RIKEN super combined cluster system and RIKEN-CCJ.
[^1]: In a further analysis, we used only events in which $e^+$ went into the left arm and $e^-$ went into the right arm because the number of events satisfying the opposite criterion comprised only 6% of the data.
[^2]: If we take internal bremsstrahlung into account, the peak positions of both the $\omega$ mesons and the $\phi$ mesons agree with the expected positions.
[^3]: In this article, neither the effect of internal bremsstrahlung nor the uncertainty of the branching fractions is considered.
[^4]: The $\rho^0/\omega$ ratio was measured by the reactions of $p + p \to \rho^0$ + charged particles and $p + p \to \omega$ + charged particles [@Blobel].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'An alternative pathway for the human brain to communicate with the outside world is by means of a brain computer interface (BCI). A BCI can decode electroencephalogram (EEG) signals of brain activities, and then send a command or an intent to an external interactive device, such as a wheelchair. The effectiveness of the BCI depends on the performance in decoding the EEG. Usually, the EEG is contaminated by different kinds of artefacts (e.g., electromyogram (EMG), background activity), which leads to a low decoding performance. A number of filtering methods can be utilized to remove or weaken the effects of artefacts, but they generally fail when the EEG contains extreme artefacts. In such cases, the most common approach is to discard the whole data segment containing extreme artefacts. This causes the fatal drawback that the BCI cannot output decoding results during that time. In order to solve this problem, we employ the Lomb-Scargle periodogram to estimate the spectral power from incomplete EEG (after removing only parts contaminated by artefacts), and Denoising Autoencoder (DAE) for learning. The proposed method is evaluated with motor imagery EEG data. The results show that our method can successfully decode incomplete EEG to good effect.'
address:
- 'Laboratory for Advanced Brain Signal Processing, Brain Science Institute, RIKEN, Saitama, 351-0198, Japan'
- 'Key Laboratory of Shanghai Education Commission for Intelligent Interaction and Cognitive Engineering, Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, 200240, China'
author:
- Junhua Li
- Zbigniew Struzik
- Liqing Zhang
- Andrzej Cichocki
title: Feature Learning from Incomplete EEG with Denoising Autoencoder
---
Brain Computer Interface,Spectral Power Estimation,Denoising Autoencoder ,Motor Imagery,Incomplete EEG
Introduction
============
The combination of advanced neurobiology and engineering creates a new pathway, namely a brain computer interface (BCI). The BCI provides a bridge connecting the human brain to the outside world [@ortiz2013brain]. This means that people do not have to rely on the conventional pathway of an intent initialized in the brain being passed to muscles through peripheral nerves, and are able to interact directly with the external environment [@wolpaw2000brain]. Due to the lack of involvement of peripheral nerves and muscles, with the aid of a BCI system, disabled people could restore their abilities of communication [@muller2008machine] and the degenerated motor function [@li2013design; @pfurtscheller2003thought]. During the past two decades, a variety of BCI systems have been created for different applications. These BCI systems are generally divided into two types: active BCI and passive BCI, according to the level of interaction with external stimuli. In the case of a passive BCI, when using a steady-state visual evoked potential (SSVEP) BCI [@bin2009online], the user may, for example, simply stare at an intended digital number shown on a screen to dial a phone number. When a steady-state flicker is replaced with an occasional flicker, a different type of BCI called P300 can be used to output letters by hierarchical selections [@muller2008machine]. Compared to the passive BCI, the active BCI is more natural. Users can express their intents whenever they want to, rather than according to a predefined timing arrangement or external cooperation, as with the passive BCI. For instance, people with paraplegia can regain movement in a wheelchair by motor imagery [@li2013design], or can control a computer cursor in virtual 2D [@mcfarland1993eeg] or 3D [@mcfarland2010electroencephalographic] environments through brain modulation. Moreover, BCI is also used to develop prostheses, with which disabled people can, for example, move an object [@muller2005eeg] or drink a cup of coffee [@hochberg2012reach]. More recently, BCI has been applied to facilitate rehabilitation [@daly2008brain; @liu2014tensor]. Besides applications for disabled people, BCI also has promising applications for healthy persons, especially in the field of entertainment. BCI is employed to control video games instead of conventional inputs such as a keyboard and joystick [@li2013competitive]. In this way, healthy people can enjoy the experience of manipulating virtual objects in a manner different from that used in daily life.\
From the application point of view, the user experience is very important. This requires smoothness in the manipulation of the BCI system. In order to meet this requirement, the BCI system needs to translate brain activities into output information continuously without any interruption. In other words, this requires all the EEG segments to be present for the decoding. If some of the EEG segments are discarded due to extreme noise contamination, the BCI cannot generate the corresponding output during that period. Hence, it would be good to be able to utilize the remaining portion of the affected EEG segment, after only removing the part directly affected by noise. In general, spectral power features are usually utilized to distinguish different motor imageries (e.g., left-hand and right-hand motor imageries) [@palaniappan2006utilizing; @pfurtscheller2000brain; @li2010bilateral; @li2012active], as they are considered to be robust for the representation of the contents of motor imageries. If the segment is complete (continuous), the Fourier transform can be well used to transform temporal data points into the spectral domain. This fails in the case of incomplete data, such as an EEG segment with a portion (or portions) of data removed (unevenly spaced). In order still to utilize such segments of EEG with arbitrary portions of data removed and provide users with an experience of smooth manipulation, we employ the Lomb-Scargle periodogram to estimate the spectral power [@lomb1976least; @stoica2009spectral], and Denoising Autoencoder (DAE) [@vincent2008extracting; @vincent2010stacked] based neural network or support vector machine (SVM) [@vapnik2000nature; @hearst1998support] to predict the classes of motor imageries. The results show that the proposed method is suitable for decoding incomplete EEG in a BCI system.
Methodology
===========
We first employed the Lomb-Scargle periodogram [@lomb1976least; @stoica2009spectral] to estimate band powers from incomplete EEG signals. Next, the extracted features were used to train an unsupervised DAE [@vincent2008extracting; @vincent2010stacked] or a supervised SVM with Radial Basis Function (RBF) kernel [@vapnik2000nature; @hearst1998support]. In the case of DAE, the mapping weights of DAE were used to initialize a neural network. After fine-tuning the weights, this trained neural network was used to recognize the classes of motor imageries. Fig. \[schematic\] illustrates the proposed method.
{width="100.00000%"}
Lomb-Scargle Periodogram
------------------------
A four-second trial is divided into 25 segments of one-second length with an overlap of 87.5%. A segment is denoted by $X$, which is $N$ by $T$ matrix, where $N$ is the number of channels, and $T$ is the number of sampling points. The spectral power of each channel time series $y({t_i})$ is estimated by the Lomb-Scargle periodogram [@lomb1976least; @stoica2009spectral]. The estimated spectral power at frequency ${\Omega_f}$ can be obtained by minimizing the following sum of difference squares:
$$\label{equation1}
\mathop {\min }\limits_{\scriptstyle\;\;\;\;a > 0\hfill\atop
\scriptstyle\phi \in [0,\;2\pi ]\hfill} \;\sum\limits_{i = 1}^T {{{(y({t_i}) - \alpha \cos ({\Omega_f}{t_i} + \phi ))}^2}}\;.$$
Let $$a = \alpha \cos \phi$$ and $$b = - \alpha \sin \phi \;.$$
We can then rewrite equation (\[equation1\]) as: $$\label{equation4}
\mathop {\min }\limits_{a,\;b} \;\sum\limits_{i = 1}^T {{{(y({t_i}) - a\cos ({\Omega_f}{t_i}) - b\sin ({\Omega_f}{t_i}))}^2}} .$$ The optimal parameters $\hat{a}$ and $\hat{b}$ can be obtained through minimizing equation (\[equation4\]) $$\left[ {\begin{array}{*{20}{c}}
{\hat{a}}\\
{\hat{b}}
\end{array}} \right] = {R^{ - 1}}r,$$ where $$R = \sum\limits_{i = 1}^T {\left[ {\begin{array}{*{20}{c}}
{\cos ({\Omega_f}{t_i})}\\
{\sin ({\Omega_f}{t_i})}
\end{array}} \right]} \left[ {\begin{array}{*{20}{c}}
{\cos ({\Omega_f}{t_i})}&{\sin ({\Omega_f}{t_i})}
\end{array}} \right],$$ and $$r = \sum\limits_{i = 1}^T {\left[ {\begin{array}{*{20}{c}}
{\cos ({\Omega_f}{t_i})}\\
{\sin ({\Omega_f}{t_i})}
\end{array}} \right]} \;y({t_i})\;.$$ The power of specific frequency ${\Omega_f}$ is then estimated with respect to optimal parameters $\hat{a}$, $\hat{b}$ as follows: $$\begin{array}{l}
\frac{1}{T}{\sum\limits_{i = 1}^T {\left( {\left[ {\begin{array}{*{20}{c}}
{\hat{a}}&{\hat{b}}
\end{array}} \right]\left[ {\begin{array}{*{20}{c}}
{\cos (\Omega{}_f{t_i})}\\
{\sin (\Omega{}_f{t_i})}
\end{array}} \right]} \right)} ^2}\\
\quad = \frac{1}{T}\left[ {\begin{array}{*{20}{c}}
{\hat{a}}&{\hat{b}}
\end{array}} \right]\;R\;\left[ {\begin{array}{*{20}{c}}
{\hat{a}}\\
{\hat{b}}
\end{array}} \right]\\
\quad = \frac{1}{T}{r^{\rm T}}({\Omega_f}){R^{ - 1}}({\Omega_f})r({\Omega_f})\;.
\end{array}$$
Similarly, the minimization of squares mentioned above is used to estimate spectral powers at all frequencies. After that, spectral estimation for one channel is completed. These steps are repeated for all channels and all segments to obtain the spectral powers. Because the frequency range of 8-30 Hz is mostly related to the motor imagery task [@li2012active], we divided this band into four subbands with a bandwidth of 5 Hz (i.e., 8-12 Hz, 13-17 Hz, 18-22 Hz, and 23-27 Hz). Subband powers were obtained by averaging spectral powers within the corresponding frequency band range for each channel. Then, subband powers (four features for each channel) for all channels were concatenated into a feature vector: $$F = {[{f_{11}},\;{f_{12}},\;{f_{13}},\;{f_{14}},\;{f_{21}},\;{f_{22}},\;{f_{23}},\;{f_{24}}, \cdots ,\;{f_{N1}},\;{f_{N2}},\;{f_{N3}},\;{f_{N4}}]^{\rm T}}\;,$$ where $N$ is the number of channels. Subsequently, features were normalized as: $${f_{qp}} = \log \left( {\frac{{{f_{qp}}}}{{\sum\limits_{i = 1}^N {\sum\limits_{j = 1}^4 {{f_{ij}}} } }}} \right)\;.$$ The normalized features were then fed into a neural network with DAE initialization, or into an SVM classifier to distinguish which class the current EEG segment belongs to.
DAE-based neural network
------------------------
For a time, neural networks were less frequently used due to the drawback that they easily became stuck in the local minima, so more use was made of SVM classifier. However, recently neural networks have regained popularity, in particular when using a pre-training strategy [@vincent2010stacked; @erhan2010does; @glorot2011domain]. In this paper, we construct a three-layer neural network with DAE initialization (A neural network with more layers might possibly achieve a better performance through in-depth feature learning).\
The power features extracted by Lomb-Scargle Periodogram was first corrupted, denoted as $\hat{f}$, by means of a stochastic mapping $\hat{f} \thicksim q_\mathrm{\mathcal{{D}}}(\hat{f}|f)$. The part enclosed by the orange rectangle in Fig. \[schematic\] shows a schematic diagram of the DAE. We set the corrupted elements to 0. Then, the corrupted features were mapped to a hidden representation (120 units) by the sigmoid function $$y = g_{1,\;\theta }(\hat{f}) = s(W \cdot \hat{f} + b).$$ Consequently, we reconstructed the uncorrupted $z$ as $$z=g_{2,\;{\theta ^\prime }}(y).$$ The objective was to train parameters $\theta = \{ W, b \}$ and $\theta^{\prime} = \{ W^{\prime}, b^{\prime} \}$ for minimization of the average reconstruction error over a training set. In other words, to find the parameters to let $z$ be as close as possible to $f$, we performed the following optimization: $$\begin{array}{l}
[{\theta^*},{\theta^{{\prime}\;*}}] = \mathop {\arg \min }\limits_{\theta ,\;{\theta^{\prime}}} \frac{1}{n}\sum\limits_{i = 1}^n {L({f^{(i)}},\;{z^{(i)}})} \\
\quad \quad \quad \quad = \mathop {\arg \min }\limits_{\theta ,\;{\theta^{\prime}}} \frac{1}{n}\sum\limits_{i = 1}^n {L({f^{(i)}},\;{g_{2,\;{\theta^{\prime}}}}({g_{1,\;\theta }}({\hat{f}^{(i)}})))},
\end{array}$$ where $L$ is a squared error loss function $L(f,z) = {\left\| {f - z} \right\|^2}$, n is the number of training samples, and ${\theta^*},{\theta^{{\prime}\;*}}$ are the optimal values of ${\theta ^{}},{\theta^{\prime}}$. Once the optimal parameters were obtained, we were able to use those parameters to initialize a neural network. A top layer was added onto the neural network. After that, the parameters were fine-tuned in a supervised way.
Evaluation Data
===============
Two different categories of data are used to prove the feasibility of the proposed method. One is the simulated data and the other is the two-class motor imagery data. We use simulated data to illustrate systematically that spectral power can be correctly estimated when the data become unevenly spaced after removing some data points from them. Further, we use real motor imagery data to demonstrate that classification accuracy does not dramatically decrease when increasing the percentage of data within the segment that has been removed, so that the proposed method is useful to process incomplete data in a BCI system.\
The simulated data were generated by mixing two sinusoidal signals, which were 3 Hz and 6 Hz, respectively. The maximal amplitude of the 3 Hz sinusoidal signal was 1.5 times that of the 6 Hz sinusoidal signal. The motor imagery data came from three subjects. Fourteen electrodes (shown with a green background in the scalp illustration in Fig. \[schematic\]) were used to record the EEG signal on the sensorimotor cortex while the subject was conducting motor imagery at a sampling rate of 250 Hz. Those electrodes were referenced at the mastoids behind the ears and grounded at AFz. Each subject participated in four sessions. Each session consisted of 15 trials, each of which was four seconds long. The subject conducted either left hand motor imagery or right hand motor imagery according to the cue shown on the computer monitor.
Results
=======
Simulated data
--------------
{width="100.00000%"}
We first evaluated the performance of the spectral power estimation on simulated data. The simulated data was mixed with two sinusoidal signals, which were 3 Hz and 6 Hz, respectively. The spectral power estimated from the complete signal, and the incomplete signals with different proportional removal of data points (from 10% to 80% with an interval of 10%) are shown in Fig. \[SimulatedPlot\]. The data points were removed at random. In order to keep the same scale over cases with different proportional data removal to facilitate comparisons between them, the powers shown in Fig. \[SimulatedPlot\] were normalized by dividing by a proportional factor (1-p, where p is the percentage of data removed). For example, the estimated power is divided by a factor of 0.7 when 30% of data points are removed from the signal. From Fig. \[SimulatedPlot\], we can see that the components at 3 Hz and 6 Hz can be well estimated in all cases with different proportions of data removal, even up to removal of 80% of data points.
Real motor imagery data
-----------------------
{width="100.00000%"}
In general, BCI encounters a common problem that there is no output when the whole segment has to be discarded due to partial noise contamination in that segment. If a method can obtain comparable recognition accuracy (the same or slightly worse) by using only the remaining portion of the segment (the portion from which noise contamination has been removed), this method is considered as an effective solution to the aforementioned problem.
{width="100.00000%"}
For real motor imagery data, two ways were used to randomly remove the partial data from the segment. One is that data points within a segment were randomly removed (see Fig. \[Point60Rem\] for an example). The other is that data blocks within a segment were randomly removed (see Fig. \[Block60Rem\] for an example). The width of the blocks removed was generated according to a normal distribution with a mean of 20 and a standard deviation of 10.
{width="100.00000%"}
{width="100.00000%"}
We used the data from the preceding session to train the SVM classifier with a RBF kernel, and tested it with the data from the following session. Two approaches were used for the evaluation of the accuracy (i.e., sliding time window accuracy and trial accuracy). Sliding time window accuracies were calculated as the number of segments classified as correct divided by the total number of segments. A trial was classified as belonging to the class to which most of the sliding time windows within that trial belonged. Then trial accuracies were obtained according to the ratio of trials classified as correct. Fig. \[AccPoint\] and Fig. \[AccBlock\] show test accuracies for the conditions of data point removal and data block removal, respectively. In general, the accuracies for all sessions of all subjects did not dramatically decrease. Trial accuracies varied more than sliding time window accuracies across different proportional sections of data removal. This is because a trial was classified as correct even if the number of sliding time windows in the trial classified as correct was only one more than the number of sliding time windows classified as incorrect. Likewise, trials with one more incorrect sliding time window than correct sliding time window were classified as incorrect. Therefore, in some cases, the trial accuracy changed greatly while the accuracy of the sliding time widows did not change much. A comparable classification accuracy could be achieved even when 80% of data were removed. High accuracies were retained no matter how many data points were removed - in the range from 10% to 80% - for subject 1, especially for sessions 2 and 3. The accuracies for 80% data removal were substantially worse than those for 70% data removal for subject 1 in the condition of block data removal. It appears that our method is relatively sensitive to data removal in block form.
Comparison between DAE and SVM
------------------------------
In this section, we show a comparison between DAE and SVM in terms of classification accuracy of sliding time windows. SVM has been widely adopted since its conception and has been successfully applied in many fields. Deep learning is a promising and burgeoning method. DAE is utilized as a building brick to compose a deep learning network. It is meaningful to illustrate the effectiveness of this for EEG feature recognition using our paradigm. The parameters used in the training are listed in Table \[Paras\]. Fig. \[AccComPoint\] shows the accuracy difference between DAE and SVM for each session of each subject under the condition of data point removal. Asterisks located above the zero horizontal line mean that the accuracy of DAE is higher than that of SVM. The bars shown on the right of each sub-plot are the average differences. The bottom right plot illustrates the overall difference averaged across all sessions of all subjects. From Fig. \[AccComPoint\], we can see that there is no clear winner - the DAE is better than the SVM in a number of sessions but turns out to be worse in other sessions. The overall average accuracy of DAE is still better than that of SVM. Fig. \[AccComBlock\] shows the accuracy comparison under the condition of block point removal. The result is similar to the condition of data point removal. The overall average accuracy of DAE is higher than that of SVM under the condition of block point removal, but the increase in accuracy of DAE compared with SVM is less than the case of data point removal.
Parameters Values
-------------------------------- --------
Corrupted fraction 0.3
Mini-batch size 25
Learning rate for pre-training 0.9
Number of pre-training epochs 20
Learning rate for fine-tuning 0.9
Number of fine-tuning epochs 50
: Parameter Settings[]{data-label="Paras"}
{width="100.00000%"}
{width="100.00000%"}
From the results of comparisons, the DAE is shown to be comparable to the SVM. However, it is possible that the DAE can outperform the SVM when more layers are used and parameters are better tuned. It is not yet clear whether the DAE can significantly exceed the SVM in terms of EEG classification, but there has been a report that stacked DAE (i.e., multiple DAEs combined together to obtain deeper learning of features) performed better than the SVM on the image benchmark dataset named MNIST [@vincent2008extracting].
Conclusion
==========
We propose the combination of the Lomb-Scargle periodogram and either SVM or DAE to distinguish incomplete EEG segments (i.e., segments from which a portion of data has been removed due to noise contamination). The results indicate that classification accuracy is not dramatically decreased when different percentages of data are removed. Therefore, the classification performance using the proposed method for incomplete segments is acceptable for a BCI application system. This means that the segment with noise contamination can still be utilized to output commands after only removing the noisy portion, instead of discarding the whole segment, as is conventionally done in BCI systems. In brief, the proposed method can achieve comparable classification performance even when most of the data points in a segment have been removed. It provides an alternative solution for the frequent problem occurring in a BCI system that there is no output when a segment is discarded.
Acknowledgments
===============
The work of Liqing Zhang was supported by the national natural science foundation of China (Grant No. 91120305, 61272251).
References {#references .unnumbered}
==========
[10]{} url \#1[`#1`]{}urlprefixhref \#1\#2[\#2]{} \#1[\#1]{}
A. Ortiz-Rosario, H. Adeli, Brain-computer interface technologies: from signal to action, Reviews in the Neurosciences 24 (5) (2013) 537–552.
J. R. Wolpaw, N. Birbaumer, W. J. Heetderks, D. J. McFarland, P. H. Peckham, G. Schalk, E. Donchin, L. A. Quatrano, C. J. Robinson, T. M. Vaughan, et al., Brain-computer interface technology: a review of the first international meeting, IEEE Transactions on Rehabilitation Engineering 8 (2) (2000) 164–173.
K.-R. M[ü]{}ller, M. Tangermann, G. Dornhege, M. Krauledat, G. Curio, B. Blankertz, Machine learning for real-time single-trial EEG-analysis: from brain–computer interfacing to mental state monitoring, Journal of Neuroscience methods 167 (1) (2008) 82–90.
J. Li, J. Liang, Q. Zhao, J. Li, K. Hong, L. Zhang, Design of assistive wheelchair system directly steered by human thoughts, International journal of Neural Systems 23 (3) (2013) 1350013.
G. Pfurtscheller, G. R. M[ü]{}ller, J. Pfurtscheller, H. J. Gerner, R. Rupp, ‘Thought’–control of functional electrical stimulation to restore hand grasp in a patient with tetraplegia, Neuroscience Letters 351 (1) (2003) 33–36.
G. Bin, X. Gao, Z. Yan, B. Hong, S. Gao, An online multi-channel SSVEP-based brain–computer interface using a canonical correlation analysis method, Journal of Neural Engineering 6 (4) (2009) 046002.
D. J. McFarland, G. W. Neat, R. F. Read, J. R. Wolpaw, An EEG-based method for graded cursor control, Psychobiology 21 (1) (1993) 77–81.
D. J. McFarland, W. A. Sarnacki, J. R. Wolpaw, Electroencephalographic (EEG) control of three-dimensional movement, Journal of Neural Engineering 7 (3) (2010) 036007.
G. R. M[ü]{}ller-Putz, R. Scherer, G. Pfurtscheller, R. Rupp, EEG-based neuroprosthesis control: a step towards clinical practice, Neuroscience Letters 382 (1) (2005) 169–174.
L. R. Hochberg, D. Bacher, B. Jarosiewicz, N. Y. Masse, J. D. Simeral, J. Vogel, S. Haddadin, J. Liu, S. S. Cash, P. van der Smagt, et al., Reach and grasp by people with tetraplegia using a neurally controlled robotic arm, Nature 485 (7398) (2012) 372–375.
J. J. Daly, J. R. Wolpaw, Brain–computer interfaces in neurological rehabilitation, The Lancet Neurology 7 (11) (2008) 1032–1043.
Y. Liu, M. Li, H. Zhang, H. Wang, J. Li, J. Jia, Y. Wu, L. Zhang, A tensor-based scheme for stroke patients’ motor imagery EEG analysis in bci-fes rehabilitation training, Journal of Neuroscience Methods 222 (2014) 238–249.
J. Li, Y. Liu, Z. Lu, L. Zhang, A competitive brain computer interface: Multi-person car racing system, in: Engineering in Medicine and Biology Society (EMBC), 2013 35th Annual International Conference of the IEEE, IEEE, 2013, pp. 2200–2203.
R. Palaniappan, Utilizing gamma band to improve mental task based brain-computer interface design, Neural Systems and Rehabilitation Engineering, IEEE Transactions on 14 (3) (2006) 299–303.
G. Pfurtscheller, C. Guger, G. M[ü]{}ller, G. Krausz, C. Neuper, Brain oscillations control hand orthosis in a tetraplegic, Neuroscience Letters 292 (3) (2000) 211–214.
J. Li, L. Zhang, Bilateral adaptation and neurofeedback for brain computer interface system, Journal of Neuroscience Methods 193 (2) (2010) 373–379.
J. Li, L. Zhang, Active training paradigm for motor imagery BCI, Experimental Brain Research 219 (2) (2012) 245–254.
N. R. Lomb, Least-squares frequency analysis of unequally spaced data, Astrophysics and Space Science 39 (2) (1976) 447–462.
P. Stoica, J. Li, H. He, Spectral analysis of nonuniformly sampled data: a new approach versus the periodogram, Signal Processing, IEEE Transactions on 57 (3) (2009) 843–858.
P. Vincent, H. Larochelle, Y. Bengio, P.-A. Manzagol, Extracting and composing robust features with denoising autoencoders, in: Proceedings of the 25th International Conference on Machine Learning, ACM, 2008, pp. 1096–1103.
P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, P.-A. Manzagol, Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion, The Journal of Machine Learning Research 11 (2010) 3371–3408.
V. Vapnik, The nature of statistical learning theory, Springer, 2000.
M. A. Hearst, S. Dumais, E. Osman, J. Platt, B. Scholkopf, Support vector machines, Intelligent Systems and their Applications, IEEE 13 (4) (1998) 18–28.
D. Erhan, Y. Bengio, A. Courville, P.-A. Manzagol, P. Vincent, S. Bengio, Why does unsupervised pre-training help deep learning?, The Journal of Machine Learning Research 11 (2010) 625–660.
X. Glorot, A. Bordes, Y. Bengio, Domain adaptation for large-scale sentiment classification: A deep learning approach, in: Proceedings of the 28th International Conference on Machine Learning (ICML-11), 2011, pp. 513–520.
**Junhua Li** received his Ph.D. degree from the Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China, in March 2013. He is currently a research scientist at the Laboratory for Advanced Brain Signal Processing, Brain Science Institute, RIKEN, Japan. His research interests include signal processing, brain computer interface, and machine learning. He has been a member of IEEE since 2013, and was a student member of IEEE in 2012.\
**Zbigniew R. Struzik** received a Master of Science in Engineering degree in technical physics from the Warsaw University of Technology, Poland, in 1986, and a Doctor degree from the faculty of Mathematics, Computer Science, Physics and Astronomy at the University of Amsterdam, the Netherlands, in 1996. From 1997 to 2003, he worked at the Centre for Mathematics and Computer Science (CWI), Amsterdam, and since 2003 has worked at the University of Tokyo, Japan, where he is currently affiliated. From 2012, his main position has been at RIKEN Brain Science Institute in Wakoshi, Japan. His scientific work contributed to the amalgamation of (multi-)fractal analysis, wavelet analysis and time series data mining. His current research interests include applications of information science and statistical physics in life sciences, complexity and emergence, time series processing and mining, and recently, analytic approaches to elucidating the nature of creative processes in art and science, in particular in neuroscience. He is on the editorial board of the Fractals Journal, the Open Medical Informatics Journal, Frontiers in Fractal Physiology, Frontiers in Computational Physiology and Medicine, Frontiers in Human Neuroscience, International Journal of Statistical Mechanics, Journal of Neuroscience Methods and Integrative Medicine International Journal. He has co-authored over one hundred journal papers and book chapters.\
**Liqing Zhang** received the Ph.D. degree from Zhongshan University, Guangzhou, China, in 1988. He was promoted in 1995 to the position of full professor at South China University of Technology. He worked as a research scientist at RIKEN Brain Science Institute, Japan from 1997 to 2002. He is now a Professor with Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China. His current research interests cover computational theory for cortical networks, brain-computer interface, perception and cognition computing model, statistical learning and inference. He has published more than 170 papers in international journals and at conferences.\
**Andrzej Cichocki** received Ph.D. and Dr.Sc. (Habilitation) degrees, in electrical engineering, from Warsaw University of Technology (Poland). He is currently the senior team leader head of the laboratory for Advanced Brain Signal Processing, at RIKEN Brain Science Institute (JAPAN).\
He is a co-author of more than 250 technical papers and 4 monographs (two of them translated to Chinese). According to a 2011 analysis, he is a co-author of one of the top 1% most highly cited papers in his field worldwide.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Motivated by a conjecture of Savage and Visontai about the equidistribution of the descent statistic on signed permutations of the multiset $\{1,1,2,2,\ldots,n,n\}$ and the ascent statistic on $(1,4,3,8,\ldots,2n-1,4n)$-inversion sequences, we investigate the descent polynomial of the signed permutations of a general multiset. We obtain a factorial generating function formula for a $q$-analog of these descent polynomials and apply it to show that they have only real roots. Two different proofs of the conjecture of Savage and Visontai are provided.'
address: 'Institut Camille Jordan, UMR 5208 du CNRS, Université Claude Bernard Lyon 1, France'
author:
- Zhicong Lin
title: On the descent polynomial of signed multipermutations
---
Introduction
============
For a sequence of positive integers $\s=\{s_i\}_{i\geq1}$, let the set of [*$\s$-inversion sequences of length $n$*]{}, $\I_n^{\s}$, be defined as $$\I_n^{(\s)}:=\{(e_1,\ldots,e_n)\in\Z^n : 0\leq e_i<s_i\,\text{for}\, 1\leq i\leq n\}.$$ The [*ascent set*]{} of an $\s$-inversion sequence $\e=(e_1,\ldots,e_n)\in\I_n^{\s}$ is the set $$\Asc(\e):=\left\{0\leq i<n : \frac{e_i}{s_i}<\frac{e_{i+1}}{s_{i+1}}\right\},$$ with the convention that $e_0=0$ and $s_0=1$. Let $\asc(\e):=|\Asc(\e)|$ be the [*ascent statistic*]{} on $\e\in\I_n^{\s}$. The $\s$-inversion sequences and the ascent statistic were introduced by Savage and Schuster in [@ss].
A [*descent*]{} in a permutation $\pi=\pi_1\pi_2\cdots\pi_n$ of a multiset with elements from $\N$ is an index $i\in\{0,1,\ldots,n-1\}$ such that $\pi_i>\pi_{i+1}$ (with the convention that $\pi_0=0$). Denote by $\Des(\pi)$ the set of descents of $\pi$ and by $\des(\pi)$ the number of descents of $\pi$. The major index of $\pi$, denoted $\maj(\pi)$, is defined as $$\maj(\pi):=\sum_{i\in\Des(\pi)}i.$$
Let $P(\{1,1,2,2,\ldots,n,n\})$ be the set of permutations of the multiset $\{1,1,2,2,\ldots,n,n\}$. The following connection between multiset permutations and special inversion sequences was shown in [@sv Theorem 3.23].
$$\sum_{\pi\in P(\{1,1,2,2,\ldots,n,n\})} t^{\des(\pi)}=\sum_{\e\in\I_{2n}^{(1,1,3,2,\ldots,2n-1,n)}}t^{\asc(\e)}$$
Now consider the signed multiset permutations. Let $P^{\pm}(\{1,1,2,2,\ldots,n,n\})$ be the set of all [*signed*]{} permutations of the multiset $\{1,1,2,2,\ldots,n,n\}$, whose elements are those of the form $\pm\pi_1\pm\pi_2\cdots\pm\pi_n$ with $\pi_1\pi_2\cdots\pi_n\in P(\{1,1,2,2,\ldots,n,n\})$. For convenient, we write $-n$ by $\overline{n}$ for each positive integer $n$. For example, $$P^{\pm}(\{1,1\})=\{1\,1,1\,\overline{1},\overline{1}\,1,\overline{1}\,\overline{1}\}.$$ Clearly, we have $$|P^{\pm}(\{1,1,2,2,\ldots,n,n\})|=\frac{(2n)!}{2^n}2^{2n}=2^n(2n)!=|\I_{2n}^{(1,4,3,8,\ldots,2n-1,4n)}|.$$ Savage and Visontai [@sv Conjecture 3.25] further conjectured the following equdistribution, which was proved very recently (and independently) by Chen et al. [@ch] using type B $P$-Partitions.
\[conj:VS\] For any $n\geq1$, we have $$\label{con:parti}
\sum_{\pi\in P^{\pm}(\{1,1,2,2,\ldots,n,n\})} t^{\des(\pi)}=\sum_{\e\in\I_{2n}^{(1,4,3,8,\ldots,2n-1,4n)}}t^{\asc(\e)}.$$
For $n\leq2$, we have $$\sum_{\pi\in P^{\pm}(\{1,1\})} t^{\des(\pi)}=1+3t=\sum_{\e\in\I_{2}^{(1,4)}}t^{\asc(\e)}$$ and $$\sum_{\pi\in P^{\pm}(\{1,1,2,2\})} t^{\des(\pi)}=1+31t+55t^2+9t^3=\sum_{\e\in\I_{4}^{(1,4,3,8)}}t^{\asc(\e)}.$$
In section \[conj:sv\], we will give a simple proof of Theorem \[conj:VS\] through verifying the recurrence formulas for both sides of Eq. . Motivated by this conjecture, we study the descent polynomial of signed permutations of a general multiset (or called general signed multipermutations for short). In section \[gen:chow\], we proved a factorial generating function formula for the $(\des,\fmaj)$-enumerator of general signed multipermutations, which generalizes a result of Chow and Gessel [@cg]. Finally, in section \[sig:simion\], we use this factorial formula to show that the descent polynomial of the signed multipermutations is real-rootedness.
Proof of Theorem \[conj:VS\] {#conj:sv}
============================
\[lem:1\] Let $$E_n(t)=\sum_{i=0}^{2n-1}E_{n,i}t^i:=\sum_{\e\in\I_{2n}^{(1,4,3,8,\ldots,2n-1,4n)}}t^{\asc(\e)}.$$ Then $$\label{rec:Ent}
E_{n+1,i}=(2i^2+3i+1)E_{n,i}+(2i(4n-2i+3)+2n+1)E_{n,i-1}+(2n+2-i)(4n-2i+5)E_{n,i-2},$$ with boundary conditions $E_{n,i}=0$ for $i<0$ or $i>2n-1$.
The following formula was established in [@ss Theorem 13] using Ehrhart theory: $$\label{eq:ehrhart}
\frac{E_n(t)}{(1-t)^{2n+1}}=\sum_{k\geq0}((k+1)(2k+1))^nt^k.$$ Thus we have $$\begin{aligned}
\frac{E_{n+1}(t)}{(1-t)^{2n+3}}&=\sum_{k\geq0}((k+1)(2k+1))^{n+1}t^k\\
&=\sum_{k\geq0}((k+1)(2k+1))^{n}(k+1)(2k+1)t^k\\
&=\sum_{k\geq0}((k+1)(2k+1))^{n}(2k(k-1)+5k+1)t^k\\
&=2t^2(E_{n}(t)(1-t)^{-2n-1})''+5t(E_{n}(t)(1-t)^{-2n-1})'+\frac{E_{n}(t)}{(1-t)^{2n+1}}.\end{aligned}$$ The recursive formula then follows from the above equation by multiplying both sides by $(1-t)^{2n+3}$ and then extracting the coefficients of $t^i$.
\[lem:2\] Let $$P_n(t)=\sum_{i=0}^{2n-1}P_{n,i}t^i:=\sum_{\pi\in P^{\pm}(\{1,1,2,2,\ldots,n,n\})} t^{\des(\pi)}.$$ Then $$\label{rec:Pni}
P_{n+1,i}=(2i^2+3i+1)P_{n,i}+(2i(4n-2i+3)+2n+1)P_{n,i-1}+(2n+2-i)(4n-2i+5)P_{n,i-2},$$ with boundary conditions $P_{n,i}=0$ for $i<0$ or $i>2n-1$.
Denote by $\P_{n,i}$ the set of signed permutation of $\{1,1,2,2,\ldots,n,n\}$ with $i$ descents. Clearly, every signed permutation in $\P_{n+1,i}$ can be obtained from one of the following three different cases:
1. from a signed permutation in $\P_{n,i}$ by inserting $\{n+1,n+1\}$, $\{\overline{n+1},\overline{n+1}\}$ or $\{n+1,\overline{n+1}\}$;
2. from a signed permutation in $\P_{n,i-1}$ by inserting $\{n+1,n+1\}$, $\{\overline{n+1},\overline{n+1}\}$ or $\{n+1,\overline{n+1}\}$;
3. from a signed permutation in $\P_{n,i-2}$ by inserting $\{n+1,n+1\}$, $\{\overline{n+1},\overline{n+1}\}$ or $\{n+1,\overline{n+1}\}$.
In case $(1)$, one can check that there are $2i^2+3i+1$ ways to insert $\{n+1,n+1\}$, $\{\overline{n+1},\overline{n+1}\}$ or $\{n+1,\overline{n+1}\}$ into each signed permutation in $\P_{n,i}$ to become a signed permutation in $\P_{n+1,i}$ (be careful when $n+1$ or $\overline{n+1}$ is inserted to the right of a signed permutation). So there are $(2i^2+3i+1)P_{n,i}$ signed permutations constructed from case $(1)$. Similarly, there are $(2i(4n-2i+3)+2n+1)P_{n,i-1}$ and $(2n+2-i)(4n-2i+5)P_{n,i-2}$ signed permutations in $\P_{n+1,i}$ constructed from cases $(2)$ and $(3)$, respectively. Those amount to the right hand side of , which completes the proof.
By Lemma \[lem:1\] and \[lem:2\], we see that $E_{n,i}$ and $P_{n,i}$ satisfy the same recurrence relation and boundary conditions, so they are equal. This finishes the proof of Theorem \[conj:VS\].
We do not see how to deduce recurrence for $E_{n,i}$ directly from its interpretation as inversion sequences.
General signed multipermutations {#gen:chow}
================================
In this section, we consider the descent polynomial on signed permutations of the general multiset $
M_{\mm}:=\{1^{m_1},2^{m_2},\ldots,n^{m_n}\}
$ for each vector $\mm:=(m_1,m_2,\ldots,m_n)\in\PP^n$. Let $P(\mm)$ and $P^{\pm}(\mm)$ denote the set of all permutations and signed permutations of the multiset $M_{\mm}$, respectively.
Recall that the $q$-shift factorial $(t;q)_n$ is defined by $(t;q)_n :=\prod_{i=0}^{n-1}(1-tq^i)$ for any positive integer $n$ and $(t;q)_0=1$. The $q$-binomial coefficient ${n\brack k}_q$ is then defined as $${n\brack k}_q:=\frac{(q;q)_n}{(q;q)_{n-k}(q;q)_k}.$$ The following is a $q$-analog of a result of MacMahon [@MM Volme 2, p. 211], whose proof can be found in [@fh:q § 7].
\[macmahon\] For every $\mm\in\PP^n$ with $m_1+\cdots+m_n=m$, we have $$\frac{\sum_{\pi\in P(\mm)}t^{\des(\pi)}q^{\maj(\pi)}}{(t;q)_{m+1}}=\sum_{k\geq0}{m_1+k\brack m_1}_q\cdots{m_n+k\brack m_n}_qt^k.$$
Our signed version is:
\[thm:signed\] For every $\mm\in\PP^n$ with $m_1+\cdots+m_n=m$, we have $$\label{sign:eq}
\frac{\sum_{\pi\in P^{\pm}(\mm)}t^{\des(\pi)}q^{\fmaj(\pi)}z^{N(\pi)}}{(t;q^2)_{m+1}}=\sum_{k\geq0}\prod_{r=1}^n\left(\sum_{i=0}^{m_r}(zq)^i{m_r-i+k\brack m_r-i}_{q^2}{i+k-1\brack i}_{q^2}\right)t^k,$$ where $\fmaj(\pi):=2\maj(\pi)+N(\pi)$ and $N(\pi)$ is the number of negative signs in $\pi$.
Setting $z=0$, we recover Theorem \[macmahon\]. It is worth noticing that Foata and Han [@fh:sw3 Theorem 1.2] has calculated (another signed version of Theorem \[macmahon\]) the factorial generating function formula for $$\sum_{\pi\in P^{\pm}(\mm)}t^{\fdes(\pi)}q^{\fmaj(\pi)}z^{N(\pi)},$$ involving the so-called [*flag descent statistic*]{} $\fdes$, $\fdes(\pi):=2\des(\pi)-\chi(\pi_1<0)$, on signed multipermutations (or words).
$$\frac{\sum_{\pi\in P^{\pm}(\{1,2,\ldots,n\})} t^{\des(\pi)}q^{\fmaj(\pi)}}{(t;q^2)_{2n+1}}=\sum_{k\geq0}[2k+1]_q^nt^k.$$
Setting $m_1=\cdots=m_n=1$ and $z=1$ in Theorem \[thm:signed\].
$$\label{eq:sign}
\frac{\sum_{\pi\in P^{\pm}(\{1,1,2,2,\ldots,n,n\})} t^{\des(\pi)}}{(1-t)^{2n+1}}=\sum_{k\geq0}((k+1)(2k+1))^nt^k.$$
Setting $m_1=\cdots=m_n=2$ and $z=q=1$ in Theorem \[thm:signed\].
Comparing with we get another proof of Theorem \[conj:VS\]. As was already noticed in [@sv], there is not natural $q$-analog of Eq. .
We will prove Theorem \[thm:signed\] by using the technique of barred permutation motivated by Gessel and Stanley [@gs]. For each $\pi=\pi_1\cdots\pi_m\in P^{\pm}(\mm)$, we call the space between $\pi_i$ and $\pi_{i+1}$ the $i$-th space of $\pi$ for $0< i< m$. We also call the space before $\pi_1$ and the space after $\pi_{m}$ the $0$-th space and the $m$-th space of $\pi$, respectively. If $i\in\Des(\pi)$, then we call the $i$-th space a descent space.
A [*barred permutation*]{} on $\pi\in P^{\pm}(\mm)$ is obtained by inserting one or more vertical bars into some spaces of $\pi$ such that there is at least one bar in every descent space of $\pi$. For example, $||\bar{1}|\bar{2}2|1||$ is a barred permutation on $\pi=\bar{1}\bar{2}21$ but $\bar{1}|\bar{2}2|1||$ is not, since $0\in\Des(\pi)$ and there is not bar in the $0$-th space (i.e. before $\pi_1=\bar{1}$).
Let $B(\mm)$ be the set of barred permutations on $P^{\pm}(\mm)$. Let $\sigma\in B(\mm)$ be a barred permutation on $\pi$ with $b_i$ bars in the $i$-th space of $\pi$, we define the weight $\wt(\sigma)$ to be $$\wt(\sigma):=t^{\sum b_i}q^{N(\pi)+2\sum i b_i}z^{N(\pi)}.$$ For example, $\wt(||\bar{1}|\bar{2}2|1||)=t^6q^{26}z^2$. Now we count the barred permutations in $B(\mm)$ by the weight $\wt$ in two different ways. First, fix a permutation $\pi$, and sum over all barred permutations on $\pi$. Then fix the number of bars $k$, and sum over all barred permutations with $k$ bars.
Fix a permutation $\pi$, a barred permutation on $\pi$ can be obtained by inserting one bar in each descent space and then inserting any number of bars in every space. So counting all the barred permutations on $\pi$ by the weight $\wt$ gives $$t^{\des(\pi)}q^{\fmaj(\pi)}z^{N(\pi)}(1+t+t^2+\cdots)(1+tq^2+(tq^2)^2+\cdots)\cdots(1+tq^{2m}+(tq^{2m})^2+\cdots),$$ which is equal to $\frac{t^{\des(\pi)}q^{\fmaj(\pi)}z^{N(\pi)}}{(t;q^2)_{m+1}}$. This shows that $$\label{left}
\sum_{\sigma\in B(\mm)}\wt(\sigma)=\frac{\sum_{\pi\in P^{\pm}(\mm)}t^{\des(\pi)}q^{\fmaj(\pi)}z^{N(\pi)}}{(t;q^2)_{m+1}}.$$
For a fix $k\geq0$, let $B_k(\mm)$ be the set of barred permutations in $B(\mm)$ with $k$ bars. Now each barred permutation in $B_k(\mm)$ can be constructed by putting $k$ bars in one line and then inserting $m_r$ integers from $\{r,-r\}$, for $1\leq r\leq n$, to the $k+1$ spaces between each two adjacency bars (including the space in the left side and the right side), with the rule that negative integers can not be inserted to the left side. Thus by the well-known interpretation (cf. [@fh:q Proposition 4.1]) of the $q$-binomial coefficient $${n+r\brack n }_q=\sum_{0\leq c_1\leq c_2\cdots\leq c_n\leq r}q^{\sum c_i},$$ we have $$\sum_{\sigma\in B_k(\mm)} \wt(\sigma)=\prod_{r=1}^n\left(\sum_{i=0}^{m_r}(zq)^i{m_r-i+k\brack m_r-i}_{q^2}{i+k-1\brack i}_{q^2}\right)t^k.$$ Summing over all $k$ in the above equation and comparing with Eq. we get .
Signed version of Simion’s result about real-rootedness {#sig:simion}
=======================================================
It was shown in [@sv Theorem 1.1] that the ascent polynomial $\sum_{\e\in\I_{2n}^{(\s)}}t^{\asc(\e)}$ has only real roots for each $\s\in\PP^n$. In view of Theorem \[conj:VS\] we have
\[real:spec2\] The polynomial $$\sum_{\pi\in P^{\pm}(\{1,1,2,2,\ldots,n,n\})} t^{\des(\pi)}$$ has only real roots for any positive integer $n$.
Simion [@si § 2] proved that the descent polynomial on the permutations of a general multiset has only real roots. We have the following signed version of Simion’s result, which generalizes the $m_1=m_2=\cdots=m_n=1$ (i.e. the type B coxeter group) case of Brenti [@br] and Corollary \[real:spec2\].
\[realroot\] The descent polynomial $$\sum_{\pi\in P^{\pm}(\mm)}t^{\des(\pi)}$$ has only real roots for every $\mm\in\PP^n$.
The key point of the proof of the above result lies in the following simple lemma. Let $\PF_1[t^i]$ be the set of all polynomials with nonnegative coefficients and let $\PF[t^i]$ be the set of all real-rooted polynomials in $\PF_1[t^i]$. Actually, $\PF[t^i]$ is the set of all polynomials in $\R[t]$ whose coefficients form a [*Póly frequency sequence*]{} (cf. [@br Theorem 2.2.4]).
\[euler\] Let $$\label{rec:ab}
\frac{F_{n}(t)}{(1-t)^{n+1}}=\sum_{k\geq0}f(k)(ak+b)t^k\quad\text{and}\quad \frac{F_{n-1}(t)}{(1-t)^{n}}=\sum_{k\geq0}f(k)t^k.$$ If $a>0, n>\frac{b}{a}$, $F_{n-1}(t)\in \PF[t^i]$ and $F_{n}(t)\in \PF_1[t^i]$, then we have $F_{n}(t)\in \PF[t^i]$.
Clearly, by we have $$\frac{F_{n}(t)}{(1-t)^{n+1}}=\sum_{k\geq0}f(k)(ak+b)t^k=b\frac{F_{n-1}(t)}{(1-t)^{n}}+a\left(\frac{F_{n-1}(t)}{(1-t)^{n}}\right)'t.$$ From the above equation we deduce that $$F_{n+1}(t)=\left((an-b)t+b\right)F_n(t)+at(1-t)F_n'(t).$$ The lemma then follows by applying a result of Brenti [@br Theorem 2.4.5], which was established through some standard argument by using Rolle’s theorem.
By setting $q=1,z=1$ in Eq. and using Binomial theorem, we have $$\label{sig:nq}
\frac{\sum_{\pi\in P^{\pm}(\mm)}t^{\des(\pi)}}{(1-t)^{m+1}}=\sum_{k\geq0}\prod_{r=1}^n\frac{(2k+1)(2k+2)\cdots(2k+m_r)}{m_r!}t^k.$$ Observe that for any permutation $\pi$ of $[n]$, $$\sum_{\pi\in P^{\pm}(\mm)}t^{\des(\pi)}=\sum_{\pi\in P^{\pm}(\pi(\mm))}t^{\des(\pi)},$$ where $\pi(\mm)=\{1^{m_{\pi(1)}},2^{m_{\pi(2)}},\ldots,n^{m_{\pi(n)}}\}$. So we can assume that $m_1\geq m_2\geq\cdots\geq m_n$ and $m_1\geq2$. Let $\mm-\e_1:=\{1^{m_1-1},2^{m_2},\ldots,n^{m_n}\}$. By Eq. we have $$\frac{\sum_{\pi\in P^{\pm}(\mm)}t^{\des(\pi)}}{(1-t)^{m+1}}=\sum_{k\geq0}f(k)\frac{(2k+m_1)}{m_1}t^k
\quad\text{and} \quad\frac{\sum_{\pi\in P^{\pm}(\mm-\e_1)}t^{\des(\pi)}}{(1-t)^{m}}=\sum_{k\geq0}f(k)t^k,$$ where $f(k)=\frac{(2k+1)\cdots(2k+m_1)\cdots(2k+1)\cdots(2k+m_{n-1})\cdots(2k+1)\cdots(2k+m_n-1)}{(m_1-1)!\cdots m_{n-1}!(m_n)!}$. Note that $m>\frac{m_1}{2}$ and $\sum_{\pi\in P^{\pm}(\mm)}t^{\des(\pi)}\in\PF_1[t^i]$. Thus by Lemma \[euler\], the polynomial $\sum_{\pi\in P^{\pm}(\mm-\e_n)}t^{\des(\pi)}$ is a polynomial in $\PF[t^i]$ implies that of $\sum_{\pi\in P^{\pm}(\mm)}t^{\des(\pi)}$. The theorem then follows by induction on $m$.
Note that the above approach is also available for Simion’s result [@si] about the real-rootedness of $\sum_{\pi\in P(\mm)}t^{\des(\pi)}$.
The descent polynomial $$\sum_{\pi\in P^{\pm}(\mm)}t^{\des(\pi)}$$ is log-concave and unimodal for each $\mm\in\PP^n$.
[99]{}
F. Brenti, Unimodal, log-concave, and P—lya frequency sequences in combinatorics, Mem. Amer. Math. Soc., 81 (1989), 1–106.
F. Brenti, q-Eulerian polynomials arising from Coxeter groups, European J. Combin., 15 (1994), 417–441.
W.Y.C. Chen, A.J.X. Guo, P.L. Guo, H.H.Y. Huang and T.Y.H. Liu, $\s$-inversion Sequences and $P$-Partitions of type B, [arXiv:1310.5313](http://arxiv.org/abs/1310.5313).
C.-O. Chow and I.M. Gessel, On the descent numbers and major indices for the hyperoctahedral group, Adv. in Appl. Math., 38 (2007), 275–301.
I.M. Gessle and R.P. Stanley, Stirling polynomials, J. Combin. Theory Ser. A, 24 (1978), 24–33.
D. Foata and G.-N. Han, Signed words and permutations, III; The MacMahon Verfahren, Sém. Lothar. Combin. 54, B54a, 20 pp.
D. Foata and G.-N. Han, [*The $q$-series in Combinatorics; Permutation Statistics*]{}, Preliminary version, May 5, 2011.
P.A. MacMahon, [*Combinatory analysis, Vol I, II*]{}, Originally published in two Volumes (1915, 1916) by Cambridge University Press, Reprinted in one Volume, 1960, Chelsea Publishing Company.
C.D. Savage and M.J. Schuster, Ehrhart series of lecture hall polytopes and Eulerian polynomials for inversion sequences, J. Combin. Theory Ser. A, 119 (2012), 850–870.
C.D. Savage and M. Visontai, The $\s$-Eulerian polynomials have only real roots, Trans. Amer. Math. Soc., to appear (2013).
R. Simion, A multi-indexed Sturm sequence of polynomials and unimodality of certain combinatorial sequences, J. Combin. Theory Ser. A, 36 (1984), 15–22.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this paper, STAR’s measurement of clan multiplicity is presented for AuAu collisions at = 7.7, 11.5, 19.6, 27, 39, 62.4 and 200 GeV, for a variety of centrality classes. The mean number of particles per clan is found to decrease with decreasing centrality. Within the same centrality class, the mean number of particles per clan exhibits a reduction between 19.6 GeV and 62 GeV, with the minimum around 27 GeV. The structure is visible for most centralities, and most prominent for central collisions.'
address: 'Physics Department, Brookhaven National Laboratory, PO BOX 5000, Upton, NY 11973-5000'
author:
- 'Aihong Tang (for the STAR Collaboration)'
title: Beam Energy Dependence of Clan Multiplicity at RHIC
---
Introduction
============
Particle distributions in high energy collisions are frequently found to be of Negative Binomial Distribution (NBD) form for both total multiplicities and multiplicities in finite pseudorapidity intervals. One satisfactory explanation of this observation is that particle are produced via cascade process, and such interpretation leads to the development of clan concept of clustering in particle production [@Giovanni86].
Let’s recall the negative binomial probability density function, f(n)= [k+n-1 n]{}(1-p)\^n p\^k, \[eq:NegBinom\] where $p \in [0,1]$, and the real number $k>0$. Its mean and variance are given by $\mu=\frac{(1-p)k}{p}$ and $\sigma^2=\frac{(1-p)k}{p^2}$, respectively. In order to interpret the wide occurrence of NBD, a parameterization alternative to the standard NBD parameterization has been proposed [@Giovanni86]. It consists of the average number of groups of particles of common ancestor, also called the average number of clans, |[N]{}=k (1+), \[eq:meanN\] and the average number of particles per clan, |[n]{}\_c=. With this set of parameterization, particle production can be regarded as an independent production of clans followed by the decay of clans into particles according to a logarithmic multiplicity distribution. This can be seen by investigating the generating functions of relevant distributions. The generating function for the logarithmic distribution is, G\_(x) = , where $z$ is set to be $z=\frac{\mu}{\mu +k}$. Because clans are independently produced, they can be described by Poisson distribution, for which the generating function is G\_(x) = e\^[|[N]{}(x-1)]{}. The final distribution is a compound distribution of logarithm and Poisson, for which the generating function is given by G\_ &=& G\_(G\_(x))\
&=& e\^[|[N]{}(G\_[(x)]{}-1)]{}\
&=& ( )\^k, by which we have recovered the generating function of NBD.
Note that with the introduction of the clan concept, the NBD, if arranged properly, yields a similarity to the grand-canonical partition function. Such analogy led to the development of clan thermodynamics [@clanThermal02; @clanThermal08]. Here, adopting the concept of clan does not necessarily imply the endorsement of clan thermodynamics. The clan parameters have been used to identify abnormalities due to phase transition [@UA5; @EMC; @NA22; @E802; @NA35; @PHENIX]. For a similar purpose, this study is to examine the averaged charged multiplicity per clan as a function of collision energy.
Data Sets and Cuts
==================
The data set consist of minimum bias events of Au+Au collisions at = 7.7, 11.5, 19.6, 27, 39, 62.4 and 200 GeV, taken by the STAR experiment. Total events used in this analysis are, from the lowest energy to the highest one, 5, 7, 40, 71, 56, 73, and 25 million, respectively. Events from energies lower than 62.4 GeV were taken as part of the Beam Energy Scan program [@BESWhitePaper] at RHIC, and events from 62.4 and 200 GeV were taken in year 2010. All events are required to have at least one Time of Flight [@TOF] hit matching to a track, a vertex within 30 centimeters to the center of STAR’s Time Projection Chamber [@TPC] in z direction, and within 2 centimeters to the beam in transverse direction. To avoid complications arising from having more than one collisions in the time window of detector readout, which usually happens when the luminosity is high, events from 62.4 GeV and 200 GeV are selected from those taken with low luminosity beam (Beam Beam Counter coincidence rate less than 25k). Tracks are required to have a minimum of 15 hits, and the ratio of number of hits to number of maximum possible hits to be greater than 0.52. The distance of the closest approach (DCA) of a track to the primary vertex is required to be less than 1 centimeter. To avoid auto correlation, particles used in the study of multiplicity distribution and those used in the centrality definition are taken from different pseudorapidity ranges, which is $|\eta|<0.5$ for the former and $|\eta|>0.5$ for the latter. In this paper the multiplicity in the latter is also called “reference-multiplicity-2" to distinguish itself from the usual reference multiplicity defined with particles in $|\eta|<0.5$. In the picture of particles being produced by clans, one is interested in produced particles only. Thus protons are removed in this study, because they are contaminated by transported ones which can introduce trivial energy dependence, in particular at low energies where transported protons become dominant. The removal of protons is achieved by requiring tracks’ energy loss per unit length ($dE/dx$) to be larger than a threshold set to 2 standard deviations below the theoretically expected value for protons at the same rigidity.
Analysis Procedure and Results
==============================
For events with a given reference-multiplicity-2, the corresponding multiplicity distribution in $|\eta|<0.5$ can be well described by NBD. This can be seen in Figure \[fig:NBD\_data\], in which the multiplicity distribution in $|\eta|<0.5$ is plotted for a selected set of reference-multiplicity-2 numbers. The curves are fits to the NBD and they describe the data very well.
![\[fig:NBD\_data\]From left to right, the multiplicity distribution in $|\eta|<0.5$ for reference-multiplicity-2 = 11, 18, 30, 48, 76, 116, 169, 242 and 289, respectively. Curves are fits to NBD. All histograms are for Au+Au collisions at = 27 GeV. ](DemoNB_27GeV_negAndPosExP.eps){width="20pc"}
The fit is applied to multiplicity distributions with the finest bin of reference-multiplicity-2, and the corresponding $\sigma^2$ and $\mu$ can be obtained. With both $\sigma^2$ and $\mu$ known, $\bar{N}$ and $\bar{n}_c$ can be calculated with the transformation mentioned above. In the top panel of Figure \[fig:fitPar\_refMult2\], $\sigma^2$, $\mu$ and $\bar{N}$ are plotted as a function of reference-multiplicity-2, and in the bottom panel, the average multiplicity for charged particles per clan ($\bar{n}_c$) is shown. Note that $\bar{N}$ and $\bar{n}_c$ can be calculated with either directly calculated $\sigma^2$ and $\mu$, shown by black symbols, or that from fitting the NBD, shown by non-black symbols. The two sets of results agree with each other well and in the follow-up analysis only the set based on fitting NBD is used, because fitting is less sensitive to contaminations from rare events with abnormally large or small multiplicity. In general $\sigma^2$ is found to be greater than $\mu$ which is a feature of NBD. With decreasing centrality (increasing reference-multiplicity-2), $\bar{N}$ keep increasing while $\bar{n}_c$ shows a moderate decrease towards unit, which means that particle production is approaching the limit of independent production (Poisson) in central collisions.
![\[fig:fitPar\_refMult2\] $\sigma^2$, $\mu$, $\bar{N}$ (top panel) and $\bar{n}_c$ (bottom panel) as a function of reference-multiplicity-2 for Au+Au collisions at = 27 GeV. See text for details.](DemoFitPar_refMult2_negAndPosExP_27GeV.eps){width="20pc"}
The obtained $\bar{N}$ and $\bar{n}_c$ are then averaged over reference-multiplicity-2 for each centrality bin, of which the boundaries are shown by mesh area in the bottom panel of Figure \[fig:fitPar\_refMult2\]. Note by performing the analysis with the finest reference-multiplicity-2 bin, one avoids the additional fluctuation from a wide multiplicity bin which can significantly bias the variance.
![\[fig:meanNClan\] Average total number of clans per event $\bar{N}$ as a function of collision energy, for various centrality classes. Boxes represents systematic error resulted from variations in STAR’s day-to-day operation during data-taking.](meanNOfClansHis_fit_rebinned.eps){width="23pc"}
![\[fig:meanClanM\] Average charged particles per clan $\bar{n}_c$ as a function of collision energy, for various centrality classes. Boxes represents systematic error resulted from variations in STAR’s day-to-day operation during data-taking.](meanClanMultHis_fit_rebinned.eps){width="23pc"}
The same procedure is applied on data from all energies and the final $\bar{N}$ and $\bar{n}_c$ as a function of collision energy is presented in Figure \[fig:meanNClan\] and Figure \[fig:meanClanM\], respectively. Not surprisingly the average total number of clans per event is found to increase either with decreasing centrality, or with increasing energy, indicating that more clans are produced when the system becomes larger and/or when there is more energy available for particle production. The average charged particles per clan shows little dependence on energy for peripheral collisions, which is expected if peripheral collisions are regarded as simple superpositions of independent binary collisions, thus they should show the same clustering property. From peripheral to central collisions, $\bar{n}_c$ exhibits an energy dependence, with a minimum at 27 GeV and an enhancement at 62 GeV. Such dependence is seen in non-peripheral collisions and is strongest in most central collisions. If there is a phase transition, one would expect a disturbance in particle production which reveals itself through the change of clustering characteristics. The feature observed is a motivation for further studies in this direction. It is worth to point out that the dependence becomes less significant when the DCA cut is released to 2 cm or 3 cm, or when protons, contaminated by transported ones, are included. Both observations (not shown) indicate that the feature is prominent in primarily produced particles, for which the clan concept applies. The apparent enhancement at 62 GeV may be affected by worse beam condition at that particular energy and needs a further investigation.
Due to the change in operation condition, for example, temperature and air pressure of TPC, the efficiency is not a constant. Nevertheless, one can correct for efficiency and recover the efficiency-corrected $\bar{N}$ and $\bar{n}_c$ by applying the following two identities [@TangWang]: &=&\
p &=& , where $\mu'$ and $p'$ are measured observables, and $\mu$ and $p$ are efficiency-corrected ones. Here $\epsilon$ is the [*detecting efficiency*]{} in general meaning, which includes both the finite detector efficiency effect and the finite detector acceptance effect, $\epsilon$ = efficiency $\times$ acceptance. With that, the efficiency corrected $\bar{N}$ and $\bar{n}_c$ can be calculated, |[N]{} &=& -k p = -k\
|[n]{}\_c &=& = . To study if the efficiency variation can cause such a feature, without loosing generality, assuming a 30% efficiency and 5% variation in year-to-year operation, the resulted variation in $\bar{n}_c$ is too small (comparable to symbol size in Figure \[fig:meanClanM\]) to explain the variation seen in data.
Summary
=======
In this paper, STAR’s measurement of average total number of clans per event, $\bar{N}$, and average charged particles per clan, $\bar{n}_c$, are studied for AuAu collisions at = 7.7, 11.5, 19.6, 27, 39, 62.4 and 200 GeV, for a variety of centrality classes. $\bar{N}$ is found to increase with decreasing centrality and/or increasing energy, and $\bar{n}_c$ is found to decrease with decreasing centrality. Within the same centrality class, $\bar{n}_c$ exhibits a reduction between 19.6 GeV and 62 GeV, with the minimum around 27 GeV. The structure is not seen in peripheral collisions (50%-70%) but is visible in middle central collisions and, becomes prominent in central collisions (top 5%). The variation of mean number of particles per clan with energy cannot be explained by variation in detecting efficiency. Whether the observed structure is connected to a phase transition needs further investigations.
References {#references .unnumbered}
==========
[99]{}
Giovanni A and Van Hove L 1986 Z. Phys. C 30 391 Giovanni A, Lupia S and Ugoccioni R 2002 Phys. Rev. D 65 094028 Brambilla M, Giovannini A and Ugoccioni 2008 Physica A 387 1110 Alner G J [*et al.*]{} \[UA5 Collaboration\] 1987 Phys. Rep. 154 247 Arneodo M [*et al.*]{} \[European Muon Collaboration\] 1987 Z. Phys. C 35 335; 1987 Erratum-[*ibid*]{}.36, 512 Adamus M [*et al.*]{}, \[EHS/NA22 Collaboration\] 1988 Z. Phys. C 37 215 Abbott T [*et al.*]{}, \[E-802 Collaboration\] 1995 Phys. Rev. C 52 2663 Bachler J [*et al.*]{}, \[NA35 Collaboration\] 1993 Z. Phys. C 57 541 Adare A [*et al.*]{}, \[PHENIX Collaboration\] 2008 Phys. Rev. C 78 044902 Aggarwal M. M. [*et al.*]{}, \[STAR Collaboration\] 2010 arXiv:1007.2613 Bonner B. [*et al.*]{}, 2003 Nucl. Instr. and Meth. in Phys. Res. A 508 181; Llope W.J. (for the STAR Collaboration), 2012 Nucl. Instr. and Meth. in Phys. Res. A s661 s110-s113 Anderson M. [*et al.*]{}, 2003 Nucl. Instr. and Meth. in Phys. Res. A 499 659 Tang A. and Wang G. 2013 Phys. Rev. C 88 024905
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
The downward closure of a language $L$ of words is the set of all (not necessarily contiguous) subwords of members of $L$. It is well known that the downward closure of any language is regular. Although the downward closure seems to be a promising abstraction, there are only few language classes for which an automaton for the downward closure is known to be computable.
It is shown here that for stacked counter automata, the downward closure is computable. Stacked counter automata are finite automata with a storage mechanism obtained by *adding blind counters* and *building stacks*. Hence, they generalize pushdown and blind counter automata.
The class of languages accepted by these automata are precisely those in the hierarchy obtained from the context-free languages by alternating two closure operators: imposing semilinear constraints and taking the algebraic extension. The main tool for computing downward closures is the new concept of Parikh annotations. As a second application of Parikh annotations, it is shown that the hierarchy above is strict at every level.
author:
- Georg Zetzsche
bibliography:
- 'bibliography.bib'
title: Computing downward closures for stacked counter automata
---
Introduction
============
Preliminaries {#sec:preliminaries}
=============
A hierarchy of language classes {#sec:hierarchy}
===============================
This section introduces a hierarchy of language classes that divides the class of languages accepted by stacked counter automata into levels. This will allow us to apply recursion with respect to these levels.
Parikh annotations {#sec:pa}
==================
Computing downward closures {#sec:dclosure}
===========================
Strictness of the hierarchy {#sec:strictness}
===========================
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In the present work, we consider a nonlinear inverse problem of identifying the lowest coefficient of a parabolic equation. The desired coefficient depends on spatial variables only. Additional information about the solution is given at the final time moment, i.e., we consider the final redefinition. An iterative process is used to evaluate the lowest coefficient, where at each iteration we solve the standard initial-boundary value problem for the parabolic equation. On the basis of the maximum principle for the solution of the differential problem, the monotonicity of the iterative process is established along with the fact that the coefficient approaches from above. The possibilities of the proposed computational algorithm are illustrated by numerical examples for a model two-dimensional problem.'
address:
- 'Nuclear Safety Institute, Russian Academy of Sciences, 52, B. Tulskaya, Moscow, Russia'
- 'North-Eastern Federal University, 58, Belinskogo, Yakutsk, Russia'
author:
- 'Petr N. Vabishchevich'
title: 'Computational identification of the lowest space-wise dependent coefficient of a parabolic equation'
---
Inverse problem ,identification of the coefficient ,parabolic partial differential equation ,two-level difference scheme 65M06 ,65M32 ,80A23
Introduction
============
Mathematical modeling of many applied problems of science and engineering leads to numerical solving inverse problems for equations with partial derivatives [@alifanov2011inverse; @lavrentev1986ill]. In the theoretical study of such problems, the main attention is given to issues of well-posedness of problems, the uniqueness of the solution and its stability.
For parabolic equations, inverse coefficient problems attract particular interest. In these problems, identification of coefficients of equations and/or their right-hand side is conducted using some additional information about the solution. It is possible to identify dependence of coefficients on time or on spatial variables [@isakov2017inverse; @prilepko2000methods]. Problems of identifying the right-hand side of the equation belong to the class of linear inverse problems. Other inverse coefficient problems are non-linear that complicate significantly their study.
Among inverse problems of coefficient identification for parabolic equations we can highlight problems of determining the dependence of the lowest coefficient (reaction coefficient) on spatial variables. As a rule, additional conditions are formulated as the solution value at the final time moment and so, in this case, we speak of the final redefinition. In a more general case, a redefinition condition is formulated as some integral time-average relation (integral redefinition). The existence and uniqueness of the solution to such an inverse problem and the well-posedness of this problem are considered in a number of works. The pioneer works [@isakov1991inverse; @prilepko1987solvability] are devoted to problems with the final redefinition in Hölder classes and are based on the Schauder principle. Later works [@kamynin2013inverse; @kamynin2010two; @prilepko1992inverse; @prilepko1993inverse] deal with problems with integral redefinition and so, they are studied in Sobolev classes.
In works [@kostin2015inverse; @prilepko1992inverse] (see also [@isakov2017inverse Theorem 9.1.4]), the existence of the solution to the inverse problem of finding the lowest coefficient of a parabolic equation is proved constructively. Namely, an iterative process is used with solving the standard initial-boundary parabolic problem at each iteration. It seems natural to implement this approach in a corresponding computational algorithm.
The standard approach to numerical solving inverse coefficient problems for partial differential equations is associated with the minimization of the residual functional using regularization procedures [@samarskii2007numerical; @vogel2002computational]. Computational algorithms are based on the employment of gradient iterative methods, where we solve initial-boundary value problems both for the initial parabolic equation and the equation that is conjugate to it. For problems of identifying the lowest coefficient of parabolic equations, which depends only on spatial variables, the optimization method in a combination with finite element approximations in space is used in the work [@chen2006solving]. Among later studies in this direction, we mention [@cao2018reconstruction; @yang2008inverse].
In the work [@liu2011lie], an iterative process for the identification of the reaction coefficient in the diffusion-reaction equation is proposed without any exact mathematical justification. For model one- and two-dimensional boundary value problems, using finite-difference approximations in space, the efficiency of this computational algorithm has been demonstrated. This approach has been also applied to some other inverse problems for parabolic equations, in particular, for identifying the highest coefficient [@liu2008lgsm].
In the present paper, we construct a computational algorithm for identifying the lowest coefficient with the final redefinition, which is based on an iterative adjustment of the reaction coefficient similarly to [@isakov2017inverse; @kostin2015inverse]. The main attention is paid to obtaining new conditions for the monotonicity of the iterative process for finding the lower coefficient of the parabolic equation, when the coefficient approaches from above. This study continues the work [@vabishchevich2017iterative], where we consider iterative methods for the approximate solving the linear inverse problem of identifying the right-hand side for the parabolic equation.
The paper is organized as follows. Statements of direct and inverse problems for the second-order parabolic equation are given in Section 2. The identification of the reaction coefficient that is independent of time is considered for the two-dimensional diffusion-reaction equation. An additional information on the solution of the equation is given at the final time moment. An iterative adjustment algorithm for the desired coefficient is investigated in Section 3. The proof of its monotonicity is based on the fulfillment of the maximum principle not only for the solution, but also for derivative of the solution with respect to time. In Section 4, we construct a computational algorithm for approximate solving the identification problem for the lowest coefficient of the parabolic equation, and a discrete problem is formulated using finite-element approximations in space and two-level time-stepping schemes. Results of computational experiments for a model boundary-value problem are represented in Section 5. The findings of the work are summarized in Section 6.
Problem formulation
===================
The inverse problem of identifying the lowest coefficient of a parabolic equation is considered. We confine ourselves to the two-dimensional case. Generalization to the 3D case is trivial. Let ${\bm x} = (x_1, x_2)$ and $\Omega$ be a bounded polygon. The direct problem is formulated as follows. We search $u({\bm x},t)$, $0 \leq t \leq T, \ T > 0$ such that it is the solution of the homogeneous parabolic equation of second order: $$\label{1}
\frac{\partial u}{\partial t}- {\rm div} (k({\bm x}) \, {\rm grad} \, u) + c({\bm x}) u = f({\bm x},t),
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T ,$$ with coefficient $0 < k_1 \leq k({\bm x}) \leq k_2$. The boundary conditions are also specified: $$\label{2}
k({\bm x}) \frac{\partial u}{\partial n} + \mu({\bm x}) u = 0,
\quad {\bm x} \in \partial\Omega,
\quad 0 < t \leq T,$$ where $\mu ({\bm x}) \geq 0, \ {\bm x} \in \partial \Omega$ and $\bm n$ is the normal to $\Omega$. The initial conditions are $$\label{3}
u({\bm x}, 0) = 0,
\quad {\bm x} \in \Omega .$$ The formulation (\[1\])–(\[3\]) presents the direct problem, where the coefficients of the equation as well as the boundary conditions are specified.
Let us consider the inverse problem, where in equation (\[1\]), the lowest coefficient $c({\bm x})$ that depends on spatial variables only is unknown. An additional condition is often formulated as $$\label{4}
u({\bm x}, T) = \psi ({\bm x}),
\quad {\bm x} \in \Omega .$$ In this case, we have the case of the final redefinition.
Conditions for the unique solvability of the inverse coefficient problem (\[1\])–(\[4\]) and its correctness in various functional classes are established, for example, in the works cited above (see [@isakov1991inverse; @prilepko1987solvability]). We focus on using the iterative process to identify the coefficient $c({\bm x})$, which has been employed, in particular, in [@isakov2017inverse; @kostin2015inverse; @prilepko1992inverse]. Let us formulate wider conditions for the monotonicity of the iterative process of defining a new initial approximation, when the desired coefficient approaches from above. In our consideration, we assume that the solution of the problem, the coefficients of the equation, and the boundary conditions are sufficiently smooth, i.e., we have all necessary derivatives with respect to the space variables and time.
On the set of functions satisfying the homogeneous boundary conditions (\[2\]), let us define the elliptic operator $\mathcal{A}$ by the relation $$\mathcal{A} u = - {\rm div} (k({\bm x}) \, {\rm grad} \, u) .$$ In this case, equations (\[1\]), (\[2\]) can be written in the compact form: $$\label{5}
\frac{\partial u}{\partial t} + \mathcal{A} u + c({\bm x}) u = f({\bm x},t),
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T .$$ Without loss of generality, we consider the inverse problem (\[3\])–(\[5\]) for the definition of the pair $(u,c)$ under a priori restrictions on the reaction coefficient: $$\label{6}
c({\bm x}) \geq 0,
\quad \bm x \in \Omega .$$ If $c({\bm x}) \geq m$ with a constant $m$, it is possible to employ the standard transition to the problem for the function $v = \exp(m t) u$.
Assume that for the right-hand side of equation (\[1\]) holds $$\label{7}
f({\bm x}, 0) = 0,
\quad \frac{\partial f}{\partial t}({\bm x}, t) > 0,
\quad \bm x \in \Omega .$$ Under these conditions, on the basis of the maximum principle for parabolic equations (see, e.g., [@friedman2008partial; @il1962linear]), the solution $u$ at the final time moment is positive, i.e. $$\label{8}
\psi ({\bm x}) > 0,
\quad \bm x \in \Omega .$$
Iterative process
=================
The inverse problem consists in evaluating the pair of functions $(u,c)$ from the conditions (\[3\])–(\[5\]) under the constraints (\[6\])–(\[8\]). The iterative process of identifying the coefficient $c({\bm x})$ is implemented as follows. It starts from specifying some initial approximation $c^0({\bm x})$. With the known $c^k({\bm x})$, $k=0,1, ...$, where $k$ is the iteration number, the direct problem is solved: $$\label{9}
\frac{\partial u^{k}}{\partial t} + \mathcal{A} u^k + c^k({\bm x}) u^k = f(\bm x,t),
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T ,$$ $$\label{10}
u^k({\bm x}, 0) = 0,
\quad {\bm x} \in \Omega .$$ A new approximation for the desired coefficient is evaluated from the equation at the final time moment $t=T$ using the redefinition (\[4\]): $$\label{11}
c^{k+1} ({\bm x}) \psi = - \frac{\partial u^{k}}{\partial t}(\bm x,T) - \mathcal{A} \psi + f(\bm x,T),
\quad {\bm x} \in \Omega .$$
In the works [@isakov2017inverse; @kostin2015inverse; @prilepko1992inverse], the initial approximation is given in the form $$\label{12}
c^0({\bm x}) = 0,
\quad {\bm x} \in \Omega .$$ In this case, the monotone approach to the required coefficient ($c^{k+1}({\bm x}) \geq c^k({\bm x})$, approaching from below) holds, if this monotonicity condition holds for $k=1$: $$\label{13}
\mathcal{A} (u^{0} - \psi) \geq 0,
\quad {\bm x} \in \Omega .$$ The condition (\[13\]) is strong enough, but it can be removed. To do this, consider the algorithm for monotone approaching the reaction coefficient $c({\bm x})$ from above.
For the initial-boundary value problem (\[3\]), (\[5\]), in assumption (\[7\]), we have $$\label{14}
u(\bm x,t) \geq 0,
\quad \frac{\partial u}{\partial t} (\bm x,t) \geq 0,
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T .$$ The non-negativity of the solution follows from the maximum principle and the non-negativity of the right-hand side ($f(\bm x,t) \geq 0$). The non-negativity of the time derivative is established similarly when considering problem for ${\displaystyle w = \frac{\partial u}{\partial t}}$. Differentiation of equation (\[5\]) by time gives $$\frac{\partial w}{\partial t} + \mathcal{A} w + c({\bm x}) w = \frac{\partial f}{\partial t} ({\bm x},t),
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T .$$ For $t=0$, from equation (\[5\]) and the first condition in (\[7\]), we get $$w({\bm x}, 0) = 0,
\quad \bm x \in \Omega .$$ From the maximum principle for this problem, it follows that $w({\bm x}, 0) \geq 0$.
In view of (\[14\]), from equation (\[5\]), for $t=T$, we obtain $$\label{15}
c({\bm x}) \psi \leq - \mathcal{A} \psi + f(\bm x,T),
\quad {\bm x} \in \Omega .$$ Thus, the inverse problem (\[3\])–(\[5\]) is considered with two-side restrictions (\[5\]) and (\[15\]) for the lowest coefficient $c({\bm x})$.
Let us consider the iterative process (\[9\])–(\[11\]) with the initial approximation $$\label{16}
c^0({\bm x}) \psi = - \mathcal{A} \psi + f(\bm x,T),
\quad {\bm x} \in \Omega .$$ To find $u^0({\bm x},t)$, we solve the problem $$\frac{\partial u^{0}}{\partial t} + \mathcal{A} u^0 + c^0({\bm x}) u^0 = f(\bm x,t),
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T ,$$ $$u^0({\bm x}, 0) = 0,
\quad {\bm x} \in \Omega .$$ For $u^0({\bm x},t)$, similarly to (\[14\]), we have $$\frac{\partial u^0}{\partial t} (\bm x,t) \geq 0,
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T .$$ In view of this, from (\[11\]), we obtain $$c^{1} ({\bm x}) \psi = - \frac{\partial u^{0}}{\partial t}(\bm x,T) - \mathcal{A} \psi + f(\bm x,T)
\leq - \mathcal{A} \psi + f(\bm x,T).$$ By (\[16\]), we arrive at $$\label{17}
c^{1} ({\bm x}) \leq c^{0} ({\bm x}) ,
\quad {\bm x} \in \Omega .$$
Let us formulate a problem for the solution difference between two adjacent iterations: $$\xi^{k}({\bm x}) = c^{k}({\bm x}) - c^{k-1}({\bm x}),
\quad w^{k}({\bm x},t) = u^{k}({\bm x},t) - u^{k-1}({\bm x},t) ,
\quad k = 1,2, ... .$$ From (\[9\]), (\[10\]), we have $$\label{18}
\frac{\partial w^{k}}{\partial t} + \mathcal{A} w^k + c^{k-1} ({\bm x}) w^k = - \xi^k({\bm x}) u^k ,$$ $$\label{19}
w^k({\bm x}, 0) = 0.$$ From (\[11\]), we get $$\label{20}
\xi ^{k+1} ({\bm x}) \psi = - \frac{\partial w^{k}}{\partial t} (\bm x,T).$$
Similarly to (\[14\]), we prove that $$\label{21}
u^k(\bm x,t) \geq 0,
\quad \frac{\partial u^k}{\partial t} (\bm x,t) \geq 0,
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T ,$$ for $k = 0,1, ...$. Considering the problem (\[18\]), (\[19\]) for $k=1$, in view of (\[17\]), on the basis of the maximum principle, we obtain $$\label{22}
w^1 (\bm x,t) \geq 0,
\quad \xi^1 (\bm x) \leq 0,
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T .$$ An analogous property of the monotonicity of the approximate solution also holds for other $k = 2,3, ... $: $$\label{23}
w^k (\bm x,t) \geq 0,
\quad \xi^k (\bm x) \leq 0,
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T .$$
The proof is by induction on $k$. For $k=1$, it is satisfied (see (\[22\])). Let us show that from the fulfillment (\[23\]) for some $k$ this holds also for $k+1$. If (\[23\]) holds, taking into account the second inequality (\[21\]), after differentiating (\[18\]) with respect to $t$, we obtain $$\frac{\partial w^k}{\partial t} (\bm x,t) \geq 0,
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T .$$ Under these conditions, directly from (\[20\]), it follows that $$\xi^{k+1} (\bm x) \leq 0,
\quad {\bm x} \in \Omega ,$$ and from (\[18\]), (\[19\]), for $k \rightarrow k+1$, we get $$w^{k+1} (\bm x,t) \geq 0,
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T .$$
Define the error of the approximate solution as follows: $$\delta c^{k}({\bm x}) = c^{k}({\bm x}) - c({\bm x}),
\quad \delta u^{k}({\bm x},t) = u^{k}({\bm x},t) - u({\bm x},t) ,
\quad k = 0,1, ... .$$ By (\[9\]), (\[10\]), we get $$\label{24}
\frac{\partial \delta u^{k}}{\partial t} + \mathcal{A} \delta u^k + c ({\bm x}) \delta u^k = - \delta c^k({\bm x}) u ,$$ $$\label{25}
\delta u^k({\bm x}, 0) = 0.$$ From (\[11\]), we have $$\label{26}
\delta c^{k+1} ({\bm x}) \psi = - \frac{\partial \delta u^{k}}{\partial t} (\bm x,T)$$ for $k = 0,1, ... $.
For $k=0$, we obtain $$\label{27}
\delta u^0 (\bm x,t) \leq 0,
\quad \delta c^0 (\bm x) \geq 0,
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T .$$ The second inequality follows immediately from (\[15\]), (\[16\]). The first inequality is established for the solution of the problem (\[24\]), (\[25\]) with $k=0$ using the maximum principle. Further, similarly to (\[23\]), on the basis of induction, the property of monotonicity is established for other $k = 1,2, ... $: $$\label{28}
\delta u^k (\bm x,t) \leq 0,
\quad \delta c^k (\bm x) \geq 0,
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T .$$
The result of our consideration is the following statement on the monotonicity of the iteration process (\[9\])–(\[11\]).
\[t-1\] The iteration process (\[9\])–(\[11\]) with the initial approximation specified by (\[16\]) is monotone and $$\label{29}
\begin{split}
u(\bm x,t) & \geq u^k (\bm x,t) \geq u^{k-1} (\bm x,t), \\
c(\bm x) & \leq c^k (\bm x) \leq c^{k-1} (\bm x) ,
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T ,
\end{split}$$ for all $k = 1,2, ... $.
If the initial condition is given in the form (\[12\]), the similar statement that $$\begin{split}
u(\bm x,t) & \leq u^k (\bm x,t) \leq u^{k-1} (\bm x,t), \\
c(\bm x) & \geq c^k (\bm x) \geq c^{k-1} (\bm x) ,
\quad {\bm x} \in \Omega,
\quad 0 < t \leq T ,
\end{split}$$ is proved under the additional condition (\[13\]).
Computational implementation
============================
It seems reasonable to recall some general points of numerical solving the inverse coefficient problem (\[1\])–(\[4\]) on the basis of the iterative adjustment of the desired reaction coefficient. The monotonicity of the iterative process is established in Theorem \[t-1\] using the maximum principle for the solution and its time derivative (\[14\]). In constructing discretizations in space and time, we need to preserve this basic property of the differential problem, i.e., an approximate solution of the problem should satisfy the maximum principle.
Special attention should be given to monotone approximations in space (approximations of the diffusion-reaction operators) and discretizations in time. The maximum principle is formulated in the most simple way (see, e.g., [@Samarskii]) for difference schemes on rectangular grids. For steady-state problems, its implementation is associated with a diagonal dominance for the corresponding matrix and non-positivity of off-diagonal elements. Some possibilities for constructing monotone approximations on general irregular grids (using the finite volume method) and the maximum principle for convection-reaction problems with anisotropic diffusion coefficients are discussed in the work [@droniou2014finite].
Discretization in time leads to additional restrictions on monotonicity. For example, a typical situation is the case, where the monotonicity of the approximate solution in two-level schemes is ensured by using a small enough step in time. Unconditionally monotone time approximations for parabolic problems occure (see, for example, [@HundsdorferVerwer2003; @Samarskii]) when using fully implicit two-level schemes (backward Euler).
Here, we focus on the application of the finite element method. Monotone approximations in space for linear finite elements can be constructed with restrictions on a computational grid (Delaunay-type mesh, see, e.g., [@huang2011discrete; @letniowski1992three]). Some additional restrictions arise (see, for instance, [@brandts2008discrete; @ciarlet1973maximum]) from the reaction coefficient. They can be removed using the standard approach based on a correction of the approximations of the coefficient at the time derivative and reaction coefficient employing lumping procedures (see, e.g., [@chatzipantelidis2015preservation; @Thomee2006]).
To solve numerically the problem (\[1\])–(\[4\]), we employ finite element approximations in space [@brenner2008mathematical; @Thomee2006]. In the Hilbert space $H = L_2(\Omega)$, we define the scalar product and norm in the standard way: $$(u,v) = \int_{\Omega} u({\bm x}) v({\bm x}) d{\bm x},
\quad \|u\| = (u,u)^{1/2} .$$ We define the bilinear form $$a(u,v) = \int_{\Omega } k \, {\rm grad} \, u \, {\rm grad} \, v \, d {\bm x} +
\int_{\partial \Omega } \mu \, u v d {\bm x} .$$ Define a subspace of finite elements $V^h \subset H^1(\Omega)$. Let $\bm x_i, \ i = 1,2, ..., M_h$ be triangulation points for the domain $\Omega$. When using Lagrange finite elements of the first order (piece-wise linear approximation), we can define pyramid function $\chi_i(\bm x) \subset V^h, \ i = 1,2, ..., M_h$, where $$\chi_i(\bm x_j) = \left \{
\begin{array}{ll}
1, & \mathrm{if~} i = j, \\
0, & \mathrm{if~} i \neq j .
\end{array}
\right .$$ For $v \in V_h$, we have $$v(\bm x) = \sum_{i=i}^{M_h} v_i \chi_i(\bm x),$$ where $v_i = v(\bm x_i), \ i = 1,2, ..., M_h$.
Let us define a uniform grid in time $$t_n=n\tau,
\quad n=0,1,...,N,
\quad \tau N=T$$ and denote $y_n = y(t_n), \ t_n = n \tau$. Define an approximate solution of the inverse problem (\[1\])–(\[4\]) as $$w_{n}(\bm x) \in V^h, \quad n=0,1,...,N,
\quad s(\bm x) \in V^h .$$ For the fully implicit scheme, the solution is evaluated from $$\label{30}
\begin{split}
& \left (\frac{w_{n+1} - w_{n}}{\tau} , v \right ) + a(w_{n+1}, v) + (s w_{n+1}, v) = (f_{n+1}, v), \\
& \qquad n=0,1,...,N-1,
\quad 0 < t \leq T,
\quad v \in V^h .
\end{split}$$ From (\[3\]), (\[4\]), we have $$\label{31}
w_{0} = 0,
\quad (w_{N}, v) = (\psi, v),
\quad v \in V^h .$$
The computational algorithm for solving the problem (\[30\]), (\[31\]) is based on the iterative method (\[9\])–(\[11\]), (\[16\]) for identifying the lowest coefficient. The calculation starts from specifying the initial approximation for the desired coefficient: $$\label{32}
(c^0({\bm x}) \psi, v) = - a(\psi, v) + (f(\bm x,T), v),
\quad v \in V^h .$$ For known $c^k({\bm x})$, we solve the direct problem for evaluating $w_{n}^k({\bm x})$: $$\label{33}
\begin{split}
& \left (\frac{w_{n+1}^k - w_{n}^k}{\tau} , v \right ) + a(w_{n+1}^k, v) + (s^k w_{n+1}^k, v) = (f_{n+1}, v), \\
& \qquad n=0,1,...,N-1,
\quad 0 < t \leq T,
\end{split}$$ $$\label{34}
w_{0}^k = 0,
\quad v \in V^h .$$ After this, the reaction coefficient is adjusted: $$\label{35}
(c^{k+1}({\bm x}) \psi, v) = - \frac{w_{N}^k - w_{N-1}^k}{\tau } - a(\psi, v) + (f(\bm x,T), v),
\quad v \in V^h .$$
If we apply (\[32\])–(\[35\]), no additional procedures are needed for monotonization.
Numerical experiments
=====================
To illustrate the capabilities of the iterative technique for solving inverse problems of identifying the lowest coefficient of parabolic equations, we present the results of numerical experiments for a test problem. Let us consider model problem (\[1\])–(\[3\]), where $$k({\bm x}) = 1,
\quad f({\bm x}, t) = 100 t \exp(-x_1),
\quad \mu({\bm x}) = 10,
\quad T = 0.25 .$$ The problem is solved in the unit square $$\Omega = \{ \bm x = (x_1, x_2) \ | \ 0 < x_1 < 1, \ 0 < x_2 < 1 \}$$ The data at the final time moment (see (\[4\])) are obtained from the solution of the direct problem with a given coefficient $c({\bm x})$.
In our case, the coefficient $c({\bm x})$ is the piecewise constant (see Fig. \[f-1\]): inside a circle of radius 0.3 with the center (0.6,0.4), we put $c({\bm x})=5$; inside the square with side 0.2 and the center (0.3,0.8), we have $c({\bm x})=1$; and otherwise, we put $c({\bm x})=0$.
(0,0) rectangle +(10,10); (6,4) circle (3); (2,7) rectangle +(2,2); (0,0) – (12,0); (0,0) – (0,12); (-0.5,-0.5) node [$0$]{}; (10,0-0.5) node [$1$]{}; (11.5,-0.5) node [$x_1$]{}; (-0.5,10) node [$1$]{}; (-0.5,11.5) node [$x_2$]{}; (6,4) node [$5$]{}; (3,8) node [$1$]{}; (2,2) node [$0$]{};
The solution of the direct problem with this coefficient $c({\bm x})$ at the final time moment (the function $u(\bm x,T)$) is used as input data for the inverse problem. In our analysis, we focus on iterative solving the identification problem after finite element discretizations in space. Because of this, we do not discuss the dependence of the accuracy of the numerical solution on approximations in space, it seems appropriate to do in a separate study. The effect of computational errors is studied via calculations on different time grids, when the input data is derived from the solution of the direct problem on more fine time grids and with higher-order approximations in time.
To solve the direct problem, we employ the time step $\tau = 1\cdot 10^{-5}$. The division into 50 intervals in each direction is used to construct the uniform spatial grid, the Lagrangian finite elements of first degree are applied. The solution at the finite time moment is shown in Fig. \[f-2\].
The inverse problem is solved using the fully implicit scheme (see (\[33\])–(\[35\])). The error of the approximate solution of the identification problem on a separate iteration is evaluated as follows: $$\varepsilon_\infty(k) = \max_{\bm x \in \Omega} |c^k(\bm x) - c(\bm x)|,$$ $$\varepsilon_2(k) = \|c^k(\bm x) - c(\bm x)\| .$$
The main issue is to evaluate an actual convergence rate for the iterative processes under the consideration. We need to recognize clearly how quickly the accuracy of the approximate solution is stabilized with increasing the iteration number. The obtained error itself depends on a time step, namely, the smaller time step, the higher accuracy of the approximate solution. Influence of the time step of the iterative process (\[33\])–(\[35\]) with the initial approximation (\[32\]) on accuracy is shown in Fig. \[f-3\]. We observe a high convergence of the iterative process and the improvement of the accuracy of the approximate solution by reducing the time step. Similar results for the iterative process with the initial approximation (\[12\]) are presented in Fig. \[f-4\].
\
\
\
\
The convergence of the approximate solution for the reaction coefficient is shown in Fig. \[f-5\] and Fig. \[f-6\]. For the iterative process with initial approximation (\[32\]), we observe a monotone convergence from above (see Fig. \[f-5\]). The iterative process with the initial approximation (\[12\]) is non-monotone. In particular, for $k=1$, on a part of the domain $\Omega$, the function $c^1({\bm x})$ is negative (see Fig. \[f-6\]).
![The solution $c({\bm x})$ for different $k$ with initial approximation (\[32\]): left — $k = 1$, center — $k = 2$, right — $k = 3$[]{data-label="f-5"}](figs/5-1.png){width="1.\linewidth"}
![The solution $c({\bm x})$ for different $k$ with initial approximation (\[32\]): left — $k = 1$, center — $k = 2$, right — $k = 3$[]{data-label="f-5"}](figs/5-2.png){width="1.\linewidth"}
![The solution $c({\bm x})$ for different $k$ with initial approximation (\[32\]): left — $k = 1$, center — $k = 2$, right — $k = 3$[]{data-label="f-5"}](figs/5-3.png){width="1.\linewidth"}
![The solution $c({\bm x})$ for different $k$ with initial approximation (\[12\]): left — $k = 1$, center — $k = 2$, right — $k = 3$[]{data-label="f-6"}](figs/6-1.png){width="1.\linewidth"}
![The solution $c({\bm x})$ for different $k$ with initial approximation (\[12\]): left — $k = 1$, center — $k = 2$, right — $k = 3$[]{data-label="f-6"}](figs/6-2.png){width="1.\linewidth"}
![The solution $c({\bm x})$ for different $k$ with initial approximation (\[12\]): left — $k = 1$, center — $k = 2$, right — $k = 3$[]{data-label="f-6"}](figs/6-3.png){width="1.\linewidth"}
Conclusion
==========
1. A nonlinear inverse problem of identifying the lowest coefficient that depends only on spatial variables is studied for a second-order parabolic equation. The solution of the parabolic equation at the final time moment is given, i.e., the final redefinition is considered.
2. An iterative process of identifying an unknown coefficient is conducted by solving the standard initial-boundary value problem at each iteration. The main result is in establishing the monotonicity of the iterative process, where the desired lower coefficient approaches from above.
3. The computational algorithm is based on standard approximations in space by linear finite elements, whereas time-stepping is implemented using the fully implicit two-level schemes.
4. Possibilities of the proposed algorithms were demonstrated by numerical solving a test two-dimensional problem.
Acknowledgements {#acknowledgements .unnumbered}
================
The work was supported by the mega-grant of the Russian Federation Government (\# 14.Y26.31.0013) and RFBR (project 17-01-00689).
[31]{} natexlab\#1[\#1]{}\[1\][`#1`]{} \[2\][\#2]{} \[1\][\#1]{} \[1\][[](http://dx.doi.org/#1)]{} \[1\][[](pmid:#1)]{} \[2\][\#2]{} , , , . , , , , () . , , , , , . , , , (). , , , , () . , , , () . , , () . , , () . , , , . , , () . , , , , . , , , , () . , , () . , , ed., , . , , () . , , , () . , , () . , , , , , . , , () . , , () . , , , () . , , , () . , , , () . , , , , , . , , , () . , , , , . , , , , . , , , , . , , () . , , , . , , , , () .
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'Tomohiro Mano$^1$ and Tomi Ohtsuki$^1 $[^1]'
bibliography:
- 'manoJPSJ19.bib'
---
[^1]: [email protected]
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
With [*ASCA*]{}, we have detected three X-ray flares from the Class I protostar YLW15. The flares occurred every $\sim$20 hours and showed an exponential decay with time constant 30–60 ks. The X-ray spectra are explained by a thin thermal plasma emission. The plasma temperature shows a fast-rise and slow-decay for each flare with $kT_{peak} \sim$ 4–6 keV. The emission measure of the plasma shows this time profile only for the first flare, and remains almost constant during the second and third flares at the level of the tail of the first flare. The peak flare luminosities $L_{X,peak}$ were $\sim$ 5–20 $\times
10^{31}$ erg s$^{-1}$, which are among the brightest X-ray luminosities observed to date for Class I protostars. The total energy released in each flare was 3–6$\times$10$^{36}$ ergs. The first flare is well reproduced by the quasi-static cooling model, which is based on solar flares, and it suggests that the plasma cools mainly radiatively, confined by a semi-circular magnetic loop of length $\sim
14\,R_\odot$ with diameter-to-length ratio $\sim 0.07$. The two subsequent flares were consistent with the reheating of the same magnetic structure as of the first flare. The large-scale magnetic structure and the periodicity of the flares imply that the reheating events of the same magnetic loop originate in an interaction between the star and the disk due to the differential rotation.
author:
- 'Yohko Tsuboi, Kensuke Imanishi and Katsuji Koyama'
- Nicolas Grosso and Thierry Montmerle
title: 'Quasi-periodic X-ray Flares from the Protostar YLW15'
---
Introduction
============
Low-mass Young Stellar Objects (YSOs) evolve from molecular cloud cores through the protostellar (ages $\sim$10$^{4-5}$ yr), the Classical T Tauri (CTTS: $\sim$ 10$^6$ yr) and Weak-lined T Tauri (WTTS: $\sim$10$^7$ yr) phases to main sequence. Protostars are generally associated with the Class 0 and I spectral energy distributions (SEDs), which peak respectively in the millimeter and infrared (IR) bands. Bipolar flows are accompanied with this phase, suggesting dynamic gas accretion. CTTSs have still circumstellar disks though they have expelled or accreted the infalling envelopes. They are associated with the Class II spectra, which peak at the near-IR. Finally, as the circumstellar disk disappears, YSOs evolve to WTTSs, associated with Class III stars. Early stellar evolution is reviewed by Shu, Adams, & Lizano (1987) and André & Montmerle (1994).
The [*Einstein*]{} Observatory discovered that T Tauri Stars (TTSs), or Class II and Class III infrared objects, are strong X-ray emitters, with the luminosities of 100–10000 times larger than solar flares. These X-rays showed high amplitude time variability like solar flares. The temperature ($\sim$1 keV) and plasma density ($n_e\sim$10$^{11}$cm$^{-3}$) are comparable to those of the Sun, hence the X-ray emission mechanism has been thought to be a scaled-up version of solar X-ray emission; i.e., magnetic activity on the stellar surface enhanced by a dynamo process (Feigelson & DeCampli 1981; Montmerle et al. 1983). X-ray and other high energy processes in YSOs are reviewed by Feigelson & Montmerle (1999).
In contrast to TTSs, Class I infrared objects are generally surrounded by circumstellar envelopes of $A_V$ up to $\sim$ 40 or more, hence are almost invisible in the optical, near infrared and even soft X-ray bands. The [*ASCA*]{} satellite, sensitive to high energy X-rays up to 10 keV which can penetrate heavy absorption, has found X-rays from Class I objects in the cores of the R CrA, $\rho$ Oph, and Orion clouds at the hard band ($>$ 2 keV) (Koyama et al. 1994; Koyama et al. 1996; Kamata et al. 1997, Ozawa et al. 1999). Even in the soft X-ray band, deep exposures with the [*ROSAT*]{} Observatory detected X-rays from YLW15 in $\rho$ Oph (Grosso et al. 1997) and CrA (Neuhäuser & Preibisch 1997).
A notable aspect of these pioneering observations was the discovery of X-ray flares from Class I stars. [*ROSAT*]{} discovered a giant flare from the protostar YLW15 with total luminosity (over the full X-ray band) of 10$^{34-36}$erg s$^{-1}$, depending on the absorption. [*ASCA*]{} observed more details of X-ray flares from protostars EL29 in the Opiuchus, R1 in the R CrA core, and SSV63E+W in the Orion, which are associated with larger $N_H$ $\simeq$ 10$^{22-23}$ cm$^{-2}$ than seen in TTSs.
All these findings led us to deduce that greatly enhanced magnetic activity, already well-established in older TTSs, is present in the earlier protostellar phase. Stimulated by these results, and to search for further examples of the protostellar activity in the X-ray band, we have performed an extensive follow-up observation of a core region in $\rho$ Oph, with several Class I X-ray sources. The follow-up observation was made with [*ASCA*]{} 3.5 years after the first observation (Koyama et al. 1994, Kamata et al. 1997). Some previously bright Class Is became dim, while other Class Is were identified as hard X-ray sources. This paper discusses the brightest hard X-ray source, YLW15, concentrating on the characteristics and implications of its peculiar time behavior: quasi-periodic hard X-ray flares. For comparison with previous results, we assume the distance to the $\rho$ Oph region to be 165 pc (Dame et al. 1987) throughout of this paper, although new Hipparcos data suggest a closer distance $d\sim120$ pc (Knude & Hog 1998).
Observation
===========
We observed the central region of $\rho$ Oph cloud with [*ASCA*]{} for $\approx$100 ks on 1997 March 2–3. The telescope pointing coordinates were $\alpha$(2000) = 16h 27.4m and $\delta$(2000) = $-$24${^\circ}$ 30$'$. All four detectors, the two Solid-state Imaging Spectrometers (SIS 0, SIS 1) and the two Gas Imaging Spectrometers (GIS 2, GIS 3) were operating in parallel, providing four independent data sets. Details of the instruments, telescope and detectors are given by Burke et al. (1991), Tanaka, Inoue, & Holt (1994), Serlemitsos et al. (1995), Ohashi et al. (1996), Makishima et al. (1996), and Gotthelf (1996).
Each of the GIS was operated in the Pulse Height mode with the standard bit assignment that provides time resolutions of 62.5 ms and 0.5 s for high and medium bit rates, respectively. The data were post-processed to correct for the spatial gain non linearity. Data taken at geomagnetic rigidities lower than 6 GV, at elevation angles less than 5$^\circ$ from the Earth, and during passage through the South Atlantic Anomaly were rejected. After applying these filters, the net observing time for both GIS 2 and GIS 3 was 94 ks.
Each of the SIS was operated in the 4-CCD/Faint mode (high bit rate) and in the 2-CCD/Faint mode (medium bit rate). However, we concentrate on the 4-CCD/Faint results in this paper since YLW15 is out of the 2-CCD field of view. The data were corrected for spatial and gain non-linearity, residual dark distribution, dark frame error, and hot and flickering CCD pixels using standard procedures. Data were rejected during South Atlantic Anomalies and low elevation angles as with GIS data. In order to avoid contamination due to light leaks through the optical blocking filters, we excluded data taken when the satellite viewing direction was within 20$^\circ$ of the bright rim of the Earth. After applying these filters, the net observing time for the 4-CCD mode was 61 ks for SIS 0 and 63 ks for SIS 1.
Towards the end of this observation, we detected an enormous flare from T Tauri star ROXs31 which is located close to YLW15 (see §3.1, source 6 in Figure 1). The peak flux of ROXs31 is 1–2 orders of magnitude larger than YLW15 (see §4.2), and its broad point spread function contaminates YLW15 during the flare. Therefore, we excluded the GIS and SIS data taken during the flare of ROXs31 in all analysis of YLW15.
Results and Analysis
====================
Images
------
Figure 1 shows X-ray images of the $\rho$ Ophiuchi Core F region in the two different energy bands (left panel: 0.7–2 keV, right panel: 2–10 keV), obtained with SIS detectors. Class I sources are indicated by crosses. Since the absolute [*ASCA*]{} positional errors can be as large as 40$''$ for SISs (Gotthelf 1996), we compared the [*ASCA*]{} peak positions to the more accurately known IR positions of two bright sources in the SIS field, ROXs21 (source 5) and ROXs31 (source 6), which are indicated by filled circles in Figure 1 left panel. To obtain the [*ASCA*]{} peak positions, we executed a two-dimensional fitting in the 0.7–2 keV band; we fitted these sources with a position-dependent point spread function in the 0.7–2 keV band and a background model. This procedure was done in the Display45 analysis software package (Ishisaki et al. 1998). The position of ROXs31 was based on the flare phase of this source, while the position of ROXs21 is based on the data before the flare of ROXs31. IR positions are provided by Barsony et al. (1997). The [*ASCA*]{} SIS positions had an average offset (weighted mean by photon counts) of $+$0.18 s in right ascension and $-$7.4$''$ in declination from the IR frame. This positional offset is corrected in Figure 1. After the boresight error correction, remaining excursions between the X-ray and IR positions are 5.5$''$ (rms), which is consistent with the SIS position uncertainty for point sources (Gotthelf 1996). We take the systematic positional error to be 5.5$''$.
From the 2–10 keV band image, we find that the X-ray fluxes from Class I protostars EL29 and WL6 (sources 3 and 4 in Figure 1, right panel) are fainter by one third and less than one third, respectively, comparing to those in the first [*ASCA*]{} observation made in August 1993 (Kamata et al. 1997). The brightest X-ray source in the 2–10 keV band is an unresolved source at $\alpha$(2000) = 16h 27m 27.0s and $\delta$(2000) = $-$24$^\circ$ 40$'$ 50$''$ in the position corrected frame. We derived this peak position by the two-dimensional fitting in 2–10 keV band. Since the statistical error is 1$''$, the overall X-ray error (including the systematic error) is $\pm6''$. The closest IR source is YLW15 with VLA position of $\alpha$(2000) = 16h 27m 26.9s and $\delta$(2000) = $-$24$^\circ$ 40$'$ 49.8$''$ ($\pm0.5''$; Leous et al. 1991), located 1.5$''$ away from the X-ray source. The next nearest source is GY263 with IR position of $\alpha$(2000) = 16h 27m 26.6s and $\delta$(2000) = $-$24$^\circ$ 40$'$ 44.9$''$ ($\pm$1.3$''$; Barsony et al. 1997). The source is located 5.5$''$ from the X-ray position on the border of the X-ray position error circle. Thus we conclude that the hard X-rays are most likely due to the Class I source YLW15.
X-ray Lightcurve of YLW15
-------------------------
We extracted a lightcurve from a $3'$ radius circle around the X-ray peak of YLW15 (see Figure 1). Before the enormous flare from T Tauri star ROXs31, which occurred in the last phase of this observation (see §2), we detected another large flare from Class II source SR24N, located about $7'$ away from YLW15 in the GIS field of view. To subtract the time variable contamination from the SR24N in the extended flux of YLW15, we selected a $3'$ radius background region (see Figure 1), equidistant from SR24N and YLW15. On the other hand, using such a background, we cannot exclude the contamination from ROXs21 (source 5 in Figure 1), which is 2 arcmin apart from YLW15. Since the X-rays from ROXs21 are dominant below 2 keV (see §3.3 and Fig.3), we examine time variability only in the hard X-ray band ($>$ 2 keV) in which the flux is dominated by YLW15.
Figure 2 (upper panel) shows the background-subtracted lightcurve in the 2–10 keV band with the sum of the SIS (SIS 0 and 1) and GIS (GIS 2 and 3) images. The lightcurve shows a sawtooth pattern with three flares. The peak fluxes of the flares become successively less luminous. Each flare exhibits a fast-rise and an exponential decay with an $e$-folding time of 31$\pm1$ ks ($\chi^2/d.o.f.$ = 61/46), 33$\pm3$ ks (80/47), and 58$_{-13}^{+24}$ ks (33/24), for the first, the second, and the third flares, respectively. We show the best fit lightcurves for the second and the third flares with dashed lines, and show the best-fit quasi-static model (see §4.1.1) for the first flare with a solid line in Figure 2 upper panel.
Time-Sliced Spectra of YLW15
----------------------------
To investigate the origin of the quasi-periodic flares, we made time-sliced spectra for the time intervals given in Figure 2. We extracted the source and background data for each phase from the same regions as those in the lightcurve analysis (see §3.2 and Figure 1). We found that all the spectra show a local flux minimum at $\approx$1.2 keV. For example, we show the spectra obtained with SISs at phases 1 and 8 in Figure 3. This suggests that the spectra have two components, one hot and heavily absorbed, the other cool and less absorbed.
Then we examined possible contamination from the bright, soft X-ray source ROXs21 (source 5 in Figure 1). We extracted the spectrum of ROXs21 from a $2'$ radius circle around its X-ray peak. We extracted the data only during phase 1, in order to be free from contamination from the flare on SR24N, which occurred during phases 4–6 (see §3.2). The background data for ROXs21 were extracted from a $2'$ radius circle equidistant from YLW15 and ROXs21 during phase 1. After the subtraction of the background, the spectrum of ROXs21 is well reproduced by an optically thin thermal plasma model of about 0.6 keV temperature with absorption fixed at $N_{\rm H} =
$1.3$\times$10$^{21}$ cm$^{-2}$ ($A_V$ = 0.6 mag; Bouvier and Appenzeller 1992) with $\chi^2/d.o.f.$ = 17/23. The flux of the soft component of YLW15 is about 30% of the flux of ROXs21, which is equal to the spill-over flux from ROXs21. Thus the soft X-ray component found in YLW15 spectra is due to contamination from the nearby bright source ROXs21.
Having obtained the best-fit spectrum for ROXs21, we fitted the spectrum of YLW15 in each phase with a two-temperature thermal plasma model. The cool component is set to the contamination from ROXs21, and the hot component is from YLW15. For YLW15, free parameters are temperature ($kT$), emission measure ($EM$), absorption ($N_{\rm H}$) and metal abundance. For ROXs21, $EM$ is the only free parameter and the other parameters are fixed to the best-fit values obtained in phase 1. We found no significant variation in $N_{\rm H}$ of YLW 15 from phase to phase, hence, we fixed the $N_{\rm H}$ to the best-fit value at phase 1. The resulting best-fit parameters of YLW15 for each time interval are shown in Table 1. The best-fit spectra of phases 1 and 8 are illustrated in Figure 3, and the time evolution of the best-fit parameters are shown in Figure 2.
[cccccccc]{} 1 & 5.6$^{+1.3}_{-1.0}$ & 0.4$^{+0.2}_{-0.1}$ & 7.2$_{-0.8 }^{+1.1}$ & 15$\pm$1 & 181/106\
2 & 4.4$^{+0.8}_{-0.7}$ & 0.4$^{+0.2}_{-0.2}$ & 3.4$_{-0.3}^{+0.4}$ & 6.0$\pm$0.7 & 191/164\
3 & 2.9$^{+0.5}_{-0.3}$ & 0.7$^{+0.5}_{-0.4}$ & 2.1$_{-0.3}^{+0.4}$ & 3.4$\pm$0.8 & 80/84\
4 & 5.3$^{+1.6}_{-1.1}$ & 0.6$^{+0.4}_{-0.4}$ & 2.7$_{-0.3}^{+0.4}$ & 5.6$\pm$0.8 & 55/52\
5 & 2.3$^{+0.4}_{-0.3}$ & 0.5$^{+0.6}_{-0.4}$ & 3.2$_{-0.7}^{+0.8}$ & 4.4$\pm$1.3 & 163/138\
6 & 1.6$^{+0.3}_{-0.3}$ & $<$1.3 & 2.3$_{-1.2}^{+1.5}$ & 2.5$\pm$1.5 & 46/40\
7 & 4.2$^{+1.5}_{-0.9}$ & $<$0.4 & 2.2$_{-0.3}^{+0.4}$ & 3.5$\pm$0.5 & 153/139\
8 & 3.0$^{+0.9}_{-0.5}$ & 0.7$^{+1.3}_{-0.5}$ & 1.7$_{-0.4}^{+0.4}$ & 2.9$\pm$0.6 & 59/65\
\[best\_fit\_parameters\]
Discussion
==========
In this observation, we detected hard X-rays ($>$ 2 keV) from the Class I source YLW15 for the first time; the source was not detected in the first [*ASCA*]{} observation executed 3.5 years later (Kamata et al. 1997). On the other hand, the Class I sources EL29 and WL6, which emitted bright hard X-rays in the first observation, became very faint (see §3.1). Then we conclude that hard X-rays from Class I protostars in the $\rho$ Oph cloud are highly variable in the time spans, and we suspect that the non-detection of hard X-rays from other Class I objects is partly due to the long-term time variability. From YLW15, we discovered a peculiar time behavior: quasi-periodic flares. We shall now discuss the relation between each flare from YLW15 and the physical conditions.
Physical Parameters of the Triple Flare
---------------------------------------
If the three intermittent flares are attributed to a single persistent flare, with the three events due to geometrical modulation, such as occultation of the flaring region by stellar rotation or orbit in an inner disk, then only the emission measure should have shown three peaks. The temperature would have smoothly decreased during the three flares. However, in our case, we see the temperature following the same pattern as the luminosity, decreasing after each jump, as indicated in Table 1 and Figure 2 middle panel. To test further, we fitted the temperatures in phases 1–8 with a single exponential decay model, and confirmed that it is rejected with 89 % confidence. We then interpret the variability of YLW15 as due to a triple flare, in which each flare followed by a cooling phase. Now, let us label phases 1–3, 4–6, and 7–8 as Flare I, II, and III, respectively. In this subsection we will use the cooling phases to estimate the physical conditions of the plasma in each flare. Details are given in the Appendix.
### Flare I
Here we assume that the plasma responsible for Flare I is confined in one semi-circular magnetic loop with constant cross section along the tube flux, with length $L$, and diameter-to-length (aspect) ratio $a$, based on the general analogy of solar-type flares. To the decay of Flare I, we apply a quasi-static cooling model (van den Oord & Mewe 1989), in which the hot plasma cools quasi-statically as a result of radiative (90 %) and conductive (10 %) losses (see detailed comment in Appendix §1). If the cooling is truly quasi-static, $T^{13/4}/EM$ shows constant during the decay (where $T$ is plasma temperature and $EM$ is the emission measure). We find for the three successive bins (phases 1–3) of Flare I that $(T/10^7\,{\rm K})^{13/4}/(EM/10^{54}\,{\rm cm}^{-3}) =
61\pm55, ~59\pm42, ~25\pm18$, which are not inconsistent with a constant value, taking into account the large error bars. Then, fitting our data with the quasi-static model, we find satisfactory values of $\chi^2_{red}$ for the count rates, temperature, and emission measure (see panel 1 in Table \[fit\]). We conclude that our hypothesis of quasi-static cooling is well verified; an underlying quiescent coronal emission is not required to obtain a good fit. The best fits are shown by solid lines in Figure 2. From the peak values of $T$ and $EM$, and the quasi-static radiative time scale, we derived loop parameters as listed in Table \[physical\_parameters\]. The detailed procedure is written in Appendix. The aspect ratio $a$ of 0.07 is within the range for solar active-region loops ($0.06 \le a_\odot \le 0.2$, Golub et al. 1980), which supports the assumed solar analogy.
### Flares II and III
The appearance of Flare II can be interpreted either by the reheating of the plasma cooled during Flare I, or by the heating of a distinct plasma volume, independent from that of Flare I. In the latter case, the lightcurve and $EM$ would be the sum of Flare I component and new flare component. The $EM$ of the new flare component can be derived by subtracting the component extrapolated from Flare I. The derived $EM$ are 1.6$\pm$0.4 (phase 4), 2.5$\pm$0.8 (phase 5), and 1.9$\pm$1.5 (phase 6). Then we fitted the lightcurve and $EM$ with the quasi-static model assuming the above two possibilities. However, in both of the cases, the model did not reproduce the lightcurve and $EM$ simultaneously, and the quasi-static cooling model cannot be adopted throughout the triple flare but Flare I.
For simplicity, we fitted the parameters in Flare II with an exponential model. The best-fit parameters are shown in Table 2, and each models for the reheating and the distinct flare assumptions are shown by the dashed and the dotted lines in Figure 2, respectively. The obtained $\chi^2_{red}$ and the timescale for each lightcurve and $EM$ were similar between the two assumptions. Therefore both of the possibilities cannot be discriminated. As for $EM$, both of the fits show no decay or very long time decay, which is not seen in the usual solar flares. The constant feature in $EM$ makes the quasi-static model unacceptable. Since we cannot derive the aspect ratio of Flare II by fitting with quasi-static cooling model, assuming the ratio derived in Flare I and that radiative cooling is dominant, we deduced the plasma parameters as shown in Table 3. Here, we derived the values assuming the reheating scenario (the former assumption). The results show that the plasma density and the volume remain roughly constant from Flare I. This supports that Flare II resulted from reheating of the plasma created by Flare I.
As for Flare III, because of the poor statistics and short observed period, the fits to lightcurve with the exponential model give no constrain between the above two possibilities. The results are shown in Table \[fit\] and Figure 2. We derived the plasma parameters of Flare III in the same way for Flare II, as shown in Table 3. These values are similar to those in Flare I and II. All these are consistent with the scenario that a quasi-periodic reheating makes the triple flare. The heating interval is $\sim$20 hour. The loop size is approximately constant through the three flares and is as large as $\sim$14 $R_\odot$. The periodicity and the large-scale magnetic structure support a scenario that an interaction between the star and the disk occurred by the differential rotation and reheated the same loop periodically (e.g., Hayashi, Shibata, & Matsumoto 1996). The detail will be presented in Paper II (Montmerle et al. 1999).
[cccccc]{} ${\cal F}_{\mathrm{I,qs}}(t)={\cal F}_{\mathrm{p}} \times \{ 1+(t-t_{\mathrm{I}})/3\tau
\} ^{-\alpha}$ & $I$ \[cntsks$^{-1}$\] & 4 & $172 \pm 5$ & $31 \pm 1$ & 0.8 (45)\
& $kT$ \[keV\] & 8/7 & $5.6 \pm 0.6$ & \[31\] & 0.5 (2)\
& $EM$ \[10$^{54}$cm$^{-3}$\] & 26/7 & $10.4 \pm 0.9$& \[31\] & 0.4 (2)\
${\cal F}_{\mathrm{II}}(t)={\cal F}_{\mathrm{p}} \times exp \{ -(t-t_{\mathrm{II}})/\tau\}$ & $I$ \[cntsks$^{-1}$\] & – & $62 \pm 3$ & $33 \pm 3$ & 1.7 (47)\
& $kT$ \[keV\] & – & $4 \pm 1$ & $50 \pm 20$ & 1.5 (1)\
& $EM$ \[10$^{54}$cm$^{-3}$\] & – & $3 \pm 1$ & $>$ 53 & 0.4 (1)\
${\cal F}_{\mathrm{I+II}}(t)={\cal F}_{\mathrm{I,qs}}(t) + {\cal F}_{\mathrm{II}}(t) $ & $I$ \[cntsks$^{-1}$\] & – & $46 \pm 3$ & $29 \pm 2$ & 1.6 (47)\
& $EM$ \[10$^{54}$cm$^{-3}$\] & – & $1.6 \pm 0.8$ & $>$ 54 & 0.6 (1)\
${\cal F}_{\mathrm{III}}(t)={\cal F}_{\mathrm{p}} \times exp \{ -(t-t_{\mathrm{III}})/\tau\}$ & $I$ \[cntsks$^{-1}$\] & – & $35 \pm 2$ & $60 \pm 20$ & 1.4 (24)\
& $kT$ \[keV\] & – & $5 \pm 3$ & 55 ($>$ 16) & $\chi^2=0$ (0)\
& $EM$ \[10$^{54}$cm$^{-3}$\] & – & $2.5 \pm 0.7$ & 71 ($>$ 23) & $\chi^2=0$ (0)\
${\cal F}_{\mathrm{I+II+III}}(t)={\cal F}_{\mathrm{I+II}}(t)+{\cal F}_{\mathrm{III}}(t)$ & $I$ \[cntsks$^{-1}$\] & – & $28 \pm 2$ & $60 \pm 20$ & 1.4 (24)\
\[fit\]
[cccccccc]{} I & 0.07$\pm$0.02 & 14$\pm$2 & 0.5$\pm$0.1 & 150$\pm$20 & 20$\pm$1 & 31$\pm$1 & 6.0$\pm$0.5\
II &\[0.07$\pm$0.02\]& 11$\pm$5 & 0.4$\pm$0.1 & 120$\pm$30 & 8$\pm$1 & 33$\pm$3 & 2.8$\pm$0.6\
III &\[0.07$\pm$0.02\]& 15$\pm$10 & 0.3$\pm$0.2 & 100$\pm$60 & 4.6$\pm$0.5 & 60$\pm20$ & 2.7$\pm$1.0\
\[physical\_parameters\]
Comparison with Other Flares
----------------------------
Among YSOs, TTSs have been known for strong X-ray time variabilities since the $Einstein$ Observatory discovered them (see §1). At any given moment, 5–10 % are caught in a high-amplitude flare with timescales of hours (Feigelson & Montmerle 1999). The most recent example is that of V773 Tau, which exhibits day-long flares with $L_{X,peak} \sim$ 2–10 $\times 10^{32}$ erg s$^{-1}$ and very high temperatures of $\sim 10^8$ K (Tsuboi et al. 1998). Other examples of bright TTS X-ray flares are P1724, a WTTS in Orion ($L_{X,peak} \sim 2 \times 10^{33}$ erg s$^{-1}$; Preibisch, Neuhaüser, & Alcalá 1995), and LkH$\alpha$92, a CTTS in Taurus (Preibisch, Zinnecker, & Schmitt 1993). These X-ray properties resemble those of RS CVn systems.
Recently, a dozen protostars have been detected in X-rays, and four of those showed evidence for flaring (RCrA core; Koyama et al. 1996, EL29; Kamata et al. 1997, YLW15; Grosso et al. 1997, and SSV63 E+W; Ozawa et al. 1999). In the “superflare” of YLW15 (Grosso et al. 1997), an enormous X-ray luminosity was recorded during a few hours. If we adopt the absorption we derived ($N_H =$ 3$\times10^{22}$ cm$^{-2}$), the absorption-corrected X-ray luminosity in 0.1–100 keV band is $L_{X,peak} \sim 10^{34}$ ergs$^{-1}$. The “triple flare” we detected in this observation does not reach the same level as the “superflare”: $L_{X,peak} = 5-20 \times 10^{31}$ erg s$^{-1}$ in the same X-ray band.
To compare the flare properties of our triple flare from YLW15 with other flare sources, including RS CVns, we selected bright flare sources as seen in Table 4. All the flare sources have a well-determined temperature using a wide range of energy band of $Tenma, Ginga$, and $ASCA$ satellites. Since the samples of YSO flares were less, we added two TTS flares; the flares on ROXs21 and SR24N detected in our observations (see §2 and §3). We analyzed them using GIS data. The densities and volumes for all the sources were derived assuming that radiative cooling is dominant.
As a result, we found that although less energetic than the “superflare”, with total energies in excess of 3–6 $\times 10^{36}$ ergs, the triple flare are on the high end of the energy distribution for protostellar flares. While the plasma densities, temperatures, and then derived equipartition magnetic fields are typical of stellar X-ray flares, the emitting volume is huge; it exceeds those in binary systems of RS CVns by a few orders of magnitude.
[cccccccccc]{} YLW15(I) & I & 0.5 & 3 & 31 & 6 & 0.7 & 6 &\
YLW15(II) & I & 0.4 & 2 & 33 & 5 & 0.3 & 3 &\
YLW15(III) & I & 0.3 & 2 & 60 & 4 & 0.2 & 3 &\
EL29 & I & 1 & 0.2 & 8 & 4 & 0.3 & 0.4 & (1)\
RCrA & I & 1 & 0.04 & 20 & 6 & 0.04 & 0.2 & (2)\
SSV63E+W & I & 1 & 2 & 10 & 6 & 2 & 4 & (3)\
SR24N & II & 5 & 0.1 & 6 & 7 & 3 & 6 &\
ROXs31 & III & 3 & 0.6 & 7 & 6 & 6 & 13 &\
ROX7 & III & 3 & 0.01 & 6 & 3 & 0.1 & 0.1 & (1)\
V773Tau & III & 3 & 0.6 & 8 & 10 & 5 & 8 & (4)\
Algol & & 1 & 0.09 & 20 & 5 & 0.1 & 0.6 & (5)\
$\pi$ Peg & & 8 & 0.006 & 2 & 6 & 0.3 & 0.2 & (6)\
AU Mic & & 6 & 0.002 & 2 & 5 & 0.06 & 0.03 & (7)\
\[comparison\]
Summary and Conclusions
=======================
In the course of a long exposure of the $\rho$ Oph cloud with [*ASCA*]{}, we found evidence for a ‘triple flare’ from the Class I protostar YLW15. This triple flare is the first example of its kind; it shows an approximate periodicity of $\sim 20$ hours. Each event shows a clear decrease in the temperature, followed by reheating, with $kT_{peak} \sim$ 4–6 keV, and luminosity $L_{X,peak} \sim$ 5–20 $\times
10^{31}$ erg s$^{-1}$. Apart from the periodicity, the characteristics of the flares are among the brightest X-ray detections of Class I protostars.
A fitting with the quasi-static model (VM), which is based on solar flares, reproduces the first flare well, and it suggests that the plasma cools mainly radiatively, having semi-circular shape with length $\sim 14\,R_\odot$ (radius of the circle $R \sim
4.5\,R_\odot$) and aspect ratio $\sim 0.07$. The minimum value of the confining field is $B \sim 150$ G. The two subsequent flares have weaker intensity than the first one but consistent with the reheating of basically the same magnetic structure as the first flare. The plasma volume is huge; a few orders of magnitude larger than the typical flares in RS CVns.
The fact that the X-ray flaring is periodic suggests that the cause for the heating is periodic, hence is linked with rotation in the inner parts of the protostar. The large size of the magnetic structure and the periodicity support the scenario that the flaring episode has originated in a star-disk interaction; differential rotation between the star and disk would amplify and release the magnetic energy in one rotation period or less, reheating the flare loop as observed in the second and third flares.
The authors thank all the members of the $ASCA$ team whose efforts made this observation and data analysis available. We also thank Prof. Eric D. Feigelson, Mr. Kenji Hamaguchi, Dr. Mitsuru Hayashi, Dr. Antonio Magazzu, Mr. Michael S. Sipior, and Prof. Kazunari Shibata for many useful comments in the course of this work. Y.T.acknowledges the award of Research Fellowship of the Japan Society for Young Scientists.
André, P., & Montmerle, T. 1994, , 420, 837
Barsony, M., Kenyon, S. J., Lada, E. A., & Teuben, P. J. 1997, , 112, 109
Bouvier, J., & Appenzeller, I. 1992, , 92, 481
Burke, B. E., Mountain, R. W., Harrison, D. C., Bautz, M. W., Doty, J. P., Ricker, G. R., & Daniels, P. J. 1991, IEEE Trans. ED-38, 1069
Chen, H., Myers, P. C., Ladd, E. F., & Wood, D. O. S. 1995, , 445, 377
Chen, H., Grenfell, T. G., Myers, P. C., & Hughes, J. D. 1997, , 478, 295
Cully, S. L., Siegmund, O. H. W., Vedder, P. W., & Vallerga, J. V. 1993, , 414, L49
Dame, T. M. et al. 1987, , 322, 706
Doyle, J. G. et al. 1991, , 248, 503
Feigelson, E. D., & DeCampli, W. M. 1981, , 243, L89
Feigelson, E. D., & Montmerle, T. 1999, , in press
Golub, L., Maxson, C., Rosner, R., Vaiana, G. S. & Serio, S. 1980, , 238, 343
Gotthelf, E. 1996, ASCANEWS (Greenbelt: NASA GSFC), 4, 31
Grosso, N., Montmerle, T., Feigelson, E. D., André, P., Casanova, S., & Gregorio-Hetem, J. 1997, , 387, 56
Hayashi, M. R., Shibata, K., & Matsumoto, R. 1996, , 468, L37
Ishisaki, Y., Matsuzaki, K., Shirahashi, A., Takahashi, T., & Itoh, R. 1998, Interactive Tool for Data Analysis for DISPLAY45 (version 1.80)
Kamata, Y., Koyama, K., Tsuboi, Y., & Yamauchi, S. 1997, , 49, 461
Knude, J. & Hog, E. 1998, aap, 338, 897
Koyama, K., Maeda, Y., Ozaki, M., Ueno, S., Kamata, Y., Tawara, Y., Skinner, S., & Yamauchi, S., 1994, , 46, L125
Koyama, K., Hamaguchi, K., Ueno, S., Kobayashi, N., & Feigelson, E. D., 1996, , 48, L87
Leous, J. A., Feigelson, E. D., André, P., & Montmerle, T., 1991, apj, 379, 683
Makishima, K., [[et al.]{}]{} 1996, , 48, 171 Montmerle, T., Koch-Miramond, L., Falgarone, E., & Grindlay, J. E. 1983, , 269, 182
Montmerle, T., Grosso, N., Tsuboi, Y., & Koyama, K. 1999, , submitted
Motte, F., André, P., & Neri, R. 1998 , 336, 150
Neuhäuser, R., Preibisch, T. 1997 , 322, L37
Ohashi, T., Ebisawa, K., Fukazawa, Y., Hiyoshi, K., [[et al.]{}]{} 1996, , 48, 157
Ozawa, H., Nagase, F., Ueda, Y., Dotani, T., & Ishida, M. 1999, , 523, L81
Preibisch, T., Zinnecker, H., & Schmitt, J. H. M. M. 1993, , 279, L33
Preibisch, T., Neuhäuser, R., & Alcala, J. M. 1995, , 304, L13
van den Oord, G. H. J. & Mewe, R. 1989, , 213, 245
Serlemitsos, P. J., [[et al.]{}]{} 1995, , 47, 105
Shu, F. H., Adams, F. C. & Lizano, S. 1987, ARA&A, 25, 23
Stelzer, B., Neuhäser, R., Casanova, S., & Montmerle, T. 1999, , 344, 154
Stern, R. A., Uchida, Y., Tsuneta, S., & Nagase, F. 1992, , 400, 321
Tanaka, Y., Inoue, H., & Holt, S. S. 1994, , 46, L37
Tsuboi, Y., Koyama, K., Murakami, H., Hayashi, M., Skinner, & S., Ueno, S. , 503, 894
Wilking, B. A., & Lada, C. J. 1983, , 274, 698
APPENDIX\
Determination of the Physical Parameters of the Flares {#appendix .unnumbered}
======================================================
1. Loop Parameters {#VM .unnumbered}
------------------
We make use of the general treatment for solar-type flares, put forward by van den Oord & Mewe (1989, hereafter VM). The decrease in the thermal energy of the cooling plasma is assumed to be caused by radiative ($\tau_{\mathrm{r}}$) and conductive losses ($\tau_{\mathrm{c}}$): its decay time is thus $1/\tau_{\mathrm{eff}}=1/\tau_{\mathrm{r}}+1/\tau_{\mathrm{c}}$. VM assumed that the flare lightcurve and temperature decrease exponentially with decay times $\tau_{\mathrm{d}}$ and $\tau_{\mathrm{T}}$, respectively. This effective cooling time is related to observed time scales by $1/\tau_{\mathrm{eff}}=7/8\tau_{\mathrm{T}}+1/2\tau_{\mathrm{d}}$. We assume here that the flare occurs in only one semi-circular loop with constant section along the tube flux, radius $R$ ($R=L/\pi$; length $L$), diameter-to-length ratio $a$, and volume $V$. VM gives an expression of $L$ versus $a$, depending on $\tau_{\mathrm{eff}}$, the temperature, the emission measure (hereafter $EM$), and the ratio $\tau_{\mathrm{r}}/\tau_{\mathrm{c}}$. Because of the assumed exponential behavior of the lightcurve and temperature, the moment at which this expression is applied is not important. The only restriction is that both the temperature and the $EM$ started to decrease.
Due to the low statistics, we have only time-sliced values of the temperature and the $EM$. Let us call $t_i$ ($t_f$) the beginning (end) of the time interval within which the temperature was estimated, we have: $\overline{T} = \int_{t_{i}}^{t_{f}}
T(t^\prime)\,dt^\prime/(t_f-t_i)$. We use this relation to find the behavior of the temperature as a function of time.
We now have a relation between $L$ and $a$, but the ratio $\tau_{\mathrm{r}}/\tau_{\mathrm{c}}$ is unknown, and even worse may change during the decay of the flare. An exception to this is when the flare volume cools quasi-statically, evolving through a sequence of [*quasi-static equilibria*]{}, where the loop satisfies scaling laws, and where $\tau_{\mathrm{r}}/\tau_{\mathrm{c}}=cst$. Due to the dependency of $\tau_{\mathrm{r}}$ and $\tau_{\mathrm{c}}$ on the temperature and the $EM$, $\tau_{\mathrm{r}}/\tau_{\mathrm{c}}=cst$ implies $T^{13/4}/EM=cst$. VM gives in that case analytical expressions for several physical quantities versus time, all depending on the quasi-static radiative time scale ($\tau_{\mathrm{r,qs}}$). $\tau_{\mathrm{r,qs}}$ can be estimated from the lightcurve which must be proportional to the radiative loss. The temperature and the $EM$ can be fitted as described above using this value of $\tau_{\mathrm{r,qs}}$, and give the peak values $kT_{\mathrm{p,qs}}$, $EM_{\mathrm{p,qs}}$ (see Col. \[4\]–\[5\] in our Table \[fit\], panel 1 for details).
Using an expression for the radiative time (eq.\[23a\] of VM), the $EM$ (eq.\[3\] of VM) and the scaling law (SL\[2\], eq.\[196\] of VM), we obtained the loop characteristics for the quasi-static model: $$a=1.38 \times (\tau_{\mathrm{r,qs}}/10\,ks)^{-1/2} \times
(kT_{\mathrm{p,qs}}/keV)^{-33/16} \times
(EM_{\mathrm{p,qs}}/10^{54}\,cm^{-3})^{1/2},
\label{eq:a}$$ $$L=1.0\,R_\odot \times (\tau_{\mathrm{r,qs}}/10\,ks) \times
(kT_{\mathrm{p,qs}}/keV)^{7/8},$$ $$n_e=4.4 \times 10^{10}\,cm^{-3} \times (\tau_{\mathrm{r,qs}}/10\,ks)^{-1}
\times
(kT_{\mathrm{p,qs}}/keV)^{3/4}.
\label{eq:n_e}$$
Using an expression for the ratio $\tau_{\mathrm{r}}/\tau_{\mathrm{c}}$ on page 252 of VM, we found $\tau_{\mathrm{r}}/\tau_{\mathrm{c}}=\mu \times f$, with the parameter $\mu$ depending only on the exponent of the temperature in the expression for the radiative loss (for temperature above 20MK, $\mu=0.18$), and the multiplicative function $f$ coming from the expression for the mean conductive energy loss (formula \[7\] of VM), which is equal to 4/7 for a loop with a constant section. Thus, $\tau_{\mathrm{r}}/\tau_{\mathrm{c}}=0.1$. [^1] In other words, the quasi-static model implies that 91$\%$ of the lost energy are radiative losses: radiation is the dominant energy loss process.
Assuming that the cooling is only radiative we used these simplified relations based on the exponential decay of the lightcurve, the temperature, and the $EM$: $$n_e=4.4\times 10^{10}\,cm^{-3}\times (\tau_{\mathrm{d}}/10\,ks)^{-1} \times
(kT_{\mathrm{p}}/keV)^{3/4},\mathrm{~for~kT>2\,keV},$$ $$L=7.4\,R_\odot \times (a/0.07)^{-2/3}\times (\tau_{\mathrm{d}}/10\,ks)^{2/3}
\times
(kT_{\mathrm{p}}/keV)^{-1/2} \times (EM_{\mathrm{p}}/10^{54})^{1/3},$$ with $kT_{\mathrm{p}}$ ($EM_{\mathrm{p}}$) the peak value of the temperature ($EM$).
2. Magnetic Field {#magnetic-field .unnumbered}
-----------------
Assuming equipartition between the magnetic pressure $B^2/8\pi$ and the ionized gas pressure $2n_e kT$, we can obtain a minimum value of the magnetic field confining the emitting plasma using: $$B=28.4\,G \times
(n_e/10^{10}\,cm^{-3})^{1/2} \times (kT/keV)^{1/2}
\label{eq:B}$$
3. Released Energy {#released-energy .unnumbered}
------------------
For estimating the energy released by the flare during its cooling phase we need the peak luminosity value of this flare. As the lightcurve must be proportional to the intrinsic luminosity, we fit the time averaged intrinsic luminosities in the 0.1–100keV band (given in Table 1) with the same model used for the lightcurve fitting. This gives the peak luminosity, $L_{\mathrm{X,\,peak}}$, and a characteristic decay time $\tau$. Thus, the total energy released by this flare is: $$E_{\mathrm{tot}} \sim 10^{35}\,erg \times (L_{\mathrm{X,\,peak}}/10^{32}\,erg\,s^{-1}) \times
(\tau/ks)
\label{eq:E}$$
[^1]: VM wrote $\tau_{\mathrm{r}}/\tau_{\mathrm{c}}=\mu=0.18$, and the analytical expression for the conductive energy loss in the quasi-static model without taking this multiplicative factor $4/7$ into account (see Table 5 of VM).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
We present a new Bayesian approach to constrain the intrinsic parameters (stellar mass, age) of the eclipsing binary system –CEP0227– in the Large Magellanic Cloud. We computed several sets of evolutionary models covering a broad range in chemical compositions and in stellar mass. Independent sets of models were also constructed either by neglecting or by including a moderate convective core overshooting ($\beta_{ov}$=0.2) during central hydrogen burning phases. Sets of models were also constructed either by neglecting or by assuming a canonical ($\eta$=0.4,0.8) or an enhanced ($\eta$=4) mass loss rate. The most probable solutions were computed in three different planes: luminosity–temperature, mass–radius and gravity–temperature. By using the Bayes Factor, we found that the most probable solutions were obtained in the gravity–temperature plane with a Gaussian mass prior distribution. The evolutionary models constructed by assuming a moderate convective core overshooting ($\beta_{ov}$=0.2) and a canonical mass loss rate ($\eta$=0.4) give stellar masses for the primary (Cepheid) –M=4.14$^{+0.04}_{-0.05}$ M$_\odot$– and for the secondary –M=4.15$^{+0.04}_{-0.05}$ M$_\odot$– that agree at the 1% level with dynamical measurements. Moreover, we found ages for the two components and for the combined system –t=151$^{+4}_{-3}$ Myr– that agree at the 5% level. The solutions based on evolutionary models that neglect the mass loss attain similar parameters, while those ones based on models that either account for an enhanced mass loss or neglect convective core overshooting have lower Bayes Factors and larger confidence intervals. The dependence on the mass loss rate might be the consequence of the crude approximation we use to mimic this phenomenon.
By using the isochrone of the most probable solution and a Gaussian prior on the LMC distance, we found a true distance modulus –18.53$^{+0.02}_{-0.02}$ mag– and a reddening value –E(B-V)= 0.142$^{+0.005}_{-0.010}$ mag– that agree quite well with similar estimates in the literature.
author:
- 'P. G. Prada Moroni, M. Gennaro, G. Bono, G. Pietrzyński, W. Gieren, B. Pilecki, D. Graczyk, I. B. Thompson'
title: 'On the evolutionary and pulsation mass of Classical Cepheids: III. the case of the eclipsing binary Cepheid CEP0227 in the Large Magellanic Cloud'
---
Introduction
============
Classical Cepheids are very popular in the recent literature, since they play a crucial role to address several open astrophysical problems. The reason is threefold. [*i)*]{}– They are the most popular primary distance indicators, since they are bright and can be easily identified. [*ii)*]{}– They are excellent tracers of intermediate-mass stars, since their evolutionary status is well defined (central helium burning, blue loops). [*iii)*]{}– They are fundamental laboratories to constrain the micro (equation of state, opacity, cross sections) and macro (mass loss, mixing, convective transport) physics adopted in both evolutionary and pulsation models.
Dating back to the seminal investigation by @hof64 [and references therein] and by @bec79, the evolutionary properties of intermediate-mass have been investigated in a series of theoretical [@sto78; @chi86; @cas90; @lan91; @bon00; @mae00; @mey02; @kar03; @pie04; @pie06; @val09; @cha10; @cas11; @nei11; @bro11a; @bro11b] and empirical [@alc99; @tes99; @broc03; @hun09a; @hun09b; @bona09; @bona10; @san09; @efr11] investigations.
One of the most interesting problem concerning evolutionary and pulsation properties of classical Cepheids is the so-called mass discrepancy problem. During the late sixties it was noticed by @chr66 [@chr70] and by @sto69 that the evolutionary masses –estimated using the comparison between isochrones and observations in the Color-Magnitude Diagram (CMD)– were almost a factor of two larger than the pulsation masses –estimated using the Period-Mass-Radius relation– of Galactic Cepheids [@fri71].
This conundrum was partially solved by @mos92 using the sets of radiative opacities released by the OPAL [@igl91] and by the Opacity Project [@sea94] groups. However, several investigations focussed on Galactic [@bon01; @cap05] and Magellanic [@bea01; @bon02; @kel02; @kel06; @kel08] Cepheids suggested that such a discrepancy was still of the order of 10-20%. A similar discrepancy was also found by @eva05a using dynamical mass estimates of Galactic binary Cepheids and by @broc03 using cluster Cepheids located in NGC 1866, a young Large Magellanic Cloud (LMC) cluster.
The Cepheid mass discrepancy problem can be addressed following two different paths.\
[*a)*]{} –[*He-core mass*]{}– current evolutionary predictions underestimate the He-core mass for intermediate-mass structures, and in turn their Mass-Luminosity (ML) ratio. This possible drawback has a substantial impact both on evolutionary and pulsation predictions, since the latter models assume an ML relation.
[*b)*]{} –[*Envelope mass*]{}– Current Cepheid masses are smaller than their main sequence (MS) progenitors, because they have lost a fraction of their initial envelope mass. The latter working hypothesis implies that the Cepheid mass discrepancy is intrinsic, i.e., it is not caused by limits in the physical assumptions adopted in constructing evolutionary and pulsation models. The input physics adopted in constructing evolutionary models of intermediate-mass stars has already been addressed in several papers [@chi86; @broc93; @sto93; @sto96; @cas04; @ber09; @val09].
Here, we briefly mention the most relevant ones affecting the He-core mass and the envelope mass.
*a) He-core mass*
[*i)*]{} –Extra-mixing– Several detailed studies based on the comparison between predicted CMDs and Luminosity Functions of NGC1866, reached opposite conclusions in favor [@bar02] and against [@tes99; @broc03] the occurrence of mild convective core overshooting. The need for a mild overshooting was also suggested by @kel06 who investigated several young clusters in the LMC and in the Small Magellanic Cloud. More recently, it has also been suggested by @cor02 that the degree of overshooting might also depend on the metal abundance, namely it increases when metal abundance decreases. [*ii)*]{} –Rotation– Evolutionary models constructed by accounting for the effects of rotation predict both an increase in Helium core mass and an enhancement in the surface abundance of both Helium and Nitrogen [@mey02]. It has also been suggested that rotation might account for the significant changes in surface chemical compositions observed in Galactic and Magellanic supergiants [@kor05; @hun09a; @hun09b]. Moreover, recent theoretical [@mae00] and empirical [@ven99] investigations indicate that the efficiency of such a mechanism might depend on the initial metal abundance, i.e., it increases when metal abundance decreases. [*iii)*]{} –Radiative Opacity– A new set of radiative opacities has been recently computed by the Opacity Project [@bad05]. The difference between old and new opacities is at most of the order of 5-10% across the Z-bump (T$\approx$ 250,000 K). To account for the mass discrepancy the increase in the opacity should be almost a factor of two. This indicates that the adopted radiative opacities have a marginal impact on the mass discrepancy problem.\
*b) Envelope mass*
[*i)*]{} –Canonical Mass Loss– The total mass of an actual Cepheid is also affected by a decrease in the envelope mass. Evolutionary models accounting for the mass loss, during Hydrogen and Helium burning phases, by means of several semi-empirical relations [@rei75; @nie90] do not solve the Cepheid mass discrepancy problem. Plausible values for the free parameter –$\eta$– give mass loss rates that are too small. This is not surprising, since current semi-empirical relations are only based on scaling arguments and they are not rooted on a robust physical basis [@sch05]. Empirical mass loss estimates based on NIR and ultraviolet emission for Galactic [@mca86; @dea88; @mar10a] and Large Magellanic Cloud [@nei08; @nei09] Cepheids cover a broad range (10$^{-10}$ – 10$^{-7}$ M$_\odot$ yr$^{-1}$). [*ii)*]{} –Enhanced Mass Loss– More recent mid-infrared observations for 29 Galactic Cepheids collected with Spitzer indicate the presence of extended emission around seven of them [@barm11]. This finding together with the interferometric detection of circumstellar shells around several Galactic Cepheids [@mer06; @mer07; @ker06; @ker08; @ker09] further supports the evidence that there is a wind associated with Cepheids possibly triggered by pulsation-driven mass loss.
A new spin concerning the occurrence of circumstellar shells around Cepheids was recently given by the detection of a large nebula around $\delta$ Cephei [@mar10b], the protype of Cepheid variables. The IR emission was detected in several MIR bands and shows a radial extent larger than $10^4$ AU (5 arcmin). The nebula shows a parabolic shape and it is aligned with the direction of the motion, thus suggesting the possible occurrence of a bow shock caused by the interaction between the mass-losing Cepheid with the interstellar material. Current estimates indicate that $\delta$ Cephei may be losing mass with a rate of $\sim$5$\times$10$^{-9}$ to $\sim$6$\times$10$^{-8}$ $M_\odot$ yr$^{-1}$. This finding has been independently confirmed by @mat11 using data collected in the radio with the Extended Very Large Array. They found an outflow velocity of $\sim$ 36 km/s and a mass loss rate of $\sim$ 10$^{-7}$–10$^{-6}$ M$_\odot$ yr$^{-1}$. The modest dust content of the outflow indicates that the wind from $\delta$ Cephei might be pulsation-driven instead of dust-driven as typical for other classes of evolved stars.
However, the stepping stone concerning the evolutionary and pulsation properties of classical Cepheids was the detection of a Cepheid in a well detached, double-lined eclipsing binary system in the Large Magellanic Cloud. The geometry and the precision of both photometric and spectroscopic data gave the opportunity to measure the mass of the Cepheid (CEP0227) with the unprecedented precision of 1% [@pietr10 hereinafter Paper I]. A similar precision was also attained for the classical Cepheid in the LMC eclipsing binary CEP1812 by [@pietr11 hereinafter Paper II].
This finding paved the road for new constraints on pulsation and evolutionary masses. Soon after, @cas11 by using evolutionary models that account for a moderate convective core overshooting found evolutionary masses agree quite well with the dynamical mass of CEP0227 both in the Radius-Age plane and in the Radius-Effective Temperature plane. More recently, @nei11 found that a combination of both moderate convective core overshooting and pulsation-driven mass loss is required to solve the Cepheid mass discrepancy.
This is the third paper of a series focussed on the evolutionary and pulsation properties of Cepheids. The new series present two novel approaches when compared with similar investigations available in the literature: [*i)*]{}– the intrinsic parameters will be estimated using a Bayesian approach in dealing with the comparison between theory and observations; [*ii)*]{}– the empirical measurements are based on a homogeneous approach for both photometry and spectroscopy. In this investigation we provide new evolutionary constraints on the binary Cepheid CEP0227.
The plan of the paper is the following. In §2 we introduce the input physics adopted to construct evolutionary models. We also discuss the physical assumptions adopted to deal with mass loss, mixing and convective transport. In this section, we also outline the assumptions adopted to compute stellar isochrones. The description of the method adopted to estimate the intrinsic parameters of the binary components is given in §3. In this section we discuss the different most probable solutions in the L–T, in the M–R and in the g–T planes. §4 deals with the new empirical constraints on both stellar mass and age we obtained in the different planes for the binary system. In §5 we address the comparison between the most probable stellar isochrone and observations in the Color-Magnitude diagram. The true distance modulus and the reddening for the combined system are also discussed. In §6 we summarize the results and outline future possible avenues of the project.
The theoretical framework
=========================
Input physics
-------------
The evolutionary tracks were computed with an updated version of the FRANEC evolutionary code [@deg08; @val09; @tog11; @dell11]. For these models, the 2006 release of the OPAL equation of state (EOS) was adopted [@rog96], together with the radiative opacity tables released in 2005 by the same Livermore group [@igl96][^1] for $\log T[K] > 4.5$ and extended at lower temperatures with the radiative opacity tables by @fer05. The conductive opacities are from @sht06 [see also @pot99]. All the opacity tables were calculated for the @asp09 [hereafter As09] heavy-element solar mixture. The nuclear reaction rates are from the NACRE compilation [@nacre], except for the $^{14}$N(p,$\gamma$)$^{15}$O, which is from the LUNA collaboration [@imb05; @bem06; @lem06] and $^{12}$C($\alpha$,$\gamma$)$^{16}$O from @ham05. These two reactions [see, e.g., @val09 and references therein] affect both the extension and the morphology of the loop performed by intermediate-mass He-burning stars in the HR diagram. Hence the most recent values should always be adopted in stellar evolutionary codes in order to provide reliable theoretical predictions.
Current calculations did not take into account He and metal diffusion, since its effect on the evolutionary properties of intermediate-mass stars is minimal. To model super-adiabatic convection, we adopted the mixing-length formalism [@boh58], in which the efficiency of the convective transport depends on the mixing-length, i.e. $l=\alpha H_p$, where $H_p$ is the pressure height-scale and $\alpha$ is a free parameter. The $\alpha$ value –1.74– was calibrated using the Standard Solar Models computed with both the same code and the same input physics, in particular using the same heavy-element mixture (As09), To account for current theoretical and empirical uncertainties, additional models with $\alpha$=1.9 were also computed. We also account for convective core overshooting, during central hydrogen burning phases, by extending the central mixed region of $l_{ov}$=$\beta H_p$ from the border of the standard convective core defined by the Schwarzschild criterion, where $\beta$ is a free parameter [@cas00; @val09]. Note that we do not account for a mechanical core overshooting during the central He-burning phase. The use of the Schwarzschild criterion to determine the convective core extension, during these phases, would lead to an unphysical discontinuity in the radiative gradient at the border of the convective core. This effect is caused by the increase in the opacity due to the conversion of He into C and O [@cas04]. To overcome the problem we model the growth of the convective core and the development of a semiconvective region following @castellani71a [@castellani71b]. Note that we neglect the convective overshooting at the base of the convective envelope during shell hydrogen burning phases.
{width="70.00000%"} \[plotP\]
{width="99.00000%"} \[fig:ploISO\]
Evolutionary tracks
-------------------
The approach we plan to adopt to estimate the intrinsic parameters of the binary components does require detailed set of evolutionary tracks and of isochrones properly covering the helium burning phases. The stellar masses of the evolutionary tracks range from 3 to 5 $M_\odot$, and the step in stellar mass is 0.05 $M_\odot$. Therefore, each set includes 42 evolutionary tracks. Unfortunately, we still lack an accurate measurement of the iron and heavy element abundances of CEP0227. Therefore, we adopted the mean metallicity of LMC Cepheids (\[Fe/H\]=-0.33$\pm$0.13 dex) based on recent high-resolution, high signal-to-noise spectra collected with UVES at ESO/VLT [@roma08]. The top, left panel of Fig. 1 shows the set of evolutionary models computed at fixed helium (Y=0.260) and metal (Z=0.006) content and stellar masses ranging from 3 to 5 $M_\odot$. The evolutionary phases plotted in this panel cover the central helium burning evolutionary phases (blue loops).
To take into account possible changes in iron abundance, due to intrinsic metallicity dispersion [@pom08; @roma08; @hun09a; @hun09b] we also computed four independent sets of evolutionary tracks with metallicity ranging from Z=0.004 to Z=0.008, while the helium content was fixed assuming an helium-to-metal enrichment ratio $\Delta Y$/$\Delta Z$=2 [@pag98; @jim03; @fly04; @gen10] and a primordial helium abundance of $Y_p$=0.2485 [see, e.g., @cyb05; @ste06; @pei07; @kom11]. Therefore, current helium values range from Y=0.256 to Y=0.265 (see top, right panel of Fig. 1).
We already mentioned that intermediate-mass stars are affected by uncertainties on the actual size of the He-core mass. To constrain the impact of the convective core overshooting during central hydrogen burning phases on the intrinsic parameters, the evolutionary tracks were computed either neglecting ($\beta_{ov}$=0) or accounting for a moderate convective core overshooting ($\beta_{ov}$=0.2, see top panels of Fig. 1). The two components of the binary system are relatively cool objects (see Table 1) with extended convective envelopes. To constrain the dependence on the efficiency of convective transport in the super-adiabatic regions the calculations were performed assuming two different mixing length parameters. The former one –$\alpha$=1.74– calibrated using the standard solar model [@dell11] and the latter one –$\alpha$=1.90– to mimic a more efficient convection (see middle panels of Fig. 1).
Recent theoretical investigations based on evolutionary models accounting for both convective core overshooting and for a pulsation driven mass loss suggested that the latter physical mechanism can play a crucial role in settling the Cepheid mass discrepancy problem [@nei11]. A detailed analysis of the dependence of the blue loops on this physical mechanism is out of the aim of this investigation. However, to constrain the dependence of the most probable solution on the mass loss rate, different sets of evolutionary models were computed accounting for mass loss after the central hydrogen exhaustion. We used the Reimers formula
$dM/dt$ = $\eta$ $\times$4 $\times$ $10^{-13}$ $\times$ L/gR \[$M_\odot$/yr\] where $\eta$ is a free parameter, L,R and g are the luminosity, the radius, and the surface gravity in solar units. Sets of models with four different $\eta$ values, namely 0, 0.4, 0.8 and 4.0, were computed. The evolutionary tracks constructed by assuming $\eta$=0.4 and 0.8 show marginal changes and in the following we will only take into account the former ones (see bottom panels of Fig. 1).
We ended up with a sample of more than 2,000 evolutionary tracks computed for this specific project.
Stellar isochrones
------------------
For each set of evolutionary tracks discussed in §2.2, we also computed a fine grid of stellar isochrones with ages ranging from a few tens to a few hundreds of Myr. To improve the accuracy of the parameter estimates (see §\[sec:parest\]) and to provide robust determinations of the confidence interval, the isochrones were further interpolated. The actual grid of isochrones has a step in age of 0.02 Myr, necessary to cover the fast evolution of the stars across the blue loop.
Subsets of isochrones for selected input parameters (see labeled values) in the HR diagram are shown in Fig.\[fig:ploISO\].
Parameter estimates {#sec:parest}
===================
The parameters that determine the evolutionary status of a star are its mass, $\mu$, its age, $\tau$ and its chemical composition, $\zeta$. The value of these parameters determine the position of a star in the HR diagram or, similarly, in any diagram were observable properties are displayed. Stellar models predict observables as a function of these three parameters, hence their values can be estimated by comparing predictions and observations. To provide robust estimates of both stellar masses and ages of the binary components we followed the Bayesian approach described in @gen12. The interested reader is referred to the quoted paper, here we only summarize the relevant points of the approach we adopted.
Bayesian approach
-----------------
Given a set of observables $\vec{q}$ the probability of the model parameters, $\vec{p} = (\mu,\tau,\zeta)$ is given by:
$$f( \vec{p} | \vec{q} ) \propto \mathcal{L}(\vec{p}|\vec{q}) f(\vec{p}) .$$
Here $\mathcal{L}(\vec{p}|\vec{q})$ is the Likelihood of the parameters $\vec{p}$ given the set of observations $\vec{q}$ and $f(\vec{p})$ is the prior distribution of the parameters.
We used three different pairs of observables, $\vec{q}$, to determine the $(\mu,\tau,\zeta)$ set of parameters of the binary components. In particular, we performed the comparison in the $\vec{q} = (\log L/L_\odot, \log T_{\mathrm{eff}}),
(M/M_\odot, R/R_\odot)$ and $(\log T_{\mathrm{eff}}, \log g)$ planes.
To estimate the best values of the stellar masses and ages the probability $f( \vec{p} | \vec{q} )$ is marginalized w.r.t the other two variables: $$\begin{aligned}
\label{eq:Gtau}
G(\tau) & = & \int \mathcal{L}(\tau,\mu,\zeta|\vec{q}) f(\tau,\mu,\zeta)\, \mathrm{d}\mu \,\mathrm{d}\zeta \\
\label{eq:Hmu}
H(\mu) & = & \int \mathcal{L}(\tau,\mu,\zeta|\vec{q}) f(\tau,\mu,\zeta)\, \mathrm{d}\tau \, \mathrm{d}\zeta\end{aligned}$$
The mode of the marginal distributions is used as the best parameter estimator [@jor05; @gen12]. Confidence intervals are defined by excluding 16% of the area under the posterior distribution on each side of the variable domains.
In our treatment the metallicity is not considered as an unknown parameter to be determined. On the contrary, we adopted the available prior information on the metallicity of Cepheids in the LMC to constrain our predictions on the stellar masses and ages. In detail, we adopted a Gaussian distribution of \[Fe/H\] with mean $<\mathrm{[Fe/H]}> = -0.33$ and standard deviation $\sigma_{\mathrm{[Fe/H]}} = 0.13$ dex. The quoted values are based on the metallicity distribution of LMC Cepheids measured by @roma08.
Given the available information on the dynamical stellar masses we adopted two different priors for the stellar mass. In one case a flat prior is adopted, defined across the entire range of stellar masses for which we computed evolutionary tracks: $M\in[3,5]M_{\odot}$. In the second case we used a prior with a Gaussian distribution with mean and standard deviations based on the dynamical mass estimates by @pietr10.
Stellar ages were estimated separately for the two binary components and for the system as a whole. The age of each star is determined from the $G(\tau)$ marginal distribution of equation (\[eq:Gtau\]). In the reasonable hypothesis that the two components of the binary system have the same age, their composite age distribution, i.e. the age of the system, can be simply derived as: $$G_{\mathrm{C}} = G_{\mathrm{P}} \times G_{\mathrm{S}} \; ;$$ where the subscripts indicate the composite (C), the primary component (P) and the secondary component (S) age distributions. In a recent investigation, @gen12 showed that the age distribution of a stellar system based on this approach is a more precise indicator of the real age when compared to single stellar ages.
The above approach was applied to each set of evolutionary models listed in Table 1, i.e. for each different assumption concerning the value of the overshooting parameter –$\beta$– the mass-loss –$\eta$– and the mixing length –$\alpha$– parameter. The posterior probabilities were computed using both the flat and the Gaussian prior on the stellar mass. Selected results of the individual solutions are given in Table 3. The solutions based on different observables are discussed in the following subsections.
The Hertzsprung-Russell Diagram
-------------------------------
As a first step to constrain the intrinsic parameters of the binary system, we applied our method in the Hertzsprung-Russell Diagram. The main advantage in using this plane is that we can estimate the intrinsic luminosities using the Stefan-Boltzmann relation. Therefore, the solution is independent of distance modulus, reddening and stellar mass. The main drawback is that the two parameters (luminosity and temperature) are correlated and more affected by measurement errors than the stellar mass and the radius (see Fig. 3). However, our method takes into account the fact that temperature and luminosity are not independent, and indeed the covariance matrix is also included in the solution [see equations (3), (4) and (5) in @gen12].
{width="99.00000%"} \[plotP\]
.
The stellar mass vs radius plane
--------------------------------
To further constrain the intrinsic parameters of the binary system we also applied the Bayesian method in the stellar mass vs radius plane. The main advantage in using this plane is that the adopted observables are independent of distance modulus and reddening and they are also independent from each other (see Fig. 4).
{width="99.00000%"} \[plotP\]
The $\log T_{eff}$ vs $\log g$ plane
------------------------------------
{width="99.00000%"} \[plotP\]
.
Finally, we applied the Bayesian method also into the effective temperature vs surface gravity plane. The main advantage in using this plane is that the adopted observables are independent of distance modulus and reddening. Moreover, the observables are independent from each other and the isochrones display a relevant difference in this age range (see Fig. 5). This plane combines three independent measurements, thus providing the most stringent test for the theory.
Constraints on the stellar mass and age of the binary system
============================================================
The most probable mass and age values that we found indicate that the mixing length parameter has a minimal impact on the estimate of these parameters. This result is not surprising, since the extent in effective temperature of the blue loop is minimally affected by the change in $\alpha$ from 1.74 to 1.90. This is the reason why in Table 3 we only included the results based on models with $\alpha$=1.74. What is clear from the values listed in Table 3 is that the solutions based on evolutionary models that do not account for the convective core overshooting and those with enhanced mass-loss ($\eta=4.0$) are in worse agreement with the measured dynamical masses, when compared with the results based on models accounting for the convective core overshooting and $\eta=0.0$ or 0.4.
In the case of a flat mass prior, the $\beta =0.0$ models predict stellar masses that are systematically larger than observed, even though the difference with the dynamical mass is typically within the 1$\sigma$ confidence interval. Similarly, the models with enhanced mass-loss, in the case of flat mass prior, tend to underestimate the mass of the primary component. These two sets of models give a good agreement with the data, only when the method is used in the M–R plane.
Comparing the results for a flat and a Gaussian mass prior, it is clear that the latter drives the solution towards the measured dynamical mass. While this might appear obvious, we will show in Sect. \[sec:bf\] that not only the mode of the mass distributions are shifted towards the measured values, but also the whole solution has a higher reliability, measured using the so-called $\mathrm{evidence}$ parameter (see §4.1). The impact of the Gaussian prior is very important also for the age results. In the Gaussian mass prior case, the age solutions for the two components are generally more precise, meaning that the uncertainty interval is narrower compared to the flat prior case. In addition to this, there is also a better agreement between the ages of the two components, and therefore, a more robust result for the composite age of the system.
For each set of models, we plotted the isochrones corresponding to the most probable composite age (indicated by a C in Table 3) in Figs. 3, 4 and 5. The three figures show the most probable isochrone as obtained by applying the method in the L–T, in the M–R and in the g–T plane. Solid and dashed lines display the most probable ages for a flat and a Gaussian mass prior, respectively.
In passing, we note that evolutionary models constructed assuming a slightly enhanced mass loss rate ($\eta$=0.8) also provide very similar solutions to the canonical ($\eta$=0.4) case. This indicates that canonical recipes to account for mass loss rate do not allow us to provide more quantitative constraints on mass loss efficiency during advanced evolutionary phases [@nei11]. However, a detailed treatment of the pulsation-driven mass loss inside the instability strip is out of the aim of this investigation.
$\mathrm{Evidence}$ of the most probable solutions {#sec:bf}
--------------------------------------------------
The confidence intervals of the different solutions provide a hint concerning the robustness of individual fits using the different observables. However the precision of the result alone migth be a deceptive indicator of the goodness of the fit between theory and data. A narrower confidence interval in the $(\mu,\tau)$ parameter space does not necessarily imply that the corresponding choice either of the observables, or of the mass prior or of the set of models is better than other possible choices.
A quantitative measurement of the relative goodness of the different solutions can be provided by the Bayes Factor. The Bayes Factor, $\mathrm{BF}_{ij}$, between two sets of models is defined as the ratio of the $\mathrm{evidence}$ of the two models. The $\mathrm{evidence}$ itself is the integral of the Likelihood, marginalized over the model parameters prior distribution [see, e.g., @bj11]. The $\mathrm{evidence}$ gives the total probability of the observed data given the particular set of models considered. Since our sets of models differ for the choice of the mixing length parameter, $\alpha$, the core overshooting parameter, $\beta$, and the mass-loss parameter, $\eta$, we introduce $\Xi_i = (\alpha_i,\beta_i,\eta_i)$ to identify a particular choice of the three parameters. The $\mathrm{evidence}$ is therefore: $$\label{eq:evid}
\mathcal{E}_i(\vec{q}) \equiv \mathcal{E}(\vec{q}|\Xi_i) = \int \mathcal{L}(\tau,\mu,\zeta|\vec{q},\Xi) f(\tau,\mu,\zeta|\Xi) \mathrm{d}\tau \, \mathrm{d}\mu \, \mathrm{d}\zeta$$
The $\mathrm{evidence}$ depends on the plane in which the comparison between data and models is performed (L–T, M–R, g–T), on the choice of the prior distribution of the model parameters (i.e., $f(\tau,\mu,\zeta|\Xi)$, in our case flat or Gaussian mass distribution and Gaussian metallicity distribution) and on the $\Xi$-set of parameters identifying that set of models (see Fig. 6). $\mathcal{E}_i(\vec{q})$ is proportional to the total probability of observing the data $\vec{q}$ under the assumption that the $\Xi_i$ model is the correct model.
In equation (\[eq:evid\]) the $\vec{q}$ observables refer to a single star. The composite $\mathrm{evidence}$ for both stars is simply: $$\mathcal{E}_i( \{ \vec{ q_{ \mathrm{P} } }, \vec{ q_{ \mathrm{S} } } \} ) = \mathcal{E}_i( \vec{ q_{\mathrm{P}} } ) \times \mathcal{E}_i( \vec{ q_{\mathrm{S}} } )$$
From equation (\[eq:evid\]), it follows that the Bayes factor is: $$\mathrm{BF}_{ij} = \frac{\mathcal{E}_i}{\mathcal{E}_j}$$
The Bayes factor does not constrain the triplet $(\tau,\mu,\zeta)$ that provides the highest posterior probability for each set of models. However, it can be used to quantify which set of models, variable plane and prior distribution choice give the best overall agreement with the data. Typically, in order to claim that a set is preferred over another, the Bayes factor should be smaller/larger than one, with either BF $< 0.1$ or BF $> 10$ [@kr95].
Table 4 gives the BF values using as a reference the set of models with $\alpha=1.74$, $\beta=0.2$ and $\eta=0.4$, together with the variables $\vec{q} = (\log g, \log T_{\mathrm{eff}})$ and with the Gaussian mass prior distribution. This reference set gives the largest $\mathrm{evidence}$ for the composite system, therefore it can be considered –within our range of models– the best set of models and the best prior distribution choice to describe the observations. The BFs listed in Table 4 were estimated by dividing the corresponding $\mathrm{evidence}$ by the $\mathrm{evidence}$ of the reference set either for the primary star (P entries), or for the secondary star (S entries) or for the composite $\mathrm{evidence}$ (C entries).
The ($\beta=0.2, \eta=0.0$) set of models gives BFs for the primary (see Fig. 6), the secondary (see Fig. 7) and the composite system that are very similar to the reference set. Therefore, these two sets give similar likely solutions and with the present data we can not distinguish between them. On the contrary, the set without overshooting, $\beta=0$, and the set with enhanced mass-loss rate, $\eta=4.0$, show very low values of the BFs, suggesting that these parameters can be excluded, hence, confirming the qualitative analysis at the beginning of this section.
It is worth noticing that, the BFs for a given $\vec{q}$ plane and a given $\Xi$ set are generally larger when the Gaussian mass prior is used. This suggests that the likelihoods themselves are already centered in a region of the parameter space which is close to the dynamical mass measurements. Therefore, the Gaussian prior enhances the posterior probabilities when compared with the flat case. This finding is also supported by the fact that several solutions obtained using a flat mass prior distribution are characterized by large confidence intervals (see top panels in Figs. 6 and 7 and the values listed in Table 3). In addition to this, the plane in which the largest BFs are found is the g–T plane. This means that the models, when the measurements of stellar mass, radius and effective temperature are combined, are still able to reproduce these three independent observables, and actually the total probability increases.
$
\begin{array}{ccc}
\includegraphics[width=0.32\textwidth]{f6_Ia.eps} &
\includegraphics[width=0.32\textwidth]{f6_Ib.eps} &
\includegraphics[width=0.32\textwidth]{f6_Ic.eps} \\
\includegraphics[width=0.32\textwidth]{f6_IIa.eps} &
\includegraphics[width=0.32\textwidth]{f6_IIb.eps} &
\includegraphics[width=0.32\textwidth]{f6_IIc.eps} \\
\label{plotP}
\end{array}$
$
\begin{array}{ccc}
\includegraphics[width=0.32\textwidth]{f7_Ia.eps} &
\includegraphics[width=0.32\textwidth]{f7_Ib.eps} &
\includegraphics[width=0.32\textwidth]{f7_Ic.eps} \\
\includegraphics[width=0.32\textwidth]{f7_IIa.eps} &
\includegraphics[width=0.32\textwidth]{f7_IIb.eps} &
\includegraphics[width=0.32\textwidth]{f7_IIc.eps}
\label{plotS}
\end{array}$
Finally, we mention that the presence of two different regions in all the probability diagrams for the secondary component (see Fig. 7) is the consequence of an intrinsic evolutionary feature during the blue loop (helium burning phases). Intermediate-mass stars soon after helium ignition (tip of the red giant branch) move, at fixed stellar mass, towards higher effective temperatures (the so-called blue tip) and then back to lower effective temperatures [@bon00 and references therein]. During these phases stellar structures with similar masses can be found in similar positions at different ages. However, the contribution of the two probability-regions to the marginal distribution is very different, with the main region contributing a much larger probability than the secondary one. The latter region only causes an increase in the skewness of the marginal distributions, without manifestation of secondary maxima.
The color-magnitude diagram
===========================
{width=".99\columnwidth"} \[plotP\]
The binary system CEP0227 offers a unique opportunity to test the plausibility of the physical assumptions adopted in evolutionary and pulsation models because the comparison between theory and observations can be performed by using observables (luminosity, radius, effective temperature, surface gravity) directly predicted by evolutionary models. This means that these observables do not need to be transformed into the observational plane. However, the Cepheid mass discrepancy problem arises from the comparison between evolutionary prescriptions and observations in the CMD. The main advantage in using this plane to estimate the evolutionary mass is that the estimate only relies on two mean magnitudes in different photometric bands. The main drawback is that the mass estimates are affected by uncertainties in the true distance modulus and in the reddening. To overcome the latter problem the mass estimates have also been performed in CMDs based on NIR bands [@bon01].
Absolute distance determination and NIR magnitudes for the binary CEP0227 are not available yet. We have already collected new NIR time series data and we plan to provide a detailed analysis of the CMD and of the color-color plane in a forthcoming investigations.
However, the solutions based on theoretical observables do allow us to constrain the possible occurrence of thorny systematic uncertainties in the bolometric corrections and in the color-temperature relations adopted to transform theoretical predictions into the observational plane. Moreover, we can also estimate the reddening of the system by assuming as a prior a Gaussian distribution on the true distance modulus. We used the theoretical solution which gives the largest $\mathrm{evidence}$ for the composite system, i.e. the 151 Myr isochrone for the $\beta=0.2, \eta=0.4$ set. This is the most probable age for the composite $G(\tau)$ in the case of the g–T plane. We transformed the isochrone into the Johnson–Kron–Cousins photometric system using the spectra provided by @ck03 based on the ATLAS9 model atmospheres (see Fig. 8). We also performed a test by using the spectra based on the atmosphere models provided by the PHOENIX code in the version by @bro05. We found that the difference –in the typical temperature, metallicity and surface gravity regime of the binary we are dealing with– is negligible, therefore in the following we adopt the evolutionary models transformed using the ATLAS9 models.
The comparison into the observational plane (CMD) between the isochrone of the most probable solution and observations provides independent estimates of both the reddening and the true distance modulus of the binary system. The reddening in the V,I-bands was fixed by adopting the @car89 extinction law. The solution was found by maximizing the probability:
$$\begin{aligned}
\label{eq:fitdmebv}
p(E(B-V),DM ) = \prod_{i = P,S} \frac{1}{2\pi\sigma_{V_i}\sigma_{(V-I)_i}} \times \\ \nonumber
\int \exp \left[ -\frac{\chi_{V_i}^2 + \chi_{(V-I)_i}^2 }{2} - \frac{(DM-DM_0)^2}{2\sigma^2_{DM_0}} \right] \mathrm{d}m
$$
where: $$\chi_{V,i} = \frac{ V_i - V(m; E(B-V), DM ) }{ \sigma_{V_i} } $$ and $$\chi_{(V-I)_i} = \frac{ (V-I)_i - (V-I)(m; E(B-V), DM ) }{ \sigma_{(V-I)_i} } $$ The index $i$ runs on the primary and on the secondary star $(P,S)$, $V_i$ and $(V-I)_i$ are the observed mean magnitudes and colors with their errors $\sigma_{V_i}$ and $\sigma_{(V-I)_i}$. The theoretical quantities are function of the mass along the isochrone and of the adopted distance modulus and reddening. The integral is performed along the isochrone integrating over the mass of the models. Given the available information on the LMC distance modulus, we used a Gaussian prior with mean and standard deviations given by $\mu = 18.45\pm0.04$ mag [@stor11].
Fig. 8 shows the 151 Myr isochrone in the V, V-I CMD, with a reddening of 0.142$^{+0.005}_{-0.010}$ mag and a distance of 18.53$^{+0.02}_{-0.02}$ mag as found by maximizing equation (\[eq:fitdmebv\]). This estimate of the reddening agrees very well with the reddening estimates based on Schlegel’s map, namely 0.15$\pm$0.03 mag [@sch98]. The same outcome applies to the LMC distance, since it agrees with similar estimates available in the literature [@stor11 and references therein]. This evidence is also supported by the fact that the plane of the LMC is also tilted along the line of sight and we are not applying any correction for the position of the binary system [@per04]. In order to constrain the sensitivity of both distance modulus and reddening on the prior we also adopted a flat prior between $\mu$=18 and 19 mag. Note that the adopted interval is significantly larger than the range of LMC distances indicated in literature and we found a true distance modulus of $\mu$=$18.57_{-0.01}^{+0.13}$ mag and a reddening of E(B-V) = $0.1275_{-0.07}^{+0.0025}$ mag. The new estimates have larger confidence intervals, but they agree quite well with the estimates based on the Gaussian prior. The above findings and the robustness of the most probable solution (see Fig. 9) provide an independent support to the current set of atmosphere models adopted to transform theory into the observational plane [@ck03].
{width=".99\columnwidth"} \[plotlast\]
Summary and final remarks
=========================
The discovery by OGLEIII of eclipsing binary systems in the Large Magellanic Cloud hosting classical Cepheids provided the unique opportunity of measure their masses with a precision ranging from 1% (CEP0227) to 1.5% (CEP1812, paper II). It goes without saying that the new measurements will allow us to constrain the precision and the plausibility of the the input physics and physical assumptions adopted in constructing both evolutionary and pulsation models. In this investigation, we addressed the evolutionary mass and the age of the binary system CEP0227. We adapted the general Bayesian approach developed by @gen12 to constrain the physical parameters of the binary components.
In order to provide robust statistical solutions, we computed several sets of evolutionary models by covering a broad range of chemical compositions (see Table 1) and stellar masses. To constrain the dependence on the free parameters currently adopted in canonical evolutionary models, independent sets of models were constructed by assuming two different values of the mixing length parameter ($\alpha$=1.74, 1.90). Moreover, to account for the possible occurrence of extra-mixing during central hydrogen burning phases, the models were constructed either by neglecting or by including a moderate convective core overshooting ($\beta_{ov}$=0.2). Finally, to account for the mass loss during evolved evolutionary phases the models were constructed either by neglecting or by including mass loss. To mimic different efficiencies in the mass loss rate, we constructed sets of models accounting either for canonical ($\eta$=0.4,0.8) or for enhanced ($\eta$=4) mass loss rates (see Table 1).
To overcome possible deceptive errors in the estimate of the physical parameters, the comparison between theory and observations was performed in three different planes: luminosity vs effective temperature ($\log$L, $\log$$T_{eff}$), mass vs radius (M,R) and surface gravity vs effective temperature ($\log$g, $\log$$T_{eff}$). We computed an independent solution (mass, age, chemical composition) in the three quoted planes for every set of evolutionary models we have constructed. Note that individual solutions were computed by adopting two different priors for the stellar mass. We adopted either a flat distribution over the entire range of stellar masses covered by computations or a Gaussian distribution anchored to the mean and the standard deviation of the dynamical mass measurements (paper I). We also adopted a Gaussian prior for the metallicity, in particular we adopted the mean and the standard deviation for LMC Cepheids recently provided by @roma08 by using high-resolution, high signal to noise spectra.
The main results we obtained in the comparison between theory and observations are the following:
- The most probable solutions are minimally affected by the adopted mixing length parameter, within the range covered by our models.
- The most probable solutions based on a Gaussian mass prior are typically more accurate than those based on the flat mass prior.
- The most stringent test to constrain the models were obtained in the g–T plane, due to the fact that the adopted parameters involve three independent observables, namely mass, radius and temperature. In this plane, the most probable solutions were obtained by using the evolutionary models that account for a moderate convective overshooting ($\beta_{ov}$=0.2) and canonical mass loss rate ($\eta$=0.4). In particular, we found that the mass of the primary (Cepheid) is M=4.14$^{+0.04}_{-0.05}$ M$_\odot$, while for the secondary is M=4.15$^{+0.04}_{-0.05}$ M$_\odot$. Note that current estimates do agree at the 1% level with dynamical measurements. We found ages for the individual components and for the combined system (t=151$^{+4}_{3}$ Myr) do agree at the 5% level, supporting once again the robustness of the solution. It is noteworthy that similar solutions are obtained by using moderate convective overshooting ($\beta_{ov}$=0.2) and no mass loss ($\eta$=0). This indicates that either a vanishing or a canonical mass loss rate play a marginal role in the Cepheid mass discrepancy. Current findings support the preliminary results concerning the evolutionary mass of CEP0227 provided by @cas11. They also suggest that evolutionary masses based on evolutionary models accounting for moderate convective core overshooting –a typical assumption in the comparison between theory and observations [@pie06]– agree quite well with the dynamical mass of Cep0227. The overall agreement among dynamical, evolutionary and pulsation masses will be addressed in a forthcoming paper in which we plan to constrain the pulsation mass using independent observables.
- The most probable solutions based on the models constructed by assuming a moderate convective overshooting ($\beta_{ov}$=0.2) and an enhanced mass loss rate ($\eta$=4) present large confidence interval when compared with the solutions based on canonical mass loss rate (see Table 3). Note that current finding should be cautiously treated. The reason is twofold. The difference between the estimated masses and the dynamical masses, taken at face value, do agree at the 1% level. We are assuming a very crude approximation to account for the mass loss, while current empirical and theoretical findings indicate that the mass loss might be driven by the pulsation instability [@nei11; @mat11]. This hypothesis is further supported by the recent results obtained by @neilson2012. They found that the observed period change of Polaris implies a mass loss rate of $\approx$ 10$^{-6}$ M$_\odot$ yr$^{-1}$, and in turn a new firm empirical argument for enhanced mass loss in Cepheids.
The above discussion concerning the formal precision of the results relies on the width of the confidence interval. However, a solution characterized by a narrow confidence interval might not be the most probable solution. The Bayes Factor between two sets of models is a more robust statistical parameter to constrain on a quantitative basis the most probable solutions.
By using the Bayes Factor the main results we obtained in the comparison between theory and observations are the following:
- The solutions with the largest $\mathrm{evidence}$ were obtained in the g–T plane by using sets of models that account for a moderate convective core overshooting ($\beta_{ov}$=0.2) and either a vanishing ($\eta$=0) or a canonical mass loss rate ($\eta$=0.4) give very similar results. This finding supports the results based on the confidence interval.
- The solutions based on sets of models that either neglect convective core overshooting or account for an enhanced mass loss rate ($\eta$=4) are characterized by very low Bayes Factors. Thus suggesting that these two sets of models give a poorer description of the observed properties of the binary system.
We also transformed the isochrone of the most probable solution ($\beta$=0.2, $\eta$=0.4, g–T plane, Gaussian mass prior) mentioned above into the observational plane using the atmosphere models provided by @ck03. By assuming a Gaussian metallicity prior distribution for the Cepheid metallicity and by assuming the Cardelli et al. (1989) reddening law, we found a true distance modulus –18.53$^{+0.02}_{-0.02}$ mag– and a reddening value –E(B-V)= 0.142$^{+0.005}_{-0.010}$ mag– that agree quite well with similar estimates available in the literature [@sch98; @per04; @stor11]. The use of a flat prior on the LMC distance, minimally affects these results. A more detailed discussion concerning new independent distances of the binary system and of the Cepheid will be addressed in a forthcoming investigation.
Intrinsic variables in eclipsing binary systems are a fundamental laboratory to constrain the input physics of both evolutionary and pulsation models. This was considered till a few years ago a broken ground for quantitative astrophysics, due to the limited number of known systems and the lack of homogeneous and precise photometric and spectroscopic data. The unprecedented photometric catalogs released by microlensing experiments (EROS, OGLEIII) together with currently underway optical (PanSTARRS-1, @stubbs; Palomar Transient Factory, @vane11; CHASE, @foerster) and NIR (VVV, @minniti) surveys and the spectrographs available at 8m class telescopes, are a very good viaticum in waiting for GAIA.
It is a pleasure to thank F. Caputo for many useful discussions concerning evolutionary and pulsation properties of classical Cepheids and M. Dell’Omodarme for his invaluable help in computer programming. We also acknowledge an anonymous referee for his/her positive opinion concerning the content of the paper and for his/her sharp and pertinent comments that helped us to improve the content and the readability of the manuscript. We gratefully acknowledge financial support for this work from the Chilean Center for Astrophysics FONDAP 15010003 and the BASAL Centro de Astrofisica y Tecnologias Afines (CATA) PFB-06/2007. Support from the Polish grants N203 387337, the IDEAS Plus, the FOCUS and the TEAM subsidies of the Fundation for Polish Science (FNP) are also acknowledged. During the preparation of this paper M. Gennaro has been a member and would like to acknowledge the “International Max Planck Research School for Astronomy and Cosmic Physics” at the University of Heidelberg, IMPRS-HD, Germany.
Alcock, C. et al. 1999, AJ, 117, 920
Angulo C., Arnould, M., Rayet, M., et al. 1999, Nuclear Physics A, 656, 3
Asplund, M.; Grevesse, N.; Sauval, A. J.; & Scott, P. 2009, ARA&A, 47, 481
Badnell, N. R. et al. 2005, MNRAS, 360, 458
Bailer-Jones, C. A. L. 2011, MNRAS, 416, 1163
Barmby, P.; Marengo, M.; Evans, N. R.; Bono, G.; Huelsman, D.; Su, K. Y. L.; Welch, D. L., & Fazio, G. G. 2011, AJ, 141, 42
Barmina, R.; Girardi, L.; Chiosi, C. 2002, A&A, 385, 847
Beaulieu, J. P., Buchler, J. R., Kollath, Z. 2001, A&A, 373, 164
Becker, S. A., & Iben, I., Jr. 1979, , 232, 831
Becker, S. A.; Iben, I., Jr.; Tuggle, R. S. 1977, , 218, 633
Bemmerer, D., Confortola, F., Lemut, A., et al. 2006, Nuclear Physics A 779, 297
Bertelli, G.; Nasi, E.; Girardi, L.; Marigo, P. 2009, A&A, 508, 355
B$\rm{\ddot{o}}$hm-Vitense, E. 1958, Zs.f.Ap., 46, 108
Bonanos, A. Z. et al. 2009, AJ, 138, 1003
Bonanos, A. Z. et al. 2010, AJ, 140, 416
Bono, G., Caputo, F., Cassisi, S., Marconi, M., Piersanti, L., & Tornambe, A. 2000, , 543, 955
Bono, G. Gieren, W. P., Marconi, M., Fouqué, P., & Caputo, F. 2001, ApJ, 563, 319
Bono, G. et al. 2002, ApJ, 565, L83
Brocato, E. & Castellani, V. 1993, ApJ, 410, 99
Brocato, E.; Castellani, V.; Di Carlo, E.; Raimondo, G.; & Walker, A. R. 2003, AJ, 125, 3111
Brott, I.; & Hauschildt, P. H. 2005, in ESA Special Publication, Vol. 576, The Three-Dimensional Universe with Gaia, ed. C. Turon, K. S. O’Flaherty, & M. A. C. Perryman, 565
Brott, I. et al. 2011a, A&A, 530, 115
Brott, I. et al. 2011b, A&A, 530, 116
Caputo, F. et al. 2005, ApJ, 629, 1021
Cardelli, J. A., Clayton, G. C., & Mathis, J. S. 1989, , 345, 245
Cassisi, S. 2004, in IAU Colloq. 193, Variable Stars in the Local Group, ed. D.W. Kurtz & K.R. Pollard (San Francisco: ASP), 489
Cassisi, S., & Salaris, M. 2011, ApJ, 728, L43
Castellani, V., Chieffi, A. & Straniero, O. 1990, ApJS, 74, 463
Castellani, V., Degl’Innocenti, S., Girardi, L., Marconi, M., Prada Moroni, P. G., & Weiss, A. 2000, A&A, 354, 150
Castellani, V., Giannone, P., & Renzini, A. 1971a, Ap&SS, 10, 340
Castellani, V., Giannone, P., & Renzini, A. 1971b, Ap&SS, 10, 355
Castelli, F., & Kurucz, R. L. 2003, Modelling of Stellar Atmospheres, 210, 20P
Charbonnel, C. & Lagarde, N. 2010, A&A, 522, 10
Chiosi, C., & Maeder, A. 1986, ARA&A, 24, 329
Christy, R. F. 1966, ApJ, 145, 340
Christy, R. F. 1970, JRASC, 64, 8
Cordier, D. et al. 2002, A&A, 392, 169
Cyburt, R. H., Fields, B. D., Olive, K. A., & Skillman, E. 2005, APh, 23, 313
Deasy, H. P. 1988, MNRAS, 231, 673
Degl’Innocenti, S., Prada Moroni, P. G., Marconi, M., & Ruoppo, A. 2008, Ap&SS, 316, 25
Dell’Omodarme, E., Valle, G., degl’Innocenti, S., Prada Moroni, P. G. 2011, A&A, submitted
Efremova, B. V. et al. 2011, ApJ, 730, 88
Evans, N. et al. 2005a, AJ, 130, 789
Ferguson, J.W., Alexander, D.R., Allard, F., et al. 2005, ApJ, 623, 585
Flynn, C. 2004, PASA, 21, 126
Fricke, K., Stobie, R. S., & Strittmatter, P. A. 1971, MNRAS, 154, 23
Foerster, F., López, N., Maza, J., Kubánek, P., Pignata, G. 2010, AdAst, 107569
Gennaro, M., Prada Moroni, P. G., Degl’Innocenti, S. 2010, A&A, 518, 13
Gennaro, M., Prada Moroni, P. G., & Tognelli, E. 2011, MNRAS, 420, 986
Hammer, J.W., Fey, M., Kunz, R. et al. 2005, Nuclear Physics A, 758, 363
Hofmeister, E., Kippenhahn, R., & Weigert, A. 1964, ZA, 60, 57
Hunter I. et al. 2009a, A&A, 496, 841
Hunter I. et al. 2009b, A&A, 504, 211
Iglesias, C. A., Rogers, F. J. 1991, ApJ, 371, 408
Iglesias, C., & Rogers, F.J. 1996, ApJ, 464, 943
Imbriani, G., Costantini, H., Formicola, A., et al. 2005, Eur. Phys. J. A, 25, 455
Jimenez, R., Flynn, C., MacDonald, J., & Gibson, B. K. 2003, Science, 299, 1552
Jorgensen, B. R., & Lindegren, L. 2005, A&A, 436, 127
Kass, R. & Raftery, A. 1995, Journal of the American Statistical Association, 90, 773
Karakas, A. I., Lattanzio, J. C. 2003, PASA, 20, 279
Keller, S. C., Wood, P. R. 2002, ApJ, 578, 144
Keller, S. C., Wood, P. R. 2006, ApJ, 642, 834
Keller, S. C., 2008, ApJ, 677, 483
Kervella, P., Merand, A., & Gallenne, A. 2009, A&A, 498, 425
Kervella, P.; Merand, A.; Perrin, G.; Coudé du Foresto, V. 2006, A&A, 448, 623
Kervella, P., Merand, A., Szabados, L., Fouqué, P., Bersier, D., Pompei, E., Perrin, G. 2008, A&A, 480, 167
Komatsu, E. et al. 2011, ApJS, 192, 18
Korn, A. J.; Nieva, M. F.; Daflon, S.; Cunha, K. 2005, ApJ, 633, 899
Langer, N. 1991, A&A, 252, 669
Lemut, A., Bemmerer, D., Confortola, F. et al. 2006, Phys. Lett. B, 634, 483
Maeder, A. & Meynet, G. 2000, ARA&A, 38, 143
Marengo, M. et al. 2010b, ApJ, 725, 2392
Marengo, M.; Evans, N. R.; Barmby, P.; Bono, G.; Welch, D. L.; Romaniello, M. 2010a, ApJ, 709, 120
Matthews, L. D., Marengo, M., Evans, N. R., & Bono, G. ApJ, in press, arXiv1112.0028
McAlary, C. W., & Welch, D. L. 1986, AJ, 91, 1209
Merand, A. et al. 2006, A&A, 453, 155
Merand, A. et al. 2007, ApJ, 664, 1093
Meynet, G., & Maeder, A. 2002, A&A, 390, 561
Minniti, D., et al. 2010, NewA, 15, 433
Moskalik, P., Buchler, J. R., Marom, A. 1992, ApJ, 385, 685
Neilson, H. R., & Lester, J. B. 2008, ApJ, 684, 569
Neilson, H. R., Ngeow, C.-C., Kanbur, S. M., & Lester, J. B. 2009, ApJ, 692, 81
Neilson, H. R., Cantiello, M., & Langer, N. 2011, A&A, 529, L9
Neilson, H. R., Engle, S. G., Guinan, E., Langer, N., Wasatonic, R. P., & Williams, D. B. 2012, ApJ, 745, L32
Nieuwenhuijzen, H. & de Jager, C. 1990, A&A, 231, 134
Pagel, B. E. J., & Portinari, L. 1998, MNRAS, 298, 747
Persson, S. E., Madore, B. F., Krzemiński, W., Freedman, W. L., Roth, M., Murphy, D. C. 2004, AJ, 128, 2239
Peimbert, M., Luridiana, V., & Peimbert, A. 2007, ApJ, 666, 636
Pietrinferni, A., Cassisi, S., Salaris, M., Castelli, F. 2004, ApJ, 612, 168
Pietrinferni, A., Cassisi, S., Salaris, M., Castelli, F. 2006, ApJ, 642, 797
Pietrzynski, G. et al. 2010, Nature, 468, 542, paper I
Pietrzynski, G. et al. 2011, ApJ, 742, L20, paper II
Pompeia, L., et al. 2008, A&A, 480, 379
Potekhin, A.Y. 1999, A&A, 351, 787
Reimers, D. 1975, [*Memories de la Societe Royal des Sciences de Liege*]{}, 8, 369
Rogers, F.J., Swenson, F.J., & Iglesias, C.A. 1996, ApJ, 456, 902
Romaniello, M. et al. 2008, A&A, 488, 731
Seaton, M. J., et al. 1994, MNRAS, 266, 805
Sanna et al. 2009, ApJ, 699, L84
Schlegel, D. J., Finkbeiner, D. P., Davis, M. 1998, ApJ, 500, 525
Schroeder, K.-P., Cuntz, M. 2005, ApJ, 630, L73
Shternin, P.S. & Yakovlev, D.G. 2006, PhRvD, 74(4), 3004
Steigman, G. 2006, IJMPE, 15, 1
Stobie, R. S. 1969, 144, 485
Storm, J., Gieren, W., Fouqué, P., Barnes, T. G., Soszyński, I., Pietrzyński, G., Nardetto, N., Queloz, D. 2011, A&A, 534, A95
Stothers, R., & Chin, C.-W. 1978, , 226, 231
Stothers, R. B. & Chin, C.-W. 1993, ApJ, 412, 294
Stothers, R. B. & Chin, C.-W. 1996, ApJ, 469, 166
Stubbs, C. W. et al. 2010, ApJS, 191, 376
Testa, V. et al. 1999, AJ, 118, 2839
Tognelli, E., Prada Moroni, P. G., & Degl’Innocenti, S. 2011, A&A, 533, 109
Valle, G.; Marconi, M.; Degl’Innocenti, S.; Prada Moroni, P. G. 2009, A&A, 507, 1541
van Eyken, J. C. 2011, AJ, 142, 60
Venn, K. A. 1999, ApJ, 518, 405
[llllllll]{} 0.004& 0.256 & 1.74 & 0.0 & 0 & 0.4 & 0.8 & 4.0\
0.005& 0.258 & 1.74 & 0.0 & 0 & 0.4 & 0.8 & 4.0\
0.006& 0.260 & 1.74 & 0.0 & 0 & 0.4 & 0.8 & 4.0\
0.007& 0.262 & 1.74 & 0.0 & 0 & 0.4 & 0.8 & 4.0\
0.008& 0.265 & 1.74 & 0.0 & 0 & 0.4 & 0.8 & 4.0\
0.004& 0.256 & 1.90 & 0.0 & 0 & 0.4 & 0.8 & 4.0\
0.005& 0.258 & 1.90 & 0.0 & 0 & 0.4 & 0.8 & 4.0\
0.006& 0.260 & 1.90 & 0.0 & 0 & 0.4 & 0.8 & 4.0\
0.007& 0.262 & 1.90 & 0.0 & 0 & 0.4 & 0.8 & 4.0\
0.008& 0.265 & 1.90 & 0.0 & 0 & 0.4 & 0.8 & 4.0\
0.004& 0.256 & 1.74 & 0.2 & 0 & 0.4 & 0.8 & 4.0\
0.005& 0.258 & 1.74 & 0.2 & 0 & 0.4 & 0.8 & 4.0\
0.006& 0.260 & 1.74 & 0.2 & 0 & 0.4 & 0.8 & 4.0\
0.007& 0.262 & 1.74 & 0.2 & 0 & 0.4 & 0.8 & 4.0\
0.008& 0.265 & 1.74 & 0.2 & 0 & 0.4 & 0.8 & 4.0\
0.004& 0.256 & 1.90 & 0.2 & 0 & 0.4 & 0.8 & 4.0\
0.005& 0.258 & 1.90 & 0.2 & 0 & 0.4 & 0.8 & 4.0\
0.006& 0.260 & 1.90 & 0.2 & 0 & 0.4 & 0.8 & 4.0\
0.007& 0.262 & 1.90 & 0.2 & 0 & 0.4 & 0.8 & 4.0\
0.008& 0.265 & 1.90 & 0.2 & 0 & 0.4 & 0.8 & 4.0\
[lll]{} Mass ($M/M_\odot$) & 4.14$\pm$0.05& 1\
…& 4.14$\pm$0.07& 1\
Radius ($R/R_\odot$) & 32.4$\pm$1.5 & 1\
…& 44.9$\pm$1.5 & 1\
$T_{eff}$ (K) & 5,900$\pm$250& 1\
…& 5,080$\pm$270& 1\
$V$(mag) & 15.940$\pm$0.001&\
…& 16.06$\pm$0.02&\
$I$(mag) & 15.246$\pm$0.001&\
…& 15.11$\pm$0.01&\
$\mu$ (mag) &18.45$\pm$0.04& 2\
$E(B-V)$(mag) & 0.15$\pm$0.03& 3\
\[Fe/H\] (dex) &-0.33$\pm$0.05& 4\
[llccccccccc]{} & & & & & & & & & &\
& & P &$115_{-9}^{+44}$ & $4.50_{-0.52}^{+0.18}$ & $165_{-19}^{+32}$ & $3.98_{-0.29}^{+0.24}$ & $167_{-22}^{+29}$ & $3.97_{-0.27}^{+0.27}$ & $137_{\it -42\tablenotemark{h}}^{+95}$ & $3.60_{-0.15}^{+1.05}$\
& & & & & & & & & &\
& & S &$96_{-6}^{+20}$ & $4.70_{-0.28}^{+0.16}$ & $143_{-11}^{+36}$ & $4.15_{-0.33}^{+0.19}$ & $147_{-15}^{+30}$ & $4.15_{-0.32}^{+0.20}$ & $148_{-12}^{+26}$ & $4.19_{-0.33}^{+0.14}$\
& & & & & & & & & &\
& & C &$107_{-10}^{+12}$ & & $154_{-13}^{+21}$ & & $150_{-9}^{+24}$ & & $137_{\it -3\tablenotemark{h}}^{+67}$ &\
& & & & & & & & & &\
& & & & & & & & & &\
& & & & & & & & & &\
& & P &$134_{-6}^{+11}$ & $4.13_{-0.05}^{+0.04}$ & $154_{-5}^{+4}$ & $4.10_{-0.02}^{+0.07}$ & $153_{-4}^{+5}$ & $4.14_{-0.05}^{+0.04}$ & $145_{\it -11\tablenotemark{h}}^{+21}$ & $4.14_{-0.07}^{+0.05}$\
& & & & & & & & & &\
& & S &$114_{-4}^{+30}$ & $4.19_{-0.09}^{+0.05}$ & $145_{-3}^{+7}$ & $4.14_{-0.07}^{+0.05}$ & $146_{-4}^{+7}$ & $4.15_{-0.08}^{+0.04}$ & $146_{-4}^{+7}$ & $4.15_{-0.07}^{+0.05}$\
& & & & & & & & & &\
& & C &$144_{-7}^{+3}$ & & $150_{-3}^{+5}$ & & $150_{-3}^{+5}$ & & $145_{-0}^{+15}$ &\
& & & & & & & & & &\
& & & & & & & & & &\
& & & & & & & & & &\
& & P &$115_{-14}^{+40}$ & $4.50_{-0.48}^{+0.25}$ & $166_{-17}^{+43}$ & $3.98_{-0.37}^{+0.21}$ & $167_{-21}^{+39}$ & $3.96_{-0.34}^{+0.26}$ & $226_{-22}^{+20}$ & $3.45_{-0.09}^{+0.30}$\
& & & & & & & & & &\
& & S &$88_{-3}^{+19}$ & $4.98_{-0.44}^{\it +0.0\tablenotemark{h}}$ & $143_{-11}^{+41}$ & $4.15_{-0.42}^{+0.20}$ & $145_{-13}^{+37}$ & $4.15_{-0.40}^{+0.21}$ & $141_{-6}^{+36}$ & $4.20_{-0.42}^{+0.15}$\
& & & & & & & & & &\
& & C &$95_{-3}^{+19}$ & & $154_{-14}^{+25}$ & & $149_{-9}^{+27}$ & & $139_{\it -52\tablenotemark{h}}^{+96}$ &\
& & & & & & & & & &\
& & & & & & & & & &\
& & P &$144_{-7}^{+5}$ & $4.13_{-0.05}^{+0.05}$ & $154_{-4}^{+5}$ & $4.12_{-0.04}^{+0.05}$ & $154_{-4}^{+4}$ & $4.13_{-0.05}^{+0.04}$ & $137_{-3}^{+6}$ & $4.15_{-0.05}^{+0.04}$\
& & & & & & & & & &\
& & S &$118_{-3}^{+32}$ & $4.17_{-0.07}^{+0.06}$ & $147_{-5}^{+9}$ & $4.15_{-0.08}^{+0.04}$ & $148_{-5}^{+8}$ & $4.15_{-0.08}^{+0.04}$ & $150_{-6}^{+7}$ & $4.14_{-0.07}^{+0.05}$\
& & & & & & & & & &\
& & C &$145_{-5}^{+4}$ & & $153_{-5}^{+3}$ & & $151_{-3}^{+5}$ & & $141_{-3}^{+11}$ &\
& & & & & & & & & &\
& & & & & & & & & &\
& & & & & & & & & &\
& & P &$134_{-6}^{+11}$ & $4.14_{-0.04}^{+0.02}$ & $154_{-4}^{+3}$ & $4.14_{-0.04}^{+0.03}$ & $153_{-3}^{+4}$ & $4.14_{-0.04}^{+0.03}$ & $145_{\it -7\tablenotemark{h}}^{+17}$ & $4.14_{-0.04}^{+0.03}$\
& & & & & & & & & &\
& & S &$148_{-34}^{\it +2\tablenotemark{h}}$ & $4.17_{-0.07}^{+0.04}$ & $146_{-3}^{+6}$ & $4.14_{-0.04}^{+0.04}$ & $146_{-3}^{+6}$ & $4.15_{-0.05}^{+0.02}$ & $146_{-3}^{+7}$ & $4.14_{-0.04}^{+0.05}$\
& & & & & & & & & &\
& & C &$146_{-7}^{+2}$ & & $150_{-2}^{+4}$ & & $150_{-2}^{+4}$ & & $145_{-0}^{+13}$ &\
& & & & & & & & & &\
& & & & & & & & & &\
& & & & & & & & & &\
& & P &$145_{-7}^{+5}$ & $4.13_{-0.05}^{+0.05}$ & $154_{-4}^{+5}$ & $4.11_{-0.04}^{+0.05}$ & $154_{-4}^{+4}$ & $4.14_{-0.05}^{+0.04}$ & $137_{-3}^{+19}$ & $4.15_{-0.08}^{+0.04}$\
& & & & & & & & & &\
& & S &$148_{-32}^{+6}$ & $4.17_{-0.07}^{+0.06}$ & $147_{-5}^{+8}$ & $4.15_{-0.08}^{+0.04}$ & $147_{-4}^{+8}$ & $4.15_{-0.08}^{+0.04}$ & $149_{-5}^{+7}$ & $4.14_{-0.05}^{+0.07}$\
& & & & & & & & & &\
& & C &$146_{-4}^{+5}$ & & $153_{-5}^{+3}$ & & $151_{-3}^{+4}$ & & $141_{-2}^{+14}$ &\
& & & & & & & & & &\
[llccccc]{} & & P & 0.395 & 0.393 & 0.369 & 0.007\
&\
& & S & 0.367 & 0.247 & 0.255 & 0.248\
&\
& & C & 0.145 & 0.097 & 0.094 & 0.002\
&\
& & P & 0.014 & 0.001 & 0.001 & $< 10^{-3}$\
&\
& & S & $< 10^{-3}$ & 0.001 & 0.001 & 0.001\
&\
& & C & $< 10^{-3}$ & $< 10^{-3}$ & $< 10^{-3}$ & $< 10^{-3}$\
&\
& & P & 0.499 & 0.567 & 0.534 & 0.011\
&\
& & S & 0.236 & 0.311 & 0.321 & 0.304\
&\
& & C & 0.117 & 0.177 & 0.171 & 0.003\
&\
&\
&\
&\
& & P & 0.654 & 0.857 & 0.825 & 0.002\
&\
& & S & 0.128 & 0.758 & 0.800 & 0.844\
&\
& & C & 0.083 & 0.650 & 0.660 & 0.001\
&\
& & P & 0.159 & 0.010 & 0.010 & $< 10^{-3}$\
&\
& & S & $< 10^{-3}$ & 0.009 & 0.010 & 0.012\
&\
& & C & $< 10^{-3}$ & $< 10^{-3}$ & $< 10^{-3}$ & $< 10^{-3}$\
&\
& & P & 0.704 & 1.017 & 1.000 & $< 10^{-3}$\
&\
& & S & 0.042 & 0.949 & 1.000 & 0.978\
&\
& & C & 0.030 & 0.965 & 1.000 & $< 10^{-3}$\
[^1]: http://opalopacity.llnl.gov/opal.html.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The ground state of a two-dimensional (2D) system of Bose particles of spin zero, interacting via a repulsive Gaussian-Core potential, has been investigated by means of Quantum Monte Carlo simulations. The quantum phase diagram is qualitatively identical to that of 2D Yukawa bosons. While the system is a fluid at all densities for weak coupling, in the strong coupling regime it transitions upon compression from a low density superfluid to a crystal, and then into a reentrant superfluid phase. No evidence of a (supersolid) cluster crystal phase is seen.'
author:
- Peter Kroiss
- Massimo Boninsegni
- Lode Pollet
title: 'Ground state phase diagram of Gaussian-Core bosons in two dimensions'
---
Introduction
============
In the absence of disorder, frustration, or an external potential the ground state of an interacting scalar Bose system is always ordered, [*i.e.*]{}, a well-defined symmetry of the Hamiltonian is spontaneously broken. Only two types of order are believed to be possible, namely crystalline order, which breaks translational symmetry, or off-diagonal long-range order (superfluid), in which case the global $U(1)$ symmetry is broken.\
Normally, only one of these two types of order is present, as a result of the competition between particle interactions, typically favoring crystallization, and quantum delocalization, promoting superfluidity. Such is the case for $^4$He, featuring either a superfluid or an insulating crystal in the ground state, depending on the external pressure.[@ceperley1995] No exceptions are known in continuous space when the interaction potential is of Lennard–Jones type. However, [*supersolid*]{} ground states, simultaneously displaying both types of order,[@boninsegni2012] have been predicted for a class of Bose systems with pair-wise interparticle potentials featuring a soft and flat repulsive core at short distances. The supersolid phase arises through the formation at high density of a cluster crystal (CC) with more than one particle per unit cell. At sufficiently low temperature, particle tunnelling across adjacent clusters establishes superfluid phase coherence throughout the whole system.[@cinti2010; @saccani2011; @boninsegni2012B]\
Cluster crystals have been extensively investigated in the context of classical soft-core systems.[@likos2001] It has been conjectured [@rossi2011] that a necessary condition for the presence of a CC phase in a soft-core system is that the Fourier transform of the potential go negative in a wave vector range close to $k \sim 1/d$, with $d$ the range of the soft core. Computer simulations of a classical two-dimensional (2D) system of particles interacting through a Gaussian-Core[@stillinger1976] potential $$\label{pot}
V(r)=\epsilon\ {\rm exp}\biggl[ -\frac{r^2}{2\sigma^2}\biggr ]$$ whose Fourier transform is positive-definite, has yielded no evidence of a CC at low temperature, [@prestipino2011] thus supporting the hypothesis of Ref. . The classical ground state is a crystal at all densities; at low temperature, equilibrium low- and high-density fluid phases exist on both sides of the crystal, with hexatic phases, characterized by the absence of positional order but by a non-vanishing orientational order parameter, separating the crystal from the fluid.\
An interesting theoretical question is to what a degree quantum-mechanical effects alter the classical phase diagram. Mean field theoretical treatments based on the Gross-Pitaevskii[@GP] equation, suggest that a negative Fourier component in the pair potential is a necessary condition for a roton instability toward crystallization.[@roton; @roton2; @roton3; @roton4] On the other hand, such an approach essentially describes a supersolid as a superfluid with a density modulation, and is therefore applicable to crystals with a very large number of particles per unit cell. If the number of particles per unit cell is only a few, it is known that Bose statistics can considerably extend the domain of existence of the CC in soft-core systems, with respect to what one would observe classically.[@cinti2014] Thus, it is conceivable that a CC phase (turning superfluid at low temperature) could be stabilized by quantum-mechanical exchanges in a system of Gaussian-Core bosons. Also, the investigation of a quantum-mechanical version of the Gaussian-Core model can offer insight into the role of those quantum fluctuations in the context of soft matter systems.[@note0]\
In this work, we present results of extensive Quantum Monte Carlo simulations at low temperature of a 2D system of spin-zero particles interacting via a Gaussian pair potential. Our study shows that, as expected, quantum effects strengthen the fluid phase, which extends all the way to temperature $T=0$ in a wide region of the quantum phase diagram. No cluster crystal and no supersolid phase is found. Indeed, superfluid and (insulating) crystal are the only two phases observed. The resulting quantum phase diagram is qualitatively identical with that of 2D Yukawa bosons,[@Magro93; @Osychenko2012] suggesting that it may generically describe all Bose systems featuring the same type of repulsive interaction at short distances, i.e., one that is strong enough to prevent the formation of clusters of particles, but not enough to stabilize the crystalline phase at high density.\
The remainder of this paper is organized as follows: in Sec. \[modset\] we describe the model of our system of interest and provide details of the calculation; in Sec. \[res\] we illustrate our results and outline our conclusions in Sec. \[conc\].
Model and Methodology {#modset}
=====================
We consider a system of $N$ spin-zero bosons of mass $m$, enclosed in a simulation box with periodic boundary conditions in both directions. The aspect ratio of the box is designed to fit a triangular solid. Particles interact via the pair potential described by Eq. \[pot\]. The many-body Hamiltonian of the system in reduced units reads $$\label{ham}
H =- \frac{\Lambda}{2} \sum_{i=1}^N \nabla_i^2 + \sum_{i<j} {\rm exp}\biggl [-\frac {1}{2}\ r_{ij}^2\biggr ],$$ where ${\bf r}_i$ is the position vector of the $i$th boson, $r_{ij}\equiv |{\bf r}_i-{\bf r}_j|$, and $\Lambda\equiv \hbar^2/(m\epsilon\sigma^2)$ is the quantum coupling constant. All lengths are expressed in units of $\sigma$, whereas $\epsilon$ sets the energy and temperature scale (Boltzmann’s constant $k_B$ is set to one). Besides $\Lambda$, the only other parameter of the system at temperature $T=0$ is the density $\rho$, or, equivalently, the (dimensionless) mean interparticle distance $r_s=1/\sqrt{\rho\sigma^2}$.\
We obtained the thermodynamic phase diagram by means of Quantum Monte Carlo simulations based on the Worm Algorithm in the continuous-space path integral representation. Because this well-established computational methodology is thoroughly described elsewhere, [@bps2006a; @bps2006b] we do not review it here. The most important aspect to be emphasized is that it enables one to compute thermodynamic properties of Bose systems at finite temperature, directly from the microscopic Hamiltonian. In particular, energetic, structural and superfluid properties can be computed in practice with no approximation.\
Technical aspects of the calculations are standard; we carried out simulations of systems comprising up to 1024 particles, using the standard form[@bps2006b] for the high-temperature imaginary-time propagator accurate up to fourth order in the time step $\tau$; the smoothness of the potential, however, allows one to obtain accurate results with the primitive form as well, with comparable CPU time. We identify the different thermodynamic phases (superfluid and crystalline) through the computation of the superfluid density, using the well-known winding number estimator, [@pollock1987] the one-body density matrix, as well as the pair correlation function. As we aim at obtaining the ground state ($T=0$) phase diagram, we performed calculations at temperatures sufficiently low not to see any changes in the values of cogent physical quantities (e.g., the energy) within the statistical uncertainties of the calculations; typically, this means $T\lesssim T^\star\equiv\Lambda/ r_s^{2}$.
![\[fig:phase\_diagram\] (Color online). Ground state phase diagram of a Bose system with Gaussian-Core interaction for different values of the parameter $\Lambda$ (see text) and mean interparticle spacing $r_s$. For low values of $\Lambda$ a crystal phase becomes increasingly stable. At low densities a superfluid phase is seen, while a reentrant superfluid phase is found for high densities. Numerical data are represented by symbols, the dashed line is a guide for the eye. ](phase_diagram.pdf){width="\linewidth"}
results {#res}
=======
Phase Diagram
-------------
Our findings are summarized in Fig. \[fig:phase\_diagram\], showing the ground state phase diagram of the system, as described by the Hamiltonian (\[ham\]), in the $(r_s$–$\Lambda)$ plane. All of the results presented here are extrapolated to the $\tau\to 0$ limit. We identify the following phases: (i) a superfluid phase at all densities for $\Lambda \gtrsim 0.03$, and in the low and high density limit for lower values of $\Lambda$, where the physics of the system is dominated by quantum delocalization and Bose statistics. All of our numerical data at finite temperature show consistency with the Berezinsky-Kosterlitz-Thouless (BKT) scenario of the superfluid transition in 2D,[@Berenzinsky1971; @KT] [*i.e.*]{}, the transition between a quasi-long-range superfluid and a normal phase at finite temperature occurs through the standard unbinding of vortex–anti-vortex pairs with charge 1. Unconventional are however the microscopic properties of the reentrant superfluid phase (see below). (ii) A crystalline (triangular) phase becomes stable for $\Lambda\lesssim 0.03$, centered around $r_s = 3$. It extends its domain of existence as $\Lambda \to 0$, as the potential energy plays an increasingly important role. We found no evidence of other phases, such as supersolid phases.
On general grounds we expect a superfluid phase at $T=0$ in the low density limit by analogy to superfluid helium: In the dilute limit, the potential energy is much smaller than the energy of zero point fluctuations, [*i.e.*]{}, the quantum pressure prevents crystallization. That a superfluid phase might also occur at high densities is different from helium: In the present case the soft-core of the potential is unable to prevent the overlap of particles at high enough density, leading to a reentrant superfluid, just as in the phase diagram of 2D Yukawa bosons.[@Magro93; @Osychenko2012]
![\[fig:correlation\_functions\] (Color online). Ground state pair correlation functions for different values of $r_s$ at $\Lambda=1/30$. Error bars are too small to be seen on the scale of the figure. While the pair correlation functions for $r_s = 4$ and $r_s = 3$ (corresponding to the superfluid and crystal phases) show hard-core separation of particles, the reentrant superfluid phase ($r_s = 1.5$) acquires a finite value at the origin. In this phase, only very weak peaks are left, rendering $g(r)$ essentially flat for $r \gtrsim 2$. ](correlation_functions.pdf){width="\linewidth"}
Insight into the structure of the various phases is offered by the pair correlation function $g(r)$, shown in Fig. \[fig:correlation\_functions\] for the crystal and both conventional and reentrant superfluids. For $r_s \ge 3$ ([*i.e.*]{}, in the crystalline and low density superfluid phases), the physics effectively mimics that of a hard-core system, characterized by a vanishing $g(r)$ at short distances, and resulting in conventionally looking pair correlation functions, as is shown in Fig. \[fig:correlation\_functions\] for $r_s=3$. The peak structure in $g(r)$ is washed out as the system is compressed and $r_s$ is reduced below 3, at which point $g(r)$ suddenly acquires a finite value at the origin, as the finite potential energy cost no longer prevents particles from overlapping. Further compression of the system into the reentrant superfluid phase has the effect of raising the value of $g(0)$, as the system approaches the behavior of a free Bose gas. Note that the first peak of $g(r)$ in the conventional superfluid phase is more pronounced than the corresponding peak of the reentrant superfluid; this is a consequence of the effective hard-core interaction between the particles. As a finite-temperature method is employed, the magnitude of the peaks of $g(r)$ decreases even in the solid phase. This is expected, as thermal fluctuations do not allow true crystalline order in two dimensions. The important observation is that the distance between neighboring peaks is constant over a large range of $r$.
![\[fig:nmu\] (Color online). Particle number as a function of grand-canonical chemical potential $\mu$ for $r_s = 1$, $T = 0.001$ and $\Lambda = 1/1000$. The initial configuration is always a solid configuration at $N= 256$ in a commensurate box. As the grand-canonical simulation allows the particle number to fluctuate, non-integer values of $N$ are possible. However, for the point of $\mu = 5.8$ (corresponding to $N = 256$), no particle number changes are observed; this happens for several values of $\mu$. The horizontal line corresponds to 256 particles. Error bars are not visible on this scale. In the inset, a zoom on the data shows regions of nearly constant particle number. ](nmu.pdf){width="\linewidth"}
![\[fig:cycle\] (Color online). For $r_s = 1$, $\Lambda = 0.001$, $T = 1/6400$ and 256 particles, the probability $P(n)$ of bosonic exchanges involving $n$ particle worldlines is shown. ](cycle.pdf){width="\linewidth"}
![\[fig:obdm\] (Color online). The equal-time one-body density matrix at $r_s = 1$ for $\Lambda = 0.002$, $T = 0.002$ and different system sizes. After a rapid initial decay, a non-integrable power-law decay is seen for big enough system sizes, rendering the phase a superfluid. The dashed line illustrates the linear regime which holds for all $r > r_*$ (see text). ](obdm.pdf){width="\linewidth"}
Reentrant superfluid phase in the limit of low $\Lambda$
--------------------------------------------------------
Next, we investigate the point $r_s = 1$, $\Lambda = 0.001$ and $T = 0.001$ which corresponds to high density and weak quantum fluctuations. Fig. \[fig:nmu\] shows the total number of particles for different grand-canonical simulations [@bps2006b] in an initially solid configuration of $N = 256$ particles for several values of the chemical potential $\mu$. While the curve suggests a linear relationship between chemical potential and particle number (implying a constant non-zero compressibility), tiny deviations of much lower compressibility can be seen. These may hint at a tendency towards insulating behavior, but this is not the case here: They are a consequence of low temperature and finite system size, similar to the observation of finite charging levels in a quantum dot. It is well known that in dilute superfluids the compressibility at very low temperatures can also be very small on small system sizes and very low temperatures due to the same mechanism. What is surprising here, is that this occurs already for temperatures of the order of the Kosterlitz-Thouless temperature. Nevertheless, a gapped solid-like structure can certainly be ruled out in the thermodynamic limit on the basis of the pair correlation function. In addition, the corresponding Green’s function at zero momentum, $G(\tau, p = 0)$, goes up with increasing system size for $|\tau| \gg 0$, [*i.e.*]{}, adding particles to the simulation becomes easier. This is one of the manifestations that this parameter regime is very difficult to simulate. In particular, the superfluid density has anomalously large autocorrelation times, which are unusual for the worm algorithm. Interestingly, bosonic particle exchanges do not suffer from the same decorrelation problem: Fig. \[fig:cycle\] shows a typical distribution of particle permutations where exchanges up to the total number of particles are reached. The distribution can already be reliably measured in early stages of the simulation without observing any superfluid response. This disagrees with the perception that macroscopic exchanges directly trigger superfluidity (which holds for dilute systems). The superfluidity of the phase is never in doubt though, as can be seen from the the one-body density matrix $n(r)$, shown for a similar point ($\Lambda = 0.002$) in Fig. \[fig:obdm\]. It experiences a weak power-law decay (setting in at a distance $r_*$) after a fast initial drop with a power less than $1/4$, demonstrating the existence of off-diagonal long-range order in the system. Although the asymptotic behavior of the curve is consistent with conventional superfluids, the low value of $n(r_*)$ at which this power-law sets in, is unusual. Comparing this curve with measurements for higher $\Lambda$, it follows that we can tune $n(r_*)$ with $\Lambda$. For increasing $\Lambda$, the winding estimator yields the correct superfluid response more and more reliably (cf. Fig. \[fig:phase\_diagram\]).
On the basis of all these observations, we can state that the system is ultimately a superfluid based on its properties for big enough system sizes. The unusual microscopics are due to the denseness witnessed in this parameter regime. Finally, we note that the behavior of $n(r)$ and $g(r)$ for sufficiently large values of $r$, as well as of the superfluid density, is remarkably similar to the observations of the superglass in Ref. . However, as we are looking for the thermodynamic ground state of the system, such a metastable state can be excluded. Hence, the claim for a superglass, as in Ref. , should only be made on the basis of additional real-time considerations. We leave for future work the static response of this phase, *i.e.*, how it responds to pinning or disorder.
Conclusion {#conc}
==========
In conclusion, a first principles numerical investigation of the phase diagram of a 2D Bose system with Gaussian-Core pair-wise interactions has yielded two different phases: A crystal and a superfluid, which also shows reentrant behavior at high densities. No supersolid or cluster crystal phases were found. This was anticipated by the positiveness of the interaction potential in Fourier space and affirms the cluster crystal conjecture of Ref. .
The reentrant superfluid phase demonstrates unexpected behavior for high particle mass: The power-law decay of the one-body density matrix sets in at large distances, where its value is already quite low. This requires big system sizes to capture the relevant length scales. Likewise, grand-canonical simulations experience deviations from non-zero compressibility for finite systems. This is complemented by the occurrence of large cycles of particle wordline permutations, independent of system size.
Whether such a system may lend itself to experimental realization is difficult to assess. Recent progress in cold atom manipulation allows one to tailor, to some degree, the interaction among atoms.[@zwerger2008] The other aspects of the system, including its detection, are already well within current technology.[@zwerger2008] The crucial part is that the Gaussian potential has no preferred length scale, unlike the softened dipolar [@moroni] or Rydberg potentials, whereas a system of Yukawa bosons [@Magro93; @Osychenko2012] shows a qualitatively similarly looking phase diagram.
We would like to thank W. Krauth, J. Nespolo, T. Pfeffer, N. Prokof’ev and B. Svistunov for fruitful discussions. This work is supported by the Excellence Cluster NIM, FP7/Marie-Curie Grant No. 321918 (FDIAGMC), FP7/ERC Starting Grant No. 306897 (QUSIMGAS), as well as by the Natural Science and Engineering Research Council of Canada. Computing support from Westgrid is gratefully acknowledged.
[99]{} D. M. Ceperley, Rev. Mod. Phys. **67**, 279 (1995). M. Boninsegni and N. Prokof’ev, Rev. Mod. Phys. [**84**]{}, 759 (2012). F. Cinti, P. Jain, M. Boninsegni, A. Micheli, P. Zoller and G. Pupillo, Phys. Rev. Lett. [**105**]{}, 135301 (2010). S. Saccani, S. Moroni and M. Boninsegni, Phys. Rev. B [**83**]{}, 092506 (2011). M. Boninsegni, J. Low Temp. Phys. [**168**]{}, 137 (2012). C. N. Likos, Phys. Rep. [**348**]{}, 267 (2001). M. Rossi, S. L. Zavattari, D. E. Galli and L. Reatto, Phys. Rev. B [**84**]{}, 052504 (2011). F. H. Stillinger, J. Chem. Phys. [**65**]{}, 3968 (1976). S. Prestipino, F. Saija and P. V. Giaquinta, Phys. Rev. Lett. [**106**]{}, 235701 (2011). L.P. Pitaevskii, Sov. Phys. JETP [**13**]{}, 451 (1961); E.P. Gross, J. Math. Phys. [**4**]{}, 195 (1963). D. A. Kirzhnits and Yu. A. Nepomnyashchii, Sov. Phys. JETP [**32**]{}, 1191 (1971). Yu. A. Nepomnyashchii, Theor. Math. Phys. [**8**]{}, 928 (1971). Yu. A. Nepomnyashchii and A. A. Nepomnyashchii, Theor. Math. Phys. [**9**]{}, 1033 (1971). C. Josserand, Y. Pomeau and S. Rica, Phys. Rev. Lett. [**98**]{}, 195301 (2007). F. Cinti, M. Boninsegni and T. Pohl, New J. Phys. [**16**]{}, 033038 (2014). This pair potential is used, for example, to model colloids in a solvent. See, for instance, Ref. . W. R. Magro and D. M. Ceperley, Phys. Rev. B [**48**]{}, 411 (1993). O. N. Osychenko, G. E. Astrakharchik, F. Mazzanti, J. Boronat, Phys. Rev. A [**85**]{}, 063604 (2012). M. Boninsegni, N. Prokof’ev and B. Svistunov, Phys. Rev. Lett. [**96**]{}, 070601 (2006). M. Boninsegni, N. Prokof’ev and B. Svistunov, Phys. Rev. E [**74**]{}, 036701 (2006). E. L. Pollock and D. M. Ceperley, Phys. Rev. B [**36**]{}, 8343 (1987). V. L. Berenzinsky, Sov. Phys. JETP **32**, 493 (1971). J. M. Kosterlitz and D. J. Thouless, J. Phys. C [**6**]{}, 1181 (1973). M. Boninsegni, N. Prokof’ev, and B. Svistunov, Phys. Rev. Lett. [**96**]{}, 105301 (2006). I. Bloch, J. Dalibard and W. Zwerger, Rev. Mod. Phys. [**80**]{}, 885 (2008). S. Moroni and M. Boninsegni, Phys. Rev. Lett. [**113**]{}, 240407 (2014).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The first-order bosonic field equations of the $D$-dimensional effective low energy theory which describes the massless background coupling of the $D$-dimensional fully Higgsed heterotic string are derived.'
author:
- |
Nejat T. Y$\i$lmaz\
Department of Mathematics and Computer Science,\
Çankaya University,\
Öğretmenler Cad. No:14,06530,\
Balgat, Ankara, Turkey.\
`[email protected]`
title: Integrated Field Equations of Heterotic Supergravities
---
Introduction
============
The ten-dimensional $\mathcal{N=}1$ supergravity [@d=10; @tani15] which is coupled to $16$ gauge multiplets with the gauge group either $O(32)$ or $E_{8}\times E_{8}$ is the low energy effective limit theory which describes the massless background coupling of the ten-dimensional heterotic string [@kiritsis]. If one chooses Abelian gauge multiplets then one obtains the maximal torus sub-theory of the ten-dimensional $O(32)$ or $E_{8}\times E_{8}$ Yang-Mills supergravity theory. In this case the full gauge group is broken down to its maximal torus subgroup $U(1)^{16}$, whose Lie algebra is the Cartan subalgebra of the non-abelian gauge groups mentioned above. This mechanism is due to the general Higgs vacuum structure of the heterotic string which causes a spontaneous symmetry breakdown. Thus in this respect upon compactification one obtains the $D$-dimensional fully Higgsed massless heterotic string coming from the maximal torus sub-theory of the ten-dimensional $O(32)$ or $E_{8}\times
E_{8}$ Yang-Mills supergravity. In either of the $O(32)$ or the $E_{8}\times E_{8}$ heterotic string theories in order to obtain the $D$-dimensional massless heterotic string in the fully Higgsed compactification only the ten-dimensional $16$ Cartan gauge fields are kept in the reduction since the non-cartan gauge fields lead to massive fields following the compactification. In other words only the Cartan gauge fields are kept since they are the only fields which will remain massless for generic values of the Wilson lines.
In [@heterotic] one can refer to the torodial compactification of the bosonic sector of the ten-dimensional $\mathcal{N=}1$ supergravity which is coupled to $16$ Abelian gauge multiplets. As we have discussed above such a reduction gives us the $D$-dimensional bosonic low energy theory of the massless sector of the $D$-dimensional fully Higgsed heterotic string. In a formalism which treats the scalar manifolds as generic $G/K$-cosets and which uses the solvable Lie algebra parametrization [@fre; @nej1; @nej2] the field equations of these bosonic theories are studied in [@het].
Starting from the bosonic field equations of the $D$-dimensional effective massless fully Higgsed heterotic string which is the $D$-dimensional heterotic supergravity in this note we derive the first-order field equations by locally integrating the second-order field equations obtained in [@het]. By integration we mean cancelling an exterior derivative on both sides of the equations. Therefore we obtain a first-order formulation of the theory. We will effectively make use of the results derived in [@consist2] which states that there exists a one-sided on-shell decoupling between the coset scalars and the gauge fields of the heterotic supergravities. For this reason to obtain the first-order field equations of the coset scalars we will adopt the general formulation of [@nej2] which works out the first-order field equations of the pure symmetric space sigma model. We will also give a brief discussion how one can make use of the first-order field equations to perform the on-shell bosonic coset construction of the $D$-dimensional heterotic supergravity.
The First-Order Field Equations {#section23}
===============================
The bosonic field content which constitutes the low energy effective Lagrangian that describes the bosonic sector of the $D$-dimensional massless background coupling of the fully Higgsed heterotic string can be given as [@het] $$\label{eq1}
\{C_{(1)}^{I},A_{(2)},\phi,\phi^{i},\chi^{\alpha}\}.$$ The fields $C_{(1)}^{I}$ are $(20-2D+16)$ one-forms, and $A_{(2)}$ is a two-form, the rest of the fields are scalars. The scalar field $\phi$ is decoupled from the rest of the scalars which are the coset ones. The coset scalars $\{\phi^{i},\chi^{\alpha}\}$ parametrize the coset manifold $$\label{eq2}
O^{\prime}(10-D+16,10-D)/O(10-D+16)\times O(10-D),$$ whose elements which are ($20-2D+16$)-dimensional real matrices in the fundamental representation satisfy $$\label{eq3}
\nu^{T}\Omega \nu=\Omega.$$ Here the $(20-2D+16)\times(20-2D+16)$ matrix $\Omega$ is $$\label{eq4}
\Omega=\left(\begin{array}{ccc}
0 & 0 & -{\mathbf{1}}_{(10-D)} \\
0 & {\mathbf{1}}_{(16)} & 0 \\
-{\mathbf{1}}_{(10-D)} & 0 & 0 \\
\end{array}\right),$$ where ${\mathbf{1}}_{(n)}$ is the $n\times n$ unit matrix. Following the notation of [@het] we use a prime in which stands for the particular representation of $O(10-D+16,10-D)$ defined through that is generated by the indefinite signature metric [^1]. We should also state that we separate $N=16$ in our expressions to emphasize the number of Abelian matter multiplets coupling to the ten-dimensional $\mathcal{N=}1$ type I supergravity which forms when coupled to these $16$ $U(1)$ vector multiplets the low energy effective limit of the ten-dimensional fully Higgsed heterotic string. The coset parametrization can be constructed in the solvable Lie algebra gauge [@fre; @nej1; @nej2; @het] as $$\label{eq5}
\nu=e^{\frac{1}{2}\phi ^{i}H_{i}}e^{\chi ^{\alpha}E_{\alpha}},$$ where $i=1,\cdots,r$ and $\alpha=1,\cdots,n$. Here $H_{i}$ are the Cartan generators and $E_{\alpha}$ are the positive root generators of the solvable Lie algebra which takes part in the Iwasawa decomposition of $o^{\prime}(10-D+16,10-D)$ [@hel]. One can find a more detailed study of the solvable Lie algebra parametrization in [@nej1; @nej2; @het]. The field equations of the bosonic fields are already derived in [@het]. They read $$\label{eq6}
\begin{aligned}
(-1)^{D}d(\ast
d\phi)&=\frac{1}{2}\sqrt{8/(D-2)}\:e^{-\sqrt{\frac{8}{(D-2)}}\phi}
\ast F_{(3)}\wedge
F_{(3)}\\
&\quad
+\frac{1}{2}\sqrt{2/(D-2)}\:e^{-\sqrt{\frac{2}{(D-2)}}\phi}\mathcal{M}_{IJ}
\ast H_{(2)}^{I}\wedge H_{(2)}^{J},\\
d(e^{-\sqrt{\frac{8}{(D-2)}}\phi}\ast F_{(3)})&=0,\\
d(e^{-\sqrt{\frac{2}{(D-2)}}\phi}\mathcal{M}^{I}_{\:\:\:J}\ast
H_{(2)}^{J})&=(-1)^{D}e^{-\sqrt{\frac{8}{(D-2)}}\phi}\Omega^{I}_{\:\:\:J} H_{(2)}^{J}\wedge\ast F_{(3)},\\
d(e^{\gamma _{i}\phi ^{i}}\ast U^{\gamma })&=\sum\limits_{\alpha
-\beta=-\gamma }N_{\alpha ,-\beta }U^{\alpha }\wedge e^{\beta
_{i}\phi ^{i}}\ast U^{\beta },\\
d(\ast d\phi ^{i})&=\frac{1}{2}\sum\limits_{\beta\in\Delta_{nc}^{+}}^{}\beta _{i}%
e^{\frac{1}{2}\beta _{j}\phi ^{j}}U^{\beta }\wedge e^{%
\frac{1}{2}\beta _{j}\phi ^{j}}\ast U^{\beta
}\\
&\quad -\frac{1}{2}(-1)^{D}e^{-\sqrt{\frac{2}{(D-2)}}\phi}\ast
H_{(2)}\wedge\nu^{T}H_{i}\nu H_{(2)},
\end{aligned}$$ where $\alpha,\beta,\gamma$ whose corresponding generators enter in the solvable lie algebra parametrization of are the elements of $\Delta_{nc}^{+}$ which is the set of non-compact positive roots of $o^{\prime}(10-D+16,10-D)$ [@nej2; @het][^2]. The field strengths of the fields $\{C_{(1)}^{I},A_{(2)},\chi^{\beta}\}$ are respectively defined as $$\label{eq7}
\begin{aligned}
H_{(2)}^{I}&=dC_{(1)}^{I},\\
F_{(3)}&=dA_{(2)}+\frac{1}{2}\:\Omega_{IJ}\: C_{(1)}^{I}\wedge
dC_{(1)}^{J},\\
U^{\alpha}&=\mathbf{\Omega}^{\alpha}_{\:\:\:\beta}d\chi^{\beta}.
\end{aligned}$$ The matrix $\mathcal{M}$ is $$\label{eq7.5}
\mathcal{M}=\nu ^{T}\nu.$$ In the above relations $\beta_{i},\gamma_{i}$ are the root vector components and $N_{\alpha,\beta}$ are the structure constants of the corresponding positive root generators of the solvable Lie algebra generated by $\{H_{i},E_{\alpha}\}$ [@nej1; @nej2; @het]. More specifically $$\label{yeni0.5}
[H_{j},E_{\gamma}]=\gamma_{j}E_{\gamma},$$ and $$\label{yeni0.6}
[E_ {\alpha},E_{\beta}]=N_{\alpha,\beta}E_{\alpha+\beta}.$$ From [@nej1; @nej2] the definition of the $n\times n$ matrix $\mathbf{\Omega}(\chi^{\beta})$ reads[^3] $$\label{yeni1}
\mathbf{\Omega}=(e^{\omega}-I)\,\omega^{-1},$$ where $\omega$ is an $n \times n$ matrix with components $$\label{yeni2}
\omega_{\beta}^{\gamma}=\chi^{\alpha}K_{\alpha\beta}^{\gamma}.$$ Here $K_{\alpha\beta}^{\gamma}$ are defined through the commutators of $\{E_{\alpha}\}$[^4] $$\label{yeni3}
[E_{\alpha},E_{\beta}]=K_{\alpha\beta}^{\gamma}E_{\gamma}.$$ We should state that we freely lower and raise indices by using various dimensional Euclidean metrics when necessary for convenience of notation. In [@het] it is shown that the second term on the right hand side of the last equation in which is compactly written in matrix form can be given as $$\label{eq8}
-\frac{1}{2}(-1)^{D}e^{-\sqrt{\frac{2}{(D-2)}}\phi}\ast
H_{(2)}\wedge\nu^{T}H_{i}\nu
H_{(2)}=(-1)^{D}\frac{\partial\mathcal{L}_{m}}{\partial\phi^{i}}.$$ Here $H_{i}$ are the ($20-2D+16$)-dimensional matrix representatives of the Cartan generators and $H_{(2)}$ is the vector of the field strengths defined in . Also $\mathcal{L}_{m}$ is the matter-scalar coupling Lagrangian [@het]. However on the other hand in [@consist2] it is proven that the expression in vanishes on-shell for the elements of the solution space indicating that the coset scalar field equations coincide with the pure sigma model ones. Therefore we can legitimately drop the second term on the right hand side of the last equation in .
As discussed in the Introduction our aim in this note is to integrate the field equations in locally. In this respect we will use the fact that locally a closed differential form is an exact one. For this reason we first introduce the dual fields $$\label{eq9}
\{\widetilde{C}^{I},\widetilde{B},\widetilde{\phi}\},$$ where $\widetilde{C}^{I}$ are $(D-3)$-forms, $\widetilde{B}$ is a $(D-4)$-form and $\widetilde{\phi}$ is a $(D-2)$-form. It is a straightforward operation that if one applies the exterior derivative on both sides of $$\label{eq10}
e^{-\sqrt{\frac{8}{(D-2)}}\phi}\ast F_{(3)}=d\widetilde{B},$$ one obtains the second equation in . Thus is our first first-order field equation. Next let us consider $$\label{eq11}
e^{-\sqrt{\frac{2}{(D-2)}}\phi}\mathcal{M}^{I}_{\:\:\:J}\ast
H_{(2)}^{J}=(-1)^{D}(d\widetilde{C}^{I}+\Omega^{I}_{\:\:\:K}C^{K}_{(1)}\wedge
d\widetilde{B}).$$ If we take the exterior derivative of both sides we get $$\label{eq12}
d(e^{-\sqrt{\frac{2}{(D-2)}}\phi}\mathcal{M}^{I}_{\:\:\:J}\ast
H_{(2)}^{J})=(-1)^{D}\Omega^{I}_{\:\:\:K}dC^{K}_{(1)}\wedge
d\widetilde{B}.$$ By using in it gives the third equation in . Finally let us take the first-order equation $$\label{eq13}
\ast
d\phi=d\widetilde{\phi}-\frac{1}{2}\sqrt{8/(D-2)}A_{(2)}\wedge
d\widetilde{B}+\frac{1}{2}\sqrt{2/(D-2)}\delta_{IJ}C^{I}_{(1)}\wedge
d\widetilde{C}^{J}.$$ If we apply the exterior derivative on both sides of we get $$\label{eq14}
d(\ast d\phi)= -\frac{1}{2}\sqrt{8/(D-2)}dA_{(2)}\wedge
d\widetilde{B}+\frac{1}{2}\sqrt{2/(D-2)}\delta_{IJ}dC^{I}_{(1)}\wedge
d\widetilde{C}^{J}.$$ By using and the above equation can be written as
\[eq15\] $$\begin{aligned}
d(\ast
d\phi)&=\frac{1}{2}\sqrt{8/(D-2)}\:e^{-\sqrt{\frac{8}{(D-2)}}\phi}
(-1)^{D}\ast F_{(3)}\wedge
F_{(3)}\notag\\
&\quad +(-1)^{D}\frac{1}{2}\sqrt{2/(D-2)}\:e^{-\sqrt{\frac{2}{(D-2)}}\phi}\mathcal{M}_{IK}
\ast H_{(2)}^{K}\wedge H_{(2)}^{I}\notag\\
&\quad
+\frac{1}{2}\sqrt{2/(D-2)}\:e^{-\sqrt{\frac{8}{(D-2)}}\phi}\Omega_{IJ}C^{I}_{(1)}\wedge
H_{(2)}^{J}\wedge
\ast F_{(3)}\notag\\
&\quad
-\frac{1}{2}\sqrt{2/(D-2)}\:e^{-\sqrt{\frac{8}{(D-2)}}\phi}\Omega_{KI}C^{K}_{(1)}\wedge
H_{(2)}^{I}\wedge \ast üF_{(3)}.\tag{\ref{eq15}}\end{aligned}$$
Since $\Omega$ is a symmetric matrix the last two terms cancel and as $\mathcal{M}$ is also a symmetric matrix this equation gives us the first equation in . The first-order formulation of the coset scalar field equations in is a straightforward task. Following our discussion above when we drop the second term on the right hand side of the last equation in we obtain the pure sigma model field equations which are the same with the ones derived for a generic coset manifold in [@nej2]. As we have remarked before this fact is a consequence of the on-shell conditions satisfied by the general solutions of the theory which are rigorously derived in [@consist2]. The first-order field equations of the general non-split [@hel] scalar coset are already derived in [@nej2]. In general the coset manifolds in are also in non-split form. Therefore we can adopt the results of [@nej2] for the scalar sectors of the heterotic supergravities. For the sake of completeness we will repeat the first-order scalar field equations of [@nej2] here. From [@nej2] we have $$\label{eq16}
\ast \overset{\rightharpoonup }{\mathbf{\Psi }}=(-1)^{D}e^{\mathbf{\Gamma }%
}e^{\mathbf{\Lambda }}\overset{\rightharpoonup }{\mathbf{A}}.$$ Here we define the $(r+n)$-dimensional column vectors $\overset{\rightharpoonup }{\mathbf{\Psi }}$ and $\overset{\rightharpoonup }{\mathbf{A}}$ whose components can be given as
\[eq17\] $$\begin{gathered}
\mathbf{\Psi}^{i}=\frac{1}{2}d\phi^{i},\quad \text{for}\quad
i=1,...,r,\quad
\mathbf{\Psi}^{\alpha+r}=e^{\frac{1}{2}\alpha_{i}\phi^{i}}\mathbf{\Omega
}_{\:\:\:\gamma}^{\alpha}d\chi^{\gamma},\quad\text{for}\quad\alpha=1,...,n,\notag\\
\notag\\
\mathbf{A}^{i}=\frac{1}{2}d\widetilde{\phi}^{i},\quad\text{for}\quad
i=1,...,r,\quad\text{and}\quad
\mathbf{A}^{\alpha+r}=d\widetilde{\chi}^{\alpha},\quad\text{for}\quad\alpha=1,...,n,
\tag{\ref{eq17}}\end{gathered}$$
where we have introduced the dual $(D-2)$-forms $\widetilde{\phi}^{i}$ and $\widetilde{\chi}^{\alpha}$. In $\mathbf{\Gamma }(\phi^{i})$ and $\mathbf{\Lambda
}(\chi^{\beta})$ are $(n+r)\times(n+r)$ matrix functions. Their components read $$\label{yeni7}
\mathbf{\Gamma }%
_{n}^{k}=\frac{1}{2}\phi ^{i}\,\widetilde{g}_{in}^{k}\quad,\quad
\mathbf{\Lambda }_{n}^{k}=\chi ^{\alpha}\widetilde{f}_{\alpha
n}^{k}.$$ The real constant coefficients $\{\widetilde{g}_{in}^{k}\}$ and $\{\widetilde{f}_{\alpha n}^{k}\}$ are already listed in [@nej1][^5]. They are
\[yeni8\] $$\begin{gathered}
\widetilde{f}_{\alpha m}^{n}=0,\quad\quad m\leq r\quad,\quad
\widetilde{f}_{\alpha ,\alpha +r}^{i}=\frac{1}{4}\alpha _{i},\quad\quad%
i\leq r,\notag\\
\notag\\
\widetilde{f}_{\alpha ,\alpha +r}^{i}=0,\quad\quad i>r\quad,\quad
\widetilde{f}_{\alpha ,\beta +r}^{i}=0,\quad\quad i\leq r,%
\quad\alpha \neq \beta ,\notag\\
\notag\\
\widetilde{f}_{\alpha ,\beta +r}^{\gamma +r}=N_{\alpha ,-\beta
},\quad\quad \alpha -\beta =-\gamma,\quad
\alpha \neq \beta,\notag\\
\notag\\
\widetilde{f}_{\alpha ,\beta +r}^{\gamma +r}=0,\quad\quad \alpha
-\beta \neq -\gamma,\quad \alpha \neq \beta,\tag{\ref{yeni8}}\end{gathered}$$
and
\[yeni9\] $$\begin{gathered}
\widetilde{g}_{im}^{n}=0,\quad\quad m\leq r\quad,\quad\widetilde{%
g}_{im}^{n}=0,\quad\quad m>r,\quad m\neq n,\notag\\
\notag\\
\widetilde{g}_{i\alpha }^{\alpha }=-\alpha _{i},\quad\quad\alpha
>r.\tag{\ref{yeni9}}\end{gathered}$$
Since as discussed in detail in [@nej1; @nej2] beside being enumerated $\alpha,\beta,\gamma,...$ correspond to the set of non-compact positive roots of $o^{\prime}(10-D+16,10-D)$ the conditions on them in and must be understood in the root sense. We should also state that likewise in [@nej2] we assume the signature of the spacetime as $s=1$. It is proven in [@nej2] that as a consequence of the dualisation of the general symmetric space sigma model the first-order field equations in correspond to the local integration of the last two equations of when the term which comes from the scalar-matter coupling Lagrangian is dropped as discussed before. Therefore we have derived the entire set of first-order field equations which are obtained by locally cancelling an exterior derivative on both sides of the equations in . Namely the equations , , , and represent the first-order formulation of the $D$-dimensional low energy massless background coupling of the fully Higgsed heterotic string which is the $D$-dimensional heterotic supergravity.
Before concluding we will present a discussion of an important application of the first-order field equations of the heterotic supergravities. In [@julia2] the locally integrated first-order bosonic field equations of the maximal and IIB supergravities are used to derive the superalgebras that lead to the complete coset constructions of the bosonic sectors of these theories. Similarly the methodology of [@julia2] can be extended to the heterotic supergravities. We will not present the complete coset construction of the heterotic supergravities here and leave it to a future work however we will discuss the outline of deriving the superalgebra of the on-shell coset construction of the heterotic supergravities. The first task in constructing the coset formalism is to assign an algebra generator to each original and dual field in the first-order equations , , , and then to propose a coset map. In our case this map becomes $$\begin{aligned}
\label{yeni23}
\nu &=exp(\frac{1}{2}\phi^{j}H_{j})exp(\chi^{m}E_{m})exp(\phi
K)exp(C_{(1)}^{I}V_{I})exp(\frac{1}{2}A_{(2)}Y)\nonumber\\
\nonumber\\&\quad\times
exp(\frac{1}{2}\widetilde{B}\widetilde{Y})exp(\widetilde{C}^{I}\widetilde{V}_{I})exp(\widetilde{\phi}
\widetilde{K})exp(\widetilde{\chi}^{m}\widetilde{E}_{m})exp(\frac{1}{2}\widetilde{\phi}^{j}\widetilde{H}_{j})\label{32}.\end{aligned}$$ The associated Cartan-form may be defined as $$\label{yeni33}
\mathcal{G}=d\nu\nu^{-1}.$$ From [@julia2] we know that in the doubled formalism coset construction the Cartan-form satisfies a twisted self-duality equation $$\label{yeni43}
\ast\mathcal{G}=\mathcal{SG},$$ with $\mathcal{S}$ being a pseudo-involution of the coset algebra of the generators introduced in . The key ingredient of the coset construction is the requirement that must give us the first-order field equations of the theory. Therefore the method of revealing the coset algebra structure is to calculate in terms of the desired structure constants, then to insert it in and finally to compare the result with the equations , , , to read the structure constants of the coset algebra.
Conclusion
==========
In this work, by locally integrating the second-order field equations which are derived in [@het] and which govern the massless sector of the $D$-dimensional fully Higgsed heterotic string namely the $D$-dimensional heterotic supergravity we have obtained the first-order field equations of the theory which contain only a single exterior derivative acting on the potentials. In these first-order field equations we have introduced dual fields which may be considered as integration constants. The dual fields are nothing but the Lagrange multipliers associated with the Bianchi identities of the field strengths when one treats these field strengths as fundamental fields instead of their potentials [@pope]. The fact which is derived in [@consist2] that as an on-shell condition the coset scalar field equations can completely be decoupled from the gauge fields provides us the usage of the first-order symmetric space sigma model field equations of [@nej2] in our formulation.
In GR the Palatini application of the Ostrogradski method [@ostro] of reducing the derivative order of second-order Lagrangians by including auxiliary fields is a vast research area in recent years especially for the f(R) theories of gravity. The Ostrogradski method have also been effectively used to obtain the first-order formulations of supergravity theories. First-order formulations of supergravities are studied to understand the superpotentials [@pot1; @pot2; @pot3] as well as the supersymmetry transformation laws [@juliasilva]. The reader may find examples of the first-order formalism of supergravity theories in various dimensions in [@juliasilva; @first1; @first2; @first3; @first4; @first5; @first6]. In the general first-order formalism method of these works the field strengths of the basic fields are also considered as independent fields and a first-order Lagrangian is constructed which gives first-order field equations in terms of the basic fields and their field strengths. When the field equations of the field strengths are substituted back in the Lagrangian one recovers the second-order formalism. In comparison with this scheme our first-order field equations of the $D$-dimensional heterotic supergravity do contain the basic fields except the graviton but on the contrary they do not include the field strengths. Instead we have introduced dual fields which may be considered as arbitrary integration constants that algebraically came into the scene as a result of abolishing an exterior derivative on both sides of the second-order field equations. Thus our approach is purely algebraic rather than being formal. We have simply reduced the degree of the field equations without increasing the number of fields to be solved and in this process arbitrary integration constants have arouse. On the other hand we have not constructed the corresponding Lagrangian which would lead to the first-order equations we have obtained. However as we have discussed above such a Lagrangian which would kinematically be different than the one that would appear within the Ostrogradski method would rather be obtained by Lagrange multiplier method that makes use of the Bianchi identities of the field strengths.
The first-order formulation of the $D$-dimensional heterotic supergravity presented in this note has two important implications. The first-order field equations play an important role in the coset construction of the supergravities [@julia2]. Thus as we have briefly discussed in the previous section the equations derived in this note can be considered to be essential ingredients of a possible coset construction of the heterotic supergravities. Secondly since the dual fields introduced in the first-order field equations can be arbitrarily varied one can make use of this fact to generate solutions. Therefore in this respect beside being first-order the integrated field equations containing parameters which can be manipulated become powerful tools in seeking solutions of the heterotic supergravities.
[99]{} E. Bergshoeff, M. de Roo, B. de Wit and P. van Nieuwenhuizen, “*Ten-dimensional Maxwell-Einstein supergravity, its currents, and the issue of its auxiliary fields* ", Nucl. Phys. **B195** (1982) 97. G. F. Chapline and N. S. Manton, “*Unification of Yang-Mills theory and supergravity in ten-dimensions* ", Phys. Lett. **B120** (1983) 105. E. Kiritsis, “*Introduction to superstring theory*", hep-th/9709062. H. Lü, C. N. Pope and K. S. Stelle, “*M-theory/heterotic duality: A Kaluza-Klein perspective*", Nucl. Phys. **B548** (1999) 87, hep-th/9810159. L. Andrianopoli, R. D’Auria, S. Ferrara, P. Fre, M. Trigiante, “*R-R scalars, U-duality and solvable Lie algebras*", Nucl. Phys. **B496** (1997) 617, hep-th/9611014. N. T. Y$\i$lmaz, “*Dualisation of the general scalar coset in supergravity theories*", Nucl. Phys. **B664** (2003) 357, hep-th/0301236. N. T. Y$\i$lmaz, “*The non-split scalar coset in supergravity theories*", Nucl. Phys. **B675** (2003) 122, hep-th/0407006. N. T. Y$\i$lmaz, “*Heterotic string dynamics in the solvable lie algebra gauge*", Nucl. Phys. **B765** (2007) 118, hep-th/0701275. N. T. Y$\i$lmaz, “*An implicit decoupling for the dilatons and the axions of the heterotic string*", Phys. Lett. **B646** (2007) 125, hep-th/0703113. S. Helgason, “**Differential Geometry, Lie Groups and Symmetric Spaces**", (Graduate Studies in Mathematics, vol. 34, American Mathematical Society, Providence, 2001). E. Cremmer, B. Julia, H. Lü and C. N. Pope, “*Dualisation of dualities II : Twisted self-duality of doubled fields and superdualities*", Nucl. Phys. **B535** (1998) 242, hep-th/9806106. C. N. Pope, “**Lecture Notes on Kaluza-Klein Theory**", (unpublished). F. J. de Urries and J. Julve, “*Ostrogradski formalism for higher-derivative scalar field theories*", J. Phys. [**A**31]{} (1998) 6949, hep-th/9802115. B. Julia and S. Silva, “*Currents and superpotentials in classical gauge invariant theories. I: Local results with applications to perfect fluids and general relativity*", Class. Quant. Grav. [**1**5]{} (1998) 2173, gr-qc/9804029. M. Henneaux, B. Julia and S. Silva, “*Noether superpotentials in supergravities*", Nucl. Phys. [**B**563]{} (1999) 448, hep-th/9904003. S. Silva, “*On superpotentials and charge algebras of gauge theories*", Nucl. Phys. [**B**558]{} (1999) 391, hep-th/9809109. B. Julia and S. Silva, “*On first order formulations of supergravities*", JHEP [**0**001]{} (2000) 026, hep-th/9911035. S. Deser and B. Zumino, “*Consistent supergravity*", Phys. Lett. [**B**62]{} (1976) 335. I. Bars and S. W. MacDowell, “*Gravity with extra gauge symmetry*", Phys. Lett. [**B**129]{} (1983) 182. A. Higuchi, “*On the new first order formalism of d = 11 supergravity*", preprint YTP 85-02 (1985). P. Fre, “*On the spinor form of first order gravity and supergravity*", preprint CALT-68-662 (1978). I. Bars and A. Higuchi, “*First order formulation and geometrical interpretation of d = 11 supergravity*", Phys. Lett. [**B**145]{} (1984) 329. R. E. Kallosh, “*Geometry of eleven-dimensional supergravity*", Phys. Lett. [**B**143]{} (1984) 373.
[^1]: On the other hand the unprimed notation $O(10-D+16,10-D)$ is used in [@het] for the usual representation of the generalized orthogonal group generated by the diagonalized indefinite signature metric $\eta=\text{diag}(-,-,...,-,+,+,...,+)$.
[^2]: We adopt the notation of [@nej2] and randomly enumerate the roots in $\Delta_{nc}^{+}$ from $1$ to $n$.
[^3]: The reader should be aware that the plain $\Omega$ with indices $I,J,K,...$ and the bold one $\mathbf{\Omega}$ with the indices $\alpha,\beta,\gamma,...$ are two different objects. We prefer using them together for the sake of conformity with the references.
[^4]: Of course by comparing with one may relate $N_{\alpha,\beta}$ to $K_{\alpha\beta}^{\gamma}$.
[^5]: In [@nej1] the indices $i,j,...$ are taken to run from $1$ to $l$ however in the present manuscript we have preferred using $r$ instead of $l$. Thus the reader may read the coefficients $\{\widetilde{g}_{in}^{k}\}$ and $\{\widetilde{f}_{\alpha n}^{k}\}$ from equations (3.8) and (3.9) of [@nej1] by replacing $l$ with $r$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
Conformal geodesics are distinguished curves on a conformal manifold, loosely analogous to geodesics of Riemannian geometry. One definition of them is as solutions to a third order differential equation determined by the conformal structure. There is an alternative description via the tractor calculus. In this article we give a third description using ideas from holography. A conformal $n$-manifold $X$ can be seen (formally at least) as the asymptotic boundary of a Poincaré–Einstein $(n+1)$-manifold $M$. We show that any curve $\gamma$ in $X$ has a uniquely determined extension to a surface $\Sigma_\gamma$ in $M$, which which we call the *ambient surface of $\gamma$*. This surface meets the boundary $X$ in right angles along $\gamma$ and is singled out by the requirement that it it be a critical point of renormalised area. The conformal geometry of $\gamma$ is encoded in the Riemannian geometry of $\Sigma_\gamma$. In particular, $\gamma$ is a conformal geodesic precisely when $\Sigma_\gamma$ is asymptotically totally geodesic, i.e. its second fundamental form vanishes to one order higher than expected.
We also relate this construction to tractors and the ambient metric construction of Fefferman and Graham. In the $(n+2)$-dimensional ambient manifold, the ambient surface is a graph over the bundle of scales. The tractor calculus then identifies with the usual tensor calculus along this surface. This gives an alternative compact proof of our holographic characterisation of conformal geodesics.
author:
- |
Joel Fine[^1], Yannick Herfray[^2]\
[*Départment de Mathématique, Université Libre de Bruxelles,* ]{}\
[*CP 218, Boulevard du Triomphe, B-1050 Bruxelles, Belgique.*]{}
bibliography:
- 'Biblio.bib'
title: An ambient approach to conformal geodesics
---
Introduction
============
The Fefferman–Graham approach to conformal geometry
---------------------------------------------------
From a mathematical perspective, holography has its roots in the ambient metric approach to conformal geometry pioneered by Fefferman and Graham [@fefferman_conformal_1985; @fefferman_ambient_2012]. We give a rapid overview here, with more details provided later in the article, for the benefit of those unfamiliar with these constructions.
Given a compact $n$-manifold $X$ and a conformal class $[h]$ of metrics with signature $(p,q)$ on $X$, Fefferman and Graham gave two different ways to extend $(X,[h])$ to genuine metrics solving two versions of Einstein’s equations, one a $(n+1)$-manifold with metric of signature $(p+1,q)$ and $\operatorname{Ric}(g)=-ng$, the other a $(n+2)$-manifold with metric of signature $(p+1,q+1)$ and $\operatorname{Ric}=0$. We focus in this introduction on the $(n+1)$-dimensional extension but in the main body of the article both extensions are used. From now on we focus on Riemannian signature $(p,q)=(n,0)$ but the results extend straightforwardly to other signatures.
The $(n+1)$-dimensional manifold $M$ is abstractly diffeomorphic to $X \times [0,1)$. Given a representative $h_0 \in[h]$ of the conformal structure on $X$, there exists a unique way to identify $M$ and $X \times [0,1)$ under which $g$ has the form $$\label{conformally-compact}
g = \frac{d \rho^2 + h(\rho)}{\rho^2}$$ Here $\rho$ is the projection onto the factor $[0,1)$ and $h(\rho)$ is a path of metrics on $X$ with $h(0) = h_0$. More correctly, this is a *formal* construction. Things are cleanest when $n$ is odd. In this case there is a formal power series expansion of $h$ in even powers of $\rho$, of the form $$h(\rho) = h_0 + h_2 \rho^2 + h_4 \rho^4 + \cdots
\label{FG-expansion}$$ The metric is Einstein in the formal sense that $\operatorname{Ric}(g)+ng = {\mathcal{O}}(\rho^{-\infty})$. (The word “formal” here means no attention is paid to convergence; ${\mathcal{O}}(\rho^{-\infty})$ means all coefficients of the corresponding power series vanish.) The expansion is canonically associated to $h_0$: the coefficients $h_{2k}$ are completely determined by the curvature of $h_0$ and its derivatives. Moreover, whilst different choices of representatives of $[h]$ lead to different expansions, the resulting metrics are *isometric*, just not in a way which is compatible with the identifications $M \cong X \times [0,1)$. It follows that the Riemannian manifold $(M,g)$ is canonically associated to the conformal manifold $(X,[h])$
When $n \geq 4$ is even, the analogous expansion can be carried out up to $O(\rho^{n-1})$ but the next step is obstructed (at least in general). The result is a metric which has $\operatorname{Ric}(g)+ng = O(\rho^n)$. (The case $n=2$ is exceptional; we momentarily ignore it here, but treat it carefully in the main body of the article.) No matter the parity of the dimension, to a conformal $n$-manifold $(X,[h])$ Fefferman and Graham produce a germ to order $n$ of a formal Riemannian Einstein metric $(M,g)$, called the *formal Poincaré–Einstein metric with conformal infinity $(X,[h])$*.
Since $(M,g)$ is canonically associated to $(X,[h])$ the Riemannian geometry of $M$ can be used to understand the conformal geometry of $X$. This is the mathematical point of view on holography and it has been extremely successful, giving rise, for example, to new conformally invariant curvature tensors and differential operators (see for example [@fefferman_q-curvature_2002]). This article applies this “Fefferman–Graham principle” to the study of *curves* in $(X,[h])$.
The ambient surface of a curve $\gamma \subset X$
-------------------------------------------------
To any curve $\gamma \colon I \to X$, we canonically associate an *ambient surface* $\Sigma \subset M$. The surface meets the boundary at right angles in the curve $\gamma$. The area of such a surface is necessarily infinite, but there is a finite renormalisation, introduced in [@graham_conformal_1999] and which has been the subject of much investigation in both the mathematics and physics literature (see, for example, [@alexakis_renormalized_2010; @ryu_holographic_2006]). Given $h_0 \in [h]$ and the expansion , one shows that the area $A(\epsilon)$ of $\Sigma$ contained in the region $\rho \geq \epsilon$ has an expansion in powers of $\epsilon$ of the form $$\label{area-expansion}
A(\epsilon) = l(\gamma)\epsilon^{-1} + \mathcal{A} + O(\epsilon)$$ for some $\mathcal{A} \in {\mathbb{R}}$ (where $l(\gamma)$ is the length of $\gamma
\subset X$ with respect to $h_0$). Crucially, this constant term $\mathcal{A}$ *does not depend on the choice of $h_0 \in [h]$*. It is called the *renormalised area of $\Sigma$*.
Our first result is the following.
\[existence-uniqueness-ambient-surface\] Let $(X,[h])$ be a conformal $n$-manifold with formal Poincaré–Einstein extension $(M,g)$. Given any embedded curve $\gamma \subset X$ there exists a unique surface ${\Sigma}\subset M$ which is a critical point for renormalised area and which meets $X$ in right angles along $\gamma$.
We call the distinguished surface of Theorem \[existence-uniqueness-ambient-surface\] the *ambient surface associated to $\gamma$* and we denote it by $\Sigma_\gamma$. We remark that, just as for the Fefferman–Graham construction, in general the ambient surface is purely formal. When $n$ is odd, it has a uniquely determined formal expansion to all orders. When $n$ is even, it exists to the same order as the Fefferman–Graham metric. For comments on the global existence question see §\[global\] below.
We say a few words about the proof of Theorem \[existence-uniqueness-ambient-surface\]. Surfaces which are critical points of renormalised area are automatically minimal. One can see this by considering variations of the surface which are compactly supported away from the boundary, and so which only affect the constant term in $A(\epsilon)$. Analysing the minimal surface equation, one sees that it has two indicial roots, $0$ and $3$. (This is done in [@alexakis_renormalized_2010] in dimension $n=2$, and in [@graham_higher_2017] in higher dimensions.) This means that in the local existence problem, one can expect to prescribe the boundary value (Dirichlet data) corresponding to the root $0$, and a second piece of boundary data, corresponding to the root $3$ and which one can interpret as Neumann data. Once both Dirichlet and Neumann data are fixed, the solution is completely determined. (The simple analogue is that given a pair of functions $\phi,\psi$ on $S^1$ there is a unique harmonic function $h$ defined on a neighbourhood of $S^1 \subset {\mathbb{R}}^2$ with $h|_{S^1} = \phi$ and $\partial_n h = \psi$.)
It turns out that the vanishing of the Neumann data is equivalent to the surface being critical for *renormalised* area, under perturbations which are not compactly supported and actually move the boundary curve. (This was shown in [@alexakis_renormalized_2010] for dimension $n=2$, and in arbitrary dimensions is shown below.) This explains why one should expect a unique critical surface to emanate from any curve $\gamma$.
We use this ambient surface $\Sigma_\gamma$ to study the conformal geometry of $\gamma \subset X$. A first point concerns the canonical parametrisations of $\gamma$. These are conformal analogues of parametrisation by arc length in Riemannian geometry, first introduced in [@bailey_conformal_1990]. Given $h \in [h]$, the curve $\gamma \colon I \to X$ is *conformally parametrised* if the following third order equation is satisfied $$h\left({\dot{\gamma}}, |{\dot{\gamma}}|^{-1}\nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|v\right) + \frac{1}{2}|v|^2 {\dot{\gamma}}- h^{-1}\left( P_h({\dot{\gamma}}) \right) \right)
=0
\label{conformal-parametrisation}$$ Here $\nabla$ denotes the Levi-Civita connection of $h$ and we use the short hand $v = \nabla_{{\dot{\gamma}}} \left( |{\dot{\gamma}}|^{-2}{\dot{\gamma}}\right)$. Meanwhile, $$P_h = \frac{1}{n-2} \operatorname{Ric}^0_h + \frac{R_h}{2n(n-1)}h$$ is the Schouten tensor of $h$ (with $R_h$ the scalar curvature and $\operatorname{Ric}_h^0$ the trace-free Ricci curvature).
One can check that a conformal parametrisation always exists, is uniquely determined up to the action of $\operatorname{PSL}(2,{\mathbb{R}})$ and, moreover, the definition is independent of the choice of $h\in [h]$ (remembering of course that $\nabla$, $v$ and $P_h$ will all change). As presented here, the condition appears completely mysterious. There is a neat interpretation in terms of tractor calculus [@bailey_thomass_1994]. Our next result gives an alternative geometric interpretation of these preferred conformal parametrisations in terms of our ambient surface. (We make connection the with the tractor point of view later.) Given a parametrised curve $u \colon I \to X$ with image $\gamma$, we say that $U \colon I \times [0,1) \to M$ extends $u$ if $U(s,0) = u(s)$ for all $s\in I$. In the following result, we treat $I \times [0,1) \subset {\mathbb{R}}\times [0,\infty)$ as a subset of the upper half-space with hyperbolic metric $g_{\mathrm{hyp}} = t^{-2}(dt^2+ ds^2)$.
\[isometric-parametrisation\]
- Given a parametrised curve $u \colon I \to X$ with image $\gamma$ there exists an extension $U \colon I \times [0,1) \to M$ parametrising $\Sigma_\gamma$ and such that $U^*(g) =g_{\mathrm{hyp}} + O(t^2)$, where ${\mathcal{O}}(t^2)$ refers to the norm of a tensor with respect to $g_{\mathrm{hyp}}$.
- There exists an extension $U$ of $u$ with $U^*(g)=g_{\mathrm{hyp}} +{\mathcal{O}}(t^3)$ if and only if $u$ is a conformal parametrisation of $\gamma$.
Next we turn to conformal geodesics. These are a distinguished class of curves in $(X,[h])$, singled out purely by the conformal structure. Fix a representative Riemannian metric $h \in [h]$; a curve $\gamma \colon I \to X$ is a conformal geodesic if it solves the third order equation: $$\label{CGE-third-order-form}
\nabla_{{\dot{\gamma}}} \ddot{\gamma} = 3 \frac{h\left({\dot{\gamma}}, \ddot{\gamma}\right)}{|{\dot{\gamma}}|^2}\ddot{\gamma} -\frac{3}{2} \frac{|\ddot{\gamma}|^2}{|{\dot{\gamma}}|^2}{\dot{\gamma}}+ |{\dot{\gamma}}|^2 h^{-1}\left(P_h({\dot{\gamma}})\right) -2 P_h\left({\dot{\gamma}},{\dot{\gamma}}\right) {\dot{\gamma}}.$$ One can check that this definition is conformally invariant: the transformation laws for $\nabla$ and $P_h$ under change of representative metric $h$ ensure that if $\gamma$ solves for $h$ it also solves it for any metric conformal to $h$.
Again, in this form appears mysterious and again there is a concise interpretation of the equation using tractor calculus (given in [@bailey_thomass_1994] and also outlined briefly in the main body of the article). Our next result is an alternative geometric interpretation of conformal geodesics from the point of view of the ambient surface. In the following, $\gamma \colon I \to X$ is a conformally parametrised curve and we choose $U \colon I \times [0,1) \to M$ to be an extension parametrising the ambient surface $\Sigma_\gamma$ which is isometric to ${\mathcal{O}}(t^3)$, as in Theorem \[isometric-parametrisation\]. Write $K$ for the second fundamental form of $\Sigma_\gamma$. We treat $|K|$ as a function on $I \times [0,1)$ via the parametrisation $U$.
The ambient surface is asymptotically totally geodesic: $|K| = {\mathcal{O}}(t^2)$. Moreover, $|K| = {\mathcal{O}}(t^3)$ if and only $\gamma$ is a conformal geodesic.
So a curve is a conformal geodesic precisely when its ambient surface is totally geodesic to higher order than one would normally expect to see.
Global questions {#global}
----------------
As we have stressed, the discussion of the ambient surface in this article is purely local. We close this introduction with a few comments on the corresponding *global* existence problem. There is great interest, both mathematically and physically, in studying genuine (as opposed to purely formal) Poincaré–Einstein metrics, i.e. Einstein metrics on a compact manifold $M$ with boundary $X$ which have the form near $X$. A central question is the Dirichlet problem: given a conformal manifold $(X,[h])$ does it arise as the infinity of a Poincaré–Einstein metric on a compact manifold $M$ with boundary $X$? In general this remains open, but there are a wealth of examples (for example [@graham_einstein_1991; @calderbank_einstein_2004; @mazzeo_gluing_2006]). One can also pose the Dirichlet problem for minimal surfaces: let $M$ be a compact manifold with boundary $X$ and let $g$ be a Poincaré–Einstein metric on the interior; given a closed curve $\gamma$ in $X$, when is it possible to fill $\gamma$ by a closed minimal surface $\Sigma$? This question was answered positively by Anderson in the case of hyperbolic space itself [@anderson_complete_1982] and Alexakis and Mazzeo showed that for hyperbolic 3-manifolds it is possible to use degree theory to count the number of minimal solutions [@alexakis_renormalized_2010]. The general problem however remains open.
A natural question in the context of this article is the following refinement of this Dirichlet problem: when is it possible to fill $\gamma$ by a closed surface $\Sigma$ which is a critical point of *renormalised* area? There is, to the best of our knowledge, only one result in this direction, due to Alexakis and Mazzeo [@alexakis_renormalized_2010]. They prove that in hyperbolic 3-space $\mathbb{H}^3$ the only surfaces which are critical points of renormalised area are the totally geodesic copies of $\mathbb{H}^2$, which fill the closed conformal geodesics on the boundary $S^3$ (with its standard conformal structure).
Clearly, curves which can be globally filled in this way are special. We can also see this from our above discussion about Dirichlet and Neumann data. Recall that in the local existence problem for minimal surfaces, we had the right to prescribe both Dirichlet data (the curve $\gamma$) *and* the Neumann data, which was set to zero by the requirement that the surface be critical for renormalised area. In the global problem however, one will determine the other, just as one can prescribe either the Dirichlet or Neumann data of a harmonic function on the disk, but not both. From this perspective, critical surfaces are the zeros of a Dirichlet-to-Neumann map. Based on this, one might optimistically hope that generically critical surfaces are isolated (the above case of $\mathbb{H}^3$ is exceptional due to the abundance of symmetries) and even, in some circumstances, for there to be finitely many of them. In any case, one might reasonably expect the critical surfaces and their corresponding boundary curves to form an interesting collection of objects, pertinent to the study of both the Riemannian geometry of $M$ and the conformal geometry of $X$.
Acknowledgements
----------------
We would like to thank Robin Graham for helpful discussions about conformal geodesics. JF was supported by ERC consolidator grant “SymplecticEinstein” 646649. YH was supported by an FNRS *chargé de recherche* fellowship.
Conformal geometry and conformal geodesics {#Sec: Conformal geometry and conformal geodesics}
==========================================
In this section we review well known facts about conformal geometry and conformal geodesics. This will serve mainly to fix the conventions that we will use in the following sections.
Conformal manifolds
-------------------
Let $X$ be a $n$-dimensional manifold. Recall that the bundle of $1$-densities $|\bigwedge|X$ is the real line bundle associated to the frame bundle of $X$ with respect to the representation $M \mapsto |det(M)|^{-1}$ of $GL(n)$. This bundle is always trivial. If $X$ is orientable 1-densities coincide with $n$-forms but on non-orientable manifolds $1$-densities (not $n$-forms) are the right type of objects needed for integration. The *bundle of scales* is $L = \left(|\bigwedge|X\right)^{-\frac{1}{n}}$. This is an oriented real line bundle over $X$ (we will only consider positive sections).
We will say that $(X, {[h]})$ is a *conformal manifold* if $X$ is equipped with a non-degenerate symmetric bilinear form ${[h]}$ with values in $L^2$, i.e ${[h]}$ is a section of $L^2 \otimes S^2 T^*X$ ($S^2$ stands for symmetric tensor product). A choice of scale $\tau \in \Gamma\left[L\right]$ then amounts to a choice of “representative” $h = \tau^{-2}{[h]}\in \Gamma\left[S^2 T^*X \right]$.
Conformal geodesics {#subsec: Conformal geodesics}
-------------------
Let $(X , {[h]})$ be a $n$-dimensional conformal manifold. We briefly review some facts about conformal geodesics when $n>2$, the particular case $n=2$ is also discussed at the end of this subsection. Our main references are [@bailey_conformal_1990], [@bailey_thomass_1994] and [@tod_examples_2012], see also [@eastwood_uniqueness_2014].
Let $\gamma {:}{\mathbb{R}}\to X$ be a curve in $(X, {[h]})$ parametrised by $s$. It is a *parametrised conformal geodesic* if and only if it satisfies the following third order differential equation[^3]: $$\label{1- Conformal Geodesic Equation}
\nabla_{{\dot{\gamma}}} \ddot{\gamma} = 3 \frac{h\left({\dot{\gamma}}, \ddot{\gamma}\right)}{|{\dot{\gamma}}|^2}\ddot{\gamma} -\frac{3}{2} \frac{|\ddot{\gamma}|^2}{|{\dot{\gamma}}|^2}{\dot{\gamma}}+ |{\dot{\gamma}}|^2 h^{-1}\left(P_h({\dot{\gamma}})\right) -2 P_h\left({\dot{\gamma}},{\dot{\gamma}}\right) {\dot{\gamma}}.$$ Here $h$ is a choice of representative for ${[h]}$ and $P_h = \frac{1}{n-2}Ric_h\big|_{0} + \frac{1}{2n(n-1)} R_h\; h$ is its Schouten tensor. One can show, that the vanishing of does not depend on this choice. See [@bailey_conformal_1990] for a proof.
It is also shown in [@bailey_conformal_1990] that the normal part of (i.e. the component which is orthognal to ${\dot{\gamma}}$) does not depend on the chosen parametrisation. We thus call (unparametrised) *conformal geodesics* those curves which satisfy the normal projection of . They can be thought as analogues of geodesics from Riemannian geometry with the normal part of corresponding to the geodesic equations.
In Riemannian geometry there is a natural arc-length parametrisation for curves (unique up to a remaining linear reparametrisation freedom). The tangential part of gives the analogue for conformal geometry: as is shown in [@bailey_conformal_1990] one can always (at least locally) choose a parametrization such that the tangential part of is satisfied. The remaining freedom in reparametrisation is then the projective linear group $$s \mapsto \frac{as+b}{cs+d}, \qquad\text{with} \begin{pmatrix}
a & b\\c & d
\end{pmatrix} \in SL(2, {\mathbb{R}}).$$ We will say that curves satisfying the tangential part of have been given a *preferred conformal parametrisation*.
For most purposes, equation is cumbersome. Introducing an auxiliary vector field $v$, it can however be rewritten as,
$$\begin{aligned}
\nabla_{{\dot{\gamma}}} {\dot{\gamma}}&= |{\dot{\gamma}}|^2 v -2 h\left( v, {\dot{\gamma}}\right) {\dot{\gamma}}\label{1- Eq: First Conformal Geodesic Eq Bis} \\
\nabla_{{\dot{\gamma}}} v &= h\left( v, {\dot{\gamma}}\right)v - \frac{1}{2}|v|^2 {\dot{\gamma}}+ h^{-1}\left( P_h({\dot{\gamma}}) \right). \label{1- Eq: Second Conformal Geodesic Eq Bis}
\end{aligned}$$
These are equations $(1)$ and $(2)$ of [@tod_examples_2012]. Solving equation for $v$ one obtains the following, even more compact, system of equations:
$$\begin{aligned}
v &=\nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|^{-2} {\dot{\gamma}}\right) \label{1- Eq: First Conformal Geodesic Eq}\\
|{\dot{\gamma}}|^{-1}\nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|v\right) &= - \frac{1}{2}|v|^2 {\dot{\gamma}}+ h^{-1}\left( P_h({\dot{\gamma}}) \right). \label{1- Eq: Second Conformal Geodesic Eq}
\end{aligned}$$
This is this form that we will mainly use in the following. It will be especially well adapted when we come to tractors but it will also appear naturally in the holographic discussion.
We now briefly discuss the $n=2$ case. Since the Schouten tensor is only well defined for $n>2$ the same is true for the conformal geodesic equations . One can however extend to the two dimensional case by making a choice of Möbius structure (see [@calderbank_mobius_2006]). A pragmatic point of view on Möbius structures is that they amounts to a choice of traceless symmetric tensor $P_0$ whose behaviour under change of representatives of the conformal class mimics that of the traceless Ricci tensor $Ric_h|_0$. In the following we assume that we have made a such a choice and all formula then straightforwardly extend to the case $n=2$ by replacing the (missing) trace-free Schouten tensor by this $P_0$. See however [@calderbank_mobius_2006], [@burstall_conformal_2010] for more details on the underlying geometry.
The ambient surface and conformal geodesics {#Sec: Holographic Prescription for Conformal Geodesics}
===========================================
The Poincaré–Einstein metric of a conformal manifold
----------------------------------------------------
Let $(X , {[h]})$ be a $n$-dimensional conformal manifold. Let $(H , g)$ be the associated Poincaré–Einstein (formal) metric defined in terms of the Fefferman-Graham expansion [@fefferman_conformal_1985], [@fefferman_ambient_2012]. By construction, $H$ is the interior of a $n+1$-dimensional compact manifold $\bar{H}$ with boundary $\partial \bar{H} = X$ and if $x {:}\bar{H} \to {\mathbb{R}}$ is boundary defining function, $X= x^{-1}(0)$, $dx|_X \neq 0$ then $\bar{g} = x^2 g$ smoothly extends to $\bar{H}$. The leading order of the expansion for $g$ is $$\label{3- FG Third Normal Form2}
g = \frac{1}{x^2} \left(dx^2 + \left[ h - x^2 P + {\mathcal{O}}\left(x^4\right) \right] \right).$$ Where $P$ is the Schouten tensor of $h$ if $n>2$, and is given by a choice of Möbius structure if $n=2$. The boundary defining function $x$ used to write down this expansion is not unique. However once we make a choice of representative $h$ one can always find $x$ such that the expansion is of the form and this fixes $x$ uniquely, see [@fefferman_conformal_1985], [@fefferman_ambient_2012]. Such boundary defining functions are called “special” or “geodesic”.
In what follows we won’t need the precise form of the expansion beyond third order, so that for the rest of this work the equation (formally) defines the metric $g$.
The ambient surface of a conformal curve {#subsec: Dual surface to a conformal curve}
----------------------------------------
We now take ${\gamma}{:}[0,1] \to X$ to be a curve in $X$ parametrised by a parameter $s$.
\[3- Def: Surface with asymptotic boundary\] We will say that an embedding ${\Sigma}{:}[0,1] \times [0,1) \to \bar{H}$ is a *surface with asymptotic boundary ${\gamma}$* if
- ${\Sigma}\left(s, 0\right)= {\gamma}(s)$,
- the image of ${\Sigma}$ intersects the asymptotic boundary at right angles, ${\Sigma}\perp X$.
Orthogonality in the second point is taken with respect to the metric $\bar{g}$. This metric is only defined up to scale, but orthogonality does not depend on the particular choice of scale. In what follows we will frequently abuse notation by using ${\Sigma}$ to denote the restriction to $[0,1]\times(0,1)$, whose image lies in the interior $H$.
Let $x$ be a choice of boundary defining function. If ${\Sigma}$ is a surface with asymptotic boundary ${\gamma}$, we take ${\Sigma}_{\epsilon}$ to be defined by ${\Sigma}_{\epsilon} \coloneqq {\Sigma}\cap \left\{ x \geq \epsilon \right\}$ and its boundary to be $\gamma_{\epsilon} = \partial {\Sigma}_{\epsilon} $. As is explained in [@graham_conformal_1999], the area of ${\Sigma}_{\epsilon}$ diverges as the length of ${\gamma}_{\epsilon}$, one can however “cut off” this diverging part to obtain the renormalised area.
\[3- Def: Renormalised Area\] The *renormalised area* $\mathcal{A}\left({\Sigma}\right)$ of a surface ${\Sigma}$ with asymptotic boundary ${\gamma}$ is defined as $$\label{3- Renormalised Area}
\mathcal{A}\left({\Sigma}\right) = lim_{\epsilon \to 0} \left( \int_{{\Sigma}_{\epsilon}} dA - \int_{{\gamma}_{\epsilon}} dl \right).$$ Where $dA$ and $dl$ are the volume forms induced by the Poincaré–Einstein metric $g$. (It might seem that $\mathcal{A}({\Sigma})$ depends on a choice of special boundary defining function, but [@graham_conformal_1999] shows that actually it does not.)
The next theorem is the main result of this subsection. It allows us to associate to a curve ${\gamma}$ in $X$ a unique ambient surface ${\Sigma}_{{\gamma}}$ in $H$.
\[3- Thrm: Holographic dual\]\
Let ${\gamma}$ be a curve in $\left(X ,[h]\right)$, let $\left(H , g \right)$ be the (formal) associated Poincaré–Einstein metric , then there is a unique (formal) surface ${\Sigma}_{{\gamma}}$ in $H$ which both is a critical point of the renormalised area and has asymptotic boundary ${\gamma}$. We will say that ${\Sigma}_{{\gamma}}$ is the ambient surface associated to the curve ${\gamma}$.
What is more, this ambient surface can be asymptotically described by the expansion,
[-9pt]{}[-0pt]{} $$\label{3- Holographic dual, expansion}
{\Sigma}_{{\gamma}} \left\{\begin{array}{ccccccccccc}
x\left(s,t\right) & =& 0 &+& t |{\dot{\gamma}}| &+& 0 &+&\frac{t^3 }{3}\left(-\frac{3}{4}|{\dot{\gamma}}|^3|v|^2 - \frac{1}{4} |{\dot{\gamma}}| A_I A^I\right) &+& {\mathcal{O}}\left(t^4\right) \\ \\
y^i\left(s,t\right) & =& {\gamma}^i(s) &+& 0 &+& \frac{t^2}{2} |{\dot{\gamma}}|^2 v^i &+& 0 &+& {\mathcal{O}}\left(t^4\right)
\end{array} \right.$$
where $v$ satisfies the first conformal geodesic equation : $$\label{3- leading minimal surface eq }
v =\nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|^{-2} {\dot{\gamma}}\right)$$ and $$\label{3- Conformal Parametrisation Def1}
-\frac{1}{2}A^{I}A_{I} \coloneqq h\left({\dot{\gamma}}, |{\dot{\gamma}}|^{-1}\nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|v\right) + \frac{1}{2}|v|^2 {\dot{\gamma}}- h^{-1}\left( P_h({\dot{\gamma}}) \right) \right)$$ where the right-hand-side of is the tangential part of the second conformal geodesic equation .
The expansion is not unique but can be invariantly defined as a choice of conformal parametrisation up to order three, that is such that the induced metric is $$\label{3- Induced metric1}
{\Sigma}^*g = f\left(s,t\right)\left[\begin{pmatrix}
1 & 0 \\0 & 1
\end{pmatrix} + {\mathcal{O}}(t^3)\right].$$
In the following we will not use the details of ${\Sigma}_{\gamma}$ beyond third order in the parameter $t$, consequently we could just as well have taken the expansion as our definition for the ambient surface.
We break down the proof of the above theorem in three lemmas from which it is a direct consequence.
\[3- Lemma: Conformal parametrisation\] Let ${\gamma}$ be a curve in $\left(X , {[h]}\right)$, let $\left(H , g \right)$ be the associated Poincaré–Einstein metric and let ${\Sigma}$ be a surface in $H$ with asymptotic boundary ${\gamma}$, then ${\Sigma}$ can always be described asymptotically by the expansion
[-5pt]{}[-0pt]{} $$\label{3- Asymptotic surface, expansion}
{\Sigma}\left\{\begin{array}{ccccccccccc}
x\left(s,t\right) & =& 0 &+& t |{\dot{\gamma}}| &+& 0 &+&\frac{t^3 }{3}\left(-\frac{3}{4}|{\dot{\gamma}}|^3|v|^2 - \frac{1}{4} |{\dot{\gamma}}| A_I A^I\right) &+& {\mathcal{O}}\left(t^4\right) \\ \\
y^i\left(s,t\right) & =& {\gamma}^i(s) &+& 0 &+& \frac{t^2}{2} |{\dot{\gamma}}|^2 v^i &+& \frac{t^3}{3}\; n^i &+& {\mathcal{O}}\left(t^4\right)
\end{array} \right.$$
where $v$ and $n$ satisfy $h\left({\dot{\gamma}}, v \right) = h\left({\dot{\gamma}}, \nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|^{-2} {\dot{\gamma}}\right) \right)$ and $h\left({\dot{\gamma}}, n\right)=0$ but are otherwise free parameters while $A_I A^I$ is given by ($v$ is however not supposed to satisfy ).
This description is not unique but can be invariantly defined as a choice of conformal parametrisation up to order three, $$\label{3- Induced metric2}
{\Sigma}^*g = f(s,t) \left[ \begin{pmatrix}
1 & 0 \\0 & 1
\end{pmatrix} + {\mathcal{O}}(t^3) \right].$$
\
Let ${\Sigma}$ be a surface in $H$ with asymptotic boundary ${\gamma}$, by definition \[3- Def: Surface with asymptotic boundary\] it must have an expansion of the form: $$\label{3- Lemma(conformal coordinate)2}
{\Sigma}\left\{\begin{array}{ccccccccccc}
x\left(s,t\right) & =& 0 &+& t x_1 &+& \frac{t^2}{2} x_2 &+& \frac{t^3}{3} x_3 &+& {\mathcal{O}}\left(t^4\right) \\ \\
y^i\left(s,t\right) & =& {\gamma}^i(s) &+& t\; \alpha(s)\; {\dot{\gamma}}^i(s) &+& \frac{t^2}{2} |{\dot{\gamma}}|^2 v^i &+& n^i &+& {\mathcal{O}}\left(t^4\right)
\end{array} \right.$$ We can now fix some of the functions in this expansion by a choice of isothermal coordinates (satisfying ). This is done by solving order by order equations and results in the expansion with $h\left({\dot{\gamma}}, v \right) = h\left({\dot{\gamma}}, \nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|^{-2} {\dot{\gamma}}\right) \right)$, $h\left({\dot{\gamma}}, n\right)=0$ and $A_I A^I$ given by .
From the detailed derivation of the above proof one deduces the exact form of the function $f$ in .
\[3- Crllr: Induced metric\] Let ${\Sigma}$ be a surface with asymptotic boundary ${\gamma}$ in the parametrisation given by lemma \[3- Lemma: Conformal parametrisation\] then the induced metric is, in the coordinate basis $\{\partial_s, \partial_t\}$, $$\label{3- Induced metric}
{\Sigma}^*g = \frac{1}{t^2}
\begin{pmatrix}
1 & 0\\ 0 & 1
\end{pmatrix}\left(1 + 0+ t^2\left(- \frac{1}{3}A_I A^I\right)\right) + {\mathcal{O}}(t).$$
The next two lemmas are generalisations to arbitrary dimension of results established in [@alexakis_renormalized_2010] for $n=2$.
We first consider variations of the renormalised area with compact support. Let ${\gamma}$ be a curve in $\left(X ,[h]\right)$, let $\left(H , g \right)$ be a Poincaré–Einstein metric of the form . Let ${\Sigma}(u)$ be any one dimensional family of surfaces in $H$ with asymptotic boundary ${\gamma}$, in particular it must have fixed boundary $\partial {\Sigma}(u) = {\gamma}$. Let ${\Sigma}(u)$ start at a surface ${\Sigma}(0) = {\Sigma}$ parametrised as in . We write ${\Sigma}(u)$ as $$\label{3- Lemma(first variation)1}
{\Sigma}(u) = {\Sigma}+ u \Phi + {\mathcal{O}}(u^2)$$ where $\Phi \in {\Gamma}[N_{{\Sigma}}]$ is a section of the normal tangent bundle to ${\Sigma}$ in $H$ (i.e a section of the orthogonal complement of $T{\Sigma}$ in $H$). Remark that since ${\Sigma}(u)$ is supposed to have asymptotic boundary ${\gamma}$ we must have $\Phi\big|_X=0$. In fact, since the surfaces $\Sigma(u)$ meet $X$ at right angles, $\Phi$ vanishes to first order along $X$.
\[3- Lemma: first variations with compact support\] The renormalised area of the surface $\Sigma$ is stationary for variations fixing the boundary if and only if the surface ${\Sigma}$ is minimal, that is $Tr K =0$, where $K$ denotes the second fundamental form of ${\Sigma}$. Setting $Tr K =0$ fixes ${\Sigma}$ (at least its formal expansion) uniquely up to the choice of the normal part of the coefficient $n$ in the expansion . In particular $Tr K$ vanishes to leading order in $t$ if and only if $v$ in satisfies the first conformal geodesic equation : $$\label{3- Minimality, leading order}
v =\nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|^{-2} {\dot{\gamma}}\right).$$
\
Since we only consider variations fixing the asymptotic boundary $\partial {\Sigma}(u) = {\gamma}$ the variation of the renormalised are coincides with the variation of the area. The surface ${\Sigma}$ is thus critical for such variations if and only if its mean curvature $Tr K$ vanishes. As was already discussed in the introduction it is known that this equation has indicial roots $0$ and $3$. (See, for example, [@alexakis_renormalized_2010] or [@graham_higher_2017].) This implies that one can formally solve for this equation order by order in the expansion and that the solution is unique up to a choice of terms at order $0$ and $3$. A direct computation with the parametrization then shows that this “free” terms are ${\gamma}$ and the normal part of $n$ (recall that $h\left({\dot{\gamma}}, n\right)=0$ by our choice of coordinates). All other terms are then uniquely constrained by solving $Tr K$ order by order. In particular one finds that $v$ must satisfy the first conformal geodesic equation , we however postpone a precise derivation of this fact to the next section (see Corollary \[3- Crllr: mean curvature\]).
We now consider the variations that do not necessarily fix the conformal boundary. Let ${\Sigma}$ be a minimal surface, $Tr K =0$, with asymptotic boundary ${\gamma}$. Let ${\gamma}(u)$ be any one dimensional family of curves in $X$ starting at ${\gamma}(0)={\gamma}$ and let ${\Sigma}(u)$ be any one dimensional family of surfaces in $H$ with asymptotic boundary ${\gamma}(u)$ and starting at ${\Sigma}(0)= {\Sigma}$. We write ${\gamma}(u)$ as $$\label{3- Lemma(first variation)2}
{\gamma}(u) = {\gamma}+ u \phi + {\mathcal{O}}(u^2)$$ where $\phi \in {\Gamma}[N_{{\gamma}}]$ is a section of the normal tangent bundle to ${\gamma}$ in $X$.
\[3- Lemma: first variation with non-compact support\] The first variation of the renormalised area with respect to this family is $$\left.\frac{d \mathcal{A}\left( {\Sigma}_t \right) }{d t}\right|_{t=0} = -\int_{{\gamma}}ds\;|{\dot{\gamma}}|^{-2} \; h\left( \phi, n\right) .$$ This vanishes if and only if the normal part of $n$ in vanishes and thus if and only if $$n=0.$$
\
Let ${\Sigma}_{\epsilon}(u) \coloneqq {\Sigma}(u) \cap \left\{ x \geq \epsilon \right\}$ and $\gamma_{\epsilon}(u) = \partial {\Sigma}_{\epsilon}(u) $ its boundary. Let $\Phi$ be a section of the normal bundle defined by . We note $dA(u)$ the volume-form induced by $g$ on ${\Sigma}(u)$. Finally we note $dl_{\epsilon}$ the volume-form induced by $g$ on ${\gamma}_{\epsilon}$.
We first remark that $$\begin{aligned}
\left.\frac{d }{d u} \left(\int_{{\Sigma}_{\epsilon}(u)} dA(u)\right)\right|_{u=0} &= \int_{{\Sigma}_{\epsilon}} \left.\frac{d}{d u} \left(dA(u)\right)\right|_{u=0} + \left.\frac{d }{d u}\left(\int_{{\Sigma}_{\epsilon}(u)} dA \right)\right|_{u=0}\end{aligned}$$ and treat each of this integrals separately. The first term is $$\begin{aligned}
\label{3- Lemma(first variation)5}
\begin{array}{lll}
\int_{{\Sigma}_{\epsilon}} \left.\frac{d}{d u} \left(dA(u)\right)\right|_{u=0}&= - \int_{{\Sigma}_{\epsilon}} tr_{g} K(\Phi) dA
\end{array}\end{aligned}$$ The second integral can be rewritten as $$\begin{aligned}
\label{3- Lemma(first variation)6}
\begin{array}{lll}
\left.\frac{d }{d u}\left(\int_{{\Sigma}_{\epsilon}(u)} dA \right)\right|_{u=0} &=\left.\frac{d }{d u}\left( \int_{\epsilon = {\Sigma}_u^* x }^{\infty}\int_{{\gamma}_{\epsilon}} dA \right)\right|_{u=0}
&=\left.\frac{d }{d u}\left(\int_{\epsilon= {\Sigma}^* x + dx(\Phi) u }^{\infty}\int_{{\gamma}_{\epsilon}} dA\right)\right|_{u=0}\\ \\
&= \int_{{\gamma}_{\epsilon}} dl \;\frac{ dx(\Phi)}{|{\Sigma}^* dx|^2}\left| \nabla x \right|
&= \int_{{\gamma}_{\epsilon}} dl \;\frac{ dx(\Phi)}{|{\Sigma}^* dx|}
\end{array}\end{aligned}$$ The second line following from the fact that in order to increase the function ${\Sigma}^*x$ by $dx(\Phi)$ one needs to follow the gradient flow $\nabla x = g^{-1} ({\Sigma}^* dx)$ during a time $\frac{dx(\Phi)}{|{\Sigma}^* dx|^2}$: $${\Sigma}^*dx\left(\frac{dx(\Phi)}{|{\Sigma}^* dx|^2} \nabla x\right) = dx(\Phi).$$ Combining and we have $$\begin{aligned}
\label{3- Lemma(first variation)7}
\left.\frac{d }{d u} \left(\int_{{\Sigma}_{\epsilon}(u)} dA(u)\right)\right|_{u=0} &= \int_{{\gamma}_{\epsilon}} dl\; \frac{dx\left(\Phi \right)}{|{\Sigma}^* dx|} - \int_{{\Sigma}_{\epsilon}} tr_{g} K(\Phi) dA.\end{aligned}$$
We now expand the above in powers of epsilon. One will need an asymptotic expansion of a generic element $\Phi$ of the normal bundle. This needs a bit of work and will be given a proof of its own (see Lemma \[3- Lemma: Normal vector fields\] and the following proof). We here only state the end result: Let $\Phi$ be a section of the normal bundle of ${\Sigma}$ in $H$ then there exist a one parameter family $\phi(t)$ of sections of the normal bundle to ${\gamma}$ in $X$ such that $$\Phi\left(s,t\right) = \phi(t) - h\left(\phi(t), t|{\dot{\gamma}}|v + t^2|{\dot{\gamma}}|^{-1}n\right) \partial_x -\frac{t^2}{2} \alpha {\dot{\gamma}}+ {\mathcal{O}}(t^3)$$ where $\alpha$ is a function of $\phi(t)$ and $v$. In particular $$dx\left(\Phi\right) = - h\left(\phi(t), t|{\dot{\gamma}}|v + t^2|{\dot{\gamma}}|^{-1}n\right) +{\mathcal{O}}(t^3)$$ Putting this with and the expansion of ${\Sigma}^*x$ given by , we get: $$\begin{aligned}
\label{3- Lemma(first variation)3}
\left.\frac{d }{d u} \Bigg(\int_{{\Sigma}(u)} dA(u)\Bigg)\right|_{u=0}
=-\frac{1}{\epsilon}\; \int_{{\gamma}} ds\; h\left(\phi , |{\dot{\gamma}}| v \right)- \int_{{\gamma}} ds |{\dot{\gamma}}|^{-2} h\left(\phi, n\right) - \int_{{\Sigma}_{\epsilon}} tr_{g} K(\Phi) dA + {\mathcal{O}}(\epsilon).\end{aligned}$$ Finally, noting that $$\left.\frac{d}{d u}\left( \int_{{\gamma}_{\epsilon}(u)} dl(u)\right)\right|_{u=0} = -\frac{1}{\epsilon}\int_{{\gamma}} ds\; h\left(\phi , \nabla_{{\dot{\gamma}}} \left(|{\dot{\gamma}}|^{-2} {\dot{\gamma}}\right) \right)$$ we can recast equation as $$\begin{aligned}
\label{3- Lemma(first variation)4}
&\left.\frac{d }{d u} \left(\int_{{\Sigma}_{\epsilon}(u)} dA(u) - \int_{{\gamma}_{\epsilon}(u)} dl(u) \right) \right|_{u=0} \phantom{\hspace{8cm}}\end{aligned}$$ $$\begin{aligned}
=- \int_{{\Sigma}_{\epsilon}} tr_{g} K(\Phi) dA - \frac{1}{\epsilon} \; \int_{{\gamma}} ds\;|{\dot{\gamma}}| \; h\bigg(\phi , \Big(v -\nabla_{{\dot{\gamma}}} \left(|{\dot{\gamma}}|^{-2} {\dot{\gamma}}\right)\Big)\bigg)- \int_{{\gamma}} ds |{\dot{\gamma}}|^{-2} h\left(\phi, n\right) + {\mathcal{O}}(\epsilon) .\nonumber \end{aligned}$$ One can show that the divergences in the two first in this expansion compensate each others so that the first variation of the renormalised area is well defined.
For variation fixing the conformal boundary i.e $\phi = 0$ we obtain we recover the variation of the area: $$\frac{d \mathcal{A}\left( {\Sigma}_u \right) }{d u}|_{u=0} = \lim_{\epsilon=0}\;\left.\frac{d }{d u} \left(\int_{{\Sigma}_{\epsilon}(u)} dA(u) - \int_{{\gamma}_{\epsilon}(u)} dl(u) \right) \right|_{u=0} = - \int_{{\Sigma}} tr_{g} K(\Phi) dA.$$ For $\phi \neq 0$ and if we suppose that ${\Sigma}$ is minimal the divergent terms vanishes as a consequence of lemma \[3- Lemma: first variations with compact support\] equation and the variation of the renormalised area is: $$\left.\frac{d \mathcal{A}\left( {\Sigma}_u \right) }{d u}\right|_{u=0} = \lim_{\epsilon=0}\;\left.\frac{d }{d u} \left(\int_{{\Sigma}_{\epsilon}(u)} dA(u) - \int_{{\gamma}_{\epsilon}(u)} dl(u) \right) \right|_{u=0} = -\int_{{\gamma}} ds |{\dot{\gamma}}|^{-2} h\left(\phi, n\right).$$ This concludes the proof of Lemma \[3- Lemma: first variation with non-compact support\].
Conformal geodesics and their ambient surfaces
----------------------------------------------
We now rephrase the results from section \[subsec: Conformal geodesics\] on curves ${\gamma}{:}I \to X$ in a conformal manifold $\left(X, [h]\right)$ in terms of their ambient surface ${\Sigma}_{{\gamma}} {:}I^2 \to H$ in the associated Poincaré–Einstein manifold $\left(H, g\right)$ as defined in the preceding subsection.
Let us first start with a remark. If ${\gamma}$ is a curve in $X$ and ${\Sigma}_{{\gamma}}$ is its ambient surface given by Theorem \[3- Thrm: Holographic dual\] then ${\Sigma}_{{\gamma}}$ is “asymptotically hyperbolic” in the sense that its Gauss curvature asymptotically behaves as $G_{{\Sigma}}= -1+ {\mathcal{O}}(t^3)$. (This can be directly obtained from the expansion of the induced metric given by Corollary \[3- Crllr: Induced metric\]).
The following theorem then answers the natural question “What does it take to make the induced metric on ${\Sigma}_{{\gamma}}$ to be as close as possible to the two-dimensional hyperbolic half-plane model?”.
\[3- Thrm: Conformal Parametrisation\]\
Let ${\gamma}$ be a curve in $X$ and let ${\Sigma}_{{\gamma}}$ be its ambient surface given by Theorem \[3- Thrm: Holographic dual\]. We suppose that ${\Sigma}_{{\gamma}}$ is given the parametrisation . The induced metric then asymptotically approximates the Poincaré half-plane metric up to second order, $$\label{3- Induced metric3}
{\Sigma}^*_{{\gamma}}g = \frac{1}{t^2} \left[
\begin{pmatrix}
1 & 0\\ 0 & 1
\end{pmatrix}\left(1 + t^2\left(- \frac{1}{3}A_I A^I\right)\right) + {\mathcal{O}}(t^3) \right].$$ (Recall the definition of $A_IA^I$.) What is more, the curve ${\gamma}$ satisfies the tangential part of the conformal geodesic equations if and only if the induced metric is $${\Sigma}^*_{{\gamma}}g = \frac{1}{t^2} \left[
\begin{pmatrix}
1 & 0\\ 0 & 1
\end{pmatrix} + {\mathcal{O}}(t^3) \right].$$ In other words a choice of preferred conformal parametrisation for ${\gamma}$ amounts to choosing the parametrisation of its ambient surface ${\Sigma}_{{\gamma}}$ which makes it looks as close as is possible to the hyperbolic half-plane model.
\
This is a nearly direct consequence of the combination of Theorem \[3- Thrm: Holographic dual\] with Corollary \[3- Crllr: Induced metric\]: the expansion is given by Corollary \[3- Crllr: Induced metric\] while Theorem \[3- Thrm: Holographic dual\] implies that for the ambient surface ${\Sigma}_{{\gamma}}$ associated to ${\gamma}$ the vector field $v$ in the parametrisation satisfies the first geodesic equation . It follows that $-\frac{1}{2}A_I A^I$ as defined by equation is the tangential part of the conformal geodesic equation whose vanishing defines a preferred conformal parametrisation.
Theorem \[3- Thrm: Conformal Parametrisation\] gives a geometric understanding of the vanishing of the *tangential* part of the conformal geodesic equation for a curve ${\gamma}$ in terms of its ambient surface ${\Sigma}_{{\gamma}}$, the next theorem does the same for the normal part.
\[3- Thrm: Conformal Geodesics\]\
Let ${\gamma}$ be a curve in $X$ and let ${\Sigma}_{{\gamma}}$ be its ambient surface given by theorem \[3- Thrm: Holographic dual\]. Generically, the norm $|K|$ of the extrinsic curvature of the ambient surface vanishes asymptotically as $$|K| = 0 + {\mathcal{O}}\left(t^2\right).$$ What is more, the curve ${\gamma}$ satisfies the normal part of the conformal geodesic equation if and only if the above norm vanishes one order higher than expected $$|K| = 0 + {\mathcal{O}}\left(t^3\right).$$
In order to prove theorem \[3- Thrm: Conformal Geodesics\] we will need the following lemmas.
\[3- Lemma: Normal vector fields\] Let ${\Sigma}$ be any surface in $H$ with asymptotic boundary ${\gamma}$. Let $\Phi \in N_{{\Sigma}}$ be a section of its normal bundle in $H$. Then there exists a one parameter family $\phi(t)$ of sections of the normal bundle to ${\gamma}$ in $X$ such that $$\begin{aligned}
\label{3- Normal vector fields}
\Phi\left(s,t\right) &= \phi(t) - h\left(\phi(t), t|{\dot{\gamma}}|v + t^2|{\dot{\gamma}}|^{-1}n\right) \partial_x \\&\quad -\frac{t^2}{2} \Big( h\left(\phi(t) , \partial_s v\right) + \left(v^i\partial_i h -2P\right)(\phi(t), {\dot{\gamma}}) \Big){\dot{\gamma}}+ {\mathcal{O}}(t^3)\nonumber\end{aligned}$$
\[3- Lemma: Normal vector fields bis\] Let ${\Sigma}$ be any surface in $H$ with asymptotic boundary ${\gamma}$ and let $\left\{ \eta_a\right\}_{a \in 1...n-1}$ be a basis of $N_{{\gamma}}$ the normal bundle to ${\gamma}$ in $X$. Then the set $\left\{ N_a\right\}_{a \in 1...n-1}$ of vector fields asymptotically obtained by taking $\phi(t)=\eta_a$ in the previous lemma gives, for any point $p\in{\Sigma}$, a basis of $N_p{\Sigma}$, the normal bundle to ${\Sigma}$ at $p$.
What is more, their inner product is $$\label{3- Normal vector field, inner product}
g\left(N_a, N_b\right) = \frac{1}{t^2}\bigg(h\left(\eta_a , \eta_ b\right) |{\dot{\gamma}}|^{-2} + 0 + {\mathcal{O}}(t^2)\bigg).$$
\
Let $\Phi \in {\Gamma}\left[N_{{\Sigma}}\right]$ be any section of the normal tangent bundle: $$\Phi = \Phi_0 + n_0 \partial_x + t\bigg( \Phi_1 + n_1 \partial_x \bigg) + \frac{t^2}{2} \bigg( \Phi_2 + n_2 \partial_x \bigg) + {\mathcal{O}}(t^3).$$ We now want to impose that the above is a vector normal to ${\Sigma}$, $$\begin{aligned}
\label{3 - ProofLemma(normal bundle)1}
g\left(\Phi , \partial_s \right)=0 & & g\left(\Phi , \partial_t \right)=0.\end{aligned}$$ We do so by solving the above equation order by order. With the expansion $$\label{3- Lemma(Normal bundle)2}
\bar{g}\left(x(s,t), y^i(s,t)\right) = dx^2 + h + t^2\left.\left(\frac{1}{2}v^i \partial_i h - |{\dot{\gamma}}|^2 P\right) \right|_{t=0} + {\mathcal{O}}(t^3).$$ of the restriction $g\big|_{{\Sigma}}$ of the metric to ${\Sigma}$ this is a direct computation and results in the expansion .
If we now take $\phi(t) = \eta_a$ in all that is left to do to prove Lemma \[3- Lemma: Normal vector fields bis\] is prove that the set of vector fields obtained in this way form a basis of the normal bundle at every point $p \in {\Sigma}$. Using the expansion of the metric , this is however straightforward to check that the inner product of vector fields of the form is and thus non-degenerate.
\[3- Lemma: Extrinsic curvature\] Let ${\gamma}$ be a curve in $X$ and let ${\Sigma}$ be any surface with asymptotic boundary ${\gamma}$ in the parametrisation of Lemma \[3- Lemma: Conformal parametrisation\]. Let $\Phi \in N_{{\Sigma}}$ be an element of the normal bundle parametrised as in equation of Lemma \[3- Lemma: Normal vector fields bis\].
Then the extrinsic curvature has, in the coordinate basis $\{\partial_s, \partial_t\}$ an expansion of the form $$\label{3- Extrinsic curvature, general expansion}
K\left(\Phi\right) = \frac{1}{t^3} \bigg[ 0 + t\; K_1 +\frac{t^2}{2}\; K_2 + {\mathcal{O}}(t^3)\bigg]$$ with
[30pt]{}[0pt]{} $$\begin{aligned}
K_1 &=
\begin{pmatrix}
h\left(\phi(t), \nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|^{-2} {\dot{\gamma}}\right) -v\right)& 0 \\ 0 & 0
\end{pmatrix}
\\ \\
K_2 &=
\begin{pmatrix}
- h\left(\phi(t), n\right) |{\dot{\gamma}}|^{-2} & h\left(\phi(t), |{\dot{\gamma}}|^{-1}\nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|v\right) + \frac{1}{2}|v|^2 {\dot{\gamma}}- h^{-1}\left( P({\dot{\gamma}}) \right)\right)
\\ \\h\left(\phi(t), |{\dot{\gamma}}|^{-1}\nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|v\right) + \frac{1}{2}|v|^2 {\dot{\gamma}}- h^{-1}\left( P({\dot{\gamma}}) \right)\right)
& h\left(\phi(t), n\right) |{\dot{\gamma}}|^{-2}
\end{pmatrix}\\\end{aligned}$$
Note that, if we assume $\phi(0) \neq 0$, the normal $\Phi$ and coordinate vectors $\{\partial_s, \partial_t\}$ all have norms diverging as one over $t$, see and .
\
This is again just an order by order computation of the different components of the extrinsic curvature, $$\begin{aligned}
K\left(\partial_s , \partial_s , \Phi \right) &= g\left( \nabla_s \partial_s , \Phi\right) \nonumber\\
K\left(\partial_t , \partial_t , \Phi \right) &= g\left( \nabla_t \partial_t , \Phi\right) \\
K\left(\partial_s , \partial_t , \Phi \right) &= g\left( \nabla_s \partial_t , \Phi\right) \nonumber\end{aligned}$$ From the expansion of $g$ one can derive the expansion of the Christoffel symbols $$\begin{aligned}
{\Gamma}_{xxx}&= \frac{1}{t^3 |{\dot{\gamma}}|^2}\left[ - \frac{1}{|{\dot{\gamma}}|} + t^2 \frac{x_3}{|{\dot{\gamma}}|^2} + {\mathcal{O}}(t^3)\right] & {\Gamma}_{ixj} &= \frac{1}{t^3 |{\dot{\gamma}}|^2}\left[ \frac{h_{ij}}{ |{\dot{\gamma}}|} + t^2\left( -\frac{x_3}{|{\dot{\gamma}}|^2}h + \frac{|{\dot{\gamma}}|}{2} v^i \partial_i h \right) + {\mathcal{O}}(t^3)\right] \nonumber\\
{\Gamma}_{xix}&= 0 & {\Gamma}_{xij} &= \frac{1}{t^3 |{\dot{\gamma}}|^2}\left[ -\frac{h_{ij}}{ |{\dot{\gamma}}|} + t^2\left( -\frac{x_3}{|{\dot{\gamma}}|^2}h - \frac{|{\dot{\gamma}}|}{2} v^i \partial_i h \right) + {\mathcal{O}}(t^3)\right] \\
{\Gamma}_{xxi}&= 0 & {\Gamma}_{ikj} &= \frac{1}{t^3 |{\dot{\gamma}}|^2}\left[ - t {\Gamma}_h{}_{ikj} + {\mathcal{O}}(t^3)\right] \nonumber\end{aligned}$$ with $x_3 = \bigg( -\frac{3}{4}|{\dot{\gamma}}|^3|v|^2 - \frac{1}{4} |{\dot{\gamma}}| A_I A^I \bigg)$. Then, since we already have the expansions and for $g\big|_{{\Sigma}}$ and the normal vectors fields this is tedious but straightforward to derive the expansion .
Associating Lemma \[3- Lemma: Extrinsic curvature\] with previous results we obtain the following corollaries:
\[3- Crllr: mean curvature\] Let ${\gamma}$ be a curve in $X$ and let ${\Sigma}$ be any surface with asymptotic boundary ${\gamma}$ in the parametrisation of Lemma \[3- Lemma: Conformal parametrisation\]. Let $\Phi \in N_{{\Sigma}}$ be an element of the normal bundle parametrised as in equation of Lemma \[3- Lemma: Normal vector fields\]. Then the mean curvature of ${\Sigma}$ is $$TrK(\Phi) = \frac{1}{t}\bigg( 0 + t \; h\left(\phi(t), \nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|^{-2} {\dot{\gamma}}\right) -v\right) + {\mathcal{O}}(t^3) \bigg).$$ In particular, ${\Sigma}$ is asymptotically minimal if and only if $v$ satisfies the first conformal geodesic equation .
This is a direct consequence of Lemma \[3- Lemma: Extrinsic curvature\] and Corollary \[3- Crllr: Induced metric\].
\[3- Crllr: extrinsic curvature\] Let ${\gamma}$ be a curve in $X$ and let ${\Sigma}_{{\gamma}}$ be its ambient surface given by Theorem \[3- Thrm: Holographic dual\]. Let $\Phi \in N_{{\Sigma}}$ be an element of the normal bundle parametrised as in equation of Lemma \[3- Lemma: Normal vector fields\]. Then the extrinsic curvature has, in the coordinate basis $\{\partial_s, \partial_t\}$, an expansion of the form $$\begin{aligned}
\label{3- Extrinsic curvature of the holographic dual}
&\; K\left(\Phi\right) = \frac{1}{t^3} \bigg[ 0
+ t^2
\begin{pmatrix}
0& 1
\\ 1
& 0
\end{pmatrix} h\left(\phi(t) \;,\; |{\dot{\gamma}}|^{-1}\nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|v\right) + \frac{1}{2}|v|^2 {\dot{\gamma}}- h^{-1}\left( P_h({\dot{\gamma}}) \right)\right)
+ {\mathcal{O}}(t^3) \bigg]\end{aligned}$$ In particular, ${\Sigma}_{{\gamma}}$ is asymptotically totally geodesic if and only if $v$ satisfies the second conformal geodesic equation or equivalently if and only if ${\gamma}$ satisfies the conformal geodesic equation .
This is a direct consequence of Lemma \[3- Lemma: Extrinsic curvature\] with Theorem \[3- Thrm: Holographic dual\].
We now conclude the proof of Theorem \[3- Thrm: Conformal Geodesics\].
\
Making use of Corollary \[3- Crllr: extrinsic curvature\] and the explicit metrics and of Lemma \[3- Lemma: Normal vector fields bis\] and Corollary \[3- Crllr: Induced metric\] we can directly compute the norm of the extrinsic curvature , $$|K|^2 = 2\;t^4\; \bigg| |{\dot{\gamma}}|^2\; \left( |{\dot{\gamma}}|^{-1}\nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|v\right) + \frac{1}{2}|v|^2 {\dot{\gamma}}- h^{-1}\left( P_h({\dot{\gamma}}) \right) \right)^{\perp} \bigg|^2 + {\mathcal{O}}(t^5)$$ where $\perp$ signals orthogonal projection in the direction normal to ${\Sigma}_{{\gamma}}$.
Thus generically $|K| = 0 + {\mathcal{O}}(t^2)$ and $|K| = 0 + {\mathcal{O}}(t^3)$ if and only if $v$ satisfies the normal part of the second conformal geodesic equation . Finally we recall that the first conformal geodesic equation is already satisfied as a result of Theorem \[3- Thrm: Holographic dual\] and that the vanishing of the tangential part of the second conformal geodesic equations amounts to a choice of parametrisation (see section \[subsec: Conformal geodesics\] and Theorem \[3- Thrm: Conformal Parametrisation\]). Altogether this proves Theorem \[3- Thrm: Conformal Geodesics\].
Ambient surface and tractors {#Sec: Ambient Metric Realisation of Conformal Geodesics and its Relation to Tractors}
============================
In section \[subsec: Dual surface to a conformal curve\] we associated to any curve ${\gamma}$ in a conformal manifold $X$ a unique surface ${\Sigma}_{{\gamma}}$ in the related Poincaré–Einstein space $H$. This surface can be taken to be asymptotically defined by the expansion of Theorem \[3- Thrm: Holographic dual\]. We then showed (see Theorem \[3- Thrm: Conformal Geodesics\]) that ${\gamma}$ is a conformal geodesics if and only if the extrinsic curvature of ${\Sigma}_{{\gamma}}$ vanishes to one order higher than generically expected.
Here we rephrase these results in terms of the ambient ((n+2)-dimensional) metric of Fefferman and Graham [@fefferman_ambient_2012]. This has two main advantages: first it will explicitly relate our constructions to the standard tractor formulation of the conformal geodesic equations [@bailey_thomass_1994]; second the use of tractor calculus will dramatically shorten the proofs so that this section can be seen as alternative and more transparent derivations of Theorem \[3- Thrm: Conformal Geodesics\]. The reason why we postponed this discussion is that it relies on the simultaneous use of tractor calculus and the ambient metric which are perhaps not as well-spread tools as those used in section \[Sec: Holographic Prescription for Conformal Geodesics\].
The ambient metric and tractors: a very brief overview
------------------------------------------------------
In this section we briefly review from [@cap_standard_2003] the relation between the ambient metric and tractor calculus but otherwise assume that the reader is familiar with standard tractor calculus, see [@bailey_thomass_1994]. We also review how the conformal geodesic equations can be elegantly written in terms of tractors.
### The ambient metric associated to $\left(X, [x]\right)$ and its Fefferman-Graham expansion {#subsubsec: The ambient metric}
Let $\left(X , [h]\right)$ be a $n$-dimensional conformal manifold. Let $\left(M , g_M\right)$ be the associated ambient (formal) metric [@fefferman_conformal_1985], [@fefferman_ambient_2012]: By construction, $M$ is $(n+2)$-dimensional with $M= (-1,1) \times L$ where $L$ is the total space of the scale bundle $L\to X$. Recall that a conformal metric ${[h]}$ is a section of $L^2 \otimes S^2T^*X$. Let $\left({\sigma}, y\right)$ be coordinates on $L$, then the pull back $g_L=\pi^*{[h]}= {\sigma}^2 h$ of ${[h]}$ gives a metric on $L$. Let ${\rho}{:}M \to {\mathbb{R}}$ be a boundary defining function, $L = {\rho}^{-1}(0)$, $d{\rho}\neq 0$. The ambient metric $g_M$ on $M$ extends $g_L$ and has asymptotic expansion $$\label{4- FG First Normal Form3}
g_M= -2 {\rho}d{\sigma}^2 -{\sigma}d{\sigma}d{\rho}+{\sigma}^2 \left[h -2{\rho}P +{\mathcal{O}}({\rho}^2)\right].$$ As in previous sections, if $n>2$ then $P$ is the Schouten tensor of $h$ while if $n=2$, $P$ corresponds to choice of Möbius structure. Since we won’t need the precise form of the expansion beyond second order, the equation can be taken to define the metric $g_M$.
The ambient metric can be related to the Poincaré–Einstein metric by making use of the change of coordinate $$\label{4- Coordinates change: Ambient to Poincare}
{\sigma}= l x^{-1},\quad {\rho}= \frac{1}{2}x^2.$$ We then have $$\label{4- FG Second Normal Form2}
g_M = -dl^2 + \frac{l^2}{x^2} \left(dx^2 + \left[ h - x^2 P + {\mathcal{O}}(x^4) \right] \right)$$ and the hypersurfaces $l=1$ identifies with the Poincaré–Einstein manifold $\left(H , g\right)$. In particular $H$ is a graph over the bundle of scale $L$ defined by $${\rho}= \frac{1}{2}\left( {\sigma}\right)^{-2}$$ (see figure \[fig:AmbientMetric\]).
![The associated Ambient ($M$) and Poincaré-Einstein ($H$) metrics to $(X, {[h]})$[]{data-label="fig:AmbientMetric"}](ambient_metric)
A crucial remark is that as ${\sigma}$ goes to infinity the bundle of scale $L$ and the Poincaré–Einstein metric $H$ both converge toward the same asymptotic boundary (identified with $X$). This somewhat fuzzy statement can be made precise by introducing a “conformal compactification” of the ambient metric generalizing Penrose’s conformal compactification of Minkowski space. Since the ambient metric is only defined in a neighbourhood of $L$ there is only one copy of $X$ (“at null-infinity of $L$”) that needs to be added as a boundary (instead of a full asymptotic null cone for Minkowski compactification).
Finally, we note for future reference, note that the unit normal to $H$ in $M$ is $$\partial_l\Big|_{l=1} = {\sigma}\partial_{{\sigma}}\Big|_{l=1} = E\Big|_{l=1}$$ (here and everywhere in the subsequent subsections $E = {\sigma}\partial_{{\sigma}}$).
### Tractors and the ambient metric {#subsubsec: Tractors and the ambient metric}
We briefly review from [@cap_standard_2003] the close relationship between tractor calculus on a conformal manifold $\left(X, {[h]}\right)$ and the associated ambient metric $\left(M,g_M\right)$.
As we just discussed, the bundle of scales $L\to M$ can be seen as a hyper-surface in the ambient manifold $\left(M, g_M\right)$ through the ambient metric construction [@fefferman_conformal_1985],[@fefferman_ambient_2012]. The bundle of scales is a ${\mathbb{R}}^+$-principal bundle and the restriction $g_M \big|_{L}$ is homogeneous of degree $2$ under this action. Now let $y \in X$ and $\pi^{-1}(y) \subset L$ be the fibre over $y$ in $L$. Let $\iota {:}\pi^{-1}(y) \to M$ be its inclusion in $M$ and consider sections of the pull-back bundle ${\Gamma}\left[\iota^* TM\right]$ which are homogeneous of degree $-1$ under the ${\mathbb{R}}^+$-action. This defines a $n+2$-dimensional vector space $\mathcal{T}_y$. Since the contraction of such sections with $g_M \big|_{L}$ is homogeneous degree zero this gives a metric on $\mathcal{T}_y$. Altogether this defines a $n+2$-dimensional metric vector bundle $\mathcal{T} \to X$ over $X$. It was shown in [@cap_standard_2003] that it canonically identifies with the standard tractor bundle. By construction, if $\iota {:}L \to M$ is the inclusion, sections of $\mathcal{T}$ are sections of $\iota^* TM$ which are homogeneous degree $-1$ under the ${\mathbb{R}}^+$-action.
It was also shown in [@cap_standard_2003] that the Levi-Civita connection $\nabla$ of the ambient metric induces a connection on the tractor bundle, as follows:
Let $Y^I$ be a section of the tractor bundle $Y^I \in {\Gamma}\left[\mathcal{T}\right]$, by construction this is sections of the pull-back bundle $\iota^* TM$ homogeneous of degree $-1$ under the ${\mathbb{R}}^+$-action. One first proves, see [@cap_standard_2003], that the Levi-Civita connection $\nabla$ of the ambient metric satisfies, $$\label{4- Identity: ambient LC1}
\nabla Y^I = \nabla\Big|_{\rho=0} Y^I + {\mathcal{O}}(\rho) Y^I$$ and $$\begin{aligned}
\label{4- Identity: ambient LC2}
\nabla_E\Big|_{\rho=0} Y^I &=0, & \nabla_Y\Big|_{\rho=0} E = Y^I
\end{aligned}$$ (recall that $E = {\sigma}\partial_{{\sigma}}$). Let $y \in{\Gamma}\left[TX\right]$ and let $\pi^{-1}(y) \in {\Gamma}\left[TL\right]$ be any lift of $y$ to $L$. From one sees that $\nabla_{\pi^{-1}(y)}\Big|_{\rho=0} Y^I$ does not depend on the choice of lift. One then show that $\nabla_{\pi^{-1}(y)}\Big|_{\rho=0} Y^I \in {\Gamma}\left[\iota^*TM\right]$ is homogeneous degree $-1$ under the ${\mathbb{R}}^+$-action and thus defines a section of $\mathcal{T}$. The Levi-Civita connection of the ambient metric thus induces a connection $D {:}\mathcal{T} \to T^*X \otimes \mathcal{T}$ on the tractor bundle as $$D_y Y^I \coloneqq \nabla_{\pi^{-1}(y)}\Big|_{\rho=0} Y^I.$$ It was proven in [@cap_standard_2003] that this connection actually coincides with the normal tractor connection.
### Tractor calculus and conformal geodesics {#subsubsec: Tractor calculus and conformal geodesics}
We here review the tractor formulation of the conformal geodesic equations, see [@bailey_thomass_1994] for the original exposition.
Let $\left(X , [h] \right)$ be a $n$-dimensional conformal manifold. We take $\gamma {:}I \to X$ to be a curve in $X$ parametrised by a parameter $s$ and ${\dot{\gamma}}= \partial_s \gamma$. Following our convention for conformal geometry, $h$ is a choice of representative for ${[h]}$ and $|{\dot{\gamma}}|^2 = h\left({\dot{\gamma}}, {\dot{\gamma}}\right)$.
Define the “position tractor” to be, $$X^{I} = \begin{pmatrix}
0 \\ 0 \\ \; |{\dot{\gamma}}|^{-1}
\end{pmatrix}$$ By subsequent differentiations of the “position tractor” one obtains the “velocity tractor”, $$\label{4- Velocity Tractor}
U^{I}= D_{{\dot{\gamma}}} X^{I} = \begin{pmatrix}
0 \\ |{\dot{\gamma}}|^{-1}\;{\dot{\gamma}}\\ \; \partial_s\left( |{\dot{\gamma}}|^{-1} \right)
\end{pmatrix}$$ and “acceleration tractor” : $$A^{I}= D_{{\dot{\gamma}}} U^{I} = \begin{pmatrix}
- |{\dot{\gamma}}| \\ |{\dot{\gamma}}|v \\ |{\dot{\gamma}}| \frac{1}{2} |v|^2 - \frac{1}{2} A_{I} A^{I} |{\dot{\gamma}}|^{-1}
\end{pmatrix}$$ where $$\begin{aligned}
v &\coloneqq \nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|^{-2} {\dot{\gamma}}\right), &
-\frac{1}{2}A^{I}A_{I} &\coloneqq h\left({\dot{\gamma}}, |{\dot{\gamma}}|^{-1}\nabla_{{\dot{\gamma}}}\left(|{\dot{\gamma}}|v\right) + \frac{1}{2}|v|^2 {\dot{\gamma}}- h^{-1}\left( P_h({\dot{\gamma}}) \right) \right).
\end{aligned}$$
It was proven in [@bailey_thomass_1994] that $A^{I}A_{I}=0$ is equivalent to the preferred conformal parametrisation of $\gamma$ and while $D_{{\dot{\gamma}}}A^{I}=0$ is equivalent to the (parametrised) conformal geodesic equations .
Making use of the identification of sections of the tractor bundle with (homogeneous degree -1) sections of $\iota^* TM$ we have $$\begin{aligned}
\label{4- Position and Velocity Tractor}
X^I\left({\sigma}, y\right) &= {\sigma}^{-1} \Big( |{\dot{\gamma}}|^{-1} E \Big), &
U^I\left({\sigma}, y\right) &= {\sigma}^{-1}\Big( |{\dot{\gamma}}|^{-1} {\dot{\gamma}}+ \partial_s \left(|{\dot{\gamma}}|^{-1}\right) E \Big)\end{aligned}$$ and $$\begin{aligned}
\label{4- Acceleration Tractor}
A^{I}\left({\sigma}, y\right) &= {\sigma}^{-1} \Bigg(
|{\dot{\gamma}}| \partial_{{\rho}} + |{\dot{\gamma}}|v + \left(|{\dot{\gamma}}| \frac{1}{2} |v|^2 - \frac{1}{2} A_{I} A^{I} |{\dot{\gamma}}|^{-1}\right) E
\Bigg)\end{aligned}$$ (recall that $E = {\sigma}\partial_{{\sigma}}$).
Tractor calculus for the ambient surfaces of a conformal geodesics
------------------------------------------------------------------
As we have reviewed, the tractor calculus has a direct interpretation in terms of the Ricci calculus of the ambient metric $(M, g_M)$. On the other hand the Poincaré–Einstein metric $\left(H,g\right)$ can always be thought as a sub-manifold of the ambient metric. The ambient metric is thus the correct arena to reconcile the description of conformal geodesics that we developed in the section \[Sec: Holographic Prescription for Conformal Geodesics\] with the standard tractor equations. This is one of the aims of this subsection. The second aim is to provide an alternative, more compact, proof of theorem \[3- Thrm: Conformal Geodesics\].
### Tractor calculus, ambient metric and conformal curves
Let ${\gamma}{:}I \to X$ a curve in $X$ parametrised by a parameter $s$. The bundle of scales $L\to X$ naturally is a ${\mathbb{R}}^+$-principal bundle and we can thus consider ${\mathcal{S}}_{{\gamma}} {:}I^2 \to L$, the unique ${\mathbb{R}}^+$-invariant lift of ${\gamma}$ to $L$, that is satisfying $$\begin{aligned}
\pi\left({\mathcal{S}}_{{\gamma}}\right) &= {\gamma}& R^*{\mathcal{S}}_{{\gamma}} = {\mathcal{S}}_{{\gamma}}.\end{aligned}$$ In coordinates, a convenient parametrisation is $$\label{4- L surface parametrisation}
{\mathcal{S}}_{{\gamma}}\left\{
\begin{array}{ccc}
{\rho}\left(s,t\right) & = & 0 \\ \\
y^i\left(s,t\right) & = & {\gamma}(s) \\ \\
{\sigma}\left(s,t\right) & = & f(t)\;|{\dot{\gamma}}|^{-1}
\end{array}
\right.$$ Since the metric induced by on $L$ is $g_L = {\sigma}^2 h$ the corresponding induced metric on ${\mathcal{S}}_{{\gamma}}$ is degenerate. The above parametrisation has the following property: taking $t=cst$ gives a lift of ${\gamma}$ in $L$ with constant length $f(t)$ with respect to $g_L = {\sigma}^2 h$. The precise value of the length is a remaining choice of parametrisation.
The coordinate vectors given by this parametrisation are $$\begin{aligned}
\partial_s &= f(t) \;U^I(s,t), &
\partial_t & = f'(t) \; X^I(s,t)\end{aligned}$$ where $X^i(s,t)$ and $U^I(s,t)$ are the position and velocity tractors evaluated at $({\sigma}= {\sigma}(s,t), y = y(s,t))$.
The above notation is convenient as covariant differentiation along $s$ with the connection given by the ambient metric is then nicely given in terms of the tractor connection: $$\begin{aligned}
\nabla_s \partial_t &= f'(t) D_s X^I = f'(t) \;U^I(s,t), &
\nabla_s \partial_s &= f(t) D_s U^I = f(t) \;A^I(s,t).\end{aligned}$$ where $A^I(s,t)$ is the acceleration tractor evaluated at $({\sigma}= {\sigma}(s,t), y = y(s,t))$. We also immediately have $\nabla_s \nabla_s \partial_s =f(t) D_s A^I$ and thus ${\gamma}$ is a conformal geodesics if and only if $\nabla_s \nabla_s \partial_s=0$.
The identification of tractor calculus along ${\gamma}$ with the differential calculus on the surface ${\mathcal{S}}_{{\gamma}}$ in $L$ considerably simplifies our task. We will see that essentially the same identification is possible for the ambient surface ${\Sigma}_{{\gamma}}$ in $H$: as $H$ and $L$ asymptotically converge to the same asymptotic boundary, differential calculus on ${\Sigma}_{{\gamma}}$ in $H$ asymptotically identifies with tractor calculus.
From now one we will take as a convenient parametrisation for ${\mathcal{S}}_{{\gamma}}$: $$\label{4- f(t) =1/t }
f(t) = \frac{1}{t}$$ With this parametrisation, when $t$ goes to zero the surface ${\mathcal{S}}_{{\gamma}}$ goes to the asymptotic boundary of $L$ (which, we recall, is identified with $X$).
### Ambient surface in the ambient metric
In section \[subsec: Dual surface to a conformal curve\] we associated to any curve ${\gamma}$ in $X$ an ambient surface ${\Sigma}_{{\gamma}}$ in $H$. The following proposition gives the interpretation of this result in terms of the ambient metric. In fact since $H$, identified with the hypersurface $l=1$ in $M$, is a graph over $L$ it is reasonable to expect that ${\Sigma}_{{\gamma}} \subset H$ is a graph over ${\mathcal{S}}_{{\gamma}} \subset L$ which is indeed the case:
\[4- Thrm: Holographic dual\]\
Let ${\gamma}$ be a curve in $X$, let $\left(M, g_M\right)$ be the ambient space asymptotically defined by . Let ${\mathcal{S}}_{{\gamma}}$ be the ${\mathbb{R}}$-invariant lift of ${\gamma}$ to $L$ with the parametrisation , .
The ambient surface ${\Sigma}_{{\gamma}}$ to ${\gamma}$ is asymptotically a distinguished graph over ${\mathcal{S}}_{{\gamma}}$ given by the following expansion, $$\label{4- Ambient dual}
{\Sigma}_{{\gamma}} (s,t) = {\mathcal{S}}_{{\gamma}}(s,t) + \frac{t}{2} \Big( A^I(s,t) +\frac{1}{3}\left(A_JA^J\right) X^I(s,t) \Big)+ {\mathcal{O}}(t^3)X^I +{\mathcal{O}}({\sigma}^{-1} t^3).$$ Where the position and acceleration tractors are given by , .
What is more, up to identifying the Poincaré–Einstein manifold $\left(H, g\right)$ with the hypersurface $l=1$, this definition coincides with the definition of Theorem \[3- Thrm: Holographic dual\]. In particular the parametrisations and coincide.
Note that ${\mathcal{O}}(t^n)X^I = {\mathcal{O}}(t^n){\sigma}^{-1} |{\dot{\gamma}}|^{-1}{\sigma}\partial_{{\sigma}} = {\mathcal{O}}(t^n)$ hence the distinction with ${\mathcal{O}}({\sigma}^{-1} t^n) = {\mathcal{O}}(t^{n+1})$. Along the way, proposition \[4- Thrm: Holographic dual\] gives us another simple geometrical interpretation of the preferred conformal parametrisation for ${\gamma}$: ${\Sigma}_{{\gamma}}$ is a graph above ${\mathcal{S}}_{{\gamma}}$ given by the acceleration tractor if and only ${\gamma}$ has been given a preferred conformal parametrisation $A_IA^I =0$.
\
We recall the expansion of ${\Sigma}_{{\gamma}}$ given by Theorem \[3- Thrm: Holographic dual\]: $${\Sigma}_{{\gamma}} \left\{\begin{array}{ccccccccccccccc}
x\left(s,t\right) & =& 0 &+& t |{\dot{\gamma}}| &+& 0 &+&\frac{t^3 }{3} x_3 &+& \frac{t^4 }{4} x_4 &+&{\mathcal{O}}\left(t^5\right) \\ \\
y^i\left(s,t\right) & =& {\gamma}^i(s) &+& 0 &+& \frac{t^2}{2} |{\dot{\gamma}}|^2 v^i &+& 0 &+& {\mathcal{O}}\left(t^4\right)\\ \\
l(s,t) & = &1
\end{array} \right.\nonumber$$ with $-\frac{1}{3}|{\dot{\gamma}}|^{-2} x_3 = \frac{1}{4}|{\dot{\gamma}}||v|^2 + \frac{1}{12} |{\dot{\gamma}}|^{-1} A_I A^I$. One can also show that if ${\Sigma}_{{\gamma}}$ is given an isothermal parametrisation then $x_4=0$. (If $n=3$ this uses the fact that we took the (free) third order coefficient in the expansion to be zero. This is however not really a choice here since from [@fefferman_conformal_1985], [@fefferman_ambient_2012] this is forced on us if we want the ambient metric to be smooth.)
Making use of the change of coordinates one readily sees that it is equivalent to .
### Extrinsic curvature in the ambient metric
We now show how the identification of tractor calculus with the asymptotic tensor calculus on ${\Sigma}_{{\gamma}}$ in the ambient metric simplifies the calculation of the extrinsic curvature of ${\Sigma}_{{\gamma}}$.
The parametrisation induces the coordinate basis $$\begin{aligned}
\label{4- Coordinate basis}
\partial_s &= \frac{1}{t} U^I +{\mathcal{O}}(t) X^I + {\mathcal{O}}({\sigma}^{-1} t), \\
\partial_t &= \Big(-\frac{1}{t} + \frac{t}{3} A_J A^J \Big)E + A^I + {\mathcal{O}}(t^2) X^I + {\mathcal{O}}({\sigma}^{-1} t^2).\nonumber\end{aligned}$$ (Again, ${\mathcal{O}}(t^n)X^I = {\mathcal{O}}(t^n){\sigma}^{-1} |{\dot{\gamma}}|^{-1}{\sigma}\partial_{{\sigma}} = {\mathcal{O}}(t^n)$ while ${\mathcal{O}}({\sigma}^{-1} t^n) = {\mathcal{O}}(t^{n+1})$.) Implicitly, here and in what follows, all tractors are evaluated at the point ${\sigma}(s,t)$ given by .
With the identities , in hand, the covariant derivatives of are just given by a few lines of computation:
\_s \_s &= \_s|\_[=0]{} U\^I +(t) X\^I + (\^[-1]{} t) = D\_s U\^I +(t) X\^I + (\^[-1]{} t)\
&= A\^I +(t) X\^I + (\^[-1]{} t) = ( - A\_J A\^J )E + \_t +(t) X\^I + (\^[-1]{} t) &
\_t \_t &= ( + A\_J A\^J )E + (- + A\_J A\^J ) \_[t]{}|\_[=0]{} E + \_[t]{}|\_[=0]{} A\^I + (t) X\^I + (\^[-1]{} t) &\
&= ( + A\_J A\^J )E + (- + A\_J A\^J ) \_t + \_[A]{}|\_[=0]{} A\^I + (t) X\^I + (\^[-1]{} t) &\
&= ( + A\_J A\^J )E + (- + A\_J A\^J ) \_t + (t) X\^I + (\^[-1]{} t) &
\_s \_t &= \_s ( A\_J A\^J )E + (- + A\_J A\^J ) \_[s]{}|\_[=0]{} E + \_[s]{}|\_[=0]{} A\^I + (t\^2) X\^I + (\^[-1]{} t\^2)\
&= \_s ( A\_J A\^J )E + (- + A\_J A\^J ) \_t + D\_s A\^I + (t\^2) X\^I + (\^[-1]{} t\^2).&
To obtain the extrinsic curvature, all is left to do is project out the directions , tangent to ${\Sigma}_{{\gamma}}$.
\[4- Prop: Covariant derivatives\]\
Let ${\gamma}$ be a curve in $\left(X,[h]\right)$. Let ${\Sigma}_{{\gamma}}$ be its ambient surface in the ambient space $\left(M, g_{M}\right)$ in the parametrisation given by Proposition \[4- Thrm: Holographic dual\]. Then its extrinsic curvature in $M$ is $$\begin{aligned}
K_M\left(\partial_s, \partial_s \right) &= \Big(\frac{1}{t^2} - \frac{1}{3} A_J A^J \Big)E +{\mathcal{O}}(t^2) E + {\mathcal{O}}({\sigma}^{-1} t) \\
K_M\left(\partial_t, \partial_t \right) &= \left(\frac{1}{t^2} + \frac{1}{3} A_J A^J \right)E + {\mathcal{O}}(t^2) E + {\mathcal{O}}({\sigma}^{-1} t) \\
K_M\left(\partial_s, \partial_t \right) &= D_s A^I +
\left( A_J A^J\right) U^I+ {\mathcal{O}}(t^2)E + {\mathcal{O}}({\sigma}^{-1} t).\label{4- Covariant derivatives}\end{aligned}$$
Since $E$ is the unit normal to $H$ in $M$ we immediately have a straightforward alternative proof to Theorem \[3- Thrm: Conformal Geodesics\]:
Suppose that ${\gamma}$ satisfies the conformal geodesic equations $D_s A^I=0$ then the extrinsic curvature of ${\Sigma}_{{\gamma}}$ in $M$ in the coordinate basis is proportional to $E$ up to order $t^2$ and therefore the extrinsic curvature of ${\Sigma}_{{\gamma}}$ in $H$ vanishes up to order $t^2$. Since the norm of the coordinate basis is of order one over $t$ and the Poincaré metric diverges in one over $t^2$ the norm of the extrinsic curvature vanishes up to order $t^2$.
The other way round, if the norm of the extrinsic curvature vanishes up to order $t^2$ then, in the coordinate basis, the extrinsic curvature of ${\Sigma}_{{\gamma}}$ in $H$ vanishes up to order $t^2$. Consequently the extrinsic curvature of ${\Sigma}_{{\gamma}}$ in $M$ is proportional to $E$ up to order $t^2$ and from Proposition \[4- Prop: Covariant derivatives\] this implies that $D_s A^I =0$ which is equivalent to the conformal geodesics equations.
[^1]: [email protected]
[^2]: [email protected]
[^3]: Here ${\dot{\gamma}}= \partial_s \gamma$ and $\ddot{\gamma} = \nabla_{{\dot{\gamma}}}{\dot{\gamma}}$
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
address: |
Institute of Theoretical Physics, Warsaw University, ul. Hoża 69, 00-681 Warsaw, Poland\
E-mail: [email protected],[email protected]
author:
- 'M. KRAWCZYK, A. ZEMBRZUSKI'
title: ' IFT/16/2001 PHOTOPRODUCTION OF THE ISOLATED PHOTON AT HERA IN NLO QCD'
---
Introduction
============
The production of the prompt photon with large transverse momentum, $p_T$, in the $ep$ collision is considered. Such reaction is dominated by events with almost real photons mediating the $ep$ interaction, $Q^2\approx 0$, so in practice we deal with the photoproduction of the prompt photon. The other name for such process is Deep Inelastic Compton (DIC) scattering (although $Q^2\approx 0$, the scattering is “deep inelastic” due to a large transverse momentum of the final photon). The photon emitted by the electron may interact with the proton partons directly or as a resolved one. Analogously, the observed final photon may arise directly from hard partonic subprocesses or from fragmentation processes, where a quark or a gluon decays into the photon.
The importance of the DIC process in the $ep$ collision for testing the Parton Model and then the Quantum Chromodynamics was studied previously by many authors [@Bjorken:1969ja]$^-~$[@Gordon:1994sm]. Measurements were performed at the HERA ep collider by the ZEUS group [@Breitweg:1997pa]$^-$[@Breitweg:2000su], also the H1 Collaboration has presented preliminary results [@h1]. In these experiments only events with isolated photons were included in the analysis, i.e. with a restriction imposed on the hadronic energy detected close to the photon. The corresponding cross sections for the photoproduction of an isolated photon and of an isolated photon plus jet were calculated in QCD in next-to-leading order (NLO) [@Gordon:1995km]$^-$[@Krawczyk:1998it]. There also exists analogous calculation for the large-$Q^2$ $ep$ collision (DIS events) [@Kramer:1998nb].
In this paper the results of the NLO QCD calculation for the DIC process with an isolated photon at the HERA $ep$ collider are presented. We consider the parton distributions in the photon and parton fragmentation into the photon as quantities of order $\alpha_{em}$. Our approach differs from the NLO approach [@Gordon:1995km]$^-$[@Gordon:1998yt] by set of subprocesses included in the analysis. The comparison of our predictions with the NLO results obtained by L.E. Gordon (LG) [@Gordon:1998yt] is presented for cross sections with kinematical cuts as in the ZEUS Collaboration measurements [@Breitweg:2000su].
The present analysis is the final, much extended and improved version of the previous one [@Krawczyk:1998it]. We show results for non-isolated final photon, and we study the influence of the isolation cut on the production rate of the photon. The role of other specific cuts applied by the ZEUS Collaboration are discussed and the comparison with data [@Breitweg:2000su] is made. We emphasize the importance of the box diagram $\gamma g\rightarrow \gamma g$, being the higher order process, in description of the data.
We study the renormalization scale dependence of the cross section in order to estimate the size of missing higher order (NNLO or higher) QCD corrections. The NLO results for the photoproduction of the isolated $\gamma$ are compared to the leading logarithm (LL) ones, and in addition the LL predictions for isolated $\gamma + jet$ final state are presented.
In the recent ZEUS analysis of the prompt photon plus jet production [@:2001aq] the intrinsic transverse momentum of partons in the proton was included in Monte Carlo simulations to improve agreement between data and predictions. This momentum is not included in our calculations.
We start with discussion on the choice of relevant diagrams defining our NLO approach to the DIC process (sec. \[sec:dic\]). The isolation of the photon is described in sec. \[sec:iso\], and the equivalent photon approximation in sec. \[sec:epa\]. In secs. \[results\] and \[th\] the results of numerical calculations are presented and compared with data [@Breitweg:2000su] and other NLO predictions [@Gordon:1998yt]. In sec. \[ll\] we show LL predictions for the photon plus jet production. Finally, sec. \[sec:sum\] summarizes our results.
The NLO calculation for $\gamma p\rightarrow\gamma X$ Deep Inelastic Compton scattering {#sec:dic}
=======================================================================================
General discussion {#sec:disc}
------------------
We start by describing processes which are (should be?) included in the NLO QCD calculations of the cross section for the DIC process, p X, \[eq:1\] where the final photon is produced with large transverse momentum, $p_T\gg\Lambda_{QCD}$. Although we will consider the process (\[eq:1\]), the problem which we touch upon is more general - it is related to different approaches to NLO calculations of cross sections for hadronic processes involving resolved photons, see [@Krawczyk:1990nq; @Krawczyk:1998it] and for more detailed discussion [@chyla].
0.5cm
The Born level contribution to the cross section for process (\[eq:1\]), i.e. the lowest order in the strong coupling $\alpha_s$ term, arises from the Compton process on the quark (fig. \[fig:born\]): qq. \[eq:2\] It gives the \[$\alpha_{em}^2$\] order contributions to the partonic cross section [^1]. At the same $\alpha_{em}^2$ order it contributes to the hadronic cross section for the process $\gamma p\rightarrow \gamma X$.
The Parton Model (PM) prediction for the DIC process (\[eq:1\]), which applies for $x_T=2p_T/\sqrt{S}\sim{\cal O}(1)$, relies solely on the Born contribution (\[eq:2\]) [@Bjorken:1969ja], namely: d\^[pX]{} =\_qdx\_p q\_p(x\_p)[d\^[qq]{}]{}, \[eq:3\] where $q_p$ is the quark density in the proton and $d\sigma^{\gamma p\rightarrow\gamma X}$ ($d\hat\sigma^{\gamma q\rightarrow \gamma q}$) stands for the hadronic (partonic) cross section. In the QCD improved PM the cross section is given by (\[eq:3\]), however with scale dependent quark densities. For semihard processes, where $x_T \ll 1$, the prediction based on the process (\[eq:2\]) only is not a sufficient approximation, and one should also consider the contributions corresponding to the collinear showers, involving hadronic-like interactions of the photon(s). There are two classes of such contributions: [*single resolved*]{} with resolved initial $or$ final photon, and [*double resolved*]{} with both the initial $and$ the final photon resolved (figs. \[fig:1res\], \[fig:2res\]). They correspond to partonic cross sections of orders \[$\alpha_{em} \alpha_s$\] (single resolved) and \[$\alpha_s^2$\] (double resolved). If one takes into account that partonic densities in the photon and the parton fragmentation into the photon are of order $\sim\alpha_{em}$, then the contributions to the hadronic cross section from these resolved photon processes are $\alpha_{em}^2 \alpha_s$ and $\alpha_{em}^2 \alpha_s^2$, respectively. Both single and double resolved contributions are included in the standard LL QCD analyses of the DIC process [@Duke:1982bj; @Aurenche:1984hc; @Gordon:1994sm].
-1.5cm
0.5cm -1.5cm
To obtain the NLO QCD predictions for the process (\[eq:1\]) the $\alpha_s$ corrections to the lowest order process (\[eq:2\]) have to be calculated leading to terms of order $\alpha_{em}^2\alpha_s$ [@Duke:1982bj; @Aurenche:1984hc; @mkcorr; @jan] (fig. \[fig:cor\]). In these $\alpha_{em}^2 \alpha_s$ contributions there are collinear singularities to be subtracted and shifted into corresponding quark densities [*or*]{} fragmentation functions. This way the single resolved photon contribution appears in the calculation of the $\alpha_s$ corrections to the Born process. It is worth noticing that in the NLO expression for the cross section there are no collinear singularities which would lead to the double resolved photon contributions. It indicates that taking into account \[$\alpha_s^2$\] subprocesses, associated with both the initial and final photons resolved, goes beyond the accuracy of the NLO calculation. This will be consistent within the NNLO approach, where $\alpha_s^2$ correction to the Born term and $\alpha_s$ correction to the single resolved terms should be included, all giving the same $\alpha_{em}^2\alpha_s^2$ order contribution to the hadronic cross sections.
The other set of diagrams is considered by some authors [@Gordon:1995km]$^-$[@Gordon:1998yt] in the NLO approach to DIC process (\[eq:1\]), due to their different way of counting the order of parton densities in the photon (and the parton fragmentation into the photon). This approach, which we will call “$1/\alpha_s$” approach, is motivated by large logarithms of $Q^2$ in the $F_2^{\gamma}$ existing already in the PM. By expressing $\ln (Q^2/\Lambda_{QCD}^2)$ as $\sim{1/\alpha_s}$ one treats the parton densities in photon as proportional to $\alpha_{em}/\alpha_s$ (see e.g. [@Fontannaz:1982et]$^-$[@Aurenche:1984hc; @Aurenche:1992sb; @Gordon:1994sm; @Gordon:1995km]$^-$[@Gordon:1998yt]). By applying this method to the DIC process, we see that the single resolved photon contribution to the hadronic cross section for $\gamma p\rightarrow\gamma X$ becomes of the same order as the Born term, namely 1=\_[em]{}\^2. The same is also observed for the double resolved photon contribution = \_[em]{}\^2.
We see that by such counting, the same $\alpha_{em}^2$ order contributions to the hadronic cross section are given by the direct Born process, single and double resolved photon processes although they correspond to quite different final states (observe a lack of the remnant of the photon in the direct process). Moreover, they constitute the lowest order (in the strong coupling constant) term in the perturbative expansion, actually the zeroth order, so the direct dependence of the cross section on the strong coupling constant is absent. If one takes into account that some of these terms correspond to the hard processes involving gluons, the lack of terms proportional to $\alpha_s$ coupling in the cross section seems to be contrary to intuition.
In the “$1/\alpha_s$” approach beside the $\alpha_s$ correction to the Born cross section, the $\alpha_s$ corrections to the single and to the double resolved photon contributions are included in the NLO calculation, since all of them give terms of the same order, $\alpha_{em}^2\alpha_s$ [@Gordon:1995km]$^-$[@Gordon:1998yt]. To summarize, the first approach starts with one basic, direct subprocess as in the PM (eq. \[eq:2\]), while the second one with three different types of subprocesses (as in the standard LL calculation). Obviously, some of NNLO terms in the first method belong to the NLO terms in the second one.
In this paper we apply the first type of NLO approach to the DIC process, however with some important NNLO terms additionally included. A comparison between our results and the results based on the other approach [@Gordon:1995km]$^-$[@Gordon:1998yt] is discussed in sec. \[results5\].
The cross section {#sec:cross}
-----------------
Below we describe our approach to the DIC process, where the parton densities in the photon and the parton fragmentation into the photon are treated as $\sim\alpha_{em}$.
In the NLO QCD calculation of the DIC process we take into account the following subprocesses:\
$\bullet$ the Born contribution (\[eq:2\]) (fig. \[fig:born\]);\
$\bullet$ the finite $\alpha_s$ corrections to the Born diagram (so called K-term) from virtual gluon exchange, real gluon emission (fig. \[fig:cor\]), and the process $\gamma g \rightarrow q \bar q \gamma$;\
$\bullet$ two types of single resolved photon contributions, with resolved initial and final photons (fig. \[fig:1res\]).
Besides the above full NLO set, we will include two terms of order $\alpha_{em}^2 \alpha_s^2$ (formally from the NNLO set): the double resolved contributions (fig. \[fig:2res\]) and the direct diagram (box) $\gamma g \rightarrow \gamma g$ [@Combridge:1980sx] (fig. \[fig:box\]), since they were found to be large [@Fontannaz:1982et]$^-$[@Aurenche:1992sb].
1.5cm
The cross section for the $\gamma p\rightarrow\gamma X$ scattering has the following form: E\_[d\^3\^[pX]{}d\^3p\_]{} = \_[b]{}dx f\_[b/p]{}(x,|[Q]{}\^2) [\_s(|Q\^2)2\^2 ]{} K\_b +\
+\_[abc]{}dx\_dx f\_[a/]{}(x\_,|[Q]{}\^2) f\_[b/p]{}(x,|[Q]{}\^2)\
D\_[/c]{}(z,|[Q]{}\^2) E\_[d\^3\^[abcd]{}d\^3p\_]{} The first term is the K-term describing the finite $\alpha_s$ corrections to the Born process, and the second one stands for the sum over all other contributions (including the Born contribution). The $f_{a/\gamma}$ ($f_{b/p}$) is a $a$ ($b$)-parton distribution in the photon (proton) while the $D_{\gamma /c}$ is a $c$-parton fragmentation function. For the direct initial (final) photon, where $a=\gamma$ ($c=\gamma$), we take $f_{a/\gamma}=\delta (x_{\gamma}-1)$ ($D_{\gamma /c}=\delta (z-1)$) (the Born contribution is obtained for $a=\gamma$, $b=q$ and $c=\gamma$). The variables $x_{\gamma}$, $x$ and $z$ stand for the fraction of the initial photon, proton, and $c$-parton momenta taken by the $a$-parton, $b$-parton, and the final photon, respectively. The renormalization scale is assumed equal to the factorization scale and is denoted as $\bar Q$.
The isolation {#sec:iso}
=============
In order to observe photons originating from a hard subprocess one should reduce backgrounds, mainly from $\pi^0$’s and $\gamma$’s radiated from final state hadrons. To achieve this, isolation cuts on the observed photon are introduced in experimental analyses. The isolation cuts are defined by demanding that the sum of hadronic energy within a cone of radius $R$ around the final photon, where the radius $R$ is defined in the rapidity and azimuthal angle space (see eq. \[eq:r\] in Appendix), should be smaller than the final photon energy multiplied by a small parameter $\epsilon$: \_[hadrons]{}E\_h<E\_.
The simplest way to calculate the differential cross section for an isolated photon, $d\sigma_{isol}$, is to calculate the difference of a non-isolated differential cross section, $d\sigma_{non-isol}$, and a subtraction term, $d\sigma_{sub}$ [@Berger:1990es]$^-$[@Gluck:1994iz; @Gordon:1995km]: d\_[isol]{}=d\_[non-isol]{}-d\_[sub]{}. The subtraction term corresponds to cuts opposite to the isolation cuts, i.e. hadrons with the total energy higher than the photon energy multiplied by $\epsilon$ should appear within a cone of radius $R$ around the final photon.
The isolation cuts are imposed only when calculating the K-term, and in contributions involving fragmentation function (resolved final photon). Other contributions arise from $2\rightarrow 2$ subprocesses with direct final photon that is isolated by definition.
In the analysis we apply the subtraction method with the subtraction term calculated in an approximate way, see [@Gordon:1994ut; @Gordon:1995km] for details. The approximation bases on the assumption that an angle $\delta$ between the final photon and a parton inside the cone of radius $R$ is small. It allows to simplify considerably the calculations and leads to compact analytical expressions for all relevant matrix elements involved in $d\sigma_{sub}$. Note that in this approximation the angle $\delta$ is simply proportional to the radius $R$: $\delta=R/cosh(\eta_{\gamma})$. The above small-$\delta$ approximation is used only in calculation of the $K$-term in the subtraction cross section $d\sigma_{sub}$, for the results see Appendix. Other contributions to $d\sigma_{sub}$, as well as $d\sigma_{non-isol}$ and all LL expressions, are obtained in an exact way.
It is worth mentioning that there is an ongoing discussion whether the conventional factorization breaks down, and whether the cross section is an infrared safe quantity for isolated photon photoproduction in $e^+e^-$ collisions (also for hadron-hadron reactions) [@Berger:1996cc; @Aurenche:1997ng]. In principle these questions could as well occur for the photoproduction of isolated photons in $ep$ collisions. However we do not deal with this problem because it arises from $2\rightarrow 3$ subprocesses in which a final quark fragmentates into a photon. We checked this explicitly and found that all singularities in $d\sigma_{sub}$ are canceled or factorized, as in $d\sigma_{non-isol}$. Therefore the considered by us cross section $d\sigma_{isol}$ is well defined (see also [@Gordon:1995km; @Gordon:1998yt; @Kramer:1998nb]).
The equivalent photon approximation {#sec:epa}
===================================
We consider the production of photons with large transverse momentum, $p_T\gg\Lambda_{QCD}$, in the $ep$ scattering, $ep\rightarrow e\gamma X$, at the HERA collider. This reaction is dominated by photoproduction events, i.e. the electron is scattered at a small angle and the mediating photon is almost real, $Q^2\approx 0$. The cross section for such processes can be calculated using the equivalent photon (Williams-Weizsäcker) approximation [@vonWeizsacker:1934sx] which relates the differential cross section for $ep$ collision to the differential cross section for $\gamma p$ collision. For the DIC scattering the approximation has the following form: d\^[epeX]{}=G\_[/e]{}(y) d\^[pX]{} dy , where $y$ is (in the laboratory frame) a fraction of the initial electron energy taken by the photon.
We apply the equivalent photon approximation and take the (real) photon distribution in the electron in the form [@Budnev:1974de]: G\_[/e]{}(y)=[\_[em]{}2]{} { [1+(1-y)\^2y]{}\
- [2y]{}(1-y-[m\_e\^2y\^2Q\^2\_[max]{}]{}) }, with $m_e$ being the electron mass. In the numerical calculations we assume $Q^2_{max}$ as 1 GeV$^2$ what is a typical value for the recent photoproduction measurements at the HERA collider.
The results and comparison with data {#results}
====================================
The results for the non-isolated and isolated photon cross sections are obtained in NLO accuracy with additional NNLO terms, as discussed in sec. \[sec:cross\]. We take the HERA collider energies: $E_e$=27.5 GeV and $E_p$=820 GeV [@Breitweg:2000su], and we consider the $p_T$ range of the final photon between 5 and 20 GeV ($x_T$ from 0.03 to 0.13). The calculations are performed in ${ \overline {\rm MS}}$ scheme with a hard (renormalization, factorization) scale $\bar{Q}$ equal $p_T$. Also $\bar{Q}=p_T/2$ and $2p_T$ are used to study the dependence of the results on the choice of $\bar{Q}$. We neglect the quark masses and assume the number of active flavors to be $N_f$=4 (and for comparison also $N_f$=3 and 5). The two-loop coupling constant $\alpha_s$ is used in the form \_s(|Q\^2)=[[4 ]{}]{} \[1-[[2\_1]{}]{} [ ]{}\] ($\beta_0=11-2/3 N_f$ and $\beta_1=51-19/3 N_f$), with $\Lambda_{QCD}$=0.365, 0.320 and 0.220 GeV for $N_f$=3, 4 and 5, respectively, as fitted by us to the experimental value of $\alpha_s(M_Z) = 0.1177$ [@Biebel:1999zt]. The $\Lambda_{QCD}^{LL}$ = 0.120 GeV for $N_f$=4 was taken in one-loop $\alpha_s$ when calculating the cross section in LL accuracy.
We use the GRV parametrizations of the proton structure function (NLO and LO) [@Gluck:1995uf], the photon structure function (NLO and LO) [@Gluck:1992ee], and the fragmentation function (NLO) [@Gluck:1993zx]. For comparison other parametrizations are also used: DO [@Duke:1982bj], ACFGP [@Aurenche:1992sb], CTEQ [@Lai:1997mg], MRST [@Martin:1998sq] and GS [@Gordon:1997pm].
As the reference we take the GRV NLO set of parton distributions [@Gluck:1995uf]$^-$[@Gluck:1993zx], $N_f = 4$, $\Lambda_{QCD} = 320$ GeV and $\bar{Q}=p_T$.
Non-isolated versus isolated photon cross section {#results1}
-------------------------------------------------
The $p_T$ distribution for the produced final photon without any cut is presented in fig. \[fig:dptincl\] where the NLO results and separately the Born term (with NLO parton densities) are shown. The cross section decreases by three orders of magnitude when $p_T$ increases from 4 GeV to 20 GeV, and obviously the most important contribution is coming from the lowest $p_T$ region. The subprocesses other than the Born one give all together contribution almost two times larger than the cross section for the Born subprocess alone.
The importance of particular contributions to the non-isolated cross section integrated over 5 GeV$<p_T<$ 10 GeV is illustrated in Table 1 (the first line). The total NLO cross section is equal to 226 pb, with individual contributions equal to: $Born=36.3\% $, $single$ $resolved=35.1\% $, $double$ $resolved=18.7\%$, $box=6.2\% $, $K$-$term$=3.9%. We see that the single resolved photon processes give a contribution comparable to the Born term. Also the double resolved photon processes are important. It is worth noticing that the overall double resolved photon cross section is build from many, relatively small, individual terms. The direct box diagram ($\gamma g\rightarrow\gamma g$) gives 17% of the Born ($\gamma q\rightarrow\gamma q$) contribution. Relatively large box contribution (although being $[\alpha_s^2]$) is such partially due to large gluonic content of the proton at small $x_p$.
Next, in fig. \[fig:Reps\] we compare the differential cross section $d\sigma /d\eta_{\gamma}$ for the non-isolated photon with corresponding predictions for the isolated photon using various values of the isolated cone variables ($\epsilon$, $R$) [^2]. The isolation cut suppresses the cross section by above 10% in the whole rapidity range. For $\epsilon$=0.1 and $R=1$ the suppression is 17-23% at rapidities $-1.5<\eta_{\gamma}\le 4$. This large effect is not too sensitive to the value of $\epsilon$: changing the value by a factor of 2 from $\epsilon =0.1$ to $\epsilon =0.2$ or to $\epsilon =0.05$ varies the results for isolated photon by about $4\%$ (fig. \[fig:Reps\]a). The dependence on $R$ is stronger but also not very large: when changing $R$ value by a factor of 2 from 1 to 0.5 the results increase by about $7\%$ (fig. \[fig:Reps\]b).
The suppression due to the isolation imposed on the photon is presented in Table 1 (the second line) for individual contributions and for the total cross section. As expected, the cross section for fragmentation processes (i.e. with resolved final photon) is strongly suppressed: after isolation it is lowered by a factor of 5. At the same time the QCD corrections to the Born diagram increase significantly, i.e. the contribution to the subtraction cross section, $d\sigma_{sub}$, due to this corrections is negative. The isolation restrictions do not modify contributions of other subprocesses since they involve photons isolated by definition. The subtraction cross section, being a sum of negative QCD corrections and fragmentation contributions, is of course positive and the total cross section for isolated final photon is lower, by $20\%$, than for the non-isolated one.
In the following we keep $R$=1 and $\epsilon$=0.1, standard values used in both theoretical and experimental analyses.
Other experimental cuts {#results1b}
-----------------------
In order to compare the results with data we consider other cuts imposed by the ZEUS group on prompt photon events at the HERA collider [@Breitweg:2000su]. The influence of the limited energy range, $0.2\le y\le 0.9$, is shown in fig. \[fig:cd\]. The cross section is strongly reduced, by 30-85%, in the positive rapidity region. At negative rapidities the change due to the $y$-cut is weaker: 5-10% at $-1.2<\eta_{\gamma}\le -0.4$ and 10-30% at other negative rapidities. Separately we show the results obtained without the box subprocess (fig. \[fig:cd\]). The box diagram contributes mainly in the rapidity region between -1 and 3. After imposing the $y$-cut its contribution is important in narrower region from -1 to 1. The influence of the $y$-cut can be read also from Table 1 (the third line). One sees e.g. that the Born contribution is reduced 3.5 times, while other ones are suppressed less, roughly by a factor of 2.
The results obtained for the isolated photon with the $y$-cut and in addition with the cut on the final photon rapidity, $-0.7\le\eta^{\gamma}\le 0.9$, are presented in the last line of Table 1. The restriction on $\eta^{\gamma}$ decreases the contributions of all subprocesses approximately by a factor of 2 (except for the double resolved contribution reduced almost 3 times).
The role of various experimental cuts is illustrated also in fig. \[fig:dxp\], this time for the $x_{\gamma}$ distribution. In particular we see that the isolation and the energy cut reduce considerably the contributions from large and medium $x_{\gamma}$, while the contributions from $x_{\gamma}$ below 0.1 are reduced less. On the other hand, the small $x_{\gamma}$ contributions are strongly, by two orders of magnitude, diminished by the photon rapidity cut. This shows that measurements at the central $\eta^{\gamma}$ region ($-0.7\le\eta^{\gamma}\le 0.9$) are not too sensitive to the small $x_{\gamma}$ values in the photon.
When calculating the QCD corrections to the Born process in the subtraction term $d\sigma_{sub}$ we used the small-$\delta$ approximation described in sec. \[sec:iso\]. Because these corrections give less than 10% of the cross section for the isolated photon production with various cuts (see the third column in Table 1), we expect that the error resulting from using the approximations is small, though we use in fact not a small value of $\delta$ ($\delta=R/cosh \eta_{\gamma}$, $R$ = 1). [^3]
The comparison with data {#results2}
------------------------
Two types of the final state were measured in the ZEUS experiment: 1) an isolated photon with $-0.7\le\eta^{\gamma}\le 0.9$ and $5\le p_T\le 10$ GeV; 2) an isolated photon plus jet with the photon rapidity and transverse momentum as above, the jet rapidity in the range $-1.5\le\eta^{jet}\le 1.8$, and the jet transverse momentum $p_T^{jet}\ge 5$ GeV.
We compare our NLO predictions with the ZEUS data from the first type of measurements [@Breitweg:2000su]. In fig. \[fig:dnNf\]a the comparison is made for the transverse momentum distribution for various $N_f$. Although the predictions tend to lie slightly below the data a satisfactory agreement is obtained for $N_f=4$. Note a large difference between the results for $N_f$=4 and 3 due to the fourth power of electric charge characterizing processes with two photons. We observe a very small contribution from the bottom quark (for $N_f$=5). The predictions are obtained in massless quark scheme and may overestimate the production rate.
Similar comparison of the NLO results with the data, now for the rapidity distribution, is shown in fig. \[fig:dnNf\]b. A good description of the data is obtained for $N_f$=4 and $N_f$=5 in the rapidity region $0.1\le\eta_{\gamma}\le 0.9$. For $-0.7\le\eta_{\gamma}\le 0.1$ our predictions lie mostly below the experimental points. This disagreement between predicted and measured cross sections is observed also for other theoretical calculations (LG) and for Monte Carlo simulations [@Breitweg:2000su] [^4]. In fig. \[fig:dnNf\]b we present separately an effect due to the box subprocess (for $N_f=4$). It is clear that the box term enhances considerably the cross section in the measured rapidity region. Its contribution to the integrated cross section is equal to 9.6%. The double resolved photon contribution is also sizeable, although roughly two times smaller than the box one, see Table 1 (fourth line). Both these $[\alpha_s^2]$ contributions improve description of the data.
The predictions obtained using three different NLO parton densities in the photon (ACFGP [@Aurenche:1992sb], GRV [@Gluck:1992ee] and GS [@Gordon:1997pm]) are presented for $N_f=4$ in fig. \[fig:dnpar\]a ($\bar{Q} = p_T$) and in fig. \[fig:dnpar\]b ($\bar{Q} = 2 p_T$) together with the ZEUS data [@Breitweg:2000su]. The results based on ACFGP and GRV parametrizations differ by less than 4% at rapidities $\eta_{\gamma} < 1$ (at higher $\eta_{\gamma}$ the difference is bigger), and both give good description of the data in the rapidity range $0.1\le\eta_{\gamma}\le 0.9$ (for $\bar{Q} = p_T$ and $\bar{Q} = 2 p_T$). For $-0.7\le\eta_{\gamma}\le 0.1$ none of the predictions is in agreement with the measured cross section.
For $\bar{Q}$=$p_T$ (fig. \[fig:dnpar\]a) the GS distribution leads to results considerably below ones obtained using ACFGP and GRV densities, especially in the rapidity region from roughly -1 to 1. This difference between the GS and other considered herein parton parametrizations is mainly due to their different treatment of the charm quark in the photon. In the GS approach the charm quark is absent for $\bar{Q}^2$ below 50 GeV$^2$. Since we take $5\leq \bar{Q} = p_T\leq 10$ GeV, and the most important contribution to the cross section arises from the lower $p_T$ region (see fig. \[fig:dnNf\]a), the $\bar{Q}^2$ value usually lies below the GS charm quark threshold. As a consequence, predictions based on GS have strongly suppressed the contribution of subprocesses involving charm from the photon - contrary to GRV and ACFGP predictions where the charm threshold is at lower $\bar{Q}^2$.
The above explanation of differences between cross sections involving GS and both GRV and ACFGP parton densities is insufficient for higher rapidities, $\eta_{\gamma} > 2$. Here the differences between the results based on particular photon parametrizations are bigger, especially when comparing predictions obtained using GRV and ACFGP ones (not shown). This is due to large differences between used parton densities at low $x_{\gamma}$, which is probed at the high rapidity region.
All the considered parton distributions give similar description of the data when the scale is changed to $\bar{Q}=2p_T$, see fig. \[fig:dnpar\]b. Here the calculation corresponds to $\bar{Q}^2$ which is always above 50 GeV$^2$ and the charm density in the GS parametrization is non-zero, as in other parametrizations.
In fig. \[fig:bins\] our predictions are compared to the ZEUS data divided into three ranges of $y$. This allows to establish that the above discussed discrepancy between the data and the predictions for $\eta^{\gamma} < 0.1$ is coming mainly from the low $y$ region, $0.2<y<0.32$. In the high $y$ region, $0.5<y<0.9$, a good agreement is obtained.
We have also studied the dependence of our results on the choice of the parton distributions in the proton and parton fragmentation into the photon (not shown). Cross sections calculated using GRV [@Gluck:1995uf], MRST (set ft08a) [@Martin:1998sq] and CTEQ4M [@Lai:1997mg] NLO parton parametrizations for the proton vary among one another by 4 to 7% at negative rapidities and less than $4\%$ at positive rapidity values. Results for the isolated final photon are also not too sensitive to the fragmentation function. For rapidity ranging from -1 to 4 the cross section obtained with DO LO [@Duke:1982bj] fragmentation function is $2-3.5\%$ lower than the cross section based on GRV NLO [@Gluck:1993zx] parametrization. Only at minimal ($\eta_{\gamma}<-1$) and maximal ($4<\eta_{\gamma}$) rapidity values this difference is larger, being at a level of $3.5-8\%$.
The theoretical uncertainties of the results and comparison with other NLO predictions {#th}
======================================================================================
As we already mentioned the predictions are obtained in massless quark scheme and may overestimate the production rate. An improved treatment of the charm quark, especially in the box contribution which is particularly sensitive to the change from $N_f=3$ to $N_f=4$, would be needed. However we do not expect that this improvement would change qualitatively our results.
We now discuss the theoretical uncertainties of our predictions related to the perturbative expansion.
The dependence on the $\bar{Q}$ scale {#results3}
-------------------------------------
In order to estimate the contribution due to missing higher order terms, the influence of the choice of the $\bar Q$ scale is studied for the $\eta_{\gamma}$ distribution. In fig. \[fig:dnQ\]a the results obtained using GRV densities with and without the $y$-cut are shown. When changing $\bar Q$ from $p_T$ to 2$p_T$ ($p_T$/2) the cross section increases (decreases) at rapidities below $\sim 1$ and decreases (increases) at higher rapidity values. Only at high rapidities (where the cross section is small), $\eta_{\gamma} > 3$, the dependence on the choice of the scale is strong, above 10% (up to 20-30% at $\eta_{\gamma}\approx 5$). In the wide kinematical region, $-2 < \eta_{\gamma} < 2$, the relative differences between results (with and without the $y$-cut) for $\bar Q = p_T$ and results for $\bar Q = 2p_T$ or $p_T/2$ are small and do not exceed 6%. Around the maximum of the cross section at rapidities $-1\le\eta_{\gamma}\le 0$ these differences are 4-6%. This small sensitivity of the results to the change of the scale is important since it indicates that the contribution from neglected NNLO and higher order terms is not significant.
Note that individual contributions are strongly dependent on the choice of $\bar{Q}$, e.g. results for the single resolved processes vary by $\pm$10-20% at rapidities $\eta_{\gamma}\le 1$. Results are much more stable only when the sum of resolved processes and QCD corrections is considered.
In fig. \[fig:dnQ\]b we present NLO results for various $\bar{Q}$ with and without the $y$-cut, however this time with no box contribution. At rapidities $\eta_{\gamma}<1$ the uncertainty due to the choice of the renormalization scale is about two times higher than for the cross section with included box diagram, so the box contribution ($\sim [\alpha_s^2]$) seems to stabilize the NLO prediction. At rapidities $\eta_{\gamma}> 2$ the relative dependence on the choice of the scale is similar for the cross section with and without the box term.
The comparison of NLO and LL predictions {#results4}
----------------------------------------
In the present calculation we include in LL accuracy the single and double resolved photon processes as well as the box diagram in addition to the Born contribution, see also [@Fontannaz:1982et]$^-$[@Aurenche:1984hc; @Gordon:1994sm] (although this is not fully consistent with the discussion in sec. \[sec:dic\]). The cross section for the $\gamma p\rightarrow\gamma X$ scattering in LL accuracy is obtained by convolution of partonic cross sections with relevant LO parton densities.
In fig. \[fig:jet\] we show the LL prediction for the isolated $\gamma$ photoproduction (dotted line) together with NLO predictions (solid line) and the Born contribution only (dot-dashed line). The highest differences between the LL and NLO cross sections are seen at the rapidity range $-0.5 < \eta_{\gamma} < 2.5$ where the LL results lie 10-20% below the NLO ones (fig. \[fig:jet\]a). For the $p_T$ distribution this difference is 10-14% in the whole presented range of the transverse momentum, $4\le p_T\le 20$ GeV (fig. \[fig:jet\]b).
We think that the observed difference between NLO and LL results together with the weak dependence on the $\bar{Q}$ scale discussed in sec. \[results3\] indicates reliability of the calculation.
The comparison with other NLO results {#results5}
-------------------------------------
As we discussed in Sec. \[sec:dic\], our NLO calculation of the DIC process differs from the “$1/\alpha_s$”-type NLO analysis presented in ref. [@Gordon:1995km]$^-$[@Gordon:1998yt], by set of diagrams included in the calculation. We do not take into account $\alpha_s$ corrections to the single and double resolved processes, which are beyond the NLO accuracy in our approach. On the other hand, we include the box diagram neglected in [@Gordon:1995km]$^-$[@Gordon:1998yt]. (The double resolved subprocesses are included in both analyses.)
We compare our results and the results of the LG calculation [@Gordon:1998yt] (using $N_f$=4 and $\bar{Q}=p_T$) for the isolated final photon ($R$=1, $\epsilon$=0.1) in the kinematical range as in the ZEUS analysis [@Breitweg:2000su] (i.e. for $-0.7\le\eta^{\gamma}\le 0.9$ and $0.2\le y\le 0.9$). First we use the GRV photon parton densities. The LG predictions for $d\sigma /dp_T$ cross section are about 20% higher than ours in the presented range of transverse momentum, $4\le p_T\le 20$ GeV. For $d\sigma /d\eta^{\gamma}$ cross section (with $5\le p_T\le 10$ GeV) the biggest differences are at $\eta^{\gamma}$=0.9 where the LG results are about 35% higher. The differences decrease towards negative rapidity values and are negligible at $-0.7\le\eta^{\gamma} < -0.5$. For $y$ range limited to low values only, 0.2 $< y <$ 0.32, the LG cross section is higher than ours by up to 20% at positive $\eta_{\gamma}$, while at negative $\eta_{\gamma}$ it is lower by up to 10%. For large $y$ values, 0.5 $< y <$ 0.9, where our predictions agree with data, the LG results are higher than ours by up to 80% (at $\eta_{\gamma}$ = 0.9).
As it was already discussed, for $\bar{Q}=p_T$ the GS photon distributions lead to results lower by 11-14% than those obtained with GRV densities at rapidities between -1 and 1 (see sec. \[results2\]). In calculations presented in [@Gordon:1995km]$^-$[@Gordon:1998yt] this difference is even twice larger.
The LG predictions [@Gordon:1998yt] obtained using the GS parametrization lie up to 20% below ours (also based on GS distributions, with $\bar{Q} = p_T$ and $N_f = 4$) at rapidities $-0.7\le\eta_{\gamma}\le 0.2$, and they are higher than ours by up to 30% for $0.2\le\eta_{\gamma}\le 0.9$ [@Gordon:1998yt; @Breitweg:2000su].
The LL prediction for $\gamma + jet$ photoproduction {#ll}
====================================================
The ZEUS Collaboration has also analyzed the prompt photon photoproduction in which in addition a hadron jet is measured [@Breitweg:1997pa; @unknown:1998uj; @:2001aq]. In fig. \[fig:jet\] we show the LL prediction for the isolated $\gamma + jet$ final state [^5] together with the predictions for the $\gamma$ alone. The following jet rapidity and transverse momentum are assumed: $-1.5\leq\eta_{jet}\leq 1.8$ and $p_T^{jet}>5$ GeV, respectively. This additional cuts for the final state imposed on jets decrease the cross section, especially at high rapidities. The LL predictions for the $\gamma +jet$ are lower than those for the $\gamma$ production by 5-10% at negative rapidities and by 10-80% at positive rapidity values. The difference between both LL results is about 10% in wide range of transverse momenta, $\sim 6\le p_T\le 20$ GeV, and only for the lower $p_T$ region, $4\le p_T\le 6$ GeV, it is higher (13-23%).
Summary {#sec:sum}
=======
Results of the NLO calculation, with NNLO contributions from double resolved photon processes and box diagram, for the isolated $ \gamma$ production in the DIC process at HERA are presented [^6]. The role of the kinematical cuts used in the ZEUS measurement [@Breitweg:2000su] are studied in detail.
The results obtained using GRV parametrizations agree with the data in shape and normalization for $p_T$ distribution. For $\eta^{\gamma}$ distribution a good description of the data is obtained for $\eta^{\gamma}>0.1$, while for $\eta^{\gamma}<0.1$ the data usually lie above the predictions. This discrepancy arises mainly from the low $y$ region, $0.2\le y\le 0.32$. The beyond NLO terms, especially a box contribution, improve the description of the data.
We have studied the theoretical uncertainty of results due to the choice of the renormalization/factorization scale: $\bar{Q} = p_T/2, p_T, 2p_T$. At high rapidities $\eta_{\gamma}> 3$, where the cross section is small, the uncertainty is 10-30%. In a wide range of rapidities, $-2\le\eta_{\gamma}\le 2$, the dependence on the $\bar{Q}$ scale is small, below 6%. Since we include some NNLO diagrams in our NLO calculation, this stability of the predictions versus the change of the scale is especially important. The week dependence on the $\bar{Q}$ scale, and not large differences between LL and NLO predictions (below 20%) allows to conclude that theoretical uncertainties of our NLO calculations for an isolated photon production in the DIC process at HERA are relatively small.
We compared our results with the LG ones based on different set of subprocesses. The cross section $d\sigma/dp_T$ obtained by LG is about 20% higher than ours (for GRV photonic parton distributions). For the cross section $d\sigma/d\eta_{\gamma}$ this difference is up to 35% at $\eta_{\gamma} = 0.9$. The highest differences are present for high $y$ values only, $0.5 < y < 0.9$, where on the other hand our predictions are in agreement with the data. At low $y$ range, $0.2 < y < 0.32$, differences between both calculations are smaller and none of them describe the data well for rapidities below 0.1.
Acknowledgements {#acknowledgements .unnumbered}
================
We would like to acknowledge P. Bussey, Sung W. Lee and L.E. Gordon for important discussions. Supported in part by Polish State Committee for Scientific Research, grants number 2P03B18410 (1998-1999), 2P03B11414 (2000) and 2P03B05119 (2000-2001), by Interdisciplinary Centre for Mathematical and Computational Modelling, Warsaw University, Grant No G16-10 (1999-2000), and by European Commission 50th framework contract HPRN-CT-2000-00149.
References {#references .unnumbered}
==========
[99]{}
J. D. Bjorken and E. A. Paschos, Phys. Rev. [**185**]{} (1969) 1975 T. Tu, C. Wu, Nucl. Phys. B [**156**]{} (1979) 493
M. Fontannaz and D. Schiff, Z. Phys. C [**14**]{} (1982) 151 D. W. Duke and J. F. Owens, Phys. Rev. D [**26**]{} (1982) 1600; Err: Phys. Rev. D [**28**]{} (1983) 1227
P. Aurenche, A. Douiri, R. Baier, M. Fontannaz and D. Schiff, Z. Phys. C [**24**]{} (1984) 309 M. Krawczyk, Acta Physica Polonica B [**21**]{} (1990) 999 A. C. Bawa, M. Krawczyk and W. J. Stirling, Z. Phys. C [**50**]{} (1991) 293; A. C. Bawa and M. Krawczyk, Phys. Lett. B [**262**]{} (1991) 492; A. C. Bawa and M. Krawczyk, Proc. “Physics at HERA”, Hamburg 1991, p. 579, IFT-17-91;
M. Krawczyk, IFT-17-92, [*Talk given at 27th Rencontres de Moriond: QCD and High Energy Hadronic Interactions, Les Arcs, France, 1992*]{}
P. Aurenche, P. Chiappetta, M. Fontannaz, J. P. Guillet and E. Pilon, Z. Phys. C [**56**]{} (1992) 589 L. E. Gordon and J. K. Storrow, Z. Phys. C [**63**]{} (1994) 581 ZEUS Coll., J. Breitweg et al., Phys. Lett. B [**413**]{} (1997) 201 ZEUS Coll., [*Subm. to the XXIXth International Conference on High-Energy Physics (ICHEP 98), Vancouver, Canada, 1998, abstract 815*]{}
ZEUS Coll., J. Breitweg et al., Phys. Lett. B [**472**]{} (2000) 175 H1 Coll., [*Subm. to the International Europhysics Conference on High Energy Physics, HEP97, Jerusalem, Israel, 1997, abstract 265*]{}
L. E. Gordon and W. Vogelsang, Phys. Rev. D [**52**]{} (1995) 58 L. E. Gordon and W. Vogelsang, hep-ph/9606457, [*Talk presented by W. Vogelsang at the Int. Workshop on Deep Inelastic Scattering and Related Phenomena, Rome, Italy, 1996*]{} L. E. Gordon, Phys. Rev. D [**57**]{} (1998) 235 M. Krawczyk and A. Zembrzuski, hep-ph/9810253, IFT 98/12, [*Contribution to the XXIXth International Conference on High Energy Physics, ICHEP’98, Vancouver, Canada, 1998, abstract 889*]{} G. Kramer, D. Michelsen and H. Spiesberger, Eur. Phys. J. C [**5**]{} (1998) 293;\
A. Gehrmann-De Ridder, G. Kramer and H. Spiesberger, Phys. Lett. B [**459**]{} (1999) 271; Eur. Phys. J. C [**11**]{} (1999) 137; Nucl. Phys. B [**578**]{} (2000) 326 \[ZEUS Collaboration\], hep-ex/0104001. J. Chýla, JHEP 0004 (2000) 007; hep-ph/9811455; hep-ph/0010140; hep-ph/0010309
M. Krawczyk, unpublished
J. Żochowski, [*“The corrections of order $\alpha_S$ in Deep Inelastic Compton Scattering”, MS Thesis, 1992*]{}
B. L. Combridge, Nucl. Phys. B [**174**]{} (1980) 243 E. L. Berger and J. Qiu, Phys. Lett. B [**248**]{} (1990) 371; E. L. Berger and J. Qiu, Phys. Rev. D [**44**]{} (1991) 2002 L. E. Gordon and W. Vogelsang, Phys. Rev. D [**50**]{} (1994) 1901 M. Glück, L. E. Gordon, E. Reya and W. Vogelsang, Phys. Rev. Lett. [**73**]{} (1994) 388 E. L. Berger, X. Guo and J. Qiu, Phys. Rev. Lett. [**76**]{} (1996) 2234; Phys. Rev. D [**54**]{} (1996) 5470; hep-ph/9610497; hep-ph/9708408 P. Aurenche, M. Fontannaz, J. P. Guillet, A. Kotikov and E. Pilon, Phys. Rev. D [**55**]{} (1997) R1124;\
S. Catani, M. Fontannaz and E. Pilon, Phys. Rev. D [**58**]{} (1998) 094025 C. F. von Weizsäcker, Z. Phys. [**88**]{} (1934) 612;\
E. J. Williams, Phys. Rev. [**45**]{} (1934) 729; E.J. Williams, Kgl. Danske Vidensk. Selskab. Mat.-Fiz. Medd. 13 (1935) N4
V. M. Budnev, I. F. Ginzburg, G. V. Meledin and V. G. Serbo, Phys. Rep. [**15**]{} (1975) 181;\
R. Nisius, Phys. Rep. [**332**]{} (2000) 165 O. Biebel, P. A. Movilla Fernandez, S. Bethke and the JADE Coll., Phys. Lett. B [**459**]{} (1999) 326 M. Glück, E. Reya and A. Vogt, Z. Phys. C [**67**]{} (1995) 433 M. Glück, E. Reya and A. Vogt, Phys. Rev. D [**45**]{} (1992) 3986; M. Glück, E. Reya and A. Vogt, Phys. Rev. D [**46**]{} (1992) 1973 M. Glück, E. Reya and A. Vogt, Phys. Rev. D [**48**]{} (1993) 116 H.L. Lai, J. Huston, S. Kuhlmann, F. Olness, J. Owens, D. Soper, W.K. Tung, and H. Weerts, Phys. Rev. D [**55**]{} (1997) 1280 A. D. Martin, R. G. Roberts, W. J. Stirling and R. S. Thorne, Eur. Phys. J. C [**4**]{} (1998) 463 L. E. Gordon and J. K. Storrow, Nucl. Phys. B [**489**]{} (1997) 405 L. E. Gordon, Nucl. Phys. B [**501**]{} (1997) 197
------------------------------------------- -------------------------- --------------------------- --------------------------------- -------------------------- ------------------------------------ ------------------------------------ ------------------------------------
\[12pt\]\[6pt\][\[pb\]]{} \[12pt\]\[6pt\][total]{} \[12pt\]\[6pt\][Born]{} \[12pt\]\[6pt\][($\alpha_S$)]{} \[12pt\]\[6pt\][box]{} \[12pt\]\[6pt\][single resolved]{} \[12pt\]\[6pt\][single resolved]{} \[12pt\]\[6pt\][double resolved]{}
\[0pt\]\[6pt\] \[0pt\]\[6pt\] \[0pt\]\[6pt\] \[0pt\]\[6pt\] \[0pt\]\[6pt\] \[0pt\]\[6pt\][initial $\gamma$]{} \[0pt\]\[6pt\][final $\gamma$]{} \[0pt\]\[6pt\]
\[12pt\]\[6pt\][non-isolated]{} \[12pt\]\[6pt\][226.2]{} \[12pt\]\[6pt\][82.1]{} \[12pt\]\[6pt\][8.7]{} \[12pt\]\[6pt\][13.9]{} \[12pt\]\[6pt\][54.7]{} \[12pt\]\[6pt\][24.6]{} \[12pt\]\[6pt\][42.2]{}
\[0pt\]\[6pt\] \[0pt\]\[6pt\] \[0pt\]\[6pt\][(36.3%)]{} \[0pt\]\[6pt\][(3.8%)]{} \[0pt\]\[6pt\][(6.1%)]{} \[0pt\]\[6pt\][(24.2%)]{} \[0pt\]\[6pt\][(10.9%)]{} \[0pt\]\[6pt\][(18.7%)]{}
\[12pt\]\[6pt\][isolated]{} \[12pt\]\[6pt\][180.4]{} \[12pt\]\[6pt\][82.1]{} \[12pt\]\[6pt\][15.2]{} \[12pt\]\[6pt\][13.9]{} \[12pt\]\[6pt\][54.7]{} \[12pt\]\[6pt\][5.12]{} \[12pt\]\[6pt\][9.37]{}
\[0pt\]\[6pt\] \[0pt\]\[6pt\] \[0pt\]\[6pt\][(45.5%)]{} \[0pt\]\[6pt\][(8.4%)]{} \[0pt\]\[6pt\][(7.7%)]{} \[0pt\]\[6pt\][(30.3%)]{} \[0pt\]\[6pt\][(2.8%)]{} \[0pt\]\[6pt\][(5.2%)]{}
\[12pt\]\[6pt\][isolated]{} \[12pt\]\[6pt\][72.33]{} \[12pt\]\[6pt\][23.6]{} \[12pt\]\[6pt\][6.33]{} \[12pt\]\[6pt\][6.54]{} \[12pt\]\[6pt\][28.2]{} \[12pt\]\[6pt\][2.34]{} \[12pt\]\[6pt\][5.29]{}
\[0pt\]\[6pt\][$y$ cut]{} \[0pt\]\[6pt\] \[0pt\]\[6pt\][(32.6%)]{} \[0pt\]\[6pt\][(8.8%)]{} \[0pt\]\[6pt\][(9.0%)]{} \[0pt\]\[6pt\][(39.0%)]{} \[0pt\]\[6pt\][(3.2%)]{} \[0pt\]\[6pt\][(7.3%)]{}
\[12pt\]\[6pt\][isolated]{} \[12pt\]\[6pt\][35.36]{} \[12pt\]\[6pt\][13.6]{} \[12pt\]\[6pt\][3.32]{} \[12pt\]\[6pt\][3.41]{} \[12pt\]\[6pt\][11.9]{} \[12pt\]\[6pt\][1.21]{} \[12pt\]\[6pt\][1.92]{}
\[0pt\]\[6pt\][$y, \eta_{\gamma}$ cuts]{} \[0pt\]\[6pt\] \[0pt\]\[6pt\][(38.5%)]{} \[0pt\]\[6pt\][(9.4%)]{} \[0pt\]\[6pt\][(9.6%)]{} \[0pt\]\[6pt\][(33.7%)]{} \[0pt\]\[6pt\][(3.4%)]{} \[0pt\]\[6pt\][(5.4%)]{}
------------------------------------------- -------------------------- --------------------------- --------------------------------- -------------------------- ------------------------------------ ------------------------------------ ------------------------------------
0.5cm
0.5cm -7cm
-0.8cm 4.6cm
4.6cm
0.5cm -7cm
0.5cm -7cm
0.5cm -7cm -7cm
0.5cm -7cm
0.5cm -7cm
[^1]: We denote the order of partonic subprocesses using square brackets.
[^2]: The positive rapidity is pointed in the proton direction.
[^3]: Calculations for the prompt photon production in $ep$ [@Gordon:1998yt] and hadron-hadron [@Gordon:1997kc] collisions were performed using space slicing method without the small $\delta$ assumption. Comparison of such results with predictions obtained in (discussed here) approximated way showed that the small $\delta$ approximation is an accurate analytic technique for including isolation effects in NLO calculations (also for $R$ = 1) [@Gordon:1998yt].
[^4]: The ZEUS Collaboration has presented recently an analysis [@:2001aq] in which an intrinsic transverse momentum of partons in the proton, $k_T$, was introduced in the PYTHIA 6.1 generator in order to improve agreement between the data and Monte Carlo predictions for an isolated photon plus jet photoproduction. The data, selected with $x_{\gamma}>0.9$, are consistent with the predictions for $<k_T> = 1.69 \pm 0.18 ^{+0.18}_{-0.20}$ GeV.
[^5]: The NLO calculation for $\gamma + jet$ photoproduction will be discussed in our next paper.
[^6]: Our fortran code is available upon request from [email protected].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'A recently introduced lattice model, describing an extended system which exhibits a [*reentrant*]{} (symmetry-breaking, second-order) noise-induced nonequilibrium phase transition, is studied under the assumption that the multiplicative noise leading to the transition is [*colored*]{}. Within an effective Markovian approximation and a mean-field scheme it is found that when the self-correlation time $\tau$ of the noise is different from zero, the transition is [*also reentrant*]{} with respect to the spatial coupling $D$. In other words, at variance with what one expects for equilibrium phase transitions, a large enough value of $D$ favors [*disorder*]{}. Moreover, except for a small region in the parameter subspace determined by the noise intensity $\sigma$ and $D$, an increase in $\tau$ usually [*prevents*]{} the formation of an ordered state. These effects are supported by numerical simulations.'
address:
- |
Departamento de Física, Facultad de Ciencias Exactas y Naturales\
Universidad Nacional de Mar del Plata\
Deán Funes 3350, 7600 Mar del Plata, Argentina.
- |
Departament de Física, Universitat de les Illes Balears and\
Instituto Mediterráneo de Estudios Avanzados, IMEDEA (CSIC-UIB)\
E-07071 Palma de Mallorca, Spain
- |
Centro Atómico Bariloche, Comisión Nacional de Energía Atómica, and Instituto Balseiro[^1]\
8400 San Carlos de Bariloche, Argentina.
author:
- 'Sergio E. Mangioni[^2] and Roberto R. Deza[^3]'
- 'Raúl Toral[^4]'
- 'Horacio S. Wio[^5]'
title: 'Nonequilibrium phase transitions induced by multiplicative noise: effects of self-correlation'
---
Introduction
============
In the last few decades we have witnessed a paradigmatic shift regarding the role of fluctuations, from the equilibrium picture of merely being a $N^{-1/2}$ perturbation on thermodynamic averages—or triggering at most phase transitions between well defined minima of the free energy [@Pa]—to lead a host of new and amazing phenomena in far from equilibrium situations. As examples we may cite noise-induced unimodal-bimodal transitions in some zero-dimensional models [@HL], stochastic resonance in zero-dimensional and extended systems [@M; @W], noise-induced spatial patterns [@ga93], noise-delayed decay of unstable states [@ag98], ratchets [@ratchets], shifts in critical points [@shifts], etc.
Recently it has been shown that a white and Gaussian [*multiplicative*]{} noise can lead an [*extended*]{} dynamical system (fulfilling appropriate conditions) to undergo a [*phase*]{} transition towards an [*ordered*]{} state, characterized by a nonzero order parameter and by the breakdown of ergodicity [@VPT]. This result—first obtained within a Curie–Weiss-like mean-field approximation, and further extended to consider the simplest correlation function approach—has been confirmed through extensive numerical simulations [@VPTK]. In addition to its [*critical*]{} nature as a function of the noise intensity $\sigma$, the newly found noise-induced phase transition has the noteworthy feature of being [*reentrant*]{}: for each value of $D$ above a threshold one, the ordered state exists only inside a window $[\sigma_1,\sigma_2]$. At variance with the known case of [*equilibrium*]{} order-disorder transitions that are induced (in the simplest lattice models) by the nearest-neighbor coupling constant $D$ and rely on the bistability of the local potential, the transition in the case at hand is led by the [*combined effects*]{} of $D$ and $\sigma$ through the nonlinearities of the system. Neither the zero-dimensional system (corresponding to the $D=0$ limit) nor the deterministic one ($\sigma=0$) show any transition.
This counterintuitive ordering role of noise has also been found afterwards in different models in the literature [@ga96; @VPAH; @ge98; @ki97; @mu97; @ga98]. In Refs.[@ga96; @VPAH], the authors study a noise-induced reentrant transition in a time-dependent Ginzburg–Landau model with both additive and multiplicative noises. Ref.[@ge98] introduces another simple model with a purely multiplicative noise, which also presents a noise-induced reentrant transition. This reference also gives evidence that the universality class of its critical behavior is that of the [*multiplicative noise*]{} [@gr96; @tu97] (see also Ref.[@ra95] for a discussion of the universality class of these models). In Ref.[@ki97], a first-order phase transition induced by noise is obtained in a system of globally coupled oscillators. A similar first-order phase transition is also found in Ref.[@mu97]. Finally, Ref.[@ga98] addresses the role of multiplicative noise in the context of phase-separation dynamics.
Although for the sake of mathematical simplicity all these references (in particular, Refs.[@VPT; @VPTK]) studied only the white-noise case (the only exception is reference [@ga98] in which colored noise in space, white in time, is considered) it is expected that, because of their nature, fluctuations coupled [*multiplicatively*]{} to the system will show some degree of time-correlation or “color” [@HL; @S; @SS; @DL; @HJ], and hence new effects may arise from this fact. For example, a reentrant behavior has been found recently as a consequence of color in a noise-induced transition [@CSW] and an ordering nonequilibrium phase transition can be induced in a Ginzburg–Landau model by varying the correlation time of the additive noise [@ga92; @ga94]. Thus motivated, we have studied the consequences of a [*finite*]{} (but still very short as compared with the “deterministic” or coarse-grained time scales) self-correlation time $\tau$ of the multiplicative noise in systems of this kind. We now recall some of the new effects that emerge in this colored-noise case, which have been briefly reported in a recent work [@MDWT]:
- Our main finding is that, as a consequence of the multiplicative character of the noise, a strong enough spatial coupling $D$ leads invariably (for $\tau>0$) to a [*disordered*]{} state, contrary to what would be expected to happen in equilibrium statistical-mechanical models.
- Another important result is that, except for large values of $\sigma$ and [*very small*]{} values of $\tau$, color has an [*inhibiting*]{} role for ordered states. Moreover, there exists a [*finite*]{} and not very large value of $\tau$ beyond which order is impossible.
These results represent a major departure from what one can expect on the basis of equilibrium thermodynamics, according to which one should tend to think that as $D\to\infty$ an [*ordered*]{} situation is favored. Whereas that “intuitive” notion remains valid if the multiplicative noise that induces the nonequilibrium ordering phase transition is white [@VPT; @VPTK], it ceases to be so for $\tau>0$. In the former case, the results could be interpreted in terms of a “freezing” of the short-time behavior by a strong enough spatial coupling: for $D/\sigma^2\to\infty$, the stationary probability distribution could be considered to be $\delta$-like, just as the initial one. In our case, an analysis of the short-time behavior of the order parameter up to first order in $\tau$ reveals that the disordering effect of $D$ can only be felt for [*nonzero*]{} self-correlation time. As $\tau$ increases, the minimum value of $D$ required to stabilize the disordered phase decreases and the region in parameter space available to the ordered phase shrinks until it vanishes. Thus, the foregoing results can only be interpreted once we recall that the ordered phase arises from the cooperation of [*two*]{} factors: the presence of spatial coupling [*and*]{} the multiplicative character of the noise (which may eventually lead to “counterintuitive” results).
It is our purpose in this work to render an explicit account of our calculation and, at the same time, to expose and to discuss the results more thoroughly. After presenting the model in section \[model\], we begin section \[approx\] by introducing the approximations needed to render the problem accessible to mathematical analysis. We resort to a mean-field approximation like in Refs.[@VPT; @VPTK] and to a “unified colored noise approximation" (UCNA) [@JH; @CWA], devised to deal with self-correlated noises. In section \[simpler\] a simplified treatment using the aforementioned approximations is given and in section \[refined\] we expose the more sophisticated approach that was actually used to obtain the phase diagrams. In this approach the UCNA is complemented with an appropriate interpolation scheme [@CWA]. In section \[result\] we expose and discuss the results obtained within the last approach, comparing the phase diagram with the ones arising from the simplified approach and (for $\tau\simeq 0$) from a perturbative expansion, and the $D$-dependence of the order parameter $m$ for nonzero $\tau$ with a numerical simulation [@MDWT]. A final discussion of the approach and its results is made in section \[conclu\].
The model {#model}
=========
The model under consideration has been introduced in Refs.[@VPT; @VPTK; @MDWT]: a $d$-dimensional extended system of typical linear size $L$ is restricted to a hypercubic lattice of $N=L^d$ points, whereas time is still regarded as a continuous variable. The state of the system at time $t$ is given by the set of stochastic variables $\{x_i(t)\}$ ($i=1,\dots,N$) defined at the sites ${\bf r}_i$ of this lattice, which obey a system of coupled ordinary stochastic differential equations (SDE) $$\label{eq:1}
\dot{x}_i=f(x_i)+g(x_i)\eta_i+\frac{D}{2d}\sum_{j\in n(i)}(x_j-x_i)$$ (throughout the paper, the Stratonovich interpretation for the SDE’s will be meant). Eqs.(\[eq:1\]) are the discrete version of the [*partial*]{} SDE which in the continuum would determine the state of the extended system: we recognize in the first two terms the generalization of Langevin’s equation for site $i$ to the case of multiplicative noise ($\eta_i$ is the [*colored*]{} multiplicative noise acting on site ${\bf r}_i$). For the specific example analyzed in Ref.[@VPT], perhaps the simplest one exhibiting the transition under analysis (see, however, Ref.[@ge98]), $$\label{eq:2}
f(x)=-x(1+x^2)^2,\hspace{2.0cm}g(x)=1+x^2.$$ The last term in Eqs.(\[eq:1\]) is nothing but the lattice version of the Laplacian $\nabla^2 x$ of the extended stochastic variable $x({\bf r},t)$ in a reaction-diffusion scheme. $n(i)$ stands for the set of $2d$ sites which form the immediate neighborhood of the site ${\bf r}_i$, and the coupling constant $D$ between neighboring lattice sites is the diffusion coefficient.
As previously stated, it is our aim in this work to investigate the effects of the self-correlation time $\tau$ of the multiplicative noise on the model system just described. To that end we must assume a specific form for the noises $\{\eta_i\}$: we choose Ornstein–Uhlenbeck noises, i.e. Gaussian distributed stochastic variables with zero mean and exponentially decaying correlations: $$\label{eq:3}
\langle\eta_i(t)\eta_j(t')\rangle=\delta_{ij}(\sigma^2/2\tau)
\exp(-|t-t'|/\tau).$$ They arise as solutions of an [*uncoupled*]{} set of Langevin SDE: $$\label{eq:4}
\tau\dot{\eta}_i=-\eta_i+\sigma\xi_i$$ where the $\{\xi_i(t)\}$ are white noises—namely, Gaussian stochastic variables with zero mean and $\delta$-correlated: $\langle\xi_i(t)\xi_j(t')\rangle=\delta_{ij}\delta(t-t')$. For $\tau\to 0$, the Ornstein–Uhlenbeck noise $\eta_i(t)$ approaches the white-noise limit $\xi^W_i(t)$ with correlations $\langle\xi^W_i(t)
\xi^W_j(t')\rangle=\sigma^2\,\delta_{ij}\,\delta(t-t')$, which was the case studied in Refs.[@VPT; @VPTK].
The approximations {#approx}
==================
The non-Markovian character of the process $\{x_i\}$ (due to the presence of the colored noise $\{\eta_i\}$) makes it difficult to study. However, some approximations have been devised which render a [*Markovian*]{} process that—whereas nicely simplifying the treatment—can still capture some of the essential features of the complete non-Markovian one. The aforementioned UCNA is one of them: in fact a very reliable one, because of its ability to reproduce the small– and large-$\tau$ limits [@JH]. By resorting to interpolation schemes, one can retrieve meaningful results in wider vicinities of these limits [@CWA].
As already declared, our approach is a mean-field-like one. The earlier we make this kind of assumptions in the calculation, the cruder the approximation will turn out to be. In order to find the phase diagram in the presence of colored noise we have made the mean-field approximation at some late stage, so enhancing the precision of the calculation. However, since this calculation is a tedious one, we shall first expose a simpler approximation which brings out most qualitative results. We aim in this way to underline the physical origin of the results presented in section IV. The differences arising from both calculations are pointed out there.
A simpler approach {#simpler}
------------------
The simpler mean-field approximation follows closely Curie–Weiss’ mean-field approach to magnetism, and consists in replacing the last term in Eqs.(\[eq:1\]) $$\label{eq:5}
\Delta_{i}\equiv\frac{D}{2d}\sum_{j\in n(i)}(x_j-x_i)$$ by $$\label{eq:6}
\bar\Delta_{i}\equiv D(\bar{x}-x_i),$$ where $\bar{x}$ is a [*parameter*]{} that will be determined self-consistently. In other words, the (short-ranged) interactions are substituted by a time– and space-independent “external” field whose value [*depends on the state*]{} of the system. Since in this approximation Eqs.(\[eq:1\]) get immediately decoupled, there is no use in keeping the subindex $i$ and we may refer to the systems in Eqs. (\[eq:1\]) and (\[eq:4\]) as if they were single equations. Hereafter, the primes will indicate derivatives with respect to $x$ (clearly $\bar\Delta'=-D$).
If we take the time derivative of Eq.(\[eq:1\]), replace first $\dot{\eta}$ in terms of $\eta$ and $\xi$ from Eq.(\[eq:4\]) and then $\eta$ in terms of $\dot x$ and $x$ from Eq.(\[eq:1\]), we obtain the following [*non-Markovian*]{} SDE: $$\label{eq:7}
\tau(\ddot{x}-\frac{g'}{g}
\dot{x}^2)=-\left(1-\tau\left[(f+\bar\Delta)'-\frac{g'}{g}
(f+\bar\Delta)\right]\right)\dot{x}+(f+\bar\Delta)+\sigma g\xi.$$ The aforementioned UCNA allows us to recover a Markovian SDE: for our particular problem it amounts, on one hand, to a usual adiabatic elimination (namely, neglecting $\ddot x$) and, on the other, to neglect ${\dot x}^2$ so that the system’s dynamics be governed by a Fokker–Planck equation [@WCSPR]. The resulting equation, being [*linear*]{} in $\dot x$ (but of course not in $x$), can be immediately solved for $\dot x$ giving $$\label{eq:8}
\dot x=Q(x;\bar{x})+S(x;\bar{x})\xi,$$ with $$\begin{aligned}
Q(x;\bar{x})&\equiv&(f+\bar\Delta)\theta,\label{eq:9}\\
S(x;\bar{x})&\equiv&\sigma g\theta,\label{eq:10}\\
\theta(x;\bar{x})&\equiv&\{1-\tau g[(f+\bar\Delta)/g]'\}^{-1}
\label{eq:11}.\end{aligned}$$ The parametric dependence of $Q(x)$ and $S(x)$ on $\bar{x}$ has been written explicitly.
The Fokker–Planck equation associated to the SDE Eq.(\[eq:8\]) is $$\label{eq:12}
\partial_t P(x,t;\bar{x})=
-\partial_x\left[R_1(x;\bar{x})P(x,t;\bar{x})\right]+
\frac{1}{2}\partial^2_x\left[R_2(x;\bar{x})P(x,t;\bar{x})\right],$$ with drift and diffusion coefficients given by [@risken]: $$\begin{aligned}
\label{eq:13}
R_1(x;\bar{x})&=&Q+\frac{1}{4}(S^2)'\\
\label{eq:14}
R_2(x;\bar{x})&=&S^2.\end{aligned}$$ The solution of the time-independent Fokker–Planck equation leads to the stationary probability density $$\label{eq:15}
P^{st}(x;\bar{x})={\cal N}^{-1}\exp{\left[\int_{0}^{x}dx'
\frac{2R_1(x';\bar{x})-\partial_{x'}R_2(x';\bar{x})}{R_2(x';\bar{x})}
\right]},$$ being ${\cal N}$ its norm. The partial-derivative notation $\partial_{x'}$ in Eq.(\[eq:15\])—as in Eqs. (\[eq:18\]) and (\[eq:19\]) below—is only a reminder of the parametric dependence of $R_1$, $R_2$ on $\bar{x}$.
The value of $\bar{x}$ arises from a self-consistency relation, once we equate it to the average value of the random variable $x_i$ in the stationary state $$\label{eq:16}
\bar{x}=\langle x\rangle\equiv\int_{-\infty}^{\infty}dx\,x\,
P^{st}(x;\bar{x})\equiv F(\bar{x}).$$ As in the known Curie–Weiss mean-field approach, the condition $$\label{eq:17}
\left.\frac{dF}{d\bar{x}}\right|_{\bar{x}=0}=1$$ allows us to find the transition line between the ordered and the disordered phases. Here also, $F(\bar{x})$ is a smooth odd function such that Eq.(\[eq:16\]) has always a root at $\bar{x}=0$ and for $\left.dF/d\bar{x}\right|_{\bar{x}=0}>1$ it has two nontrivial roots which differ only in sign. The condition Eq.(\[eq:17\]) thus reads: $$\label{eq:18}
{\cal N}^{-1}\int_{-\infty}^{\infty}dx\:x\int_0^x dx'\exp
{\left[\int_{0}^{x'}dx''\:\frac{2R_1-\partial_{x''}R_2}{R_2}\right]}
\partial_{\bar{x}}\left.\left(\frac{2R_1-\partial_{x'}
R_2}{R_2}\right)\right|_{\bar{x}=0}=1,$$ where $$\label{eq:19}
\left.{\cal N}=\int_{-\infty}^{\infty}dx\exp{\left[\int_{0}^{x}dx'
\frac{2R_1-\partial_{x'}R_2}{R_2}\right]}\right|_{\bar{x}=0}.$$ Eqs. (\[eq:18\]) and (\[eq:19\]) must be solved numerically in order to find the lines in parameter space ($\sigma$,$\tau$,$D$) that separate ordered ($\bar{x}\neq 0$) from disordered ($\bar{x}=0$) phases. The results of this calculation will be shown in section V. Next, we introduce a more refined approach in which the mean-field approximation is made at a later stage in the calculation.
A more refined approach {#refined}
-----------------------
As we shall see, the relations obtained with this more sophisticated approach are similar to Eqs. (\[eq:15\]) through (\[eq:19\]), but with different expressions for the functions $R_1(x;\bar{x})$ and $R_2(x;\bar{x})$. The idea here is to [*introduce first*]{} the UCNA, without yet resorting to the mean-field approximation. In the following, $\Delta_i$ has the same meaning as in Eq.(\[eq:5\]) and, as it occurred previously with $\bar\Delta'$, it satisfies $\Delta_i'=-D$. Note however that whereas for $f_i\equiv f(x_i)$ and $g_i\equiv g(x_i)$ the prime has the meaning of a total derivative with respect to $x_i$, for $\Delta_i'$ and all the functions involving them, its meaning is really that of a [*partial*]{} derivative with respect to $x_i$. Proceeding as before, i.e. taking the time derivative of Eqs.(\[eq:1\]) and using Eqs. (\[eq:1\]) and (\[eq:4\]) to eliminate the $\eta$’s in favor of the $x$’s and $\xi$’s, we obtain the following system of (non Markovian) SDE’s $$\label{eq:20}
\tau(\ddot{x}_i-\frac{g'_i}{g_i}\dot{x}_i^2)=-\left[1-\tau
g_i\left(\frac{f_i+\Delta_i}{g_i}\right)'\right]{\dot
x}_i+(f_i+\Delta_i)+\sigma g_i\xi_i+\frac{D\tau}{2d}\sum_{j\in
n(i)}{\dot x}_j.$$
The UCNA proceeds here through the neglect of ${\ddot x}_i$ and of $({\dot x}_i)^2$, so retrieving a [*linear*]{} equation in the $\dot
x$’s (but of course not in the $x$’s), which can be rewritten as $$\begin{aligned}
\label{eq:21}
{\dot x}_i&=&
\left[1-\tau g_i\left(\frac{f_i+\Delta_i}{g_i}\right)'\right]^{-1}
\left[(f_i+\Delta_i)+\sigma g_i\xi_i+\frac{D\tau}{2d}
\sum_{j\in n(i)}{\dot x}_j\right]\nonumber\\
&=&[Q_i+S_i\xi_i]+\frac{\beta_i}{2d}\sum_{j\in n(i)}{\dot x}_j.\end{aligned}$$ Here the quantities $$\begin{aligned}
\theta_i&\equiv&\left[1-\tau
g_i\left(f_i+\Delta_i/g_i\right)'\right]^{-1},\label{eq:22}\\
Q_i&\equiv & (f_i+\Delta_i) \theta_i,\label{eq:23}\\
S_i&\equiv & \sigma g_i\theta_i,\label{eq:24}\\
\beta_i&\equiv&D\tau\theta_i\label{eq:25}\end{aligned}$$ have been defined in order to simplify the notation. Note that although only the dependence upon $x_i$ has been made explicit in this notation, these quantities depend also (through $\Delta_i$) on the values $x_j$ at the neighboring sites.
Now, assuming the lattice to be isotropic, we apply to this set a [*mean-field-like approximation*]{} (but not yet the main one) consisting in [*replacing*]{} in all the functions appearing in Eq.(\[eq:21\]) the $2d$ neighbors $x_j$ of the variable $x_i$ [*by a common value*]{} $y_i$. Through this procedure one reduces to [*two*]{} the number of different coupled SDE’s: one for $x\equiv
x_i$ and another for its nearest neighbor variable $y\equiv y_i$. These are $${\dot a}=h_a+g_{ab}\,\xi_b,\label{eq:26}$$ where a sum over the values $x,y$ is implied for the indices $a,b$, and the noise variables satisfy $\langle\xi_a(t)\xi_b(t')\rangle=
\delta_{ab}\,\delta(t-t')$. If, similarly as before, we define $$\begin{aligned}
\theta(x,y)&=&\left[1-\tau g(x)\frac{\partial}{\partial x}
\left(\frac{f(x)+D(y-x)}{g(x)}\right)\right]^{-1},\label{eq:27}\\
Q(x,y)&=&[f(x)+D(y-x)]\theta(x,y),\label{eq:28}\\
S(x,y)&=&\sigma g(x)\theta(x,y),\label{eq:29}\\
\beta(x,y)&=&\tau D\theta(x,y),\label{eq:30}\\
A(x,y)&=&[1-\beta(x,y)\beta(y,x)]^{-1}=A(y,x),\label{eq:31}\end{aligned}$$ and write ${\bar a}=y$ if $a=x$ and viceversa, then the explicit forms of the functions in Eq.(\[eq:26\]) are $$\label{eq:32}
h_a=A(x,y)\,[Q(a,{\bar a})+\beta(a,{\bar a})Q({\bar a},a)]$$ and $$\begin{aligned}
g_{ab}&=&A(x,y)\,S(a,{\bar a})\qquad\qquad\mbox{ if }b=a,
\label{eq:34}\\
&=&A(x,y)\,\beta(a,{\bar a})S({\bar a},a)\quad\mbox{ if }b={\bar a}.
\label{eq:36}\end{aligned}$$ The bivariate Fokker–Planck equation associated to Eqs.(\[eq:26\]) is $$\label{eq:38}
\partial_t P=-\partial_a(R_aP)
+\frac{1}{2}\partial_a\partial_b(R_{ab}P),$$ where $P=P(x,y;t)$. According to standard techniques [@risken], the drift and diffusion coefficients are given respectively by $$\begin{aligned}
R_a(x,y)&=&h_a+\frac{1}{2}g_{bc}\partial_b(g_{ac}),\label{eq:39}\\
R_{ab}(x,y)&=&g_{ac}g_{bc}.\label{eq:40}\end{aligned}$$ Since the denominators occurring in these equations may become zero for some values of $x$ or $y$, we resort to an [*interpolation procedure*]{} (analogous to the one used in Refs.[@CSW; @CWA]) consisting in replacing the expression (\[eq:31\]) for $A(x,y)$ by $$\label{eq:41}
A(x,y)=\frac{1-\beta(x,y)\beta(y,x)}{1+\beta(x,y)^2\beta(y,x)^2}.$$ Since $\beta(x,y)$ is proportional to $\tau$, it follows that the expression in Eq.(\[eq:31\]) coincides with the interpolated one, Eq.(\[eq:41\]), as $\tau\to 0$ and $\tau\to\infty$ (the latter limit meaning indeed “$\tau$ comparable with the ‘deterministic’ time scales”).
By integrating the bivariate Fokker–Planck Eq.(\[eq:38\]) with respect to $y$ we obtain a single-variable equation, which in the stationary case reads $$\label{eq:42}
0=-\partial_x(R_1 P^{st})+\frac{1}{2}\partial^2_x(R_2 P^{st}),$$ being $R_1 P^{st}$ and $R_2 P^{st}$ functions of $x$ only: $$\begin{aligned}
R_1 P^{st}&\equiv&\int_{-\infty}^{\infty}dy\,P^{st}(x,y)\,
R_x(x,y),\label{eq:43}\\
R_2 P^{st}&\equiv&\int_{-\infty}^{\infty}dy\,P^{st}(x,y)\,
R_{xx}(x,y).\label{eq:44}\end{aligned}$$
Here it is when we resort to the [*main mean-field-type approximation*]{}, resembling the Curie–Weiss’ type of approach used in Ref.[@VPT]: [*assuming*]{} $P^{st}(x,y)\approx
P^{st}(x)\delta(y-\bar{x})$, $R_1$ and $R_2$ in Eqs.(\[eq:43\]) and (\[eq:44\]) are approximated by $$\begin{aligned}
R_1&=&R_x(x;\bar{x}),\label{eq:45}\\
R_2&=&R_{xx}(x;\bar{x}).\label{eq:46}\end{aligned}$$ In this way, from the stationary [*joint*]{} probability density function $P^{st}(x,y)$ we retrieve an [*effective*]{} single-variable one $P^{st}(x;\bar{x})$ whose expression in terms of $R_1(x;\bar{x})$ and $R_2(x;\bar{x})$ arising from Eq.(\[eq:42\]) is the same as in Eq.(\[eq:15\]). The value of $\bar{x}$ follows again from a self-consistency relation like Eq.(\[eq:16\]). The procedure to find the phase diagram is the same as in the foregoing subsection and the explicit expression of the condition $\left.dF
/d\bar{x}\right|_{\bar{x}=0}=1$ is given by Eqs. (\[eq:18\]) and (\[eq:19\]), this time in terms of the corresponding $R_1$, $R_2$ given by Eqs. (\[eq:45\]) and (\[eq:46\]).
As discussed in Ref.[@MDWT], although the kind of approximation leading to Eq.(\[eq:41\]) is not of a perturbative nature, it has provided sound results in the cases analyzed heretofore [@CSW; @CWA]. Nonetheless, for the sake of comparison we have also adapted to the case of a multiplicative noise a known perturbative procedure [@ST]. Within this context, the expressions for $R_1(x;\bar{x})$ and $R_2(x;\bar{x})$ come from a consistent small-$\tau$ expansion of a Fokker–Planck equation (Eqs.(\[eq:12\]) or (\[eq:38\])): $$\begin{aligned}
R_1(x;\bar{x})&=&(f+\bar\Delta)+\sigma^2 x
\{g[1+\tau(f+\bar\Delta)']+\tau(f+\bar\Delta)\},\label{eq:47}\\
R_2(x;\bar{x})&=&\{\sigma g[1+\tau(f+\bar\Delta)']\}^2.\label{eq:48}\end{aligned}$$
The results {#result}
===========
Phase diagram
-------------
In the following we shall describe the results obtained through the numerical solution of Eqs. (\[eq:18\]) and (\[eq:19\]) in the more refined approach, i.e. with $R_1$ and $R_2$ as prescribed by Eqs. (\[eq:45\]) and (\[eq:46\]). We shall also compare these results with the ones arising from Eqs. (\[eq:13\]) and (\[eq:14\]), and with a perturbative expansion for small $\tau$. Figures \[mdtw1\] to \[mdtw3\] are respectively the projections onto the $\sigma$–$D$, $\tau$–$\sigma$ and $\tau$–$D$ planes, of cuts of the boundary separating the ordered and disordered phases performed at fixed values of the remaining parameters.
Let us first concentrate on Fig.\[mdtw1\]: it corresponds to Fig.1 in Ref.[@MDWT], but is the result of an improved calculation based on the more refined mean-field approach described in Sect.\[refined\]. The novelty is that, at least for $\tau$ not too small, it is now evident that the region available to the ordered phase is [*bounded*]{}. The noteworthy aspects are the following:
1. \[a\] As in the white-noise case $\tau=0$ (Refs.[@VPT; @VPTK]), the ordering phase transition is [*reentrant with respect to $\sigma$*]{}: for a range of values of $D$ that depends on $\tau$, ordered states can only exist within a window $[\sigma_1,\sigma_2]$. The fact that this window shifts to the right [*for small $\tau$*]{} means that, for fixed $D$, color [*destroys*]{} order just above $\sigma_1$ but [*creates*]{} it just above $\sigma_2$.
2. \[b\] For fixed $\sigma>1$ and $\tau\ne 0$, ordered states exist [*only within a window*]{} of values for $D$. Thus the ordering phase transition is [*also reentrant with respect to $D$*]{}. For $\tau$ small enough the maximum value of $D$ compatible with the ordered phase increases rather steeply with $\sigma$, reaching a maximum around $\sigma\sim 5$ and then decreases gently. For $\tau\geq 0.1$ it becomes evident (in the ranges of $D$ and $\sigma$ analyzed) that the region sustaining the ordered phase is [*closed*]{}, and shrinks to a point for a value slightly larger than $\tau=0.123$.
3. \[c\] For fixed values of $\sigma>1$ and $D$ larger than its minimum for $\tau=0$, the system [*always*]{} becomes disordered for $\tau$ large enough. The maximum value of $\tau$ consistent with order altogether corresponds to $\sigma\sim 5$ and $D\sim 32$. In other words, ordering is possible [*only*]{} if the multiplicative noise inducing it has short memory.
4. \[d\] The fact that the region sustaining the ordered phase finally shrinks to a point means that even for that small region in the $\sigma$–$D$ plane for which order is induced by color, a further increase in $\tau$ destroys it. In other words, the phase transition is [*also reentrant with respect to $\tau$*]{}. For $D$ large enough there may exist even [*two*]{} such windows.
Some of the features just described become more evident by looking at Fig.\[mdtw2\]:
- the existence of a maximum correlation time consistent with ordering for each value of $D$ (occurring for an optimal value of $\sigma$) (\[c\]);
- the ordering ability (\[a\]) of a very small amount of color for $\sigma>\sigma_2(D)$ ($\sigma_2(\tau,D)$ increases very rapidly at first);
- the reentrance with respect to $\tau$ and even the occurrence of a [*double*]{} reentrance for $D$ large enough (\[d\]).
Fig.\[mdtw3\] represents another way of seeing the reentrance with respect to $D$ for constant $\sigma$ (large enough) and the fact that there exists a maximum $\tau$ consistent with order for each value of $\sigma$ (being it largest for $\sigma\sim 5$). The scarce dependence of $D$ on $\tau$ in the lower branch—as well as its almost linear dependence on $\sigma$—is easily understood by looking at the rightmost branch of Fig.\[mdtw1\].
In Figs. \[mdtw4\] to \[mdtw6\] we compare the results just shown—obtained as we said using the the more refined approach of Sect.\[refined\]—with the ones arising from the simpler one (Sect.\[simpler\]). Figure \[mdtw4\] corresponds to Fig.\[mdtw2\], whereas Figs. \[mdtw5\] and \[mdtw6\] focus respectively on the $\tau=0.03$ and $\tau=0.015$ curves in Fig.\[mdtw1\]. Not only does the simpler approach (grossly) overestimate the size of the ordered region but also—as one may infer from Figs. \[mdtw4\] and \[mdtw5\]—it seems to predict unbounded ordered regions.
Figure \[mdtw6\] corresponds to a rather small value of $\tau$, so that a comparison with the results obtained by using expressions (\[eq:47\]) and (\[eq:48\]) makes sense. For $\sigma$ and $D$ small enough the curves almost coincide. As it is well known, the simultaneous consideration of small $\tau$ and large $\sigma$ cannot be done independently [@SS; @DL; @HJ]. In the present case, a similar effect arises when we consider large values of $D$, as we discuss below. Hence, in order to consider larger values of $\sigma$ and $D$, one should take [*extremely low*]{} values of $\tau$. As Fig.\[mdtw6\] shows, already for $\tau=0.015$ there is an apparent discrepancy between the perturbative results and the mean-field ones even for not so large values of $\sigma$ and $D$. The noteworthy fact is that the perturbative expansion also tends to indicate the existence of a reentrance with respect to $D$.
Order parameter
---------------
The order parameter in this system is $m\equiv|\bar{x}|$ namely, the positive solution of the consistency equation (see Eq.(\[eq:16\]) in Sect.\[simpler\]). In Fig.\[mdtw7\] we plot $m$ vs. $\sigma$ for $D=20$ and different values of $\tau$. Consistently with what has been discussed in (\[a\]) and (\[c\]) and shown in Fig.\[mdtw1\], we see that as $\tau$ increases the window of $\sigma$ values where ordering occurs shrinks until it disappears. One also notices that at least for this $D$, the value of $\sigma$ corrresponding to the maximum order parameter varies very little with $\tau$. Figure \[mdtw8\] is a plot of $m$ vs. $\tau$ for $D=45$ and two values of $\sigma$ ($\approx 7.07$ and $\approx 8.94$) that illustrates the case of double reentrance in $\tau$.
Since the previous results have been obtained in the mean-field approximation, we have also performed numerical simulations in order to have an independent check of the predictions. As a representative example—corresponding to the phenomenon (\[b\]) above (the destruction of the ordered phase by an increasing coupling constant $D$)—we plot jointly in Fig.\[mdtw9\], for rather small values of $\sigma$ ($=2$) and $\tau$ ($=0.01$), the $D$-dependence of the order parameter as predicted by our mean-field approach and as resulting from a numerical integration of the original SDE’s, Eqs.(\[eq:1\]). We have taken three different lattice sizes in order to assess finite-size effects. As we see, the numerical simulations [*do*]{} predict the disordering for large enough $D$, and even the maximum ordering occurs in a region which is consistent with the mean-field prediction. This comparison warns us, however, that the mean-field approximation can severely underestimate the size of the ordered region.
The [*short-time evolution*]{} of $\langle x\rangle$ can be obtained multiplying Eq.(\[eq:12\]) by $x$ and integrating: $$\label{eq:49}
\frac{d\langle x\rangle}{dt}=\int_{-\infty}^{\infty} dx
R_1(x;\bar{x})P(x,t;\bar{x}).$$ Let us assume an initial condition such that at early times $P(x,t\sim
0;\bar{x})=\delta(x-\bar{x})$. Equating $\bar{x}=\langle x\rangle$ as before, we obtain $$\label{eq:50}
\frac{d\langle x\rangle}{dt}=R_1(\bar{x},\bar{x})$$ (again, we can use for $R_1$ the result Eq.(\[eq:13\]) of the simple approximation or the more elaborate one given by Eq.(\[eq:45\])). The numerical solution of Eq.(\[eq:50\]) has an initial [*rising*]{} period (it is initially [*unstable*]{}) reaching very soon a maximum and tending to zero afterwards.
For $D/\sigma^2\to\infty$, Eq.(\[eq:50\]) results to be valid also in the [*asymptotic regime*]{} since $P^{st}(x)=\delta(x-\bar{x})$ [@VPTK]. In Ref.[@VPTK] an equivalent equation is obtained in the $\tau=0$ case for [*both*]{} limits ($D=0$ [*and*]{} $D/\sigma^2\to\infty$) being there interpreted in terms of a “freezing” of the short-time behavior. According to this criterion, in the $D/\sigma^2\to\infty$ limit the system undergoes a second-order phase transition [*if*]{} the corresponding zero-dimensional model presents [*a linear instability in its short-time dynamics*]{}, i.e. if after linearizing Eq.(\[eq:50\]): $$\label{eq:51}
\langle\dot x\rangle=-\alpha\langle x\rangle$$ one finds that $\alpha<0$. We then see that the trivial (disordered) solution $\langle x\rangle=0$ is stable only for $\alpha>0$. For $\alpha<0$ other stable solutions with $\langle x\rangle\ne 0$ appear, and the system develops order through a genuine [*phase*]{} transition. In this case, $\langle x\rangle$ can be regarded as the [*order parameter*]{}. In the white noise limit $\tau=0$ this is known to be the case for sufficiently large values of the coupling $D$ and for a window of values for the noise amplitude $\sigma\in[\sigma_1,\sigma_2]$.
We discuss now how the stability of the ordered phase is altered by nonzero values of $\tau$. If we linearize Eq.(\[eq:50\]) using the expression of $R_1(\bar{x},\bar{x})$ from Eq.(\[eq:13\]), we obtain $$\label{eq:52}
\alpha=
\frac{(1+\tau+\tau D)^2-\sigma^2(1-3\tau+2\tau D)}{(1+\tau+\tau D)^3}.$$ If we use instead Eq.(\[eq:45\]), the result can be written exclusively in terms of $\tau D$: $$\label{eq:53}
\alpha=1-\sigma^2\frac{B(\tau D)}{A(\tau D)}$$ with $$\begin{aligned}
A(x)&=&1+5x+8x^2+3x^3-3x^4-x^5+2x^6+x^7,\label{eq:54}\\
B(x)&=&1+3x+x^2-5x^3-3x^4+3x^5+x^6,\label{eq:55}\end{aligned}$$ so the instability occurs at $\sigma^2=A/B$. Now, the ratio $B/A$ keeps always below 1 in the positive range, has a minimum of $\sim
0.05$ at $\tau D\sim 1.09$ and a maximum of $\sim 0.36$ at $\tau D\sim
2.33$. If we consider e.g. the conditions in Fig.\[mdtw5\] (namely $\tau=0.03$) and take $D$ large enough so that $\tau D>2.33$, one can see that Eq.(\[eq:53\]) approximates the left boundary better than Eq.(\[eq:52\]) does. In the limit $\tau\ll 1$ (but still finite) Eq.(\[eq:53\]) can be approximated to the expression reported in Ref.[@MDWT], namely $$\label{eq:56}
\alpha=\frac{1+\tau D-\sigma^2}{1+\tau D}.$$ It is worthwile to stress the fact (evident from Eqs. (\[eq:52\]), (\[eq:53\]) and (\[eq:56\])) that the value of $D$ has no effect on the [*location*]{} of the instability by itself, but only through the combination $\tau D$: according to Eq.(\[eq:56\]) the stable phase is the disordered one ($\langle x\rangle=0$) for $1+\tau
D>\sigma^2$ (since $\alpha>0$) and the ordered one ($\langle
x\rangle\neq 0$) for $1+\tau D<\sigma^2$. In summary, whereas the noise intensity $\sigma$ has a destabilizing effect on the disordered phase, as soon as $\tau\neq 0$ the spatial coupling $D$ tends to [*stabilize*]{} it. For $\tau=0$ the last effect is not present, being then $\sigma>1$ the condition for ordering [@VPT; @VPTK]. Considering that the effect of even a tiny correlation is enhanced by $D$, we can understand the abrupt change in slope (from negative to positive) shown in Fig.\[mdtw1\] as soon as $\tau\neq 0$. Note the approximately inverse relation between $\tau$ and $D$ for fixed $\sigma$ on the upper branches of Fig.\[mdtw3\], even when Eq.(\[eq:51\]) is strictly valid for $D/\sigma^2\to\infty$.
Conclusions {#conclu}
===========
This work has focused on the effects of a self-correlation in the multiplicative noise, on the reentrant noise-induced phase transition introduced in Ref.[@VPT]. Whereas in a recent Letter we have reported the most relevant results [@MDWT], it has been our purpose in the present work to expose in more detail the techniques and the approximations employed. We also discuss more thoroughly the results, adding new figures and enriching the contents of others.
Through the use of the UCNA we recovered a Markovian behavior for the system, and through an interpolation scheme similar to the one introduced in Ref.[@CWA] we resolved indeterminacies in the equations describing it. We stress that the equation resulting from this interpolation scheme are [*exact*]{} in the limits $\tau=0$ and $\tau\to\infty$. In addition to the fact that the interpolation scheme has been already applied with success in other works [@CSW; @CWA], the goodness of our approximations for small but nonetheless finite values of $\tau$ has been checked against a standard perturbative expansion [@ST] (adapted for multiplicative noise). It is worth to emphasize the fact that these approximations are, so far, the [*only*]{} tool available for an analytical treatment of this essentially non-Markovian problem.
The main result is that for $\tau\neq 0$, the order established as a consequence of the multiplicative character of the noise can be [*destroyed*]{} by a strong enough spatial coupling. Figure \[mdtw1\] shows that for given $\tau$ (0.03) and $\sigma$, the ordered phase can only exist between definite values of $D$. In particular, the upper bound on $D$ decreases roughly as $\tau^{-1}$ for given $\sigma$.
The foregoing result can be understood by recalling that the ordered phase arises as a consequence of the [*collaboration*]{} between the multiplicative character of the noise and the presence of spatial coupling. The disordering effect of $D$ arises [*only*]{} when $\tau\ne 0$ (the results in Ref.[@VPTK]—rightly interpreted in terms of a “freezing” of the short-time behavior by a strong enough spatial coupling—are thus consistent with ours). As $\tau$ increases, the minimum value of $D$ required to stabilize the disordered phase decreases rapidly, and the region in parameter space available to the ordered phase shrinks until it disappears.
The example worked throughout this paper shows vividly the fact that the conceptual inheritance from equilibrium thermodynamics (though often useful) is not always applicable. The equilibrium-thermodynamic lore would induce us to think that as $D\to\infty$ an ordered situation is favored [@Pa]. Although this is certainly true for the Curie-Weiss model (since in that case the deterministic potential is itself bistable and an increase of spatial coupling has the effect of rising the potential barrier between the stable states), it is [*not*]{} in the case we are dealing with, since the deterministic potential is [*monostable*]{}. Hence, it is the combined effects of the multiplicative noise [*and*]{} the spatial coupling that induce the transition.
As a summary, whereas one might say that the value of Refs.[@VPT; @VPAH; @VPTK] is that they tell experimentalists where [*not*]{} to look for a noise-induced phase transition—namely, in those systems which are prone to exhibit a usual (zero-dimensional) noise-induced transition, and for too large noise intensity—the present work tells moreover that, due to the consideration of the more realistic colored noise source, [*an ordered phase is not to be found for large values of the spatial coupling either*]{}. Though the specific choice of the forms for the functions $f(x)$ and $g(x)$ may appear to some as physically unmotivated, is up to our knowledge the [*simplest*]{} one exhibiting this phenomenon. We believe nonetheless that the phenomenon is robust and transcends the specific choice made in this work.
The next obvious step is to consider a finite correlation length in the lattice model, which requires to go beyond the mean field approach. This problem is being presently studied. 1.cm [**Acknowledgments:**]{} R. T. acknowledges financial support from DGICyT, project numbers PB94-1167 and PB97-0141-C02-01. H. S. W, R. R. D. and S. E. M. acknowledge financial support from CONICET, project number PIP-4953/96, and from ANPCyT, project number .
[99]{}
R. K. Pathria, [*Statistical Mechanics*]{}, 2nd. ed.(Butterworth-Heinemann, Oxford, 1996); J. J. Binney, N. J. Dowrick, A. J. Fisher and M. E. J. Newman, [*The Theory of Critical Phenomena*]{} (Clarendon Press Oxford, 1993).
W. Horsthemke and M. Malek-Mansour, Z. Phys. B [**24**]{}, 307 (1976); L. Arnold, W. Horsthemke and R. Lefever, Z. Phys. B [**29**]{}, 367 (1978); W. Horsthemke and R. Lefever, [*Noise-Induced Transitions: Theory and Applications in Physics, Chemistry and Biology*]{} (Springer-Verlag, 1984).
F. Moss, in [*Some Problems in Statistical Physics*]{}, G. Weiss, ed. (SIAM, Philadelphia, 1993); L. Gammaitoni, P. Hänggi, P. Jung and F. Marchesoni, Rev. Mod. Phys. [**70**]{}, 223 (1998).
H. S. Wio, Phys. Rev. E [**54**]{}, R3075 (1996); H. S. Wio and F. Castelpoggi, in Proc. Conference UPoN’96, C. R. Doering, L. B. Kiss and M. Shlesinger, eds. (World Scientific, in press); F. Castelpoggi and H. S. Wio, Europhys. Lett. [**38**]{}, 91 (1997).
J. García-Ojalvo, A. Hernández-Machado and J. M. Sancho, Phys. Rev. Lett. [**71**]{}, 1542 (1993).
N. V. Agudov, Phys. Rev. E [**57**]{}, 2618 (1998).
M. O. Magnasco, Phys. Rev. Lett. [**71**]{}, 1472 (1993); R. Bartussek, P. Hänggi and J. G. Kissner, Europhys. Lett., 459 (1994); M. M. Millonas and M. I. Dykman, Phys. Lett. A [**185**]{}, 65 (1994).
C. Meunier and A. D. Verga, J. Stat. Phys. [**50**]{}, 345 (1988); H. Zeghlache, P. Mandel and C. Van den Broeck, Phys.Rev. A [**40**]{}, 286 (1989); R. C. Buceta and E. Tirapegui, in [*Instabilities and Nonequilibrium Structures III*]{}, E. Tirapegui and W. Zeller, eds. (Kluwer, 1991), p.171.
C. Van den Broeck, J. M. R. Parrondo and R. Toral, Phys. Rev. Lett. [**73**]{}, 3395 (1994).
C. Van den Broeck, J. M. R. Parrondo, R. Toral and R. Kawai, Phys. Rev. E [**55**]{}, 4084 (1997).
J. García-Ojalvo, J. M. R. Parrondo, J. M. Sancho and C. Van den Broeck, Phys. Rev. E [**54**]{}, 6918 (1996).
C. Van den Broeck, J. M. R. Parrondo, J. Armero and A. Hernández-Machado, Phys. Rev. E [**49**]{}, 2639 (1994).
W. Genovese, M. A. Muñoz and J. M. Sancho, Phys.Rev. E [**57**]{}, R2495 (1998).
S. Kim, S. H. Park and C. S. Ryn, Phys. Rev. Lett., 1616 (1997).
R. Müller, K. Lippert, A. Kühnel and U. Behn, Phys. Rev. E [**56**]{}, 2658 (1997).
J. García-Ojalvo, A. Lacasta, J. M. Sancho and R. Toral, to appear in Europhys. Lett. (1998).
G. Grinstein, M. A. Muñoz and Y. Tu, Phys. Rev.Lett. [**76**]{}, 4376 (1996).
Y. Tu, G. Grinstein and M. A. Muñoz, Phys. Rev.Lett. [**78**]{}, 274 (1997).
S. Ramaswamy, R. Pandit and R. Lahiri, Phys. Rev.Lett. [**75**]{}, 4786 (1995); C. Van den Broeck, J. M. R. Parrondo and R. Toral, Phys. Rev. Lett. [**75**]{}, 4787 (1995).
A. V. Soldatov, Mod. Phys. Lett. B [**7**]{}, 1253 (1993).
J. M. Sancho and M. San Miguel, in [*Noise in Nonlinear Dynamical Systems*]{}, F. Moss and P. V. E. McClintock, eds.(Cambridge U. Press, 1989), p.72.
M. Dykman and K. Lindenberg, in [*Some Problems in Statistical Physics*]{}, G. Weiss, ed. (SIAM, Philadelphia, 1993), p.41.
P. Hänggi and P. Jung, in [*Advances in Chemical Physics*]{}, vol. LXXXIX, I. Prigogine and S. A. Rice, eds. (Wiley, 1995), p.239.
F. Castro, A. D. Sánchez and H. S. Wio, Phys. Rev.Lett. [**75**]{}, 1691 (1995).
J. García-Ojalvo, J. M. Sancho and L. Ramírez-Piscina, Phys. Lett. A [**168**]{}, 35 (1992).
J. García-Ojalvo and J. M. Sancho, Phys. Rev. E [**49**]{}, 2769 (1994).
S. Mangioni, R. Deza, H. S. Wio and R. Toral, Phys.Rev. Lett. [**79**]{}, 2389 (1997).
P. Jung and P. Hänggi, Phys. Rev. A [**35**]{}, 4464 (1987); L. H’walisz, P. Jung, P. Hänggi, P. Talkner and L. Schimansky-Geier, Z. Phys. B [**77**]{}, 471 (1989); P. Hänggi, Chem. Phys. [**180**]{}, 157 (1994).
F. Castro, H. S. Wio and G. Abramson, Phys. Rev. E [**51**]{}, (1995).
H. S. Wio, P. Colet, M. San Miguel, L. Pesquera and M. Rodríguez, Phys. Rev. A [**40**]{}, 7312 (1989).
H. Risken, [*The Fokker–Planck Equation*]{} (Springer-Verlag, 1989).
A. Schlenzle and T. Tél, Phys. Rev. A [**32**]{}, 596 (1985).
\#1\#2[0.46]{}
\#1\#2[0.4]{}
\#1\#2[0.4]{}
\#1\#2[0.4]{}
\#1\#2[0.4]{}
\#1\#2[0.4]{}
\#1\#2[0.4]{}
\#1\#2[0.4]{}
\#1\#2[0.4]{}
[^1]: From Comisión Nacional de Energía Atómica and Universidad Nacional de Cuyo
[^2]: [email protected]
[^3]: [email protected]
[^4]: [email protected], WWW http://formentor.uib.es/$\sim$raul
[^5]: Member of CONICET, Argentina, ICTP Associate, E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Extensive microscopic molecular dynamics simulations have been performed to study the effects of short-chain alcohols, methanol and ethanol, on two different fully hydrated lipid bilayer systems in the fluid phase at $323~\mathrm{K}$. It is found that ethanol has a stronger effect on the structural properties of the membranes. In particular, the bilayers become more fluid and permeable: Ethanol molecules are able to penetrate through the membrane in typical time scales of about $200~\mathrm{ns}$ whereas for methanol that time scale is considerably longer, at least of the order of microseconds. We find good agreement with NMR and micropipette studies. We have also measured partitioning coefficients and the rate of crossing events for alcohols, i.e., typical time scale it takes for a molecule to cross the lipid bilayer and to move from one leaflet to the other. For structural properties, two-dimensional centre of mass radial-distribution functions indicate the possibility for quasi long-range order for ethanol–ethanol correlations in contrast to liquid-like behaviour for all other combinations.'
author:
- Michael Patra
- Emppu Salonen
- Emma Terama
- Ilpo Vattulainen
- Roland Faller
- 'Bryan W. Lee'
- Juha Holopainen
- Mikko Karttunen
title: 'Under the influence of alcohol: The effect of ethanol and methanol on lipid bilayers'
---
Introduction {#secIntroduction}
============
It is well known that even small changes in the composition of cell membranes can strongly affect the functioning of intrinsic membrane proteins, such as ion and water channels, which regulate the chemical and physical balance in cells [@mazzeo88; @cantor03]. Such changes may occur due to the introduction of short-chain alcohols, or other anaesthetics, at membrane surfaces. Although anaesthetics are being used every single day in hospitals around the world, the molecular level mechanisms of general anaesthesia remain elusive, see e.g. [@cantor97; @eckenhoff01; @tang:2002]. The same applies to the effect of alcohols on biological systems. @Klemm:1998 provides a good review of the topic.
Another aspect to the effect of alcohols appears in a more applied context. In the process of producing alcoholic beverages, wine in particular, yeasts like *saccharomyces cervisiae* have to sustain high ethanol concentrations without losing their viability. However, in about $10\,\%$ of all wine fermentations the industry encounters so-called stuck fermentations [@bisson02; @Silveira:2003]. There is no satisfactory understanding of this effect. Some models propose that an effect very similar to general anaesthesia is responsible for rendering the yeast cells dormant [@cramer02]. It has been suggested that high alcohol concentrations change the membrane structure and force transmembrane proteins into unfavourable conformations. In these conformations proteins cannot fulfil their functions and thus the yield drops dramatically.
In addition to the above aspects, there are other important issues as well. In particular, in cellular systems such as bacteria and yeast, the toxicity of ethanol has been suggested to be due to its interaction with membranes [@ly02; @Silveira:2003; @ly-pre] and the consequent general effects such as changes in mechanical properties, permeability and diffusion. Changes in such generic membrane properties may affect the functions of proteins and binding sites due to changes in lateral pressure [@eckenhoff01], or, if the membrane becomes more permeable, changes in the electrostatic potential may affect signalling. These effects are not to be mixed up with the toxicity due to metabolic products such as acetaldehyde from consumption of ethanol – the cause of poisoning commonly known as hangover.
We concentrate on the effects of ethanol and methanol on structural properties of membranes. It is quite surprising that despite a vast number of clinical and biochemical studies, there has been very few computational investigations of the effect of short-chain alcohols, or other anaesthetics, on membranes. The only simulational studies of bilayers and ethanol are, to the authors’ knowledge, the one by @Feller:2002 who used molecular dynamics simulations of ethanol and POPC (palmitoyloleoylphosphatidylcholine) lipid bilayers and NMR to study the molecular level interactions in these systems, and the article by @bwlee04 discussing alcohol–membrane systems briefly. Direct comparison of our results with Feller et al. is not meaningful since their study was performed using a different ensemble close to the gel state at $283~\mathrm{K}$ whereas here we are in the biologically relevant fluid phase at $323~\mathrm{K}$. For methanol–bilayer systems there exists to the authors’ knowledge only one computational article [@Bemporad:2004].
For anaesthetics the situation is slightly better. @tang:2002 used molecular dynamics simulations to study molecular level mechanisms of general anaesthesia using halothane as a specific anaesthetic. They concluded that the global effects of anaesthetics, i.e., due to generic interaction mechanisms, are important and lead to modulations in the functions of channels and/or proteins. These conclusions are also supported by the fact that the same anaesthetics are effective for humans and a variety of animals. Similar conclusions for halothane interactions with bilayers have been pointed out by @koubi00 [@koubi01]. The importance of generic effects has also been indicated in recent experimental studies of the effect of ethanol on *Oenococcus oeni* cells [@Silveira:2003]. Although the shortage of simulational studies may be due to the high computational demands of these systems, it is still surprising since computer simulations can provide detailed information about fundamental molecular level mechanisms.
In this article we study the effect of two short-chain alcohols, ethanol and methanol, on two different lipid membranes consisting of either pure DPPC (dipalmitoylphosphatidylcholine) or POPC. Methanol is a small solute having a single hydrophilic hydroxyl group whereas ethanol possesses an additional hydrophobic carboxyl group. DPPC and POPC share the same headgroup but one of the tails of POPC has a double bond and is two carbon atoms longer, whereas DPPC has only single bonds in its chains, see Fig. \[figChemie\]. We have studied these systems under fully hydrated conditions using microscopic molecular dynamics. $50~\mathrm{ns}$ trajectories for each of the four combinations of lipid and alcohol allow us to gather high statistical accuracy.
Phospholipid bilayers can be considered as a first approximation to understand the behaviour of cell membranes under the influence of alcohol, and much information can be extracted from such systems. The simulations show that ethanol is able to pass through the bilayer much more easily than methanol. This can be explained by the hydrophobic nature of the carbon “tail” of ethanol, making passing through the hydrophobic tail regions of lipid bilayers easier. In addition, ethanol molecules condense near the interface region between lipids and the surrounding water, i.e., there is a sharply increased density of ethanol near the interface region, while for methanol only a moderate increase of the density is seen near the interface region. These effects are very pronounced for DPPC bilayers, and only slightly weaker for POPC bilayers. This has far reaching implications for the basic properties of bilayers.
The rest of this article is organised as follows. In the next section we describe the model and the simulation details. Then, in Sec. \[secResults\], we present the results from the simulations. Section \[secDiscussion\] contains a discussion and conclusions.
Model and simulation details {#secModel}
============================
We have simulated lipid bilayer systems consisting of either 128 DPPC or 128 POPC molecules (i.e., 64 lipids in each leaflet). For the lipids we used a previously validated united atom model [@tieleman:96a]. The DPPC simulations are based on the final structure of a $100~\mathrm{ns}$ run of a DPPC bilayer that is fully hydrated by 3655 water molecules. The configuration is available online[^1]. The simulations by @patra:03aa were run using the same parameters as here (details below), and the $100~\mathrm{ns}$ run used in this study is a continuation of a $50~\mathrm{ns}$ study [@patra:03aa; @falck:04a]. For the POPC simulations, such an initial structure had to be generated first. We started with a fully hydrated POPC bilayer [@tieleman:99a] and simulated it for $10~\mathrm{ns}$. The final structure of that simulation run was used as a starting point for the POPC simulations reported here.
In order to add the alcohol molecules, the simulation box was first extended in $z$-direction such that an empty volume was created. A total of $90$ ethanol (methanol) molecules were randomly inserted in the empty volume, and the remaining space was filled with water. The total number of water molecules then amounted to 8958 for the DPPC systems and 8948 for the POPC systems, or, in other words, 1 mol% alcohol on a lipid free basis. The small difference between DPPC and POPC systems is due to the different lateral extensions of the bilayers.
The force field parameters for bonded and non-bonded interaction were taken from @berger:97a, available online[^2]. Partial charges were taken from @tieleman:96a, available online for both DPPC [^3] and POPC [^4]. As is seen in the chemical structures in Fig. \[figChemie\], DPPC and POPC are identical up to a single pair of CH-groups, connected by a double bond instead of a single bond in the *sn*–2 chain of POPC, and the two additional CH$_2$ groups at the end of that chain. This similarity is reflected in the force fields, which are identical up to the modelling of the four affected atoms. Ethanol and methanol were modelled using the Gromacs force field parameters [@lindahl:01a] which are identical with the exception of the added CH$_2$ group for ethanol. Thus, differences observed between the two lipids or the two alcohols do not originate from differences in their respective force field parameterisations but are due to the physics and/or chemistry of those components. For water the Simple Point Charge (SPC) model [@berendsen:81a] was used.
![Structures of POPC (top) and DPPC (bottom). They are identical with the exception of the *sn*–2 chain which is two carbons longer and contains one double bond for POPC.[]{data-label="figChemie"}](fig01.eps){width="3.2in"}
The simulations were performed with the Gromacs package [@lindahl:01a]. The lipids, water molecules and alcohols were separately coupled to a heat bath at temperature $T=323~\mathrm{K}$ using the Berendsen thermostat [@berendsen:84a] with a coupling time constant of $0.1~\mathrm{ps}$. All the bond lengths were constrained to their equilibrium values by the Lincs algorithm [@hess:97a]. Pressure was controlled using the Berendsen barostat [@berendsen:84a] with a time constant of $1~\mathrm{ps}$. The pressure coupling was used semi-isotropically such that height of the box ($z$ direction) and the cross sectional area ($x y$-plane) were allowed to vary independently of each other. Lennard-Jones interactions were cut off at a distance of $1.0~\mathrm{nm}$ and the time step was set to $2~\mathrm{fs}$. Long-range electrostatics were updated every 10-th time step \[the twin-range scheme [@kessel; @patra:03b] was used\], and handled by the particle-mesh Ewald (PME) algorithm [@essman:95a]. For DPPC bilayers it has been shown that replacing PME by the computationally cheaper cut-off scheme leads to pronounced artifacts [@patra:03b; @patra:03aa].
The systems were simulated for a total of $50~\mathrm{ns}$ each. After $20~\mathrm{ns}$, the samples had equilibrated, and the remaining $30~\mathrm{ns}$ were used for data collection. Equilibration was determined by monitoring the area per lipid as described in the next section. For completeness, we also present results for pure DPPC and POPC bilayers. In particular the latter ones are important since many of the results have so far only been published based on simulations using a cutoff for handling electrostatics. In addition to the above systems, we also performed a control simulation with a dehydrated system containing only 10 water molecules per lipid. This was done in order to see if dehydration has a direct effect on the properties but no significant effects were found.
The simulations took a total of 16 000 CPU hours using an IBM Cluster 1600 (Power4 processors).
Simulation results {#secResults}
==================
Before presenting a systematic summary of our results, we give a quick overview of the basic properties of these systems. Alcohol molecules have a tendency to collect in or near the bilayer (Sec. \[secMassDensity\]). This tendency is stronger for ethanol than for methanol as confirmed by a partition analysis (Sec. \[secPartitioning\]). Ethanol is able to form hydrogen bonds with the lipids in the bilayer (Sec. \[secBinding\]), and these hydrogen bonds reduce the order parameter of the lipid hydrocarbon tails. The combination of all this results in an easy penetration of ethanol through the bilayer. In contrast, no hydrogen bonds or penetration was observed for methanol.
In this paper we use the following colour code for all figures. Curves for systems containing [ethanol are drawn in red]{}, curves for [methanol in green]{} and pure lipid systems [without alcohol in blue]{}.
System dimensions {#secDimensions}
-----------------
The area per lipid is one of the most important quantities characterising lipid bilayer systems and it can also be used to monitor equilibration during a simulation run. The time evolutions of the area per lipid in the systems studied here are shown in Fig. \[figAreaPerLipid\]. The average areas per lipid, $\langle A \rangle$, obtained in our simulations are listed in Table \[tabArea\].
![Temporal behaviour of the area per lipid $A(t)$ for a DPPC bilayer (left) and a POPC bilayer (right). The colour of the line marks whether the lipid bilayer has been simulated in the presence of ethanol (red), of methanol (green), or of no alcohol (blue). This is our standard colour code employed throughout this paper.[]{data-label="figAreaPerLipid"}](fig02.eps){height="3.8cm"}
System Average area per lipid
----------------- ------------------------------------
DPPC (pure) $(0.655 \pm 0.002) ~\mathrm{nm}^2$
DPPC + ethanol $(0.699 \pm 0.002) ~\mathrm{nm}^2$
DPPC + methanol $(0.693 \pm 0.004) ~\mathrm{nm}^2$
POPC (pure) $(0.677 \pm 0.003) ~\mathrm{nm}^2$
POPC + ethanol $(0.699 \pm 0.003) ~\mathrm{nm}^2$
POPC + methanol $(0.693 \pm 0.003) ~\mathrm{nm}^2$
: Average area per lipid for all systems studied in this work. A weak effect of the alcohols is visible. The error estimates have been computed from block averaging and extrapolating to large block sizes.[]{data-label="tabArea"}
For a pure DPPC bilayer we obtain $\langle A_\mathrm{DPPC}\rangle = 0.655~\mathrm{nm}^2$ agreeing well with previous simulations and experiments, see Ref. [@patra:03b] and references therein. For pure POPC we obtain $\langle A_\mathrm{POPC}\rangle = 0.677~\mathrm{nm}^2$ in agreement with previous computational studies [@chiu99; @pasen03] and slightly larger than the results from x-ray diffraction studies [@pabst:00a; @pabst:00b]. For POPC, the difference to x-ray diffraction results $\langle A_\mathrm{POPC}\rangle \approx 0.61~\mathrm{nm}^2$ may be due to differences in trans-gauche conformational changes.
As seen from Table \[tabArea\], the presence of alcohol has a small but non-vanishing effect on the area per lipid. The number of water molecules per lipid molecule plays only a minor role as was verified by an additional simulation of DPPC with ethanol and a reduced amount of water. Interestingly, ethanol and methanol have almost the same effect on the area per lipid.
System $d_1~[\mathrm{nm}]$ $V_1~[\mathrm{nm^3}]$ $d_2~[\mathrm{nm}]$ $V_2~[\mathrm{nm^3}]$
--------------- --------------------- ----------------------- --------------------- -----------------------
DPPC (pure) $2.02\pm0.05$ $1.32\pm0.03$ $2.02\pm0.05$ $1.32\pm0.03$
with ethanol $1.84\pm0.02$ $1.28\pm0.02$ $1.90\pm0.02$ $1.33\pm0.02$
with methanol $1.93\pm0.06$ $1.34\pm0.04$ $1.95\pm0.06$ $1.35\pm0.04$
POPC (pure) $1.96\pm0.04$ $1.33\pm0.03$ $1.96\pm0.04$ $1.33\pm0.03$
with ethanol $1.87\pm0.02$ $1.31\pm0.02$ $1.93\pm0.03$ $1.35\pm0.02$
with methanol $1.94\pm0.02$ $1.35\pm0.01$ $1.96\pm0.02$ $1.36\pm0.01$
: The thickness $d$ of a leaflet (the bilayer thickness is twice that value) and the corresponding volume per lipid using the two definitions given in Eq. (\[eq:vols\]). The error estimate for $d$ has been computed by cutting the analysis part of the trajectory in two parts and by applying Eq. (\[eq:vols\]) separately to both of them. []{data-label="tabVolu"}
While the definition of the area per lipid is straightforward, the same is not true for the volume occupied by a lipid. The precise definition of the volume $V$ (or the thickness $d$) of a membrane is non-trivial as discussed at length by @armen98. Here, we chose an operational definition based on local mass density. Other definitions, e.g., employing the electron density, are equally possible.
Below, we give the two definitions we used to compute the thickness. If $\rho_{\text{lipid}}$, $\rho_{\text{water}}$ and $\rho_{\text{alcohol}}$ are the mass densities of the three components, the effective thickness of a single leaflet can be defined by
$$\begin{gathered}
d_1 = \frac{1}{2} \int \frac{\rho_{\text{lipid}}}{\rho_{\text{lipid}}
+ \rho_{\text{water}} + \rho_{\text{alcohol}} } \mathrm{d}z \;,\\
d_2 = \frac{1}{2} \int \frac{\rho_{\text{lipid}}}{\rho_{\text{lipid}}
+ \rho_{\text{water}} }
\mathrm{d}z\;.\end{gathered}$$
\[eq:vols\]
These two definitions differ in their treatment of the alcohol volume fraction and give the same thickness for pure lipid bilayers. After defining the thickness, the volume is simply $V=d \langle A \rangle$ with $\langle A \rangle$ being the average area per lipid. The results using both of the above definitions are summarised in Table \[tabVolu\].
The thicknesses we obtained for pure POPC agree very well with recent x-ray diffraction studies of @Vogel:2000a and computer simulation studies of @gullingsrud04 who obtained $3.9~\mathrm{nm}$ and $3.92~\mathrm{nm}$, respectively, for the total bilayer thickness $2\,d$. For DPPC the thickness and volume are a few percent larger than the experimental results [@Nag00]. Using the electron density to define the thickness would have led to similar results, see Sec. \[secElectronDensity\].
A comparison of Tables \[tabArea\] and \[tabVolu\] shows that the addition of ethanol or methanol to a bilayer expands its surface slightly while the thickness decreases such that the volume per lipid does not change significantly. This is as assumed since the main effect of the addition of alcohol is a reduction of the surface tension of the water phase. This is in agreement with observations from a DPPC–halothane system [@tu98]. We will return to this issue in Sec. \[secDiscussion\].
![Distribution of the individual areas of the lipids as determined by two-dimensional Voronoi tessellation for DPPC (left) and POPC (right).[]{data-label="figAreaVoronoi"}](fig03.eps){height="3.8cm"}
![Autocorrelation function for the individual areas of the lipids as determined by two-dimensional Voronoi tessellation for DPPC (left) and POPC (right).[]{data-label="figAreaCorrelation"}](fig04.eps){height="3.8cm"}
![Distribution of the Voronoi areas of the alcohols with DPPC (left) and POPC (right). Only alcohol molecules located close to the bilayer interface are included in the analysis.[]{data-label="figVoronoiAlcohol"}](fig05.eps){height="3.8cm"}
To complement the average area per lipid $\langle A\rangle$ measurements, we have also computed the area probability distribution $P(A)$ by Voronoi tessellation. By definition, Voronoi tessellation measures the area that is closer to a given molecule than to any other one. The Voronoi approach does not uniquely specify which point should be used to represent the entire molecule. We used the centre-of-mass position of each lipid, projected onto the $x y$-plane. Other choices are also possible, such as the position of the *sn*–3 carbon which gives a better indication of the backbone of the lipid whereas the centre-of-mass describes the entire lipid.
The resulting distributions $P(A)$ are shown in Fig. \[figAreaVoronoi\]. The mean of that distribution is, by construction, identical to the average area per lipid as shown in Table \[tabArea\], and thus does not contain any additional information. For this reason, not the plain distribution $P(A)$ but rather the re-scaled distribution $P(A/\langle A\rangle)$ is shown in Fig. \[figAreaVoronoi\]. Plotting the result in this way shows that alcohol does not influence the Voronoi distribution in any way that would not be captured already by the average area per lipid.
It is also possible to compute the autocorrelation time of the Voronoi areas. This time gives an indication of how quickly the geometry of the bilayer changes locally whereas the fluctuations in the size of the simulation box seen in Fig. \[figAreaPerLipid\] are related to global changes of the geometry. The results are shown in Fig. \[figAreaCorrelation\]. The faster decay for the systems with ethanol suggests that the bilayer might become more fluid but care should be taken in drawing conclusions as the differences between the curves are rather small.
Next, we perform similar Voronoi tessellation for the alcohol molecules inside the bilayer interface region. The precise definition of that region turned out not to be critical, and we included all alcohol molecules within the range $0.7~\mathrm{nm}<z<1.7~\mathrm{nm}$ from the centre of the bilayer. (Our choice for this range is motivated by the results to be discussed later in Sec. \[secMassDensity\].) The variable number of molecules forbids a proper calculation of the correlation time for the areas assigned to each alcohol, though, and thus we present only the distribution $P(A)$ in Fig. \[figVoronoiAlcohol\]. Since there are fewer methanol molecules close to the bilayer than there are ethanols (see Fig. \[figDensity1\]), the average area per methanol is larger than the average area per ethanol.
Mass density {#secMassDensity}
------------
![Mass density profiles across the bilayer for DPPC (left) and POPC (right). The density of the alcohol has been scaled by a factor of $10$, and the colour code is the same as above.[]{data-label="figDensity1"}](fig06.eps){height="3.8cm"}
The mass density profiles across the bilayer are shown in Fig. \[figDensity1\]. For each analysed simulation frame, the system was first translated such that the centre of the bilayer was located at $z=0$. Particles with $z<0$ were mirrored to $z>0$ to reduce statistical error. The masses of the hydrogen atoms were accounted for in the calculation. Due to the low density of alcohol, its curve is scaled by a factor of $10$ in the figure.
![Mass density profiles across the bilayer. The densities given are the O(xygen) of alcohol as well as N(itrogen) and P(hosphorus) of the lipid, and scaled to give a maximum of unity. Left for DPPC, right for POPC.[]{data-label="figDensity2"}](fig07.eps){height="3.8cm"}
![Density profile across the whole bilayer. The lipid component is divided into the contributions from the two separate leaflets, providing a measure of interdigitation. The alcohol component has been suppressed in the figure. Left for DPPC, right for POPC.[]{data-label="figDensity3"}](fig08.eps){height="3.8cm"}
Additional information can be gained by considering separately the two charged groups in the lipid headgroups, namely the phosphate (P) and the choline (N) group (cf. Fig. \[figChemie\]). In addition to this, the oxygen atom of the alcohols is included in Fig. \[figDensity2\]. We could not find a direct comparison for the mass density but the observations from the computer simulations of @Feller:2002 are consistent with our results for the mass density. Figure \[figDensity2\] shows that the alcohol molecules have a strong tendency to accumulate below the bilayer–water interface layer (approximately given by the location of the phosphate and choline groups), and that this tendency is stronger for ethanol than for methanol. We will return to this issue in the partitioning analysis in Sec. \[secPartitioning\] and the membrane penetration analysis in Sec. \[secCrossings\].
The density of the lipid is decreased near the centre of the bilayer. This phenomenon is known as lipid trough and means that the two leaflets are repelling each other. Still, the tails of the lipids from one leaflet are able to penetrate into the other leaflet; this is known as interdigitation [@Loebbcke:1995]. To analyse this, in Fig. \[figDensity3\] we plot the density throughout the whole bilayer, i.e., the positions of all atoms are not folded into a single leaflet. We show separately the density of the lipids belonging to the upper and the lower leaflet of the bilayer. It is easily seen that the tails of the lipids can penetrate up to $0.5~\mathrm{nm}$ into the other leaflet, and the degree of interdigitation is largely independent of the presence of alcohol.
Electron density {#secElectronDensity}
----------------
Electron densities provide information about the structure of bilayers along the normal to the bilayer plane similar to the mass densities. Experimentally, x-ray diffraction provides a means to access this quantity, the measurements yielding information of the total electron density profile.
Figure \[figEleDensity\] shows the total electron densities in different cases. For the pure lipid bilayers, the curves show the typical behaviour, namely a maximum associated with the electron dense areas in the headgroup, i.e., the phosphate groups, and the minimum at the bilayer centre – the so-called methyl trough [@Nag00]. Experimentally, an electron density profile contains different information than a mass density profile since the chemical composition at depth $z$ is not known directly. (For computer simulations this problem does not exist.)
For the pure systems our results agree well with experiments [@nagle96]. The only x-ray diffraction study in a related system containing ethanol or methanol that we are aware of is that of @adachi. Unfortunately a direct comparison is not meaningful since that study was done in the gel phase using multilamellar vesicles as compared to the fluid-like phase and planar bilayer system studied here.
![Electron density profiles in the studied systems with DPPC (left) and POPC (right).[]{data-label="figEleDensity"}](fig09.eps){height="3.8cm"}
Hydrogen bonding of alcohol to lipids {#secBinding}
-------------------------------------
As was shown in Sec. \[secMassDensity\], the alcohol molecules have a tendency to be located inside the bilayer, and this tendency is stronger for ethanol than for methanol. The alcohol molecules are not located directly at the water–membrane interface but rather further inside the bilayer. For the simulations with ethanol, a direct visual inspection of the atom positions shows that ethanol molecules are located close to the ester oxygens of the lipids, see Fig. \[figZeichnung\].
![A DPPC molecule together with two ethanol molecules. The ethanols are located close to the ester oxygen. The DPPC molecule is drawn as rods whereas the ethanols are drawn in a spacefilling representation. To aid the eye, the ethanols are coloured blue-ish.[]{data-label="figZeichnung"}](fig10.eps){width="7cm"}
This visual conclusion is confirmed by a hydrogen bonding analysis. In such an analysis, possible donors and acceptors are identified by their chemical properties, and a hydrogen bond is then assumed to exist whenever two such atoms and an additional hydrogen atom fulfil certain geometric conditions. (The distance between a hydrogen and an acceptor has to be smaller than $0.25~\mathrm{nm}$, and the angle between acceptor, hydrogen and donor has to be smaller than $60$ degrees.)
DPPC POPC
----------------- -------- --------
bound ethanols $72.9$ $71.6$
bound lipids $59.7$ $59.2$
hydrogen bonds $74.1$ $72.8$
lifetime \[ns\] $1.20$ $1.15$
: Results of the hydrogen bonding analysis for the DPPC and POPC bilayers with ethanol. (The systems consist of 128 lipid molecules and 90 alcohol molecules.) No hydrogen bonds between methanol and lipids were found.[]{data-label="tabHydrogen"}
The hydrogen bonding analysis shows that the majority of the ethanols are involved in hydrogen bonds with lipids whereas not a single hydrogen bond between a methanol and a lipid molecule was found in our simulations. The results are summarised in Table \[tabHydrogen\]. Many lipids are involved in more than one hydrogen bond which comes as no surprise since they possess an ester oxygen in each of their two chains. Comparison with the lifetime data for ethanol in Table \[tabHydrogen\] with NMR experiments [@holte97:a] shows excellent agreement. In their experiments Holte and Gawrisch reported the lifetimes to be around $1~\mathrm{ns}$ while we obtained $1.20~\mathrm{ns}$ for the ethanol lipid hydrogen bonds. We are not aware of any such experiments for methanol.
The number of alcohol molecules involved in hydrogen bonds is best compared against the total number of alcohol molecules located inside the bilayer. The latter number is relatively ill-defined but from Fig. \[figDensity1\] one can easily compute that for ethanol-containing systems only of the order of $10$ ethanol molecules out of the approximately $70$ inside the bilayer are not involved in hydrogen bonds. For comparison, in the methanol systems there are of the order of $20$ methanol molecules inside the bilayer and none of them is involved in hydrogen bonds (but cf. Sec. \[secCrossings\]). These numbers show that there indeed is a significant difference between ethanol and methanol.
Hydrogen bonding analysis offers a well-defined criterion to decide whether a given lipid is interacting strongly with an alcohol molecule or not. This will be used in the following sections to study separately the two lipid populations, lipids bound to an alcohol and lipids not bound to an alcohol.
Radial-distribution functions
-----------------------------
![Radial-distribution function between the oxygen of alcohol, on the one hand, and the phosphorus and the nitrogen atoms in the headgroup as well as the ester oxygens in the lipid tail, on the other hand. The oxygen–oxygen curve has been scaled by a factor of $1/4$, i.e., in reality the RDF peaks at a value four times as large as displayed in the figure.[]{data-label="figRdf1"}](fig11.eps){height="3.8cm"}
{height="3.8cm"}
In addition to the mass density profiles, valuable information may be gained from radial-distribution functions (RDFs). The radial-distribution functions $g(r)$ give the probability of finding two particles at a mutual distance $r$ once geometric and density factors have been scaled out.
Figure \[figRdf1\] shows the RDFs between the oxygen of the alcohol and different charged groups inside the lipid. Some of these groups were depicted already in the mass density profile in Fig. \[figDensity2\] but whereas there only the vertical distance between particles was considered, the RDF considers the real three-dimensional distance between them.
While the mass density profile showed that on the average ethanol molecules prefer to reside $0.5~\mathrm{nm}$ below the lipid headgroups, the RDF shows that the three-dimensional preferred distance is only $0.38~\mathrm{nm}$. This is no contradiction but is easily understood by the observation (cf. Fig. \[figMassBinding\] a bit further down) that lipid molecules without an attached ethanol molecule are sticking out of the bilayer more than those with an attached ethanol. This is captured only by the RDFs but not by the mass density profile.
The radial-distribution functions for the different systems look quite similar – with one exception: The RDF between the alcohol and the ester group is peaked at a much smaller distance for ethanol than it is for methanol. This is in agreement with the results of the hydrogen bonding analysis.
By studying the mutual RDFs of the choline and/or phosphate groups, it is possible to detect phase transitions of the bilayer. Within error margins, these RDFs are not dependent on the presence of alcohol, and for space reasons we do not show them here as they are identical to the RDFs published previously [@patra:03aa].
We have also studied two-dimensional radial-distribution functions of entire molecules, i.e., the molecules’ centres-of-masses were projected onto the $x y$-plane and radial-distribution functions were then computed. The results are shown in Fig. \[figRdf3\]. The mutual radial-distribution functions of the lipids exhibit a very soft core as lipids are able to wrap around each other. No dependence on lipid type or presence of alcohol was observed.
The RDF between alcohol and lipid is qualitatively different for ethanol and methanol. For an ethanol, there is a large probability for it to be at the same $x$-$y$ position as the (centre-of-mass of the) lipid. This reflects the hydrogen bonding of ethanol close to the centre of the lipid. This bonding is absent for methanol, and consequently then $g(r)\to 0$ for $r\to 0$. No significant dependence on the kind of lipid is observed.
![Deuterium order parameters, computed from Eq. (\[sOrderDefCD\]), for DPPC (solid line) and POPC (dashed line). For DPPC, the average over the two tail chains is displayed while for POPC only the saturated *sn*–1 chain is shown. For numbering of carbon atoms, refer to Fig. \[figChemie\].[]{data-label="figOrderParam"}](fig13.eps){height="3.8cm"}
The alcohol–alcohol radial-distribution functions have a very different character: For methanol with DPPC, the first peak is very distinct but the correlations decay soon after. For methanol with POPC an additional peak is observed. (This is the only curve with a significant difference between DPPC and POPC. We cannot offer a convincing explanation for this.) For ethanol, the behaviour shows almost quasi long-range order. The reason for this ordering is not clear and further experiments would be needed to study this in detail.
Order parameters
----------------
Ordering of the lipid acyl chains is typically characterised using the order parameter tensor $$S_{\alpha\beta} =
\frac{1}{2}
\left\langle
3 \cos\theta_{\alpha}\,\cos\theta_{\beta} -
\delta_{\alpha\beta}
\right\rangle \;,
\label{sOrderDef}$$ where $\alpha,\beta = x,y,z$, and $\theta_{\alpha}$ is the angle between the $\alpha^{\text{th}}$ molecular axis and the bilayer normal ($z$-axis). The order parameter is then computed separately for all carbons along the acyl chain. Since lipid bilayer systems possess symmetry with respect to rotations around the $z$-axis, the relevant order parameter is the diagonal element $S_{zz}$. To relate $S_{zz}$ to the experimentally relevant deuterium order parameter $$S_{\mathrm{CD}} = \frac{2}{3} S_{xx} +
\frac{1}{3} S_{yy} \;,
\label{sOrderDefCD}$$ we use the symmetry and write $S_{xx} = S_{yy}$, and $S_{xx} + S_{yy} + S_{zz} = 0$. Using these relations we have $S_{\mathrm{CD}} = - S_{zz} / 2$. To allow comparison with experimental data, we present our results in terms of $\lvert S_{\mathrm{CD}}\rvert$.
![Order parameter of DPPC (left) and POPC (right) in the presence of ethanol. For DPPC, both the *sn*–1 and *sn*–2 chains are shown whereas for POPC only the *sn*–1 chain is depicted. The order parameter has been computed separately for lipids that are bound to at least one alcohol molecule (solid lines), and for lipids that are not bound to any alcohol (dashed lines).[]{data-label="figOrderParam2"}](fig14.eps){height="3.8cm"}
Since our simulations employ a united atom model, no explicit information about the hydrogen positions is available and they must be reconstructed assuming a perfect tetrahedral arrangement. The results are shown in Fig. \[figOrderParam\]. Reconstruction of the hydrogen positions means that for the outermost carbon atoms of the tail no order parameter can be constructed. This also includes positions where a sequence of carbon atoms connected by single bonds ends in a carbon atom having a double bond. It is not a problem to compute the order parameter for a chain of atoms connected by double bonds, there just is a problem connecting such a chain to a chain of atoms connected by single bonds. For this reason we show the order parameter for both chains of DPPC but for POPC we restrict ourselves to the saturated *sn*–1 chain, see Fig. \[figChemie\].
![Distribution of the angle $\theta$ between the P–N vector and the bilayer normal for DPPC (left) and POPC (right). []{data-label="figOrderHead"}](fig15.eps){height="3.8cm"}
The results for the order parameter for all the cases are shown in Fig. \[figOrderParam\]. We find that methanol increases ordering of the lipid acyl chains close to the glycerol group, while the effect of ethanol is strongest below the glycerol group, around the centre of the hydrocarbon tails. These results are fully consistent with the mass density profiles in Fig. \[figDensity1\].
![Distribution of the angle between the P–N vector and the bilayer normal for DPPC (left) and POPC (right) in the presence of ethanol. The distribution is separated into lipids that are hydrogen bonded respectively not hydrogen bonded to an ethanol molecule.[]{data-label="figOrderHead2"}](fig16.eps){height="3.8cm"}
As discussed in the introduction, there is, to our knowledge, only one other, related membrane–alcohol simulation study [@Feller:2002]. In that the temperature was 40 degrees lower. Although direct comparison is not possible, the order parameters in that study and ours are in qualitative agreement. It is noticeable that the effect of alcohols on the average order parameter is small – this is consistent with the fact that the volume occupied by a lipid changes only slightly.
Additional insight can be gained by combining the order parameters with the hydrogen bonding analysis from Sec. \[secBinding\]. This combination allows the computation of the order parameter depending on whether the lipid forms a hydrogen bond with an alcohol molecule or not. The result in Fig. \[figOrderParam2\] shows that binding of ethanol slightly decreases the order of the tails.
In addition to the deuterium order parameter for the acyl chains, it is possible to study the ordering of headgroups in a similar fashion. To do that, we have chosen the angle of the P–N vector with respect to the bilayer interface plane and have computed its distribution. The result is shown in Fig. \[figOrderHead\]. The P–N vector has a significantly higher tendency of being in the bilayer plane ($\theta=90\,^{\circ}$) than of sticking out of it. The computed distribution is only slightly dependent on the presence of alcohol.
Again, we can obtain additional information for the ethanol systems if the angular distribution is separated into the distributions of lipids that are hydrogen bonded to an alcohol, and those that are not. The result is shown in Fig. \[figOrderHead2\]. For DPPC the angular distribution is not influenced at all by hydrogen bonding whereas for POPC the influence is small.
To better explain the results presented in this section, we return to the mass density profiles. We analysed the positions of the phosphate and the choline groups in the heads of the lipids as well as the lipids’ centre-of-mass positions, separated into lipids that are hydrogen bonded or not hydrogen bonded to an ethanol. The data in Fig. \[figMassBinding\] shows that lipid molecules are shifted towards the centre of the bilayer by approximately $0.2~\mathrm{nm}$ if they are bonded to an ethanol molecule. This is in agreement with the reduced order parameter as that is normally associated with a thinner bilayer. The vertical distance between the choline and the phosphate group remains unchanged, in agreement with the distribution of the P–N angle.
![Distribution of the position of the phosphate group (P), the choline group (N) and the centre-of-mass of the entire lipid (COM), left for DPPC, right for POPC. Results are shown as solid (dashed) lines for lipid molecules bound (unbound) to an ethanol molecule. Please note that in this figure the colours do not mark the kind of alcohol present.[]{data-label="figMassBinding"}](fig17.eps){height="3.8cm"}
Orientation of the water dipole
-------------------------------
![Water orientation, as described by the mean cosine of the angle of the water dipole moment with respect to the bilayer normal.[]{data-label="waterDipole"}](fig18.eps){height="3.8cm"}
Ordering of the water dipole in the vicinity of the bilayer–water interface is described by calculating the time averaged projection of the water dipole unit vector $\vec{\mu}(z)$ onto the interfacial normal $\vec{n}$, $$P(z) = \langle \vec{\mu}(z) \cdot \vec{n} \rangle
= \langle \cos \theta(z) \rangle \;,$$ where $z$ is the $z$-component of the centre-of-mass of the water molecule and vector $\vec{n}$ points away from the bilayer centre along the $z$-coordinate.
The data for all the studied cases are shown in Fig. \[waterDipole\]. For pure bilayers the results are in agreement with previous studies, e.g. [@patra:03b]. When either methanol or ethanol is added, the water dipole becomes less oriented, i.e., the addition of alcohol slightly reduces the amount of ordering. For pure bilayers, and for bilayers with added methanol, the minimum remains at the same distance, at about $1.8~\mathrm{nm}$ from the bilayer centre. For added ethanol the minimum shifts to a smaller distance, to about $1.6~\mathrm{nm}$. This is a reflection of the fact that ethanol leads to a slightly larger area per lipid and thus a thinner bilayer (see Sec. \[secDimensions\]).
Electrostatic potential
-----------------------
To obtain the electrostatic potential across the bilayer the average charge density profile was first computed such that the centre of the bilayer was separately aligned to $z=0$ for each simulation frame. Then, the electrostatic potential was determined by integrating the charge density twice with the initial condition $V(z = 0) = 0$.
The electrostatic potentials for all studied cases are shown in Fig. \[figPotential\]. For pure DPPC the electrostatic potential was determined to be $-589~\mathrm{mV}$ in agreement with previous studies [@tieleman:96a; @patra:03b]. For pure POPC we obtain $-507~\mathrm{mV}$.
![Electrostatic potential through the bilayer for DPPC (top) and POPC (bottom).[]{data-label="figPotential"}](fig19.eps){height="4.75cm"}
The addition of alcohol leads only to small changes in the electrostatic potential. This is what is expected from the results presented so far. The only way in which the membrane potential could be changed significantly would be by re-arrangement of the P–N angle of the headgroup, resulting in a change of the dipole moment of the lipid headgroup. No such re-arrangement was observed, cf. Fig. \[figOrderHead\].
On a superficial level this may seem to be in contradiction to some previous suggestions that the narcotic effects of alcohols are mainly due to a change of the electrostatic potential. We would like to point out, however, that even though the direct effect of alcohol to the potential is small, this does not exclude secondary effects which may lead to a significant change in the electrostatic potential. We return to this issue in Sec. \[secDiscussion\].
Partitioning coefficients {#secPartitioning}
-------------------------
To measure whether alcohols prefer to stay close to the membrane or rather to be in the water phase, we computed the partitioning coefficients. They provide a thermodynamic quantity characterising a molecule’s tendency to choose its environment.
The partitioning coefficient $K_{\text{p}}$ is defined as $$K_{\text{p}} = X_{\text{bilayer}} / X_{\text{water}}\;,
\label{partcoeff}$$ where $X_{\text{bilayer}}$ and $X_{\text{water}}$ are the moles of solute per kg of solvent [@westh:01a]. Equation (\[partcoeff\]) can be reformulated in terms of the molar concentrations $n$ and the molar masses $m$ as $$K_{\text{p}} = \frac{( 1 - n^{\text{water}}_{\text{alcohol}})
n^{\text{bilayer}}_{\text{alcohol}}
m_{\text{water}}}{
(1 - n^{\text{bilayer}}_{\text{alcohol}})
n^{\text{water}}_{\text{alcohol}}
m_{\text{lipid}}
} \;.$$ Here, $n^{\text{a}}_{\text{b}}$ means the molar concentration of component $b$ in the phase $a$.
------------------------ ------------------------------------- ------------------------------------- ----------------
system $n^{\text{water}}_{\text{ethanol}}$ $n^{\text{lipid}}_{\text{ethanol}}$ $K_{\text{p}}$
\[0.8mm\] DPPC–ethanol $(3.98\pm0.06)\cdot 10^{-4}$ $0.403\pm0.001$ $41.6\pm0.6$
POPC–ethanol $(3.34\pm0.04)\cdot 10^{-4}$ $0.405\pm0.001$ $48.3\pm0.9$
DPPC–methanol $(4.53\pm0.01)\cdot 10^{-3}$ $0.278\pm0.001$ $2.07\pm0.02$
POPC–methanol $(4.72\pm0.02)\cdot 10^{-3}$ $0.272\pm0.001$ $1.87\pm0.02$
------------------------ ------------------------------------- ------------------------------------- ----------------
: Results of the partitioning analysis.[]{data-label="tabPartitioning"}
One might be tempted to count the number of alcohol molecules in the lipid phase and in the water phase using some functional definition of where the lipid bilayer interface is located. This approach, however, gives incorrect results as the relevant quantity is not the number of alcohol molecules *inside* the membrane but rather the number of alcohol molecules *influenced* by the membrane.
Starting point are the mass densities $\rho_{\text{water}}$ and $\rho_{\text{alcohol}}$ far from the bilayer in the water phase. These are available from Fig. \[figDensity1\], and knowledge of the molar masses gives $n_{\text{water}}$. Multiplying $n_{\text{water}}$ by the total number of water molecules in the simulation box gives the number of alcohol molecules that would be there if there was no lipid bilayer. The remaining alcohol molecules thus must be “brought in” by the lipid bilayer.
Our results are summarised in Table \[tabPartitioning\]. Although experimental measurements of partitioning coefficients have been performed [@Rowe:1998; @Westh+Trandum:1999; @trandum00], we were not able to find experimental data to make exact quantitative comparisons. However, the experimental data for DMPC [@trandum00] with ethanol is in qualitative agreement with our results. They are also in agreement with the general trends that $K_{\text{p}}$ increases as the lipid tails or the alcohols become longer. As we have studied the system at single temperature we cannot comment on the temperature dependence of the partitioning coefficients.
Crossing events {#secCrossings}
---------------
![$z$-positions of all alcohol molecules as a function of time for DPPC (left) and POPC (right). Ethanol (red) is able to penetrate into the bilayer (located at $z=0$) much better than methanol (green). Crossing events of ethanol are seen while they are completely absent for methanol.[]{data-label="figAlleZ"}](fig20.eps){height="3.8cm"}
{height="3.8cm"}
Next, we analyse the penetration of alcohol through the membrane. A quick overview can be obtained by plotting the $z$-component of the positions of the alcohol molecules. (The positions of all atoms have to be translated for every simulation frame such that the centre of the bilayer does not move). The result is shown in Fig. \[figAlleZ\]. The density (in space and time) of alcohol molecules is large in most parts of the diagram such that it is difficult to visually identify individual trajectories.
It is easily seen from Fig. \[figAlleZ\] that the density of alcohol molecules is reduced in the centre of the bilayer. In addition, the gap in the centre of the bilayer is smaller for ethanol than it is for methanol. This is in agreement with the mass density profiles presented in Sec. \[secMassDensity\]. It is also evident that there is a significant number of events where an ethanol molecule is crossing from one leaflet to the other while no such events are seen for methanol molecules. (Crossing events cannot be inferred from the mass density since these events happen so fast that the resulting mass density of alcohol in the centre of the bilayer is negligible.)
In the simulations, it is directly known which atoms form the lipid molecules of the upper leaflet of the bilayer, and which atoms form the lower leaflet, and which atoms belong to water molecules. The atom nearest to the alcohol molecule then determines in which of the three phases a given alcohol molecule is located at any given moment. In addition, it is also relevant whether an alcohol molecule is hydrogen bonded to some lipid molecule, cf. Sec. \[secBinding\].
When all these pieces of information are combined, one arrives at data as shown in Fig. \[figTrajExamples\] in which we depict a few selected ethanol molecules within a DPPC bilayer. It is seen that, while the alcohol is inside the bilayer, it is hydrogen bonded most of the time. The bonding does not persist for the entire duration of a simulation but there are short breaks in between. This is in agreement with the hydrogen bond lifetime of order $1~\mathrm{ns}$ in Table \[tabHydrogen\] while ethanol molecules can stay inside the bilayer much longer than this. Whenever the hydrogen bond is broken, it can either be re-formed shortly afterwards, or the alcohol molecule can try to move to some other place. From the figure it is seen that an alcohol molecule can move to the opposite leaflet of the bilayer but that not all such attempts are successful, i.e., the alcohol molecule may be reflected back.
-------------- -------- ------------- -------- -------------
System number time \[ps\] number time \[ps\]
DPPC–ethanol 30 325 123 245
POPC–ethanol 21 375 101 225
-------------- -------- ------------- -------- -------------
: Number of successful and unsuccessful crossing events, respectively, within $40~\mathrm{ns}$ of trajectory. In addition, the mean time spent in the crossing process is given.[]{data-label="tabCrossings"}
Using the collected information, each alcohol molecule is at any given moment in one of five different states (water phase, upper leaflet, upper leaflet hydrogen-bonded, lower leaflet and lower leaflet hydrogen-bonded; we will discuss the methanol-containing systems a bit further down). It is of little interest if an alcohol molecule is “scratching” at the surface of the bilayer – rather it is important whether the alcohol molecule reaches the part of the leaflet where it can form a hydrogen bond.
![View from the top onto part of a POPC bilayer with methanol. The lipids are coloured green, methanol is coloured blue and a few selected water molecules are shown in red. It is easily seen than the methanol is located in a cavity together with a few water molecules.[]{data-label="figTopView"}](fig22.eps){width="7cm"}
This immediately gives functional definitions for different kinetic events that can be used for an automatic analysis. A successful crossing event from the upper to the lower leaflet is for example given by a sequence “upper leaflet hydrogen-bonded $\to$ {upper leaflet} $\to$ {lower leaflet} $\to$ lower leaflet hydrogen-bonded” where the curly braces mean that this step may also be skipped. Similarly, an unsuccessful crossing from top to bottom would be “upper leaflet hydrogen-bonded $\to$ {upper leaflet} $\to$ lower leaflet $\to$ {upper leaflet} $\to$ upper leaflet hydrogen-bonded”. Other criteria are constructed similarly.
Table \[tabCrossings\] shows the results of the analysis for crossings of ethanol molecules between the two leaflets. A simple calculation shows that ethanol molecules are able to move from one leaflet to the other on a time scale of $130~\mathrm{ns}$ for DPPC and $180~\mathrm{ns}$ for POPC. The number of unsuccessful crossing attempts outnumbers the number of successful attempts by a factor of $4$, thereby demonstrating that the hydrophobic tails of the lipid pose a significant barrier to ethanol not only from the outside of the leaflet but also from the inside. For methanol we did not find any crossing events in our simulations, implying that the corresponding time scale must be at least of the order of microseconds.
A closer study of the systems containing methanol is hindered by a problem that is not obvious from any of the data presented so far. As Fig. \[figTopView\] shows, methanol is virtually never really inside the bilayer, i.e., located such that it no longer has direct contact with the bulk water phase – in all of our data for DPPC and POPC with methanol, we found only a single methanol molecule that had actually lost contact with water.
![Distribution $P(t)$ of the time $t$ for which an alcohol molecule stays inside the membrane, left for DPPC, right for POPC. Due to limited statistics, we cannot plot $P(t)$ directly but are limited to the cumulative probability $\int P(t') d t'$.[]{data-label="figSurvivalTime"}](fig23.eps){height="3.8cm"}
The above observation is also able to explain why no hydrogen bonds are formed between methanol and the lipids: Water is simply too energetically favourable a binding partner for methanol, or, in other words, the surface tension of water is too high for methanol to leave the water phase. The concept of an alcohol being inside the membrane thus does not apply – topologically, methanols are always located outside the membrane. Rather we need to introduce the concept of a methanol being located in a sufficiently deep well. This can be quantified by counting the number of atoms belonging to lipids within a certain distance around some particular methanol molecule. This number will be much larger if the methanol is inside such a well. (We use the criterion that the number of atoms belonging to lipid molecules within $0.6~\mathrm{nm}$ is larger than $50$.)
Using that functional definition, we are able to treat ethanol and methanol containing systems on a similar footage. While there are no crossing events for methanol, another interesting question still arises, namely the dynamics of alcohol exchange between the membrane and the water phase. Quantitatively, the interesting quantity is the time $t$ between an alcohol molecule entering the membrane and its subsequent leaving it again. Our results are shown in Fig. \[figSurvivalTime\]. Since there are only of the order of $200$ ($1000$) events for ethanol (methanol), the statistics is insufficient to compute the probability distribution $P(t)$. Rather, we present the cumulative probability $\int P(t') d t'$, i.e., the probability that an alcohol stays inside the membrane no longer than some time $t$, since this quantity can be computed without binning the data point. ($P(t)$ follows, in principle, by differentiation of the depicted curves.) It is seen from the figure that the dynamics is much faster for methanol than for ethanol. This comes as no surprise since methanol is not really inside the bilayer – it does not need to cross the bilayer interface but only needs to deform it (to create a well).
Discussion {#secDiscussion}
==========
This study is, to the best of our knowledge, the first detailed computational study characterising the behaviour of lipid bilayers (POPC and DPPC) under the influence of methanol and ethanol. The other existing molecular dynamics study of ethanol and POPC [@Feller:2002] concentrated on the comparisons with an NMR study under different conditions (at 10 degrees Celsius close to the gel phase and using an $NVT$ ensemble). To obtain detailed information about alcohol–membrane interactions, it is thus important to study the biologically important fluid phase.
Let us first discuss the area per lipid and bilayer thickness. The increase in the area per lipid is larger for DPPC bilayers (about $7\,\%$ for ethanol and $6\,\%$ for methanol) as compared to POPC systems ($5\,\%$ for ethanol and $3\,\%$ for methanol). This compares well with the recent micropipette studies of @ly02 who used SOPC vesicles under slightly different conditions ($20$ vol.-$\%$ ethanol at room temperature). They observed $9\,\%$ increase in the area per lipid and $8\,\%$ decrease in the thickness of the bilayer. Here, we obtained a decrease of $7\,\%$–$10\,\%$ (ethanol and DPPC) and $1\,\%$–$4\,\%$ (ethanol POPC) in thickness depending on the definition used, see Eq. (\[eq:vols\]).
As a purely structural effect, it is clear that the membrane thus becomes more permeable to small molecules due to its increased area per lipid. The differences between DPPC and POPC are most likely due to the slightly longer *sn*–2 chain of POPC and the double bond in it. In addition to the effects captured by the average area per lipid, steric constraints seem to make the POPC bilayer less susceptible to penetration of small solutes.
Furthermore, in a recent study @chanturiya:1999 proposed that the penetration of alcohols inside the bilayer, and their binding at it, and the resulting decrease in bending rigidity is a feasible pathway for promoting fusion of cells. Although it is not possible to probe this directly by current computer resources, our observations support the possibility of such a mechanism.
It has been suggested that the preferred location of ethanol close to the membrane dehydrates it [@Klemm:1990; @Klemm:1998; @holte97:a]. This should show in the water dipole orientation data, Fig. \[waterDipole\]. The relatively small changes in it, and in the electrostatic potential across the membrane, suggest that indirect effects, such as receptor blocking, may be more important in producing changes in these quantities.
To characterise thermodynamic properties, we have measured the partitioning coefficients. Our estimates of $K_\mathrm{p}$ for the different alcohol-lipid combinations are given in Table \[tabPartitioning\]. Due to the lack of experimental data, no quantitative comparison could be made but qualitatively experiments and simulations are in agreement.
Short-chain alcohols have an amphiphilic character and it has been known for long that addition of each new CH$_2$ group – adding a CH$_2$ group on methanol gives ethanol and so on – has a strong effect on the interactions with membranes. This is indicated by the well known Traube’s rule [@traube1891; @adamson97] which states that the addition of a new CH$_2$ group leads to a decrease in surface tension. In other words, short-chain alcohols have a strong effect on membrane properties and the effect depends on both the length of the hydrophobic part of the alcohol and on concentration. This has also been observed in recent experiments [@ly02; @ly-pre].
Our data for methanol and ethanol supports these conclusions. Methanol does not penetrate through the lipid tail region which is easily understood by the hydrophobic nature of the lipid tails which are repelling methanol as it is more polar than ethanol. As a second effect, methanol rarely reaches the tail region as each methanol molecule moves together with a small cluster of water molecules when it is trying to enter the membrane. This means that, on one hand, methanol is pulled back into the water phase by this, and, on the other hand, not a single small methanol molecule but a significantly larger dressed particle, or a small cluster, would need to penetrate the membrane.
The analysis of crossing events, i.e., how often the molecules travel through the membrane, showed that ethanol is able to penetrate the membrane easily whereas for methanol not a single crossing event was observed. This confirms the interpretation given above. It is difficult to compare these results directly with experiments but the possibility of such crossing events has been proposed on the basis of NMR studies [@holte97:a]. The results presented here are, to our knowledge, the first detailed analysis of crossing events. Further experiments would be needed in order to better characterise the situation as the system here is a simple model system and the relevance of these results to biological systems, in particular yeasts, needs to be better studied. The only such a study we were able to find uses NMR and *Z. mobilis* [@Schobert:1996] but direct comparison is not possible due to the different experimental setup.
In the introduction we briefly discussed general anaesthesia and membrane–protein interactions induced by the addition of anaesthetics, such as small alcohols. This was observed in a recent experiment [@Brink:2004] where the potassium channel KcsA was observed to dissociate due to the changes in lateral membrane pressure induced by small alcohols. Here, we have characterised simple membrane–alcohol systems. The detailed characterisation presented here is essential for extensive simulational studies of membrane–protein–anaesthetic systems. From our results it is obvious that the changes in pure membranes are subtle but the effects of those changes to, e.g., embedded proteins may be significant [@cantor97; @eckenhoff01; @koubi01; @tang:2002; @Brink:2004]. This is also supported by recent experiments using enflurane and DPPC [@hauet03]. Similar conclusions have been drawn by @tu98 for the interaction of halothane with bilayers. As pointed out by Hauet et al., there are various intriguing questions regarding small molecules and anaesthesia. These questions are related to interactions between membranes and small molecules and computer simulations give a direct access to study them.
We are grateful to Ole G. Mouritsen, Amy Rowat, Margie Longo, and John Crowe for fruitful discussions. This work has, in part, been supported the European Union through Marie Curie fellowship program No. HPMF–CT–2002–01794 (M.P.), the Academy of Finland through its Centre of Excellence Program (E.S., E.T. and I.V.), and the Academy of Finland Grant Nos. 54113, 00119 (M.K.), 80246 (I.V.), and 202598 (E.T.). We would like to thank the Finnish IT Centre for Science and the HorseShoe (DCSC) supercluster computing facility at the University of Southern Denmark for computer resources.
Adachi, T., 2000. A new method for determining the phase in the x-ray diffraction structure analysis of phosphatidylcholine/alcohol. 17:93–97.
Adamson, A. W. and A. P. Gast, 1997. Physical Chemistry of Surfaces. Wiley-Interscience, New York, 6th edn.
Armen, R. S., O. D. Uitto, and S. E. Feller, 1998. Phospholipid component volumes: [D]{}etermination and application to bilayer structure calculations. 75:734–744.
Bemporad, D., J. W. Essex, and C. Luttmann, 2004. Permeation of small molecules through a lipid bilayer: [A]{} computer simulation study. 108:4875–4884.
Berendsen, H. J. C., J. P. M. Postma, W. F. van Gunsteren, A. DiNola, and J. R. Haak, 1984. Molecular dynamics with coupling to an external bath. 81:3684–3690.
Berendsen, H. J. C., J. P. M. Postma, W. F. van Gunsteren, and J. Hermans, 1981. Interaction models for water in relation to protein hydration. In B. Pullman, editor, [*Intermolecular Forces*]{}, pages 331–342. Reidel, Dordrecht.
Berger, O., O. Edholm, and F. Jahnig, 1997. Molecular dynamics simulations of a fluid bilayer of dipalmitoylphosphatidylcholine at full hydration, constant pressure, and constant temperature. 72:2002–2013.
Bisson, L. F. and D. E. Block, 2002. Ethanol tolerance in [*saccharomyces*]{}. In M. Ciani, editor, [*Biodiversity and Biotechnology of Wine Yeasts*]{}, pages 85–98. Research Signpost, Kerala.
Cantor, R., 1997. The lateral pressure profile in membranes: [A]{} physical mechanism of general anesthesia. 36:2339–2344.
Cantor, R. S., 2003. Receptor desensitization by neurotransmitters in membranes: [A]{}re neurotransmitters the endogenous anesthetics? 42:11891–11897.
Chanturiya, A., E. Leikina, J. Zimmerberg, and L. V. Chernomordik, 1999. Short-chain alcohols promote an early stage of membrane hemifusion. 77:2035–2045.
Chiu, S. W., E. Jakobsson, S. Subramanian, and H. L. Scott, 1999. Combined [M]{}onte [C]{}arlo and molecular dynamics simulation of fully hydrated dioleyl and palmitoyl-oleyl phosphatidylcholine lipid bilayers. 77:2462–2469.
Cramer, A. C., S. Vlassides, and D. E. Block, 2002. Kinetic model for nitrogen-limited wine fermentations. 77:49–60.
da Silveira, M. G., E. A. Golovina, F. A. Hoekstra, F. M. Rombouts, and T. Abee, 2003. Membrane fluidity adjustments in ethanol-stressed [*oenococcus oeni*]{} cells. 69:5826–5832.
Eckenhoff, R. G., 2001. Promiscuous ligands and attractive cavities. [H]{}ow do the inhaled anesthetics work? 1:258–268.
Essman, U., L. Perela, M. L. Berkowitz, H. L. T. Darden, and L. G. Pedersen, 1995. A smooth particle mesh [E]{}wald method. 103:8577–8592.
Falck, E., M. Patra, M. Karttunen, M. T. Hyv[ö]{}nen, and I. Vattulainen, 2004. Lessons of slicing membranes: Interplay of packing, free area, and lateral diffusion in phospholipid/cholesterol bilayers. in print.
Feller, S. E., C. A. Brown, D. T. Nizza, and K. Gawrisch, 2002. Nuclear overhauser enhancement spectroscopy cross-relaxation rates and ethanol distribution across membranes. 82:1396–1404.
Gullingsrud, J. and K. Schulten, 2004. Lipid bilayer pressure profiles and mechanosensitive channel gating. 86:3496–3509.
Hauet, N., F. Artzner, F. Boucer, C. Grabielle-Madelmont, I. Cloutier, G. Keller, P. Leiseur, D. Durand, and M. Paternostre, 2003. Interaction between artificial membrane and enflurane, a general volatile anesthetic: [DPPC]{}–enflurane interaction. 84:3123–3137.
Hess, B., H. Bekker, H. J. C. Berendsen, and J. G. E. M. Fraaije, 1997. : A linear constraint solver for molecular simulations. 18:1463–1472.
Holte, L. L. and K. Gawrisch, 1997. Determining ethanol distribution in phospholipid multilayers with [MAS]{}-[NOESY]{} spectra. 36:4669–4674.
Kessel, A., D. Tieleman, and N. Ben-Tal., 2004. Implicit solvent model estimates of the stability of model structures of the alamethicin channel. 33:16–28.
Klemm, W. R., 1990. Dehydration. 7:49–59.
Klemm, W. R., 1998. Biological water and its role in the effects of alcohol. 15:249–267.
Koubi, L., M. Tarek, S. Bandyopadhyay, M. L. Klein, and D. Scharf, 2001. Membrane structural perturbations caused by anesthetics and nonimmobilizers: [A]{} molecular dynamics investigation. 81:3339–3345.
Koubi, L., M. Tarek, M. L. Klein, and D. Scharf, 2000. Distribution of halothane in a dipalmitoylphosphatidylcholine bilayer from molecular dynamics calculations. 78:800–811.
Lee, B. W., R. Faller, A. K. Sum, I. Vattulainen, M. Patra, and M. Karttunen, 2004. Structural effects of small molecules on phospholipid bilayers investigated by molecular simulations. in print.
Lindahl, E., B. Hess, and D. van der Spoel, 2001. 3.0: A package for molecular simulation and trajectory analysis. 7:306–317.
Löbbcke, L. and G. Cevc, 1995. Effects of short-chain alcohols on the phase behavior and interdigitation of phosphatidylcholine bilayer membranes. 1237:59–69.
Ly, H. V., D. E. Block, and M. L. Longo, 2002. Interfacial tension effect on lipid bilayer rigidity, stability, and area/molecule: [A]{} micropipet aspiration approach. 18:8988–8995.
Ly, H. V. and M. L. Longo, 2004. The influence of short-chain alcohols on interfacial tension, mechanical properties, area/molecule, and permeability of fluid lipid bilayers. .
Mazzeo, A. R., J. Nandi, and R. A. Levine, 1988. Effects of ethanol on parietal cell membrane phospholipids and proton pump function. 254:G57–G64.
Nagle, J. F. and S. Tristram-Nagle, 2000. Structure of lipid bilayers. 1469:159–195.
Nagle, J. F., R. Zhang, S. Tristram-Nagle, W. Sun, H. I. Petrache, and R. M. Suter, 1996. X-ray structure determination of fully hydrated [[*L*]{}]{}$_\alpha$ phase dipalmitoylphosphatidylcholine bilayers. 70:1419–1431.
Pabst, G., M. Rappolt, H. Amenitsch, S. Bernstorff, and P. Laggner, 2000a. X-ray kinematography of temperature-jump relaxation probes the elastic properties of fluid bilayers. 16:8994–9001.
Pabst, G., M. Rappolt, H. Amenitsch, and P. Laggner, 2000b. Structural information from multilamellar liposomes at full hydration: Full $q$-range fitting with high quality x-ray data. 62:4000–4009.
Pasenkiewicz-Gierula, M., T. Róg, J. Grochowski, P. Serda, R. Czarnecki, T. Librowski, and S. Lochyński, 2003. Effects of carane derivative local anesthetic on a phospholipid bilayer studied by molecular dynamics simulation. 85:1428–1258.
Patra, M., M. Karttunen, M. Hyvönen, E. Falck, P. Lindqvist, and I. Vattulainen, 2003. Molecular dynamics simulations of lipid bilayers: Major artifacts due to truncating electrostatic interactions. 84:3636–3645.
Patra, M., M. Karttunen, M. T. Hyvönen, E. Falck, and I. Vattulainen, 2004. Lipid bilayers driven to a wrong lane in molecular dynamics simulations by subtle changes in long-range electrostatic interactions. 108:4485–4494.
Rowe, E. S., F. Zhang, T. W. Leung, J. S. Parr, and P. T. Guy, 1998. Thermodynamics of membrane partitioning for a series of n-alcohols determined by titration calorimetry: [R]{}ole of hydrophobic effects. 7:2430–2440.
Schobert, S. M., B. E. Chapman, P. W. Kuchel, R. M. Wittig, J. Grotendorst, P. Jansen, and A. A. [de Graaf]{}, 1996. Ethanol transport in *Zymomonas mobilis* measured by using in vivo nuclear magnetic resonance spin transfer. 178:1756–1761.
Tang, P. and Y. Xu, 2002. Large-scale molecular dynamics simulations of general anesthetic effects on the ion channel in the fully hydrated membrane: The implication of molecular mechanisms of general anesthesia. 99:16035–16040.
Tieleman, D., M. Sansom, and H. Berendsen, 1999. Alamethicin helices in a bilayer and in solution: Molecular dynamics simulations. 76:40–49.
Tieleman, D. P. and H. J. C. Berendsen, 1996. Molecular dynamics simulations of a fully hydrated dipalmitoylphosphatidylcholine bilayer with different macroscopic boundary conditions and parameters. 105:4871–4880.
Trandum, C., P. Westh, K. J[ø]{}rgensen, and O. G. Mouritsen, 2000. A thermodynamic study of the effects of cholesterol on the interaction between liposomes and ethanol. 78:2486–2492.
Traube, I., 1891. Ueber die [C]{}apillarit[ä]{}tsconstanten organischer [S]{}toffe in w[ä]{}sserigeb [L]{}ösungen. 265:27–55.
Tu, K., M. Tarek, M. L. Klein, and D. Scharf, 1998. Effects of anesthetics on the structure of a phospholipid bilayer: [M]{}olecular dynamics investigation of halothane in the hydrated liquid crystal phase of dipalmitoylphosphatidylcholine. 75:2123–2134.
, E., V. Chupin, J. A. Killian, and B. de Kruijff, 2004. Small alcohols destabilize the [KcsA]{} tetramer via their effect on the membrane lateral pressure. 43:5937–5942.
Vogel, M., C. M[ü]{}nster, W. Fenzl, and T. Salditt, 2000. Thermal unbinding of highly oriented phospholipid membranes. 84:390–393.
Westh, P. and C. Trandum, 1999. Thermodynamics of alcohol–lipid bilayer interactions: [A]{}pplication of a binding model. 1421:261–272.
Westh, P., C. Trandum, and Y. Koga, 2001. Binding of small alcohols to a lipid bilayer membrane: [D]{}oes the partitioning coefficient express the net affinity? 89:53–63.
[^1]: http://www.softsimu.org/downloads.shtml
[^2]: http://moose.bio.ucalgary.ca/Downloads/files/lipid.itp
[^3]: http://moose.bio.ucalgary.ca/Downloads/files/dppc.itp
[^4]: http://moose.bio.ucalgary.ca/Downloads/files/popc.itp
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We explore the radial velocity performance benefits of coupling starlight to a fast-scanning interferometer and a fast-readout spectrometer with zero readout noise. By rapidly scanning an interferometer we can decouple wavelength calibration errors from precise radial velocity measurements, exploiting the advantages of lock-in amplification. In a Bayesian framework, we investigate the correlation between wavelength calibration errors and resulting radial velocity errors. We construct an end-to-end simulation of this approach to address the feasibility of achieving 10 cm $\rm s^{-1}$ radial velocity precision on a typical Sun-like star using existing, 5-meter-class telescopes. We find that such a precision can be reached in a single night, opening up possibilities for ground-based detections of Earth-Sun analog systems.'
author:
- 'Rebecca Jensen-Clem, Philip S. Muirhead, Michael Bottom, J. Kent Wallace, Gautam Vasisht, John Asher Johnson'
bibliography:
- 'ms.bib'
title: 'Attaining Doppler Precision of 10 cm $\rm s^{-1}$ with a Lock-In Amplified Spectrometer'
---
Introduction {#sec:intro}
============
Recent results from NASA’s *Kepler* mission indicate that Sun-like stars are teeming with rocky exoplanets [@Koch2010; @Borucki2011; @howard_planet_2012; @Batalha2013; @Petigura2013]. According to statistics derived from *Kepler’s* discoveries, and consistent with prior exoplanet searches, ground-based radial velocity and transit surveys targeting nearby stars are approaching the performance edge of discovering hundreds of exoplanets suitable for detailed atmospheric studies [e.g. @Deming2009; @Howard2010; @dressing_occurrence_2013; @berta_constraints_2013]. NASA’s future Transiting Exoplanet Survey Satellite (TESS) will likely find those rocky planets that transit nearby stars in short-period orbits [@ricker_transiting_2014]. TESS will determine the radii of exoplanets transiting nearby stars. With precise radial velocity measurements of the TESS discoveries, the exoplanet masses, mean densities and surface gravities can be determined, thereby constraining their interior structures and atmospheric characteristics.
The future of exoplanet science is therefore promising; however, the precision needed to detect nearby rocky planets with radial velocities (RVs), and measure the masses of those found to transit with TESS, is daunting. For example, the Earth introduces an 8.9 cm $\rm s^{-1}$ semi-amplitude Doppler reflex motion on the Sun. To date, the smallest semi-amplitude radial velocity measured on a star is 51 cm $\rm s^{-1}$ on $\alpha$ Centauri Bb, indicating the presence of an exoplanet with a minimum mass of 1.13 M$_{\oplus}$, and a semi-major axis of 0.04 AU [@dumusque_earth-mass_2012]. The planet’s discovery has yet to be confirmed by independent instruments or techniques, so $\alpha$ Centauri Bb is widely regarded as an exoplanet candidate. Nevertheless, the measurement indicates the state-of-the-art in stellar radial velocity precision. The detection of $\alpha$ Centauri Bb requires both astrophysical and instrumental noise corrections. Known astrophysical noise sources such as stellar oscillation modes, granulation, and activity signals can be individually removed through careful choices of exposure times and observing cadence, as well as modeling each effect separately [@Dumusque2011]. @pepe_harps_2011 describe how instrumental noise sources in HARPS, such as changes to temperature, pressure, and illumination, are addressed through high cadence, long time baseline observations of nearby, slowly rotating stars with little known activity. To this end, HARPS observed HD 85512 for 7.5 years, obtaining 185 RV data points and discovering a 3.5 M$_{\oplus}$ planet. The standard deviation of the data residuals after subtracting the model of the planet was measured to be 0.75 m $\rm s^{-1}$. While this value could encompass uncharacterized astrophysical noises, 0.75 m $\rm s^{-1}$ is taken to be representative of the instrumental noise in HARPS. In order to reduce the state-of-the-art RV semi-amplitude from 51 cm $\rm s^{-1}$ ($\alpha$ Centauri Bb) to 8.9 cm $\rm s^{-1}$ (Earth-Sun analog), these instrumental noise sources must be understood and eliminated.
Eliminating the effects of instrumental noise on radial velocity measurements requires ever more precise wavelength calibration techniques. Historically, iodine absorption cells and thorium-argon (ThAr) lamps have been the principal wavelength calibration tools. Starlight passing through an iodine gas cell results in a dense series of iodine absorption lines superimposed onto the stellar spectrum. This method has been implemented on HIRES, where planet semi-amplitudes as small as 1.89 m $\rm s^{-1}$ have been detected [HD 156668b, @howard_nasa-uc_2011]. The small number of absorption lines at red and near-IR wavelengths, however, limits the use of iodine cell calibration for redder stars. Furthermore, the technique requires complex modeling of the combined iodine/stellar spectrum, which in turn requires high signal to noise data [@lovis_new_2007]. As an alternative to the iodine cell approach, fiber-fed spectrographs have used ThAr lamps as a simultaneous wavelength reference [@baranne_elodie:_1996]. Starlight and ThAr lamp light are fed into the spectrograph through separate fibers such that their spectra are simultaneously recorded onto the detector. While ThAr provides more lines in the near-IR than iodine, the non-uniformity in line spacing and long-term variability of ThAr sources limits the technique.
Recently, laser frequency combs (LFCs) have provided large numbers of equally spaced, stable lines over a wide range of wavelengths [e.g. @murphy_high-precision_2007; @Steinmetz2008]. LFCs fix the phase of standing waves inside a laser cavity such that the waves periodically interfere constructively, producing bursts of light. The time between bursts can be accurately controlled using an atomic clock. In the frequency domain, the laser therefore produces a series of equally spaced spectral lines.
An experimental LFC installed on HARPS in 2010 has yielded unprecedented RV stability over short timescales [@wilken_spectrograph_2012]. The HARPS spectrograph is fed by two multimode fibers, or channels; one channel is coupled to starlight, and the other to light from the LFC. In [@wilken_spectrograph_2012], the HARPS team fed LFC light through both fibers, and differenced the two channels to measure the instrumental drift. By optimally binning 20-30s exposures for 4 minutes, the limiting RV precision due to instabilities between the two channels was found to be 2.5 cm $\rm s^{-1}$. Between November of 2010 and January of 2011, however, synchronous drift in the channels induced a standard deviation of 34 cm $\rm s^{-1}$ between data taken in the two time periods. While planned changes to the environmental conditions surrounding HARPS made this 34 cm $\rm s^{-1}$ drift measurement a slight underestimate of the long term stability, the two channels nevertheless experienced long term drifts larger than the 2.5 cm $\rm s^{-1}$ short term limiting RV precision. These drifts are thought to be due to uncontrolled, long timescale instrumental changes.
An innovative approach to acquiring Doppler measurements of stars was proposed by @Erskine2003. By introducing a Michelson or Mach-Zender interferometer into the optical path before a traditional spectrometer, @Erskine2003 suggested high Doppler precision could be achieved with instrumental fluctuations minimized. The technique, called externally dispersed interferometry (EDI), requires that the optical path difference in the interferometer be modulated, either spatially or temporally. He suggested both a spatial modulation, by tilting one of the interferometer cavity mirrors, and a temporal modulation, by physically moving one of the cavity mirrors. The phases of the resulting fringe patterns, whether temporal or spatial, are highly sensitive to changes in the Doppler velocity of the target star. As noted by @Erskine2003, the modulation of a spectrum into a fringe pattern provides a robust way of eliminating systematic effects, such as fixed-pattern noise, in the resulting radial velocity measurements.
Since then, multiple groups have implemented EDI systems. @Ge2006 and @mahadevan_measuring_2008 describe an instrument called the Exoplanet Tracker (ET) which uses a tilted cavity mirror without temporal modulation. @Vaneyken2010 describe the theory and implementation of ET and a similar instrument called the Multi-object APO Radial Velocity Exoplanet Large-area Survey (MARVELS). @Hajian2007 describe the dispersed Fourier Transform Spectrometer (dFTS), designed to acquire high-resolution spectra rather than measure precise radial velocities, also achieved by @Erskine2003b. And finally, @muirhead_precise_2011 describe the TripleSpec Exoplanet Discovery Instrument (TEDI), which used temporal rather than spatial modulation to map out the fringes.
A particularly interesting result from @muirhead_precise_2011 was the achievement of 30 m $\rm s^{-1}$ of RMS radial velocity performance without any calibration of the spectrometer point-spread function. Typically, precise radial velocity spectrometers involve a significant amount of effort to stabilize the spectrometer PSF because uncalibrated asymmetries in the PSF are degenerate with wavelength calibration errors, and result in spurious radial velocity measurements on the target star. @muirhead_precise_2011 show that the reduced importance of wavelength calibration is a direct result of the radial velocity measurement being encoded in the phase of a sinusoid varying within each pixel, rather than the change in flux across several pixels, as is the case for traditional spectroscopy as well as for ET and MARVELS, which use spatial modulation across pixels rather than temporal modulation within a pixel.
However, a significant challenge to temporal modulation techniques involves effects from readout noise. The robustness against errors in wavelength calibration is due to modulation taking place on short timescales, and the PSF fluctuations taking place over longer time scales. PSF fluctuations or wavelength calibration errors that occur over the same timescale as the modulation are not removed. Fast scanning is therefore desired. Readout noise in large format detectors limits the speed at which one can scan and readout the detector in the spectrometer. In @muirhead_precise_2011, the authors took 30 second exposures to ensure that the dominant noise source was photon noise, rather than detector readout noise.
Recent developments in large format, high frame rate and low-readout noise detectors motivate studies of temporal modulation at very high frame rates. Electron multiplying charge-coupled devices (EMCCDs) and complementary metal-oxide-semiconductor (CMOS) detectors can be manufactured and operated to have inherently low readout noise ($<$ 1 electron per pixel) and the ability to expose and readout at very high frame rates ( 100 megapixels per second). With a high enough gain enabled, EMCCDs have effectively zero readout noise. These detectors are becoming commonplace in astronomical instruments [e.g. @Dhillon2001; @baranec_rise_2013], motivating a study of the advantages of fast-modulation EDI for ultra-precise Doppler velocimetry.
In this paper, we show that the act of modulating a spectrum in time effectively decouples wavelength calibration errors from radial velocity measurements. Therefore, the nature of the radial velocity measurement is fundamentally different than that obtained from conventional spectroscopy, where wavelength calibration completely determines the resulting radial velocity measurement. Instead, the radial velocity precision relies upon precise knowledge of the interferometer positions[^1]. We augment previous theoretical studies of the EDI technique [e.g. @Erskine2003; @Vaneyken2010] by specifically introducing and recovering wavelength calibration errors into simulated data. We also investigate the effects of interferometer position errors on radial velocity precision. Finally, we simulate the performance of a time-varying interferometric system on a medium-sized telescope to determine the feasibility of an Earth-Sun analog detection.
Theory {#sec:theory}
======
Following the discussion in @Erskine2003 and also @muirhead_precise_2011, we derive the relationship between a temporally modulated spectrum and a radial velocity measurement. We use this relationship to demonstrate the decoupling of wavelength calibration errors and RV measurements.
The intensity of the stellar spectrum recorded by the detector depends on both the wavelength and the interferometer delay. The interferometer delay is assumed to vary sinusoidally with time, therefore acting as a sinusoidal transmission comb. For a wave number $\nu$, and delay $\tau$, the measured intensity is given by $$I_{\nu, \tau}=[S_{\nu}(1+\cos{(2 \pi \tau \nu)})]*R_{\nu}
\label{eqn:intensity}$$ where $S_\nu$ is the intrinsic stellar spectrum, $(1+\cos{(2 \pi \tau \nu)})$ represents the transmission comb, $R_{\nu}$ is the spectrograph line spread function, and $*$ represents the convolution. While most practical interferometer configurations will result in two output beams, we assume here that the beams have been combined (for example, by placing a spectrograph and detector at each output and combining the intensities in post-processing).
The interferometer delay $\tau$ can be expressed as the sum of a constant bulk delay, $\tau_0$, and a smaller time-varying phase shift $\Delta \tau$. Taking $\tau = \tau_0 + \Delta \tau$, and applying trigonometric identities, the measured intensity can be re-written as: $$I_{\nu, \tau_0, \Delta \tau} = A_{\nu}+B_{\nu}\cos{(2 \pi \Delta \tau \nu)} - C_{\nu}\sin{(2 \pi \Delta \tau \nu)}.
\label{eqn:decomposed_intensity}$$ The coefficients $A_{\nu}$, $B_{\nu}$, and $C_{\nu}$ are defined as $$A_{\nu} = S_{\nu}*R_{\nu},
\label{eqn:A}$$ $$B_{\nu} = [S_{\nu}cos(2 \pi \tau_0 \nu)]*R_{\nu},
\label{eqn:B}$$ $$C_{\nu} = [S_{\nu}\sin{(2 \pi \tau_0 \nu)}]*R_{\nu},
\label{eqn:C}$$ where $A_{\nu}$ is the non-modulated spectrum, while $B_{\nu}$ and $C_{\nu}$ describe the interference between the stellar spectrum and the transmission comb. These coefficients can be determined by varying the interferometer delay $\Delta \tau$, and observing the intensity $I_{\nu, \tau_0, \Delta \tau}$. A data set, or “scan" is a list of intensities at a range of $\Delta \tau$ and $\nu$ values.
In order to compare the fitted coefficients between two scans, $B_{\nu}$ and $C_{\nu}$ are combined into a “complex visibility:" $$B_{\nu} - i C_{\nu} = [S_{\nu}e^{-i 2 \pi \tau_0 \nu}]*R_{\nu}
\label{eqn:comvis}$$ The complex visibility is used to compare two scans, where one has been Doppler shifted by $\Delta \nu = (\Delta RV /c) \nu$. $\Delta \nu$ and $1/\tau_0$ are assumed to be small compared to a resolution element of the spectrograph. By applying the Fourier convolution and shift theorems, the Doppler shifted “epoch" complex visibility $B_{\nu}^{1} - i C_{\nu}^{1}$ and the unshifted “template" visibility $B_{\nu}^{0} - i C_{\nu}^{0}$ are related by an exponential phase: $$\label{eqn:compare}
\begin{split}
B_{\nu}^{1} - i C_{\nu}^{1} & = [B_{\nu}^{0} - i C_{\nu}^{0}]e^{-i 2 \pi \tau_0 \Delta \nu}\\
& = [B_{\nu}^{0} - i C_{\nu}^{0}]e^{-i 2 \pi \tau_0 (\Delta RV)\nu/c}.
\end{split}$$
For clarity, the discussion above applies to small radial velocity shifts only, due to the use of the Fourier shift theorem. On-sky observations, however, will be affected by the Earth’s motion relative to the barycenter of the solar system, producing RV shifts on the order of 10 km $\rm s^{-1}$. @muirhead_precise_2011 describe the detailed derivation of the relationship between the epoch and template complex visibilities, arriving at the following generalization of Equation \[eqn:compare\]:
$$B_{\nu}^{1} - i C_{\nu}^{1} = e^{i 2 \pi \tau_0 \Delta \nu} \left[ (B_{\nu}^{0} - i C_{\nu}^{0} )e^{-i 2 \pi \tau_0 \Delta \nu} \right ]_{\nu \rightarrow \nu + \Delta \nu}
\label{eqn:compare_general}$$
Therefore, a Doppler shift changes the phase of the complex visibility (for example, an Earth-Sun analog would result in a phase shift of $3.4\times 10^{-5}$ radians). A wavelength calibration error re-assigns the phase values to a different wavelength grid without changing the value of the phase itself. The radial velocity measurements are therefore decoupled from wavelength calibration changes; instead, the RV precision relies upon precise knowledge of the interferometer positions, described by $\Delta \tau$ and $\tau_{0}$. Section \[sec:sim\_noise\] describes the effects of wavelength calibration and interferometer position errors in detail.
Simulation Architecture {#sec:sim}
=======================
![A Horizontal cut through Figure \[fig:smile\] reveals the simulated stellar spectrum. Note that for ease of viewing, Figure \[fig:smile\] shows the mean subtracted fringes.[]{data-label="fig:smile_cut"}](f1b-eps-converted-to.pdf){width="50.00000%"}
In order to quantify instrumental contributions to radial velocity precision, we construct an end-to-end simulation of the radial velocity reconstruction process described in Section \[sec:theory\].The flux entering the instrument is modeled using a simulated solar spectrum [@coelho_spectral_2007]. We multiply the stellar spectrum by the interferometer comb, which is modeled as a sinusoidal transmission comb as a function of wavenumber $\nu$ and interferometer delay $\tau$ (see Equation \[eqn:intensity\]). The result is therefore a series of spectra multiplied by transmission combs of different interferometer delays.
The product of the stellar spectrum and interferometer transmission comb is convolved with the spectrograph instrument profile (IP). The spectrograph IP is modeled as a normal distribution with a width corresponding to a spectrograph resolution of R=10000. The resolution was chosen to be low in order to avoid resolving the interferometer comb; in order for the phase of the interference fringes to correspond to a radial velocity change, the comb must remain unresolved. The sampling is assumed to be 4 pixels per resolution element. The result is a stellar spectrum recorded at each interferometer step (Figure \[fig:smile\_cut\]). The final data product can be represented by a three dimensional map of interferometer delay, wavelength, and observed intensity (Figure \[fig:smile\]).
It is clear from the axes of Figure \[fig:smile\] that wavelength calibration and radial velocity information are separated; the y-axis represents time (the time-varying position of the interferometer), while the x-axis represents detector pixels (the wavelength information introduced by the spectrograph). A radial velocity shift will therefore result in a vertical shift as the phase of the interference fringes changes, whereas a change in wavelength calibration will result in a global horizontal shift as the spectrum’s wavelength is re-assigned.
![A simulated mean subtracted spectrum vs. wavelength and $\Delta \tau$ for SNR = 500. The x-axis corresponds wavelength, or pixels on the spectrometer detector. The y-axis corresponds to changes in the optical path difference of the interferometer, and therefore represents time as the interferometer optical path difference is modulated temporally. The regions of high fringe contrast correspond to absorption lines in the stellar spectrum. An error in the wavelength calibration is fundamentally different from a change in the radial velocity of the target star, as long as the wavelength calibration error does not occur during an interferometer scan.[]{data-label="fig:smile"}](f1-eps-converted-to.pdf){width="50.00000%"}
Each vertical cut through Figure \[fig:smile\] represents the intensity as a function of phase delay at a given wavenumber described by Equation \[eqn:decomposed\_intensity\]. Noise is added to the intensity at each phase delay by adding values drawn from a Gaussian distribution with a mean of zero and a standard deviation given by the spectrum’s mean divided by the desired signal to noise ratio. For clarity, we refer to this as Poisson noise, but our simplifications (Gaussian statistics with a constant standard deviation) should be noted. The $A_\nu$, $B_\nu$, and $C_\nu$ coefficients are then fit to each vertical cut, defining the complex visibilities (Equation \[eqn:comvis\]).
In order to measure a radial velocity shift, two data sets, each represented by a figure like Figure \[fig:smile\], are produced: a template spectrum and an epoch spectrum with an RV shift and wavelength calibration error. The template spectrum is assumed to have infinite signal to noise, due to either a long on-sky integration or the use of a simulated spectrum. The template spectrum is multiplied by a complex phasor as in Equation \[eqn:compare\_general\], containing a test RV shift, and is interpolated onto a new wavelength grid to represent the effect of the wavelength calibration error. This modified template spectrum and the originally shifted epoch spectrum are then compared using a chi-squared test that treats the real and imaginary parts of the complex visibilities separately:
$$\frac{-\chi^2}{2} = \displaystyle\sum\limits_{\nu} \left( B_{\nu}^{s} - B_{\nu}^{1} \right )^{2} + \left( C_{\nu}^{s} - C_{\nu}^{1} \right )^{2}
\label{eqn:loglike}$$
where $B_{\nu}^{1}$ and $C_{\nu}^{1}$ describe the epoch spectrum, and $B_{\nu}^{s}$ and $C_{\nu}^{s}$ describe the modified template spectrum. Because the likelihood function is defined as $$L = e^{\frac{-\chi^2}{2}},
\label{eqn:like}$$ Equation \[eqn:loglike\] is considered to be log($L$). We compute log($L$) on a grid of radial velocity and wavelength calibration error test points, and find the points that minimize Equation \[eqn:loglike\] by parabola fitting. By computing many such solutions for independent realizations of the noise, we construct histograms of radial velocity and wavelength calibration errors solutions that are well described by Gaussian statistics. We define the standard deviation of 100 radial velocity solutions to be the “radial velocity precision." By fitting a 2D Gaussian distribution to a 2D histogram of several thousand radial velocity and wavelength calibration error solutions, we compute error ellipses demonstrating the relationship between the two parameters (Figure \[fig:contours\]). In contrast to traditional RV reconstruction methods, the contours show an elliptical shape, demonstrating that the RV and wavelength calibration error measurements are not highly correlated.
![Simulated likelihood contours describing the radial velocity and wavelength calibration error reconstructions for injected values of 1000 m $\rm s^{-1}$ and 500 m $\rm s^{-1}$ respectively, for SNR=100 and a spectral bandwidth of $\Delta \lambda = 88\mbox{\AA}$. In contrast to traditional RV reconstruction methods, the contours show an elliptical shape, demonstrating that the RV and wavelength calibration error measurements are not highly correlated.[]{data-label="fig:contours"}](f2-eps-converted-to.pdf){width="50.00000%"}
------------------------------- --------- -------- -----------------------------------------------------------------------
Parameter Value Units Description
Spectrograph resolution 10,000 This is chosen to be small to avoid resolving the interferometer comb
Pixels per resolution element 4.0 pixels Slightly exceeds Nyquist sampling
Bandpass 507-595 nm $V$-band FWHM
Stellar model Coelho N/A G star spectrum from Coelho et al 2007
Scan step size 0.012 $\mu$m 20$\times$(Nyquist sampling)
Scan stroke 2.5 $\mu$m 5 fringe cycles
Bulk Delay 2.0 cm
------------------------------- --------- -------- -----------------------------------------------------------------------
![Radial velocity precision is plotted against the signal to noise ratio for Poisson noise limited measurements, (no instrumental or astrophysical noise sources). Under these conditions, an SNR of about 94 per phase step, for 200 phase steps, is required to reach a precision of 10 cm $\rm s^{-1}$. The RV precisions shown in this plot were calculated using a spectral bandwidth of $\Delta \lambda = 88\mbox{\AA}$, and were divided by $\sqrt{10}$ to reflect the RV precisions associated with the full $V$-band bandwidth of $880\mbox{\AA}$. []{data-label="fig:snr"}](f3-eps-converted-to.pdf){width="50.00000%"}
![The radial velocity precision is shown to be approximately constant over 10 km $\rm s^{-1}$ ($\sim 0.2 \mbox{\AA}$) wavelength calibration errors for SNR = 10, 100, and 1000. In contrast, the radial velocity precision is proportional to the wavelength calibration error in conventional spectroscopy. The RV precisions shown in this plot were calculated using a spectral bandwidth of $\Delta \lambda = 88\mbox{\AA}$, and were divided by $\sqrt{10}$ to reflect the RV precisions associated with the full $V$-band bandwidth of $880\mbox{\AA}$. []{data-label="fig:wav"}](f4-eps-converted-to.pdf){width="50.00000%"}
![The RV precision is plotted against the standard deviation of the phase step error. The horizontal dotted lines represent the Poisson limited RV precision for each SNR. In the phase step error limited regions, the RV precisions approach the limiting $SNR=\infty$ condition, shown by the black line. For SNR = 100, $\Delta \tau$ must be known to less than 1 nm to reach the Poisson limited regime. Because the phase step errors are assumed to be uncorrelated, the total number of phase steps affects the final RV precision. In this simulation, we chose 200 steps (Table \[tbl:assump1\]). The RV precisions shown in this plot were calculated using a spectral bandwidth of $\Delta \lambda = 88\mbox{\AA}$, and were divided by $\sqrt{10}$ to reflect the RV precisions associated with the full $V$-band bandwidth of $880\mbox{\AA}$. []{data-label="fig:phase_steps"}](f5-eps-converted-to.pdf){width="50.00000%"}
Error Budget {#sec:sim_noise}
============
The simulation described in Section \[sec:sim\] is used to analyze the affects of Poisson noise, wavelength calibration errors, and interferometer position errors. The assumptions used in the simulation are listed in Table \[tbl:assump1\]. We also note that all radial velocity and wavelength calibration error solutions discussed in this section were generated using a spectral bandwidth of $\Delta \lambda = 88\mbox{\AA}$, while the FWHM of the $V$-band filter is $\Delta \lambda = 880\mbox{\AA}$. We chose this smaller bandwidth to accommodate the available computing resources. Because the number of spectral lines increases with the square root of the bandwidth, however, we can represent the $\Delta \lambda = 880\mbox{\AA}$ radial velocity precision by dividing the $\Delta \lambda = 88\mbox{\AA}$ radial velocity precision by $\sqrt{10}$. The text and figures below refer to radial velocity precisions that have been modified by this factor.
Poisson Noise
-------------
In the absence of all instrumental or astrophysical noise sources, the radial velocity precision decreases as $1/SNR$ (Figure \[fig:snr\]). Under these conditions, a signal to noise ratio of about 94 per spectrograph resolution element is required to reach a radial velocity precision of 10 cm $\rm s^{-1}$. Sections \[sec:wav\_cal\]-\[sec:int\_errors\] below describe how instrumental noise sources cause the RV precision to deviate from this ideal case.
Wavelength Calibration Errors {#sec:wav_cal}
-----------------------------
Figure \[fig:wav\] shows that the radial velocity precision changes by mm/s per km/s of wavelength calibration error for high signal to noise ratios. The black line in Figure \[fig:wav\] represents the 1:1 correspondence between wavelength calibration and radial velocity using conventional RV reconstruction methods. We have therefore demonstrated that by temporally varying a stellar spectrum, the dependence of the radial velocity precision on a consistent, correct wavelength solution is reduced.
Interferometer Position Errors {#sec:int_errors}
------------------------------
This robustness against wavelength calibration errors comes at the cost of precise knowledge of the interferometer position. The interferometer position is described by a constant “bulk" offset and a much smaller time varying phase shift ($\tau_{0}$ and $\Delta \tau$, respectively, in Equation \[eqn:decomposed\_intensity\]).
Equation \[eqn:compare\_general\] shows that errors in estimation of the bulk delay produce proportional errors in the RV precision. Therefore, the same error in a bulk delay measurement will result in poorer RV precision for stars with larger radial velocities. In order to calculate the smallest required bulk delay measurement error, we must consider the target stars with the largest radial velocities. Barycentric motion produces radial velocity differences of up to 60 km $\rm s^{-1}$. In order to reach an RV precision of 10 cm $\rm s^{-1}$ with a typical bulk delay of $2\,$cm, the maximum bulk delay measurement error is $\left ( \frac{10 \, \mbox{cm}}{60 \, \mbox{km}}\right )2 \, \mbox{cm} = 33 \,$nm. State-of-the-art piezo electric stages are capable of sub-nm RMS measurement accuracy (e.g. @samuele_experimental_2007), so we conclude that the bulk delay can be adequately measured to provide 10 cm s$^{-1}$ precision on the stars with the largest radial velocities.
--------------------------------------------- --------- --------- -----------------------------------------------------------------------------------------------------------------------
Parameter Value Units Comments
Bandpass 507-595 nm $V$-band FWHM
Sky to detector throughput 7.4 $\%$ Representative of a single mode fiber coupled to a $30\%$ Strehl input beam and a spectrograph with $30\%$ efficiency
Interferometer position control requirement 1 nm Conservative requirement based on 0.4nm state-of-the-art piezoelectric parameters
Stellar visual magnitude 8.5 $m_{v}$ Representative of the mode of the visual magnitudes of the 100 brightest G stars
Minimum SNR per resolution element 8 Based on minimum SNR to avoid the read noise limited regime
--------------------------------------------- --------- --------- -----------------------------------------------------------------------------------------------------------------------
Errors in the phase step estimations, however, impose more strict metrology requirements. Figure \[fig:phase\_steps\] shows the radial velocity precision as a function of phase step error. Normally distributed errors with standard deviations represented by the values on the x-axis were added to the true phase delays. The horizontal regions of the SNR = 10, 100, and 1000 plots show the Poisson limited regime, while the sloped regions represent the phase step error limited regime. The black line represents SNR = $\infty$. For a single SNR = 100 scan, the phase steps must to known to sub-nanometer precision to remain in the Poisson limited regime. Because the phase step position errors are expected to be normally distributed, however, taking multiple scans will improve the RV precision by the square root of the number of scans. In this way, the SNR of the combined scan is increased while the errors due to interferometer position errors are reduced. Section \[sec:sim\_results\] describes this approach in detail.
The Feasibility of an Earth-Sun Analog Measurement {#sec:sim_results}
==================================================
We have shown that the modulation of a stellar spectrum in time decouples wavelength calibration errors from RV measurements, while coupling interferometer position errors to RV measurements. Given these new constraints, we now address the feasibility of reaching a radial velocity precision of 8.9 cm $\rm s^{-1}$ on a G-type star using an exiting telescope. Our assumed bandpass, throughput, interferometer control requirement, and target star magnitude are summarized in Table \[tbl:extra\_assumps\] and described in detail below.
Operating in the visible ($\approx 500-600\,$nm) while observing Sun-like stars has the dual advantages of covering the peak of the G star blackbody function while reducing contamination due to telluric lines compared to the infrared. A visible bandpass, however, may come at the cost of throughput. In order to minimize the effects of interferometer position errors, calibration and star light should be common path in the interferometer and spectrograph. One possible approach would be to feed the instrument with a single mode fiber, containing both starlight and light from a calibration source (for example, a frequency stabilized laser). In this way, modulation of the calibration light, imprinted on the science data itself, will measure the changes in the optical path difference of the interferometer.
The single-mode fiber coupling efficiency is determined primarily by the Strehl ratio of the associated adaptive optics system. Single-mode fibers are inherently diffraction limited, and require diffraction-limited beams delivered from a telescope in order to have any consequential coupling: seeing-limited telescopes will not couple nearly enough starlight for a Doppler survey of nearby stars [@shaklan_coupling_1988]. Recently, however, advanced adaptive optics systems have reported significant Strehl ratios at visible wavelengths. For example, the Robo-AO system on the Palomar 60-inch telescope has achieved a Strehl ratio of $18\%$ in i-band [@baranec_robo-ao:_2012]. The Palm-3000 adaptive optics system on the Palomar 200-inch telescope is designed to achieve high Strehl ratios at visible wavelengths. Palm-3000 is predicted to achieve $30\%$ Strehl in V-band, equivalent to a 95 nm RMS wavefront error [@dekany_palm-3000:_2013]. Researchers at the University of Arizona recently commissioned a visible-light adaptive optics system (VisAO) on the 6.5-m Magellan Clay Telescope, demonstrating a $43\%$ Strehl ratio, or 149 nm RMS wavefront error, in Y-band [@close_diffraction-limited_2013]. For the purposes of this study, we estimate the total sky-to-detector throughput to be $7.4\%$ by multiplying the ideal fiber coupling efficiency of $82\%$ by a Strehl ratio of $30\%$, based on the performance of Palm-3000 at visible wavelengths, and a spectrograph efficiency of $30\%$.
As described in Section \[sec:theory\], instrumental noise varying more slowly than the interferometer’s phase steps is rejected. For maximum noise rejection, the interferometer should therefore scan as quickly as possible. The interferometer’s speed is then limited by detector frame rates and read noise. The lowest noise detectors at optical wavelengths, CCDs, have significant readout times of seconds to minutes. This limits the rate at which we can modulate an optical signal to the point that lock-in amplification is effectively pointless with most CCD detectors. However, the development of high-frame rate, low read noise EMCCDs and sCMOS detectors opens up the possibility of using lock-in amplification at optical wavelengths. Assuming a detector read noise of $2 e^{-}$ per pixel, gain = 1, and four pixels per resolution element, the detector must measure 64 photons per resolution element (SNR=8) to remain well within the photon noise limited regime. The exposure time necessary to reach this minimum SNR will place an upper limit on the scanning speed.
Taking a series of scans can reduce normally distributed noise due to interferometer position errors. The number of scans is chosen such that the desired RV precision is achieved under realistic position error conditions. State-of-the-art piezoelectric stages can be controlled to less than 1 nm RMS (e.g. @samuele_experimental_2007 controlled the Physik Instrumente P-752 flexure stage to 0.4 nm rms), so we choose a 1 nm control requirement. In order to reach 10 cm $\rm s^{-1}$ with phase step errors of 1 nm RMS and bulk delay errors of $10\%$ $\tau_{0}$ , we require 139 scans, each with SNR = 8.
We now choose a target star magnitude and telescope size to determine the observing time required to take 139 such scans. To choose a representative target star magnitude, we constructed a histogram of the 100 brightest G stars. We choose $m_{v} = 8.5$ as a representative magnitude from this sample.
The Palomar 200“ telescope would take about 0.074 seconds to reach SNR=8 per resolution element, given the assumptions in Tables \[tbl:assump1\] and \[tbl:extra\_assumps\]. For 139 scans, each with 200 $0.074\,$s phase steps, the 200” telescope would require about 0.57 hours to reach a radial velocity precision of 10 cm $\rm s^{-1}$. It is therefore possible to reach the radial velocity precision necessary to detect Earth-like planets around Sun-like stars using existing, 5-meter-class telescopes.
Conclusions {#sec:conclusion}
===========
The current state-of-the-art radial velocity instruments are limited by their ability to maintain their system’s wavelength solution in the presence of slowly varying instrumental noise. Lock-in amplification, however, can suppress such long timescale noise sources, while decoupling the effects of wavelength calibration errors from radial velocity precision. The simulations presented in this paper indicate that lock-in amplified, externally-dispersed interferometry is a possible path forward to reach the radial velocity precision necessary to detect Earth-Sun analog systems on existing, medium-sized telescopes.
This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant No. DGE-1144469. P.S.M. acknowledges support for his work from the Hubble Fellowship Program, provided by NASA through Hubble Fellowship grant HST-HF-51326.01-A awarded by the STScI, which is operated by the AURA, Inc., for NASA, under contract NAS 5-26555. P.S.M, G.V., and J.K.W. were supported by the Directors Research and Development Fund and Caltech-JPL President and Director’s fund at the California Institute of Technology/Jet Propulsion Laboratory. M.B. is supported by a National Space Technology Research Fellowship. J.A.J. is supported by generous grants from the David and Lucile Packard Foundation and the Alfred P. Sloan Foundation. JKW and GV are supported by the We would like to thank Lynne Hillenbrand for useful discussions.
[^1]: A possible implementation could involve a combination of fine pathlength control with precision PZTs and fine sensing using a co-propagating laser with exceptional wavelength stability.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We present a relation which connects the propagator in the radial (Fock-Schwinger) gauge with a gauge invariant Wilson loop. It is closely related to the well-known field strength formula and can be used to calculate the radial gauge propagator. The result is shown to diverge in four-dimensional space even for free fields, its singular nature is however naturally explained using the renormalization properties of Wilson loops with cusps and self-intersections. Using this observation we provide a consistent regularization scheme to facilitate loop calculations. Finally we compare our results with previous approaches to derive a propagator in Fock-Schwinger gauge.'
---
------------------------------------------------------------------------
------------------------------------------------------------------------
March 27, 1996 TPR–95–31, hep-th/9604015
------------------------------------------------------------------------
------------------------------------------------------------------------
[**RADIAL PROPAGATORS AND WILSON LOOPS**]{}
Stefan Leupold[^1]
Institut für Theoretische Physik, Universität Regensburg,
D-93040 Regensburg, Germany
Heribert Weigert[^2]
School of Physics and Astronomy, University of Minnesota,
Minneapolis, MN 55455, USA
0.3cm
Introduction
============
While perturbation theory for gauge fields formulated in covariant gauges is very well established [@pascual] many aspects of non-covariant gauges are still under discussion. In principle one expects physical quantities to be independent of the chosen gauge. However this might lead to the naive conclusion that a quantum theory in an arbitrary gauge is simply obtained by inserting the respective gauge fixing term and the appropriate Faddeev-Popov ghosts in the path integral representation and reading off the Feynman rules. Unfortunately it is not so easy to obtain the correct Feynman rules, i.e. a set of rules yielding the same results for observable quantities as calculations in covariant gauges. Prominent examples are formulations in temporal and axial gauges. Such gauge choices are considered since one expects the Faddeev-Popov ghosts to decouple. However problems even start with the determination of the appropriate free gauge propagators. Temporal and axial gauge choices yield propagators plagued by gauge poles in their momentum space representation. These are caused by the fact that such gauge conditions are insufficient to [*completely*]{} remove the gauge degrees of freedom. The correct treatment of such poles can cause ghost fields to reappear [@cheng], can break translational invariance [@cara] or both [@leroy]. While these problems seem to be “restricted” to the evaluation of the correct gauge propagators and ghost fields, the necessity of introducing even new multi-gluon vertices appears in the Coulomb gauge [@christ]. These additional vertices are due to operator ordering problems which are difficult to handle in the familiar path integral approach. They give rise to anomalous interaction terms at the two-loop level [@doust] and cause still unsolved problems with renormalization at the three-loop level [@taylor].
In this article we are interested in the radial (Fock-Schwinger) gauge condition $$x_\mu A^\mu(x) = 0 \,. \label{eq:fsgaugecond}$$ It found widespread use in the context of QCD sum-rules (e.g. [@shif]). There it is used as being more or less synonymous to the important field strength formula $$\label{eq:fsformula}
A_\mu^{\mbox{\scriptsize rad}}(x) = \int\limits^1_0 \!\! ds \, s x^\nu
F_{\nu\mu}(sx)$$ which enormously simplifies the task of organizing the operator product expansion of QCD n-point functions in terms of gauge invariant quantities by expressing the gauge potential via the gauge covariant field strength tensor. It was introduced long ago [@fock], [@schwing] and rediscovered several times (e.g. [@cron]).
Only a few efforts have been made to establish perturbation theory for radial gauge. The main reason for this is that the gauge condition breaks translational invariance since the origin (in general an arbitrary but fixed point $z$, c.f. (\[eq:arbz\])) is singled out by the gauge condition. Thus perturbation theory cannot be formulated in momentum space as usual but must be set up in coordinate space.
The first attempt to evaluate the free radial propagator was performed in [@kumm]. Later however the function $\Gamma_{\mu\nu}(x,y)$ presented there was shown to be not symmetric [@moda]. Moreover it could not be symmetrized by adding $\Gamma_{\nu\mu}(y,x)$ since the latter is not a solution of the free Dyson equation. It was even suspected in [@moda] that it might be impossible to find a symmetric solution of this equation in four-dimensional space, due to the appearance of divergences even on the level of the [*free*]{} propagator when one uses the field strength formula to derive a free propagator. Indeed we agree with this statement in principle, but we will present an explanation for this problem and a way to bypass it. Other approaches to define a radial gauge propagator try to circumvent the problem (e.g. [@menot]) by sacrificing the field strength formula as given in (\[eq:fsformula\]) which was one of the main reasons the gauge became popular in non-perturbative QCD sum rule calculations [@shif] in the first place. If we are not prepared to do so we are forced to understand the origin of the divergences that plague most of the attempts to define even free propagators in radial gauges and see whether they can be dealt with in a satisfying manner.
In Section \[sec:radgaugecond\] we will make the first and decisive step in this direction by exploring the completeness of the gauge condition (\[eq:fsgaugecond\]) and its relation to the field strength formula and developing a new representation of the gauge potentials via link operators.
In Section \[sec:radprop\] we use this information to relate the divergences encountered in some of the attempts to define radial propagators to the renormalization properties of link operators. We find that even free propagators in radial gauge may feel remnants of the renormalization properties of closed, gauge invariant Wilson loops. Surprising as this seems to be superficially it is not impossible however if we recall that the inhomogeneous term in the gauge transformation has an explicit $1/g$ factor in it. As a result we are able to define a regularized radial propagator using the field strength formula and established regularization procedures for link operators.
Section \[sec:radcalc\] will be devoted to demonstrate the consistency of our approach by calculating a closed Wilson loop using our propagator and relating the steps to the equivalent calculation in Feynman gauge.
In Section \[sec:radren\] we obtain an explicitly finite version of our propagator by completing the renormalization program developed for link operators before we summarize and compare our results to other approaches in the literature in \[sec:radsum\] and shortly discuss the next steps in the program of establishing a new perturbative framework in radial gauges which – although the steps to be performed are quite straightforward – we will postpone for a future publication.
In the following we work in a $D$-dimensional Euclidean space. The vector potentials are given by $$A_\mu(x) \equiv A^a_\mu(x) \,t_a$$ where $t_a$ denotes the generators of an $SU(N)$ group in the fundamental representation obeying $$[t_a,t_b] = i f_{abc} \,t^c$$ and $$\mbox{tr}(t_a t_b) = {1\over 2} \,\delta_{ab} \,.$$ In general the radial gauge condition with respect to $z$ reads $$(x-z)_\mu A^\mu(x) =0 \,. \label{eq:arbz}$$ For simplicity we take $z=0$ after Section \[sec:radgaugecond\]. The results nevertheless can be easily generalized to arbitrary values of $z$.
The gauge condition revisited {#sec:radgaugecond}
=============================
Before we can go ahead and tackle the problem of divergences in the radial gauge propagator we have to establish a clearer picture of the uniqueness of the gauge condition we are about to implement. After all, if we do not succeed to fix the gauge completely we might be naturally confronted with divergences – if not at the free level then later in perturbative calculations. They would be a simple consequence of the incompleteness of the gauge fixing and the zero modes of the propagator which would then necessarily be present. This point has caused a still continuing discussion for the case of axial gauges (e.g. [@leroy]) but is only briefly mentioned in the context of radial gauges (e.g. [@azam]).
Readers who are not interested in the discussion of (in)completeness of radial gauge conditions might skip the following considerations without getting lost and start reading again after eq. (\[eq:gaugetrans\]).
To clarify the question whether the gauge condition (\[eq:arbz\]) is sufficient to completely fix the gauge degrees of freedom we have to catalogue the gauge transformations $U[B](x)$ which transform an arbitrary vector potential $B$ into the field $A$ satisfying (\[eq:arbz\]). A gauge condition is complete if $U[B](x)$ is uniquely determined up to a global gauge transformation. In other words, we want to find all solutions of $$\label{eq:transformations}
(x-z)_\mu\ U[B](x)\left[B^\mu(x) -{1\over i g}
\partial^\mu\right]U[B]^{-1}(x) = 0 \,.$$ It is easily checked that we have an infinite family of such solutions which can all be cast in the form of a product of two gauge transformations of the form $$\label{eq:gaugesol}
U[B](x) = V(z(x)) U[B](z(x),x) \,.$$ Here $$\label{eq:gaugetranssol}
U[B](z(x),x) = {\cal P} \exp i g \int^{z(x)}_x \!d\omega_\mu
B^\mu(\omega)$$ is a link operator whose geometric ingredients are parameterized via its endpoints $x$ and $z(x)$ and the straight line path $\omega$ between them, ${\cal P}$ denotes path ordering.
In particular $z(x)$ is the point where a straight line from $z$ through $x$ and a given closed hyper-surface around $z$ intersect. Since there is a unique relation between these points and the hyper-surface we will also refer to the hyper-surface itself by $z(x)$. This geometry is illustrated in Fig. \[fig:sphericalgeom\].
=
Both the detailed form of $z(x)$ and the local gauge transformation $V(x)$ are completely unconstrained as long as $(x-z).\partial^x z(x)
= 0$. In short, they parameterize the residual gauge freedom not eliminated by (\[eq:fsgaugecond\]). Note that while $V(x)$ is completely arbitrary the solutions (\[eq:gaugesol\]) ask only for its behavior at the given hyper-surface $z(x)$. The simplest and most intuitive choice for $z(x)$ is a spherical hyper-surface around $z$. Introducing the appropriate spherical coordinates it becomes obvious that $V(z(x))$ parameterizes gauge transformations which purely depend on the angles. Clearly the radial gauge condition (\[eq:arbz\]) cannot fix the angular dependence of any gauge transformation in (\[eq:transformations\]).
To eliminate the residual gauge freedom one has to impose a condition which is stronger than (\[eq:arbz\]) and suffices to pin down $V(x)$ up to a global transformation. A possible choice for such a gauge fixing would be the condition $$\begin{aligned}
\label{eq:compgaugefix}
\Box \left(\int_{z(x)}^x \!d\omega.A(\omega) + \int\!d^4\!y \ {1\over
\Box}(z(x),y)\ \partial^y.A(y)\right) \equiv 0\end{aligned}$$ which in addition to the vanishing of the radial component of the gauge potential also implements a covariant gauge on the hyper-surface $z(x)$. Such a gauge for arbitrary $z(x)$ would immediately force us to introduce ghosts into the path integral. Moreover the field strength formula would also be lost as we will illustrate below.
There is one exception to these unwanted modifications however, which may be implemented by contracting the closed surface $z(x)$ to the point $z$. Then the influence of $V(x)$ becomes degenerate with a global transformation and the gauge is completely fixed. Incidentally this is also the only case which entails the field strength formula. To see this we use $$\begin{aligned}
\lefteqn{ \delta U(x,z) = i g \;\Bigg\{ A_\mu(x)U(x,z)dx^\mu
- U(x,z)A_\mu(z)dz^\mu + } \nonumber \\
& & - \int^1_0\!\!ds\, [U(x,w_x)F_{\mu\nu}(w_x)U(w_x,z)]
{dw_x^\mu\over ds}
\left ( {dw_x^\nu\over dx^\alpha}dx^\alpha
+ {dw_x^\nu\over dz^\alpha}dz^\alpha
\right )
\Bigg\} \quad , \nonumber \\
\label{eq:udiff}\end{aligned}$$ (see e.g. [@bralic], [@ElGyuVa86]) to differentiate the link operators in the gauge transformation (\[eq:gaugesol\]) in order to find an expression for the radial gauge field: $$\begin{aligned}
\label{eq:gaugetrans}
A^{\mbox{\scriptsize rad}}_\mu(x) & = & U[A](z,x)\left[A_\mu(x) -{1\over i g}
\partial^x_\mu\right]U[A](x,z)
\nonumber \\ & = &
\int\limits^1_0 \!\! ds \, s {d\omega^\nu\over d s}
\,U[A](z,\omega)\,F_{\nu\mu}(\omega) \,U[A](\omega,z)
\nonumber \\ & = &
\int\limits^1_0 \!\! ds \, s {d\omega^\nu\over d s}
\,F_{\nu\mu}^{\mbox{\scriptsize rad}}(\omega) \,.\end{aligned}$$ This is nothing but (\[eq:fsformula\]) for arbitrary $z$ (note that in this case $\omega = \omega(s)$ is simply given by $\omega(s) = z +
(x-z) s$.) This simple result is only true since $\partial_\mu z(x)
\equiv \partial_\mu z = 0$. For general $z(x)$ there would be an additional term in the above formula reflecting the residual gauge freedom encoded in $V(z(x))$.
This sets the stage for a further exploration of the radial gauge in a context where we can be sure of having completely fixed the gauge in such a way that the field strength formula is guaranteed to be valid. Before we go on to studying the consequences the above has for the implementation of propagators we will introduce yet another representation of the gauge field in this particular complete radial gauge – this time solely in terms of link operators.
From now on we will assume the reference point $z$ to be the origin, but it will always be straightforward to recover the general case without any ambiguities. We will also suppress the explicit functional dependence of link operators on the gauge potential $A$ for brevity.
Let us start with a link operator along a straight line path $$U(x,x') = {\cal P}\exp\left[
i g\int\limits^1_0 \!\!d\omega_\mu A^\mu(\omega)
\right]$$ where now $ \omega(s):= x'+s(x-x')\, $. According to (\[eq:udiff\]) we have $$\partial_\mu^x \, U(x,x') = i g \left[ A_\mu(x) - \!\!\int\limits^1_0
\!\! ds \, s \, {d\omega^\nu\over d s}\ U(x,\omega)\,F_{\nu\mu}(\omega)
\,U(\omega,x) \right] U(x,x')
\label{eq:difstr}$$ which can be used to express the vector potential in terms of the link operator $$\lim_{x'\to x} \partial_\mu^x \, U(x,x') = i g \, A_\mu(x) \,.$$ In the case at hand the fact that $U(0,x) = 1$ in any of the $x.A(x) = 0$ gauges allows us to introduce a new gauge covariant representation $$A^{\mbox{\scriptsize rad}}_\mu(x) =
{1\over i g} \lim_{x'\to x} \partial_\mu^x
\left[ U(0,x)\,U(x,x')\,U(x',0) \right] \label{eq:gaucov}$$ for the Fock-Schwinger gauge field. It is easy to see that this is indeed equivalent to the field strength formula as given in (\[eq:fsformula\]) and consequently satisfies the same complete gauge fixing condition (i.e. (\[eq:compgaugefix\]) for $z(x) \to
z$): $$\begin{aligned}
A^{\mbox{\scriptsize rad}}_\mu(x)
&=&
{1\over i g} \lim_{x'\to x} \partial_\mu^x
\left[ U(0,x)\,U(x,x')\,U(x',0) \right]
\nonumber\\ &=&
{1\over i g} \lim_{x'\to x}
\left[ \partial_\mu^x \,U(0,x) \,U(x,x') \,U(x',0) \right.
\nonumber \\ && \hskip 2cm \left.
+ U(0,x) \,\partial_\mu^x \,U(x,x') \,U(x',0)\right]
\nonumber\\ &=&
{1\over i g} \,\partial_\mu^x \,U(0,x) \,U(x,0)
+ U(0,x) \,A_\mu(x)\,U(x,0)
\nonumber\\ &=&
\int\limits^1_0 \!\! ds \, s x^\nu F^{\mbox{\scriptsize rad}}_{\nu\mu}(sx)
\label{eq:fistr}\end{aligned}$$ where the last step uses (\[eq:difstr\]), mirroring the relations in (\[eq:gaugetrans\]) for $z=0$.
The Radial Gauge Propagator {#sec:radprop}
===========================
Having established the complete gauge fixing we are interested in, it is now straightforward to devise expressions for the propagator as a two-point function. According the above we know that $$\begin{aligned}
\label{eq:propdef}
\lefteqn{ \langle A_\mu(x) \otimes A_\nu(y)\rangle_{\mbox{\scriptsize rad}} }
\nonumber \\& = & \lim_{x'\to x \atop y'\to y}\partial_\mu^x
\partial_\nu^y \, \langle U(0,x)\,U(x,x')\,U(x',0)\otimes
U(0,y)\,U(y,y')\,U(y',0) \rangle \nonumber \\ & = & \int\limits^1_0
\!\! ds \int\limits^1_0 \!\! dt\, sx^\alpha \, ty^\beta \, \langle
U(0,sx)\,F_{\alpha\mu}(sx) \,U(sx,0)\otimes
U(0,ty)\,F_{\beta\nu}(ty) \,U(ty,0) \rangle \ \ .
\nonumber \\ &&\end{aligned}$$
Since we are in a fixed gauge it makes sense to perform a multiplet decomposition and for instance extract the singlet part of this propagator. The latter reduces to the free propagator in the limit $g
\to 0$.
We define $$\mbox{tr}\,\langle A_\mu(x) A_\nu(y) \rangle
=\mbox{tr}(t_a t_b)
\underbrace{\langle A^a_\mu(x) A^b_\nu(y) \rangle^{\mbox{\scriptsize
singlet}}}_{{\textstyle
=:\delta^{ab} D_{\mu\nu}(x,y)}} = {N^2 -1\over 2} \,
D_{\mu\nu}(x,y)$$ to extract $$\begin{aligned}
\lefteqn{\langle A^a_\mu(x) A^b_\nu(y) \rangle^{\mbox{\scriptsize
singlet}}_{\mbox{\scriptsize rad}}
= \delta^{ab} {2\over N^2 -1}
\mbox{tr}\, \langle A_\mu(x) A_\nu(y) \rangle_{\mbox{\scriptsize
rad}} = \delta^{ab} {2\over N^2 -1} }
\label{propa} \\
&& \times {1\over (ig)^2} \lim_{x'\to x \atop
y'\to y}\partial_\mu^x \partial_\nu^y \, \mbox{tr}\,\langle
U(0,x)\,U(x,x')\,U(x',0)\,U(0,y)\,U(y,y')\,U(y',0) \rangle
\,.\nonumber\end{aligned}$$ Obviously $$\begin{aligned}
W_1(x,x',y,y') := {1\over N} \, \mbox{tr}\,\langle
U(0,x)\,U(x,x')\,U(x',0)\,U(0,y)\,U(y,y')\,U(y',0) \rangle
\label{w1def}\end{aligned}$$ is a gauge invariant Wilson loop. Its geometry is illustrated in Fig. \[fig:faech\].
On the other hand, using the second expression in (\[eq:propdef\]) we have an equivalent representation for the singlet part of radial gauge propagator via the field strength formula: $$\begin{aligned}
\lefteqn{\langle A^a_\mu(x) A^b_\nu(y) \rangle_{\mbox{\scriptsize
rad}}^{\mbox{\scriptsize singlet}}
= \delta^{ab} {2\over N^2 -1} }
\label{profi}\\ && \times
\int\limits^1_0 \!\! ds \int\limits^1_0 \!\! dt\, sx^\alpha \,
ty^\beta \, \mbox{tr}\,\langle U(0,sx)\,F_{\alpha\mu}(sx) \,U(sx,0)
\,U(0,ty)\,F_{\beta\nu}(ty) \,U(ty,0) \rangle \,. \nonumber\end{aligned}$$ Modanese [@moda] tried to calculate the free radial gauge propagator from (\[profi\]) in a $D$ dimensional space-time.[^3] Unfortunately one gets a result which diverges in the limit $D\to 4$.
Since we have performed a complete gauge fixing (at least on the classical level), this comes as a surprise since we certainly do not expect zero mode problems to come into the game as a possible explanation and consequently a way out. Does this mean we are trapped at a dead end or is there another explanation for this seemingly devastating discovery?
Before we try to answer this question let us briefly recapitulate how this divergence makes its appearance: Since the right hand side of (\[profi\]) is gauge invariant we can choose an arbitrary gauge to calculate it. For simplicity we take the Feynman gauge with its free propagator $$\langle A_\mu^a(x) A_\nu^b(y) \rangle_{\mbox{\scriptsize Feyn}}
= \delta^{ab}\, D_{\mu\nu}^{\mbox{\scriptsize Feyn}}(x,y)
= -{\Gamma(D/2-1) \over 4\pi^{D/2}}
\,g_{\mu\nu} \,\delta^{ab}\, [(x-y)^2]^{1-D/2} \,.\label{feynman}$$ Using the free field relations $ U(a,b) = 1 $ and $ F_{\mu\nu}=
\partial_\mu A_\nu - \partial_\nu A_\mu $ we get (for more details see Appendix \[appprop\]) $$\begin{aligned}
\lefteqn{\langle A^a_\mu(x) A^b_\nu(y) \rangle^0_{\mbox{\scriptsize
rad}}=} \nonumber\\&& = -{\Gamma(D/2-1) \over 4\pi^{D/2}}
\,\delta^{ab} \int\limits^1_0 \!\! ds \int\limits^1_0 \!\! dt\,
sx^\alpha \, ty^\beta \nonumber\\&& \phantom{=}\times \left(
g_{\mu\nu} \partial_\alpha^{sx}\partial_\beta^{ty} +
g_{\alpha\beta}\partial_\mu^{sx}\partial_\nu^{ty} -
g_{\alpha\nu}\partial_\mu^{sx}\partial_\beta^{ty} -
g_{\mu\beta}\partial_\alpha^{sx}\partial_\nu^{ty} \right)
[(sx-ty)^2]^{1-D/2} \nonumber\\ && = -{\Gamma(D/2-1) \over
4\pi^{D/2}} \,\delta^{ab} \, \bigg( g_{\mu\nu} [(x-y)^2]^{1-D/2}
\nonumber\\ && \phantom{mmm} -\partial_\mu^x \int\limits^1_0 \!\! ds
\, x_\nu\,[(sx-y)^2]^{1-D/2} -\partial_\nu^y \int\limits^1_0 \!\! dt
\, y_\mu \,[(x-ty)^2]^{1-D/2} \nonumber\\ && \phantom{mmm}
+\partial_\mu^x \partial_\nu^y \underbrace{\int\limits^1_0 \!\! ds
\int\limits^1_0 \!\! dt\, x\cdot y \,
[(sx-ty)^2]^{1-D/2}}_{\textstyle \sim {\textstyle 1\over
\textstyle 4-D}} \bigg) \,.
\label{divprop}\end{aligned}$$ Thus the radial gauge propagator is singular for arbitrary arguments $x$ and $y$ — with one remarkable exception: It is easy to see that it vanishes for $x=0$ or $y=0$. This is simply a consequence of our task to preserve the field strength formula (\[eq:fsformula\]) which forces the vector field to vanish at the origin (in general at the reference point $z$).
The observation that the radial gauge propagator as calculated here diverges in four-dimensional space raises the question, whether it is perhaps impossible to formulate a quantum theory in radial gauge. This would suggest that the radial gauge condition – in the form that facilitates the field strength formula – is inherently inconsistent (“unphysical”) in contrast to the general belief that it is “very physical” since it allows to express gauge variant quantities like the vector potential in terms of gauge invariant ones. To answer this question we have to understand where this divergence comes from. In the following we will see that for this purpose the complicated looking Wilson loop representation (\[propa\]) is much more useful than the field strength formula (\[profi\]). (Note, however, that the result for the free propagator (\[divprop\]) of course will be the same.)
It is well-known that Wilson loops need renormalization to make them well-defined (see e.g. [@korrad] and references therein). The expansion of an arbitrary Wilson loop $$W(C) = {1\over N} \, \mbox{tr}\,\left\langle {\cal P}\exp\left[
ig\oint_C dx^\mu A_\mu(x) \right] \right\rangle$$ in powers of the coupling constant is given by $$\begin{aligned}
\lefteqn{
W(C) = 1 + {1\over N} \sum\limits_{n=2}^\infty (ig)^n
\oint_C \! dx_1^{\mu_1} \ldots \oint_C \! dx_n^{\mu_n}}
\nonumber \\ && \times
\Theta_C(x_1 > \cdots > x_n) \,\mbox{tr}\,
G_{\mu_1 \ldots \mu_n}(x_1,\ldots,x_n) \label{ptexp}\end{aligned}$$ where $\Theta_C(x_1 > \cdots > x_n)$ orders the points $x_1,\ldots,x_n$ along the contour $C$ and $$G_{\mu_1 \ldots \mu_n}(x_1,\ldots,x_n) :=
\left\langle A_{\mu_1}(x_1) \cdots A_{\mu_n}(x_n) \right\rangle$$ are the Green functions.
In general Wilson loops show ultraviolet singularities in any order of the coupling constant. If the contour $C$ is smooth (i.e. differentiable) and simple (i.e. without self-intersections) the conventional charge and wave-function renormalization — denoted by ${\cal R}$ in the following — is sufficient to make $W(C)$ finite. We refer to [@regul] for more details about renormalization of regular (smooth and simple) loops.
In our example we must apply the renormalization operation ${\cal R}$ to $W_1$ as given in (\[w1def\]). This yields $$\tilde W_1(x,x',y,y';g_R,\mu,D) = {\cal R} W_1(x,x',y,y';g,D)
\label{roper}$$ where $W_1(x,x',y,y';g,D)$ is a regularized expression calculated in $D$ dimensions and $\mu$ is a subtraction point introduced by the renormalization procedure ${\cal R}$. For the purpose of the present work the only important relation is $$g_R = \mu^{(D-4)/2} g + o(g^3) \,. \label{grug}$$
While the operation ${\cal R}$ is sufficient to make regular loops well-defined, new divergences appear if the contour $C$ has cusps or self-intersections. The renormalization properties of such loops are discussed in [@brandt] and [@brandt2]. While the singularities of regular loops appear at the two-loop level (order $g^4$ in (\[ptexp\])) cusps and cross points cause divergences even in leading (non-trivial) order $g^2$.
Since $W_1$ is indeed plagued by cusps and self-intersections a second renormalization operation must be carried out to get a renormalized expression $W_1^R$ from the bare one $W_1$: According to [@brandt] each cusp is multiplicatively renormalizable with a renormalization factor $Z$ depending on the cusp angle. In our case we have four cusps with angles $$\begin{aligned}
\alpha &:=& \angle (x-x',-x) \,, \\
\alpha' &:=& \angle (x',x-x') \,, \\
\beta &:=& \angle (y-y',-y) \,, \\
\beta' &:=& \angle (y',y-y') \,.\end{aligned}$$ The cross point at the origin introduces a mixing between $W_1$ and $$\begin{aligned}
\lefteqn{W_2(x,x',y,y') :=}\nonumber\\
&&\left\langle
{1\over N} \,\mbox{tr}\left[ U(0,x)\,U(x,x')\,U(x',0)\right] \,
{1\over N} \,\mbox{tr}\left[ U(0,y)\,U(y,y')\,U(y',0)\right]
\right\rangle \,.\end{aligned}$$ Again the divergences appearing here are functions of the angles $$\left.
\begin{array}{lcl}
\gamma_{xx'} &:=& \angle (-x,x') \\
\gamma_{yy'} &:=& \angle (-y,y') \\
\gamma_{xy} &:=& \angle (-x,-y) \\
\gamma_{x'y'} &:=& \angle (x',y') \\
\gamma_{x'y} &:=& \angle (x',-y) \\
\gamma_{xy'} &:=& \angle (-x,y')
\end{array}
\right\} \vec\gamma \,.$$ The renormalized Wilson loop $W_1^R$ is given by $$\begin{aligned}
\lefteqn{W_1^R(x,x',y,y';g_R,\mu,
\bar C_\alpha,\bar C_{\alpha'},\bar C_\beta,\bar C_{\beta'},
\bar C_{\vec\gamma})}
\nonumber\\ &&
= \lim_{D\to 4}
\, Z(\bar C_\alpha,g_R,\mu;D) \, Z(\bar C_{\alpha'},g_R,\mu;D)
\, Z(\bar C_\beta,g_R,\mu;D) \, Z(\bar C_{\beta'},g_R,\mu;D)
\nonumber\\ && \phantom{=\lim}
\times \left[ Z_{11}(\bar C_{\vec\gamma},g_R,\mu;D) \,
\tilde W_1(x,x',y,y';g_R,\mu,D)
\right. \nonumber\\ && \phantom{=\lim \times [} \left.
{}+Z_{12}(\bar C_{\vec\gamma},g_R,\mu;D) \,
\tilde W_2(x,x',y,y';g_R,\mu,D)
\right]
\nonumber\\ &&
=: \lim_{D\to 4} \bar W_1(x,x',y,y';g_R,\mu,
\bar C_\alpha,\bar C_{\alpha'},\bar C_\beta,\bar C_{\beta'},
\bar C_{\vec\gamma};D) \label{barw}\end{aligned}$$ where the second renormalization procedure introduces new subtraction points $\bar C_\sigma$ (c.f. [@korrad] and [@brandt] for more details). Of course different renormalization procedures are possible and so the $Z$ factors are not unique. We will return to this point in Section \[sec:radren\] where we specify a renormalization operation which is appropriate for our purposes.
The observation that Wilson loops with cusps and/or cross points show additional divergences has an important consequence for our radial gauge propagator as given in (\[propa\]): Even the free propagator needs renormalization! This provides a natural explanation for the fact that a naive calculation of this object yields an ultraviolet divergent result [@moda]. Note that the usual divergences of Wilson loops which are removed by ${\cal R}$, like e.g. vertex divergences, appear at $o(g^4)$ and thus do not contribute to the free part of the radial gauge propagator, while the cusp singularities indeed contribute since they appear at $o(g^2)$ and affect the free field case due to the factor $1/g^2$ in (\[propa\]).
Now we are able to answer the question whether the radial gauge is “unphysical” or “very physical”. It is just its intimate relation to physical, i.e. gauge invariant, quantities which makes the gauge propagator — even the free one — divergent. One might cast the answer in the following form: [*The propagator diverges because of — and not contrary to — the fact that the radial gauge is “very physical”*]{}.
Consequently the next questions are:
- Is there any use for a divergent expression for the free propagator? Especially: Can we use it to perform (dimensionally regularized) loop calculations?
- Can one find a renormalization program which yields a finite radial gauge propagator?
In the next Section we will perform a one loop calculation of a Wilson loop using the radial gauge propagator (\[divprop\]) and compare the dimensionally regularized result with a calculation in Feynman gauge.
In Section \[sec:radren\] we will explicitly demonstrate that the renormalization program for link operators carries over and allows to derive a finite result for the radial propagator and contrast its properties and use to the regularized version.
Calculating a Wilson Loop in Radial Gauge {#sec:radcalc}
=========================================
We choose the path $$\ell : z(\sigma) = \left\{
\begin{array}{lclcl}
\sigma x &,& \sigma\in [0,1] &,& x\in{\mathbb R}^D \\
w(\sigma -1) &,& \sigma\in [1,2] &,& w(0)=x,\, w(1)=y \\
(3-\sigma)\,y &,& \sigma\in [2,3] &,& y\in{\mathbb R}^D
\end{array}
\right. \label{looppath}$$ It is shown in Fig. \[fig:drop\]. The line $w(\sigma-1)$ is supposed to be an arbitrary curve connecting $x$ and $y$.
First we will perform the calculation of this Wilson loop in Feynman gauge. Using (\[feynman\]) we get in leading order of the coupling constant $$\begin{aligned}
W(\ell) &=&
{1\over N} \, \mbox{tr}\,\left\langle {\cal P}\exp\left[
ig\oint_\ell dz^\mu A_\mu(z) \right] \right\rangle
\nonumber \\ &\approx&
1 + (ig)^2 {N^2 -1\over 2N}
\int\limits^3_0 \!\! d\sigma \!\int\limits^3_0 \!\! d\tau \,
\Theta(\sigma-\tau)\,\dot z^\mu(\sigma) \, \dot z^\nu(\tau) \,
D_{\mu\nu}^{\mbox{\scriptsize Feyn}}(z(\sigma),z(\tau))
\nonumber \\ &=&
1 + (ig)^2 {N^2 -1\over 2N} {1\over 2}
\underbrace{\int\limits^3_0 \!\! d\sigma \!\int\limits^3_0 \!\! d\tau \,
\dot z^\mu(\sigma) \, \dot z^\nu(\tau) \,
D_{\mu\nu}^{\mbox{\scriptsize Feyn}}(z(\sigma),z(\tau))}
_{\textstyle =: I_f} \,. \label{feynres}\end{aligned}$$ To get rid of the $\Theta$-function we have exploited the symmetry property of two-point Green functions $$D_{\mu\nu}^{\mbox{\scriptsize Feyn}}(x,y)
= D_{\nu\mu}^{\mbox{\scriptsize Feyn}}(y,x) \,.$$ Decomposing the contour $\ell$ according to (\[looppath\]) we find that the Feynman propagator in (\[feynres\])) connects each segment of $\ell$ with itself and with all the other segments. Thus $I_f$ is given by $$I_f = \sum\limits_{A = 1}^3 \sum\limits_{B = 1}^3\,(A,B)$$ where $(A,B)$ denotes the contribution with propagators connecting loop segments $A$ and $B$ (c.f. Fig. \[fig:drop\]), e.g. $$\begin{aligned}
(1,2) &=&
\int\limits^1_0 \!\! d\sigma \!\int\limits^1_0 \!\! d\tau \,
x^\mu \, \dot w^\nu(\tau) \,
D_{\mu\nu}^{\mbox{\scriptsize Feyn}}(\sigma x,w(\tau))
\nonumber\\ &=&
-{\Gamma(D/2-1) \over 4\pi^{D/2}}
\int\limits^1_0 \!\! d\sigma \!\int\limits^1_0 \!\! d\tau \,
x^\mu \, \dot w_\mu(\tau) \, [(\sigma x-w(\tau))^2]^{1-D/2} \,.\end{aligned}$$
Next we will evaluate the same Wilson loop in radial gauge. Clearly the first and the third part of the path do not contribute if the radial gauge condition $x_\mu A^\mu(x) = 0$ holds. We insert the free propagator $$\begin{aligned}
\langle A^a_\mu(x) A^b_\nu(y) \rangle^0_{\mbox{\scriptsize rad}}
=: \delta^{ab} D^0_{\mu\nu}(x,y) \end{aligned}$$ from (\[divprop\]) into $$\begin{aligned}
W(\ell) &=&
{1\over N}\, \mbox{tr}\,\left\langle {\cal P}\exp\left[
ig\int\limits_0^1 \!\! d\sigma \,\dot w_\mu(\sigma) A^\mu(w(\sigma))
\right]\right\rangle
\nonumber\\ &\approx&
1 + (i g)^2 {N^2 -1\over 2N}
{1\over 2}
\underbrace{\int\limits^1_0 \!\! d\sigma \!
\int\limits^1_0 \!\! d\tau \,
\dot w^\mu(\sigma)\, \dot w^\nu(\tau) \,
D_{\mu\nu}^0(w(\sigma),w(\tau))}
_{\textstyle =: I_r}
\label{wilir}\end{aligned}$$ and observe that $$\dot w_\mu(\sigma) \,\partial^\mu_{w(\sigma)} = {d\over d\sigma} \,.$$ Thus the integral in (\[wilir\]) reduces to $$\begin{aligned}
\lefteqn{I_r =
-{\Gamma(D/2-1) \over 4\pi^{D/2}}
\Bigg[
\int\limits^1_0 \!\! d\sigma \int\limits^1_0 \!\! d\tau \,
\dot w_\mu(\sigma)\, \dot w^\mu(\tau) \,
[(w(\sigma)-w(\tau))^2]^{1-D/2}}
\nonumber\\&& \phantom{-{\Gamma(D/2-1) \over 4\pi^{D/2}}}
{}+ \int\limits^1_0 \!\! ds \int\limits^1_0 \!\! dt\,
\Big(
w_\mu(1)\,w^\mu(1)\,[(sw(1)-tw(1))^2]^{1-D/2}
\nonumber\\&&
\phantom{-MM(D/2-1)\int\limits^1_0\!\! ds\int\limits^1_0\!\!dt\,}
{}+ w_\mu(0)\,w^\mu(0)\,[(sw(0)-tw(0))^2]^{1-D/2}
\nonumber\\&&
\phantom{-MM(D/2-1)\int\limits^1_0\!\! ds\int\limits^1_0\!\!dt\,}
{}-w_\mu(1)\,w^\mu(0)\,[(sw(1)-tw(0))^2]^{1-D/2}
\nonumber\\&&
\phantom{-MM(D/2-1)\int\limits^1_0\!\! ds\int\limits^1_0\!\!dt\,}
{}-w_\mu(0)\,w^\mu(1)\,[(sw(0)-tw(1))^2]^{1-D/2}
\Big)
\nonumber\\&& \phantom{-{\Gamma(D/2-1) \over 4\pi^{D/2}}}
{}- \int\limits^1_0 \!\! ds \int\limits^1_0 \!\! d\tau \,
\dot w_\mu(\tau)
\Big(
w^\mu(1) [(sw(1)-w(\tau))^2]^{1-D/2}
\nonumber\\&&
\phantom{-MMMM(D/2-1)\int\limits^1_0\!\! ds\int\limits^1_0\!\!dt\,}
{}- w^\mu(0) [(sw(0)-w(\tau))^2]^{1-D/2}
\Big)
\nonumber\\&& \phantom{-{\Gamma(D/2-1) \over 4\pi^{D/2}}}
{}- \int\limits^1_0 \!\! dt \int\limits^1_0 \!\! d\sigma \,
\dot w_\mu(\sigma)
\Big(
w^\mu(1) [(w(\sigma)-tw(1))^2]^{1-D/2}
\nonumber\\&&
\phantom{-MMMM(D/2-1)\int\limits^1_0\!\! ds\int\limits^1_0\!\!dt\,}
{}- w^\mu(0) [(w(\sigma)-tw(0))^2]^{1-D/2}
\Big)
\Bigg]
\label{lengthy}\\&&
= (2,2) + (3,3) + (1,1) + (3,1) + (1,3) + (3,2) + (1,2) + (2,3) +
(2,1) \,. \nonumber\end{aligned}$$ A careful analysis of (\[lengthy\]) shows that it exactly coincides with the Feynman gauge calculation. This is expressed in the last line where we have denoted which parts of the loop are connected by the Feynman gauge propagator to reproduce (\[lengthy\]) term by term. Thus using the radial gauge propagator as given in (\[divprop\]) yields the same result as the calculation in Feynman gauge. Finally this regularized expression has to be renormalized. This can be performed without any problems according to [@brandt]. Since we are not interested in the Wilson loop itself but in the comparison of the results obtained in radial and Feynman gauge, we will not calculate the renormalized expression for $W(\ell)$.
However a qualitative discussion of the renormalization properties of $W(\ell)$ is illuminating. By construction $W(\ell)$ has at least a cusp at the origin. (Other cusps are possible at $x$ or $y$ or along the line parameterized by $w$, but are not important for our considerations.) To give the right behavior of the Wilson loop the calculation of $W(\ell)$ in an arbitrary gauge must reproduce the cusp singularity. Usually the parameter integrals in the vicinity of the cusp do the job. For gauge choices where the propagator do not vanish in the vicinity of the origin this is automatically achieved. Let us assume for a moment that it is possible to construct a [*finite*]{} radial gauge propagator obeying the field strength formula (\[eq:fsformula\]) and therefore have trivial gauge factors along radial lines. Of course this is nothing but saying that there are no contributions form parts 1 and 3 of the loop, i.e. in the vicinity of the origin. Since the propagator is assumed to be finite, there are no singular integrals corresponding to the cusp at the origin. Thus a finite radial gauge propagator cannot reproduce the correct behavior of the Wilson loop. In turn we conclude that [*a singular radial gauge propagator is mandatory*]{} to get the right renormalization properties of Wilson loops.
However, as we will demonstrate in the next Section, the renormalization procedure for Wilson loops can be used to devise a consistent renormalization program for the radial gauges considered here. We will apply it to write down a finite version of the free radial propagator. The generalization to higher orders in perturbation theory is straightforward. According to our considerations given above we shall show that the renormalized, thus finite version of the free radial propagator is not suitable as an input to perturbative calculations.
The Renormalized Free Propagator {#sec:radren}
================================
We define the renormalized radial gauge propagator by $$\langle A^a_\mu(x) A^b_\nu(y)
\rangle_R^{\mbox{\scriptsize singlet}} :=
\lim_{D\to 4} \delta^{ab} {2N\over N^2 -1}
{1\over (ig_R)^2} \mu^{D-4}
\lim_{x'\to x \atop y'\to y}\partial_\mu^x \partial_\nu^y \,
\bar W_1(x,x',y,y';D) \label{propren}$$ where we have suppressed most of the other variables on which $\bar
W_1$ depends (see (\[barw\])).
From now on we will concentrate on the calculation of the renormalized free propagator $\langle A^a_\mu(x) A^b_\nu(y) \rangle_R^0$. The details of the renormalization program are presented in Appendix \[app:renprog\]. Of course the procedure is closely connected to the renormalization of cusp singularities of Wilson loops. The result is $$\begin{aligned}
\lefteqn{\langle A^a_\mu(x) A^b_\nu(y) \rangle_R^0 =}
\\ &&
=\lim_{D\to 4} \left(
\delta^{ab} \mu^{D-4} \partial_\mu^x \partial_\nu^y
\left(
{1\over 4\pi^2} {1\over 4-D} (\pi-\gamma_{xy})\cot\gamma_{xy}
\right)
+\langle A^a_\mu(x) A^b_\nu(y) \rangle^0_{\mbox{\scriptsize rad}}
\right)
\nonumber\end{aligned}$$
Before discussing some properties of the renormalized free propagator we shall show that the counter term $$C_{\mu\nu}^{ab}(x,y) :=
\delta^{ab} \mu^{D-4} \partial_\mu^x \partial_\nu^y
\left(
{1\over 4\pi^2} {1\over 4-D} (\pi-\gamma_{xy})\cot\gamma_{xy}
\right)$$ exactly cancels the divergence of the propagator (\[divprop\]), i.e. that $\langle A^a_\mu(x) A^b_\nu(y) \rangle_R^0$ really is finite. To this end we use some technical results derived in Appendix \[siint\]. The divergent part of the propagator (\[divprop\]) is given by $$U_{\mu\nu}^{ab}(x,y):=-{\Gamma(D/2-1) \over 4\pi^{D/2}} \,\delta^{ab} \,
\partial_\mu^x \partial_\nu^y
\int\limits^1_0 \!\! ds \int\limits^1_0 \!\! dt\, x\cdot y \,
[(sx-ty)^2]^{1-D/2} \,.$$ Using (\[appi2\]) and (\[i2neq\]) we find $$\begin{aligned}
U_{\mu\nu}^{ab}(x,y) &=&
{\Gamma(D/2-1) \over 4\pi^{D/2}} \,\delta^{ab} \,
\partial_\mu^x \partial_\nu^y \, I_2(x,-y)
\nonumber\\ &=&
{\Gamma(D/2-1) \over 4\pi^{D/2}} \,\delta^{ab} \,
\partial_\mu^x \partial_\nu^y
\left({1\over 4-D} \,(\pi-\gamma_{xy}) \cot(\pi-\gamma_{xy}) \;
+\mbox{finite} \right)
\nonumber\\ &=&
-{1 \over 4\pi^2} \,\delta^{ab} \,
\partial_\mu^x \partial_\nu^y
\left({1\over 4-D} \,(\pi-\gamma_{xy}) \cot\gamma_{xy} \right) \;
+\mbox{finite}\end{aligned}$$ and thus $$U_{\mu\nu}^{ab}(x,y) + C_{\mu\nu}^{ab}(x,y) = \mbox{finite} \,.$$
Note that if one tries to guess a finite expression like $\langle
A^a_\mu(x) A^b_\nu(y) \rangle_R^0$ one would have to introduce a scale $\mu$ by hand without interpretation. In our derivation this scale appears naturally as the typical renormalization scale of the ${\cal
R}$ operation.
The counter term $C_{\mu\nu}^{ab}(x,y)$ has some interesting properties. It is symmetric with respect to an exchange of all variables and it obeys the gauge condition $$x^\mu C_{\mu\nu}^{ab}(x,y) = 0 = C_{\mu\nu}^{ab}(x,y) \,y^\nu \,.
\label{courad}$$ Thus $\langle A^a_\mu(x) A^b_\nu(y) \rangle_R^0$ is finite in the limit $D\to 4$ but still can be interpreted as a gluonic two-point function which fulfills the radial gauge condition $$x^\mu \,\langle A^a_\mu(x) A^b_\nu(y) \rangle_R^0 =0 \,.$$ However the counter term $C_{\mu\nu}^{ab}(x,y)$ and thus also $\langle A^a_\mu(x) A^b_\nu(y) \rangle_R^0$ is ill-defined at the origin and hence conflicts with the field strength formula (\[eq:fsformula\]). Note that the regularized propagator in contrast to the renormalized propagator is well defined and vanishes if one of its arguments approaches zero, as pointed out after eq. (\[divprop\]). We therefore conclude that we may use the regularized propagator in perturbative calculations and can be ensured to preserve relations like the field strength formula or eqs. (\[eq:fistr\]) or (\[eq:propdef\]) throughout the calculation. Although the counterterms are not well defined at the reference point itself – a property the radial gauge propagator simply inherits from renormalizing the cusp singularity of the underlying Wilson line – physical (gauge invariant) quantities are not affected, they are rendered finite and unambiguous.
Summary and Outlook {#sec:radsum}
===================
In this article we have shown how to calculate the radial gauge propagator in a $D$-dimensional space using Wilson loops. As discovered in [@moda] the free propagator diverges in four-dimensional space. We were able to explain this singular behavior by studying the properties of associated Wilson loops. Furthermore we have shown that the free propagator, in spite of being divergent in four dimensions, can be used for perturbative calculations in a (dimensionally) regularized framework and that the result for a gauge invariant quantity agrees with the calculation in Feynman gauge. Finally we have presented a renormalization procedure for the radial gauge propagator and calculated the explicit form of the renormalized free propagator. We have pointed out that any version of the radial propagator which is finite in four-dimensional space at least cannot reproduce the correct renormalization properties of Wilson loops with cusps at the reference point $z$.
It is instructive to compare the radial gauge propagators as presented here with other approaches: As discussed in Section \[sec:radgaugecond\] the radial gauge condition (\[eq:fsgaugecond\]) does not completely fix the gauge degrees of freedom. Thus the field strength formula $$A_\mu(x) = \int\limits^1_0 \!\! ds \, s x^\nu F_{\nu\mu}(sx) \label{fistrag}$$ is not the only solution of the system of equations[^4] $$\left\{\begin{array}{l}
x_\mu A^\mu(x) =0 \,, \\
F_{\mu\nu}(x) =
\partial^x_\mu A_\nu(x) - \partial^x_\nu A_\mu(x) \,.
\end{array} \right.$$ One might add a function [@moda] $$A_\mu^0(x) = \partial^x_\mu f(x)$$ to (\[fistrag\]) where $f$ is an arbitrary homogeneous function of degree 0. However any $A_\mu^0(x)$ added in in order to modify (\[fistrag\]) is necessarily singular at the origin. Hence regularity at the origin may be used as a uniqueness condition [@cron]. If we relax this boundary condition other solutions are possible, e.g. $$\bar A_\mu(x) = -\int\limits^\infty_1 \!\! ds \, s x^\nu F_{\nu\mu}(sx)
\label{fistrinf}$$ where we must assume that the field strength vanishes at infinity. While (\[fistrag\]) is the only solution which is regular at the origin, (\[fistrinf\]) is regular at infinity. Ignoring boundary conditions for the moment one can construct a radial gauge propagator by [@menot] $${1\over 2} \left( G_{\mu\nu}(x,y) + G_{\nu\mu}(y,x) \right)
\label{menprop}$$ with $$G_{\mu\nu}(x,y) :=
-\int\limits^1_0 \!\! ds \, s x^\alpha
\int\limits^\infty_1 \!\! dt \, t y^\beta
\left\langle F_{\alpha\mu}(sx) F_{\beta\nu}(ty) \right\rangle \,.$$ It turns out that this propagator is finite in four dimensions. However the price one has to pay is that boundary conditions are ignored and thus the object “lives” in the restricted space ${\mathbb R}^4 \setminus\{0\}$ and not in ${\mathbb R}^4$ anymore. In our approach we insist on the field strength formula (\[fistrag\]) widely used in operator product expansions [@shif] and on the regular behavior of vector potentials at the origin [@cron]. One might use the propagator (\[menprop\]) to calculate the $g^2$-contribution to the Wilson loop on the contour (\[looppath\]). It is easy to check that the result differs from the one obtained in (\[wilir\]), (\[lengthy\]). Clearly this is due to the fact that (\[menprop\]) is ill-defined at the origin.
In the above, all calculations were performed in Euclidean space. In Minkowski space Wilson loops show additional divergences if part of the contour coincides with the light cone [@korkor]. Thus we expect the appearance of new singularities also for the radial propagator, at least if one or both of its arguments are light-like. Further investigation is required to work out the properties of the radial gauge propagator in Minkowski space.
To formulate perturbation theory in a specific gauge the knowledge of the correct free propagator is only the first step. In addition one has to check the decoupling of Faddeev-Popov ghosts in radial gauge which is suggested by the algebraic nature of the gauge condition. However the still continuing discussion about temporal and axial gauges might serve as a warning that the decoupling of ghosts for algebraic gauge conditions is far from being trivial (c.f. [@cheng], [@leroy] and references therein). To prove (or disprove) the decoupling of ghosts in radial gauge we expect that our Wilson loop representation of the propagator is of great advantage since it yields the possibility to calculate higher loop contributions in two distinct ways: On the one hand one might use the Wilson loop representation to calculate the full radial propagator up to an arbitrary order in the coupling constant. The appropriate Wilson loop can be calculated in any gauge, e.g. in a covariant gauge. On the other hand the radial propagator might be calculated according to Feynman rules. Since the results should coincide this might serve as a check for the validity and completeness of a set of radial gauge Feynman rules.
[**Acknowledgments:** ]{}
HW wants to thank Alex Kovner for his invaluable patience in his role as a testing ground of new ideas. SL thanks Professor Ulrich Heinz for valuable discussions and support. During this research SL was supported in part by Deutsche Forschungsgemeinschaft and Bundesministerium für Bildung, Wissenschaft, Forschung und Technologie. HW was supported by the U.S. Department of Energy under grants No. DOE Nuclear DE–FG02–87ER–40328 and by the Alexander von Humboldt Foundation through their Feodor Lynen program.
Derivation of the Free Radial Propagator {#appprop}
========================================
The free radial propagator derived form the field strength formula shows a divergence in $D=4$, as already indicated in section \[sec:radprop\], eq. (\[divprop\]). Here we give the details of the algebra leading to this conclusion.
The following relations summarize the steps carried out in the calculation below: $$\begin{aligned}
&& \hskip 2cm
x_\mu \partial_x^\mu = \vert x\vert \,\partial_{\vert x\vert} \,,
\\
T_{\mu\nu}(x,y) &:=&
x^\alpha y^\beta \left(
g_{\mu\nu} \partial_\alpha^{x}\partial_\beta^{y}
+ g_{\alpha\beta}\partial_\mu^{x}\partial_\nu^{y}
- g_{\alpha\nu}\partial_\mu^{x}\partial_\beta^{y}
- g_{\mu\beta}\partial_\alpha^{x}\partial_\nu^{y}
\right)
\\ &=&
g_{\mu\nu} \partial_{\vert x\vert} \partial_{\vert y\vert}
\vert x\vert \,\vert y\vert
-\partial_\mu^x \,x_\nu \,\partial_{\vert y\vert} \vert y\vert
-\partial_\nu^y \,y_\mu \,\partial_{\vert x\vert} \vert x\vert
+\partial_\mu^x \partial_\nu^y \,x\cdot y \,. \nonumber $$ Introducing $\hat x :=x/\vert x\vert$ and $u=s\vert x\vert$, we have for arbitrary $f$: $$\begin{aligned}
&& \hskip 2cm sx_\alpha \,\partial_\beta^{sx} f(sx) = x_\alpha \,\partial_\beta^x
f(sx) \,,
\\ &&
\partial_{\vert x\vert}\int\limits_0^1 \!\! ds \, \vert x\vert f(sx)
= \partial_{\vert x\vert} \int\limits_0^1 \!\! ds \, \vert x\vert
f(s\vert x\vert \hat x) = \partial_{\vert x\vert}
\int\limits_0^{\vert x\vert} \!\! du \, f(u \hat x) = f(\vert x\vert
\hat x) = f(x) \,. \nonumber \\ &&\end{aligned}$$
We get $$\begin{aligned}
\lefteqn{{}-{4\pi^{D/2} \over \Gamma(D/2-1)} \,
\langle A^a_\mu(x) A^b_\nu(y) \rangle^0_{\mbox{\scriptsize rad}}=}
\nonumber\\&&
= \delta^{ab}
\int\limits^1_0 \!\! ds \int\limits^1_0 \!\! dt\,
T_{\mu\nu}(sx,ty) \,[(sx-ty)^2]^{1-D/2}
\nonumber\\ &&
= \delta^{ab} \,T_{\mu\nu}(x,y)
\int\limits^1_0 \!\! ds \int\limits^1_0 \!\! dt\,
[(sx-ty)^2]^{1-D/2}
\nonumber\\ &&
= \delta^{ab}
\left(
g_{\mu\nu} \partial_{\vert x\vert} \partial_{\vert y\vert}
\vert x\vert \,\vert y\vert
-\partial_\mu^x \,x_\nu \,\partial_{\vert y\vert} \vert y\vert
-\partial_\nu^y \,y_\mu \,\partial_{\vert x\vert} \vert x\vert
+\partial_\mu^x \partial_\nu^y \,x\cdot y
\right)
\nonumber\\&& \phantom{=}\times
\int\limits^1_0 \!\! ds \int\limits^1_0 \!\! dt\,
[(sx-ty)^2]^{1-D/2}
\nonumber\\ &&
= \delta^{ab}
\bigg(
g_{\mu\nu} [(x-y)^2]^{1-D/2}
\nonumber\\ && \phantom{mmm}
-\partial_\mu^x \int\limits^1_0 \!\! ds \,x_\nu \, [(sx-y)^2]^{1-D/2}
-\partial_\nu^y \int\limits^1_0 \!\! dt \,y_\mu \, [(x-ty)^2]^{1-D/2}
\nonumber\\ && \phantom{mmm}
+\partial_\mu^x \partial_\nu^y
\underbrace{\int\limits^1_0 \!\! ds \int\limits^1_0 \!\! dt \, x\cdot y \,
[(sx-ty)^2]^{1-D/2}}_{\textstyle \sim {\textstyle 1\over \textstyle 4-D}}
\bigg) \,.\end{aligned}$$ The divergent part of the double integral in the last line can be found in Appendix \[siint\]. At the moment however the exact form of the divergence is not important.
Renormalization Program for the Free Propagator {#app:renprog}
===============================================
In Section \[sec:radren\] we discussed the effect of renormalization on the free radial propagator. Here derive in detail the appropriate renormalization procedure starting form the renormalization properties of Wilson lines. Only a few of the many possible renormalization constants will contribute to the final result.
Since in the relation between the propagator and the appropriate Wilson loop (\[propa\]) a factor $1/g^2$ is involved all quantities especially all the $Z$’s and $W$’s of eq. (\[barw\]) have to be calculated up to $o(g_R^2)$. We have $$\begin{aligned}
\tilde W_i &=& 1 + (ig_R)^2 \delta\tilde W_i + o(g_R^4) \qquad
(i=1,2)\,, \\ Z &=& 1 + (ig_R)^2 \delta Z + o(g_R^4) \,,\\ Z_{11}
&=& 1 + (ig_R)^2 \delta Z_{11} + o(g_R^4) \,,\\ Z_{12} &=& 0 +
(ig_R)^2 \delta Z_{12} + o(g_R^4)\end{aligned}$$
yielding $$\begin{aligned}
\lefteqn{\langle A^a_\mu(x) A^b_\nu(y) \rangle_R^0 =
\lim_{D\to 4} \delta^{ab} {2N\over N^2 -1} \mu^{D-4}}
\label{freepr} \\ && \times
\lim_{x'\to x \atop y'\to y}\partial_\mu^x \partial_\nu^y \,
\left[
\begin{array}{c}
\delta Z(\bar C_\alpha) +\delta Z(\bar C_{\alpha'})
+\delta Z(\bar C_\beta) +\delta Z(\bar C_{\beta'}) \\
\quad
+\delta Z_{11} +\delta Z_{12} +\delta\tilde W_1
\end{array}
\right] \,.
\nonumber \end{aligned}$$ Using the fact that up to $o(g_R^2)$ the two quantities $W_1$ and $\tilde W_1$ are essentially the same[^5] we find $$\begin{aligned}
\lefteqn{\lim_{D\to 4} \delta^{ab} {2N\over N^2 -1} \mu^{D-4}
\lim_{x'\to x \atop y'\to y}\partial_\mu^x \partial_\nu^y \,
\delta\tilde W_1}
\nonumber\\ &&
= \lim_{D\to 4} \delta^{ab} {2N\over N^2 -1}
{1\over (ig)^2}\lim_{x'\to x \atop y'\to y}\partial_\mu^x
\partial_\nu^y \,
\left(1+(ig)^2 \delta W_1 \right)
\nonumber\\ &&
= \lim_{D\to 4} \delta^{ab} {2N\over N^2 -1}
{1\over (ig)^2}\lim_{x'\to x \atop y'\to y}\partial_\mu^x
\partial_\nu^y \,
W_1 \,\Big\vert_{g=0}
\nonumber\\ &&
=\lim_{D\to 4}
\langle A^a_\mu(x) A^b_\nu(y) \rangle^0_{\mbox{\scriptsize rad}} \,.
\label{naipa}\end{aligned}$$ To get the $\delta Z$’s we must calculate $\delta\tilde W_1$ which is straightforward using (\[w1def\]) and (\[roper\]). We only need the Feynman propagator (\[feynman\]) to get $$\begin{aligned}
\lefteqn{\delta\tilde W_1 =
-\mu^{4-D} {N^2 -1\over 2N} {\Gamma(D/2-1) \over 4\pi^{D/2}}}
\nonumber\\ &&
\left[
\left(
\vert x' \vert^{4-D} +\vert x-x' \vert^{4-D} + \vert x \vert^{4-D} +
\vert y' \vert^{4-D} +\vert y-y' \vert^{4-D} + \vert y \vert^{4-D}
\right) I_1
\right.\nonumber\\ &&
{}+ I_2(x',x-x') + I_2(x-x',-x) + I_2(-x,x')
\nonumber \\ &&
{}+ I_2(y',y-y') + I_2(y-y',-y) + I_2(-y,y')
\nonumber \\ &&
{}- I_2(x',-y') + I_2(x',-y) + I_2(y',-x) - I_2(x,-y)
\nonumber \\ &&
{}- I_3(y',-x',y-y') + I_3(y',-x,y-y') - I_3(x'-y',x-x',y'-y)
\nonumber \\ &&
{}- I_3(x',-y',x-x') + I_3(x',-y,x-x')
\Big] \label{dw1l}\end{aligned}$$ with $$I_1:= \int\limits_0^1 \!\! ds \int\limits_0^1 \!\! dt \,\Theta(s-t) \,
{1 \over [(s-t)^2]^{D/2-1}} \,,$$ $$I_2(p,q):= \int\limits_0^1 \!\! ds \int\limits_0^1 \!\! dt \,
{p\cdot q \over [(sp+tq)^2]^{D/2-1}} \,,$$ and $$I_3(m,p,q):= \int\limits_0^1 \!\! ds \int\limits_0^1 \!\! dt \,
{p\cdot q \over [(m+sp+tq)^2]^{D/2-1}} \,.$$ In the following we are interested only in the divergent parts of these integrals. The integrals $I_1$ and $I_2$ are calculated in Appendix \[siint\]. The results are $$I_1 = -{1\over 4-D} + \mbox{finite} \label{inti1}$$ and $$I_2(p,q) = {1\over 4-D} \, \gamma \cot\gamma + \mbox{finite}
\label{inti2}$$ where $\gamma$ is the angle between $p$ and $q$. The integral $I_3$ is finite as long as $m\neq 0$.
To specify the renormalization factors $Z$ we choose the minimal subtraction scheme $K^{\mbox{\scriptsize MS}}_\gamma$ as described in [@korrad]. In dimensional regularization all the divergences are given by sums of pole terms. We define every $Z$ factor to be given just by the respective sum. The important property of this renormalization scheme is that the $Z$ factors depend on the angles only and not on the length of the loop or of any part of the loop. Using (\[inti1\] and (\[inti2\])) the $Z$ factors can be read off from (\[dw1l\]) (c.f. [@brandt2]): $$\begin{aligned}
\delta Z(\bar C_\alpha)
&=&
{N^2 -1\over 2N}{1\over 4\pi^2} {1\over 4-D} \,
(\alpha \cot\alpha -1) \,,
\\
\delta Z(\bar C_{\alpha'})
&=&
{N^2 -1\over 2N}{1\over 4\pi^2} {1\over 4-D} \,
(\alpha' \cot\alpha' -1) \,,
\\
\delta Z(\bar C_\beta)
&=&
{N^2 -1\over 2N}{1\over 4\pi^2} {1\over 4-D} \,
(\beta \cot\beta -1) \,,
\\
\delta Z(\bar C_{\beta'})
&=&
{N^2 -1\over 2N}{1\over 4\pi^2} {1\over 4-D} \,
(\beta' \cot\beta' -1) \,,
\\
\delta Z_{11}
&=&
{N^2 -1\over 2N}{1\over 4\pi^2} {1\over 4-D}
\left[
(\gamma_{x'y} \cot\gamma_{x'y} -1) +
(\gamma_{xy'} \cot\gamma_{xy'} -1)
\right] \,, \nonumber \\ &&
\\
\delta Z_{12}
&=&
{N^2 -1\over 2N}{1\over 4\pi^2} {1\over 4-D}
\left[
\gamma_{xx'} \cot\gamma_{xx'} + \gamma_{yy'} \cot\gamma_{yy'}
\right.\nonumber\\ && \left.
{}- (\pi-\gamma_{x'y'})\cot(\pi-\gamma_{x'y'})
- (\pi-\gamma_{xy})\cot(\pi-\gamma_{xy})
\right] \,.\end{aligned}$$ Now we exploit the fact that only one of the angles, namely $\gamma_{xy}$, depends on $x$ [*and*]{} $y$. All the other ones depend only on $x$ or $y$ separately, or on none of them. This simplifies (\[freepr\]) drastically: $$\begin{aligned}
\lefteqn{\langle A^a_\mu(x) A^b_\nu(y) \rangle_R^0 =
\lim_{D\to 4} \delta^{ab} {2N\over N^2 -1} \mu^{D-4}
\lim_{x'\to x \atop y'\to y}\partial_\mu^x \partial_\nu^y \,
\left( \delta Z_{12} +\delta\tilde W_1 \right)}
\\ &&
=\lim_{D\to 4} \left(
\delta^{ab} \mu^{D-4} \partial_\mu^x \partial_\nu^y
\left(
{1\over 4\pi^2} {1\over 4-D} (\pi-\gamma_{xy})\cot\gamma_{xy}
\right)
+\langle A^a_\mu(x) A^b_\nu(y) \rangle^0_{\mbox{\scriptsize rad}}
\right)
\nonumber\end{aligned}$$ where we have used (\[naipa\]) to get the last expression.
Some Important Integrals {#siint}
========================
The integrals $I_1$ and $I_2$ played an important part in the renormalization procedure of appendix \[app:renprog\] and determine the divergences of the naive free radial propagator introduced in section \[sec:radprop\]. They are discussed in detail below.
To calculate $$I_1:= \int\limits_0^1 \!\! ds \int\limits_0^1 \!\! dt \,\Theta(s-t) \,
{1 \over [(s-t)^2]^{D/2-1}}$$ we introduce the substitution $$g=s-t \quad,\quad h=s+t$$ to get $$\begin{aligned}
I_1 &=&
{1\over 2} \int\limits_0^1 \!\! dg \int\limits_g^{2-g} \!\! dh \, g^{2-D}
= \int\limits_0^1 \!\! dg \, (1-g) \, g^{2-D} \nonumber\\
&=& {\Gamma(2) \Gamma(3-D) \over \Gamma(5-D)} = {1\over (4-D)(3-D)} \,.
\label{i1erg}\end{aligned}$$
For the calculation of $$I_2(p,q):= \int\limits_0^1 \!\! ds \int\limits_0^1 \!\! dt \,
{p\cdot q \over [(sp+tq)^2]^{D/2-1}} \,. \label{appi2}$$ we have to distinguish the two cases $p \neq \alpha q$ where the only divergence that appears is for $s=t=0$ and $p = \alpha q$ with an additional divergence at $s=t\alpha$. Here we will only need the former.
As a first step it is useful to separate off the divergence at the origin by the substitution $$\lambda = s+t \quad,\quad x=s/\lambda \,.$$ This yields $$\begin{aligned}
I_2(p,q) &=& \left( \int\limits_0^{1/2} \!\! dx
\int\limits_0^{1/(1-x)} \!\!\!\! d\lambda \, + \int\limits_{1/2}^1
\!\! dx \int\limits_0^{1/x} \!\! d\lambda \right) \lambda^{3-D}
{p\cdot q \over [(xp+(1-x)q)^2]^{D/2-1}} \nonumber\\ &=&
\int\limits_0^{1/2} \!\! dx \,{(1-x)^{D-4} \over 4-D} \, {p\cdot q
\over [(xp+(1-x)q)^2]^{D/2-1}} \nonumber\\ && +\int\limits_{1/2}^1
\!\! dx \,{x^{D-4} \over 4-D} \, {p\cdot q \over
[(xp+(1-x)q)^2]^{D/2-1}} \label{xint}\,.\end{aligned}$$ As long as $p \neq -q$ holds there are no divergences in the $x$-integration since $$u(x):=xp+(1-x)q$$ never vanishes. We introduce the angle between $p$ and $q$ $$\cos\gamma := {p\cdot q \over \vert p \vert \, \vert q\vert }$$ and the substitution [@korrad] $$e^{2i\psi} = {x\vert p\vert +(1-x)\vert q\vert e^{i\gamma} \over
x\vert p\vert +(1-x)\vert q\vert e^{-i\gamma}} \label{psisub}\,.$$ Note that $\psi$ is nothing but the angle between $p$ and $u(x)$. To perform this substitution in (\[xint\]) we need $$x = \vert q\vert \sin(\gamma-\psi) /N(\psi) \quad,\quad
1-x = \vert p \vert \sin\psi / N(\psi) \,,$$ $$[u(x)]^2 = p^2 q^2 \sin^2\gamma /[N(\psi)]^2
\quad\mbox{and} \quad
{d\psi \over dx} =
- {[N(\psi)]^2 \over \vert p\vert\,\vert q\vert \sin\gamma }$$ with $$N(\psi) := \vert p \vert \sin\psi + \vert q\vert \sin(\gamma-\psi)
\,.$$ In addition it is useful to introduce $$\psi' :=\psi(x=1/2)$$ which is the angle between $p$ and $p+q$ (cf. Fig. \[fig:pqangl\]).
Using all that we end up with $$\begin{aligned}
\lefteqn{
I_2(p,q) = \int\limits_\gamma^{\psi'}\!\! d\psi \,
{-\vert p\vert\,\vert q\vert \sin\gamma \over N^2}
\left({\vert p \vert \sin\psi \over N}\right)^{D-4}
\left({N^2 \over p^2 q^2 \sin^2\gamma} \right)^{D/2-1} {p\cdot q \over 4-D}}
&&\nonumber\\
&& +\int\limits_{\psi'}^0\!\! d\psi \,
{-\vert p\vert\,\vert q\vert \sin\gamma \over N^2}
\left({ \vert q\vert \sin(\gamma-\psi) \over N}\right)^{D-4}
\left({N^2 \over p^2 q^2 \sin^2\gamma} \right)^{D/2-1} {p\cdot q \over 4-D}
\nonumber \\
&=& {- \cos\gamma \sin^{3-D}\gamma\over 4-D} \left( \vert
q\vert^{4-D} \int\limits_\gamma^{\psi'}\!\! d\psi \,
\sin^{D-4}\psi + \vert p\vert^{4-D} \int\limits_{\psi'}^0\!\!
d\psi \, \sin^{D-4}(\gamma-\psi) \right) \nonumber\\
&=& {- \cos\gamma \sin^{3-D}\gamma \over
4-D} \left( \vert q\vert^{4-D}
\int\limits_\gamma^{\psi'}\!\! d\psi \, \sin^{D-4}\psi + \vert
p\vert^{4-D} \int\limits_\gamma^{\gamma-\psi'}\!\! d\psi \,
\sin^{D-4}\psi \right) \nonumber\\ &=& {1\over 4-D} \,\gamma
\cot\gamma \; +\mbox{finite}
\label{i2neq}\,. \end{aligned}$$
P. Pascual and R. Tarrach, “QCD: Renormalization for the Practitioner”, Lecture Notes in Physics, Vol. 194 (Springer, Berlin, 1984). H. Cheng and E.-C. Tsai, Phys. Rev. Lett. [**57**]{} (1986) 511. S. Caracciolo, G. Curci, and P. Menotti, Phys. Lett. [**B113**]{} (1982) 311. J.-P. Leroy, J. Micheli, and G.-C. Rossi, Z. Phys. [**C36**]{} (1987) 305. N.H. Christ and T.D. Lee, Phys. Rev. [**D22**]{} (1980) 939. P. Doust, Ann. Phys. [**177**]{} (1987) 169. P.J. Doust and J.C. Taylor, Phys. Lett. [**B197**]{} (1987) 232. M.A. Shifman, Nucl. Phys. [**B173**]{} (1980) 13. V.A. Fock, Sov. Phys. [**12**]{} (1937) 404. J. Schwinger, Phys. Rev. [**82**]{} (1952) 684. C. Cronström, Phys. Lett. [**B90**]{} (1980) 267. W. Kummer and J. Weiser, Z. Phys. [**C31**]{} (1986) 105. G. Modanese, J. Math. Phys. [**33**]{} (1992) 1523. P. Menotti, G. Modanese, and D. Seminara, Ann. Phys. [**224**]{} (1993) 110. M. Azam, Phys. Lett. [**B101**]{} (1981) 401. N.E. Bralic, Phys. Rev. [**D22**]{} (1980) 3090. H.-Th. Elze, M. Gyulassy, and D. Vasak, Nucl. Phys. [**B276**]{} (1986) 706. G.P. Korchemsky and A.V. Radyushkin, Nucl. Phys. [**B283**]{} (1987) 342. V.S. Dotsenko and S.N. Vergeles, Nucl. Phys. [**B169**]{} (1980) 527. R.A. Brandt, F. Neri, and M.-A. Sato, Phys. Rev. [**D24**]{} (1981) 879. R.A. Brandt, A. Gocksch, M.-A. Sato, and F. Neri, Phys. Rev. [**D26**]{} (1982) 3611. I.A. Korchemskaya and G.P. Korchemsky, Phys. Lett. [**B287**]{} (1992) 169.
[^1]: [email protected]
[^2]: [email protected]
[^3]: In fact he discussed the Abelian case but this makes no difference for free fields.
[^4]: For simplicity we discuss the QED case here. Aiming at an expression for the free gauge propagator this is no restriction of generality. For non-Abelian gauge groups c.f. [@azam].
[^5]: Only a factor $\mu^{D-4}$ comes in since $g_R$ as given in (\[grug\]) is dimensionless in contrast to $g$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We consider the homogenization of parabolic equations with large spatially-dependent potentials modeled as Gaussian random fields. We derive the homogenized equations in the limit of vanishing correlation length of the random potential. We characterize the leading effect in the random fluctuations and show that their spatial moments converge in law to Gaussian random variables. Both results hold for sufficiently small times and in sufficiently large spatial dimensions $d\geq{{\mathfrak m}}$, where ${{\mathfrak m}}$ is the order of the spatial pseudo-differential operator in the parabolic equation. In dimension $d<{{\mathfrak m}}$, the solution to the parabolic equation is shown to converge to the (non-deterministic) solution of a stochastic equation in the companion paper [@B-CMP-2-08]. The results are then extended to cover the case of long range random potentials, which generate larger, but still asymptotically Gaussian, random fluctuations.'
author:
- 'Guillaume Bal [^1]'
bibliography:
- '../../bibliography.bib'
title: Homogenization with large spatial random potential
---
#### keywords:
Homogenization theory, partial differential equations with random coefficients, Gaussian fluctuations, large potential, long range correlations
#### AMS:
35R60, 60H05, 35K15.
Introduction {#sec:intro}
============
Let ${{\mathfrak m}}>0$ and $P(D)$ the pseudo-differential operator with symbol $\hat p(\xi)=|\xi|^{{\mathfrak m}}$. We consider the following evolution equation in dimension $d\geq{{\mathfrak m}}$: $$\label{eq:parabeps}
\begin{array}{rcll}
\Big({\dfrac{\partial{}}{\partial{t}}} + P(D) - \dfrac{1}{{\varepsilon}^\alpha} q\big(\dfrac x{\varepsilon}\big)
\Big) u_{\varepsilon}(t,x) &=&0,\qquad & x\in\Rm^d,\quad t>0, \\
u_{\varepsilon}(0,x) &=& u_0(x), \quad& x\in\Rm^d .
\end{array}$$ Here, $u_0\in L^2(\Rm^d)$ and $q(x)$ is a mean zero stationary Gaussian process defined on a probability space $(\Omega,\mathcal
F,{\mathbb P})$. We assume that $q(x)$ has bounded and integrable correlation function $R(x)={\mathbb E}\{q(y)q(x+y)\}$, where ${\mathbb E}$ is the mathematical expectation associated with ${\mathbb P}$, and bounded, continuous in the vicinity of $0$, and integrable power spectrum $(2\pi)^d\hat
R(\xi)=\int_{\Rm^d} e^{-i\xi\cdot x} R(x)dx$ in the sense that $\int_{\Rm^d\backslash B(0,1)} \hat R(\xi) |\xi|^{-{{\mathfrak m}}}d\xi<\infty$. The size of the potential is constructed so that the limiting solution as ${\varepsilon}\to0$ is different from the unperturbed solution obtained by setting $q=0$. The appropriate size of the potential is given by $$\label{eq:epsalpha}
{\varepsilon}^{\alpha} = \left\{
\begin{matrix}
{\varepsilon}^{\frac{{{\mathfrak m}}}2}|\ln{\varepsilon}|^{\frac12} & d={{\mathfrak m}}, \\
{\varepsilon}^{\frac{{{\mathfrak m}}}2} & d>{{\mathfrak m}}.
\end{matrix}\right.$$
The potential is bounded ${\mathbb P}$-a.s. on bounded domains but is unbounded ${\mathbb P}$-a.s. on $\Rm^d$. By using a method based on the Duhamel expansion, we nonetheless obtain that for a sufficiently small time $T>0$, the above equation admits a weak solution $u_{\varepsilon}(t,\cdot)\in
L^2(\Omega\times\Rm^d)$ uniformly in time $t\in (0,T)$ and $0<{\varepsilon}<{\varepsilon}_0$.
Moreover, as ${\varepsilon}\to0$, the solution $u_{\varepsilon}(t)$ converges strongly in $L^2(\Omega\times\Rm^d)$ uniformly in $t\in (0,T)$ to its limit $u(t)$ solution of the following homogenized evolution equation $$\label{eq:parab}
\begin{array}{rcll}
\Big({\dfrac{\partial{}}{\partial{t}}} + P(D) -\rho
\Big) u(t,x) &=&0,\qquad & x\in\Rm^d,\quad t>0, \\
u(0,x) &=& u_0(x), \quad & x\in\Rm^d ,
\end{array}$$ where the effective (non-negative) potential is given by $$\label{eq:rho}
\rho = \left\{
\begin{array}{ll}
\,\,c_d \hat R(0) & d={{\mathfrak m}}, \\[0mm]
{\displaystyle\int}_{\Rm^d} \dfrac{\hat R(\xi)}{|\xi|^{{\mathfrak m}}} d\xi \quad & d>{{\mathfrak m}}.
\end{array}\right.$$ Here, $c_d$ is the volume of the unit sphere $S^{d-1}$. We denote by $\mathcal G^\rho_t$ the propagator for the above equation, which to $u_0(x)$ associates $\mathcal G^\rho_t u_0 (x) = u(t,x)$ solution of .
We assume that the non-negative (by Bochner’s theorem) power spectrum $\hat R(\xi)$ is bounded by $f(|\xi|)$, where $f(r)$ is a positive, bounded, radially symmetric, and integrable function in the sense that $\int_1^\infty r^{d-1-{{\mathfrak m}}}f(r)dr <\infty$. Then we have the following result.
\[thm:convergence\] There exists a time $T=T(f)>0$ such that for all $t\in (0,T)$, there exists a solution $u_{\varepsilon}(t)\in
L^2(\Omega\times \Rm^d)$ uniformly in $0<{\varepsilon}<{\varepsilon}_0$. Moreover, let us assume that $\hat R(\xi)$ is of class $\mathcal
C^\gamma(\Rm^d)$ for some $0\leq \gamma\leq 2$ and let $u(t,x)$ be the unique solution in $L^2(\Rm^d)$ to . Then, we have the convergence results $$\label{eq:L2error}
\begin{array}{rcl}
\|(u_{\varepsilon}-{{\mathfrak u}}_{\varepsilon})(t)\|_{L^2(\Omega\times\Rm^d)}
&\lesssim& {\varepsilon}^{\frac\beta2}
\|u_0\|_{L^2(\Rm^d)}, \\[2mm]
\|({{\mathfrak u}}_{\varepsilon}-u)(t)\|_{L^2(\Rm^d)} &\lesssim& {\varepsilon}^{\gamma\wedge\beta}
\|u_0\|_{L^2(\Rm^d)},
\end{array}$$ where $a\lesssim b$ means $a\leq Cb$ for some $C>0$, $a\wedge b=\min(a,b)$, where ${{\mathfrak u}}_{\varepsilon}(t,\cdot)$ is a deterministic function in $L^2(\Rm^d)$ uniformly in time, and where we have defined $$\label{eq:beta}
{\varepsilon}^\beta = \left\{
\begin{array}{lc}
|\ln{\varepsilon}|^{-1} & d={{\mathfrak m}}, \\
{\varepsilon}^{d-{{\mathfrak m}}} & {{\mathfrak m}}<d<2{{\mathfrak m}}, \\
{\varepsilon}^{{{\mathfrak m}}}|\ln{\varepsilon}| \quad& d=2{{\mathfrak m}}, \\
{\varepsilon}^{{{\mathfrak m}}} & d> 2{{\mathfrak m}}.
\end{array} \right.$$ The Fourier transform ${{\mathfrak U}}_{\varepsilon}(t,\xi)$ of the deterministic function ${{\mathfrak u}}_{\varepsilon}(t,x)$ is determined explicitly in below.
Note that the effective potential $-\rho$ is non-positive. The theorem is valid for times $T$ such that $4T\rho_f<1$, where $\rho_f$ is defined in lemma \[lem:bound\] below by replacing $\hat R(\xi)$ by $f(|\xi|)$ in the definition of $\rho$ in .
The error term $u_{\varepsilon}-u$ is dominated by deterministic components when ${\varepsilon}^{\gamma\wedge \beta}\gg{\varepsilon}^{\frac{d-2\alpha}2}$ and by random fluctuations when ${\varepsilon}^{\gamma\wedge\beta}\ll{\varepsilon}^{\frac{d-2\alpha}2}$. In both situations, the random fluctuations may be estimated as follows. We show that $$\label{eq:u1eps}
u_{1,{\varepsilon}}(t,x) = \dfrac{1}{{\varepsilon}^{\frac{d-2\alpha}2}}
\big(u_{\varepsilon}-{\mathbb E}\{u_{\varepsilon}\}\big)(t,x),$$ converges weakly in space and in distribution to a Gaussian random variable. More precisely, we have
\[thm:fluct\] Let $M$ be a test function such that its Fourier transform $\hat
M\in L^1(\Rm^d)\cap L^2(\Rm^d)$. Then we find that for all $t\in
(0,T)$ $$\label{eq:conv}
(u_{1,{\varepsilon}}(t,\cdot),M) \xrightarrow{\,{\varepsilon}\to0\,}
{\displaystyle\int}_{\Rm^d} {\mathcal M}_t(x) \sigma dW_x, \qquad
\mathcal M_t(x) =
{\displaystyle\int}_0^t \mathcal G^\rho_s M(x) \mathcal G^\rho_{t-s} u_0(x) ds,$$ where convergence holds in the sense of distributions, $dW_x$ is the standard multiparameter Wiener measure on $\Rm^d$ and $\sigma$ is the standard deviation defined by $$\label{eq:sigmavar}
\sigma^2 := (2\pi)^d\hat R(0) = {\displaystyle\int}_{\Rm^d} {\mathbb E}\{q(0)q(x)\} dx.$$
This shows that the fluctuations of the solution are asymptotically given by a Gaussian random variable, which is consistent with the central limit theorem.
We observe a sharp transition in the behavior of $u_{\varepsilon}$ at $d={{\mathfrak m}}$. For $d<{{\mathfrak m}}$, the following holds. The size of the potential that generates an order $O(1)$ perturbation is now given by (see the last inequality in lemma \[lem:bound\]) $${\varepsilon}^\alpha = {\varepsilon}^{\frac d2}.$$ Using the same methods as for the case $d\geq{{\mathfrak m}}$, we may obtain that $u_{\varepsilon}(t)$ is uniformly bounded and thus converges weakly in $L^2(\Omega\times\Rm^d)$ for sufficiently small times to a function $u(t)$. The problem is addressed in [@B-CMP-2-08], where it is shown that $u(t)$ is the solution to the stochastic partial differential equation in Stratonovich form $$\label{eq:stoch}
{\dfrac{\partial{u}}{\partial{t}}} + P(D) u + u\circ \sigma {\dfrac{d{W}}{d{x}}}=0,$$ with $u(0,x)=u_0(x)$ and $\frac{dW}{dx}$ d-parameter spatial white noise “density”. The above equation admits a unique solution that belongs to $L^2(\Omega\times\Rm^d)$ locally uniformly in time. Stochastic equations have also been analyzed in the case where $d\geq{{\mathfrak m}}$ (i.e., $d\geq2$ when $P(D)=-\Delta$), see [@Hu-PA-02; @NR-JFA-97]. However, our results show that such solutions cannot be obtained as a limit in $L^2(\Omega\times\Rm^d)$ of solutions corresponding to vanishing correlation length so that their physical justification is more delicate. In the case $d=1$ and ${{\mathfrak m}}=2$ with $q(x)$ a bounded potential, we refer the reader to [@PP-GAK-06] for more details on the above stochastic equation.
The above theorems \[thm:convergence\] and \[thm:fluct\] assume short range correlations for the random potential. Mathematically, this is modeled by an integrable correlation function, or equivalently a bounded value for $\hat R(0)$. Longer range correlations may be modeled by unbounded power spectra in the vicinity of the origin, for instance by assuming that $\hat R(\xi)=h(\xi)\hat S(\xi)$, where $\hat
S(\xi)$ is bounded in the vicinity of the origin and $h(\xi)$ is a homogeneous function of degree $-{{\mathfrak n}}$ for some ${{\mathfrak n}}>0$. Provided that $d>{{\mathfrak m}}+{{\mathfrak n}}$ so that $\rho$ defined in is still bounded, the results of theorems \[thm:convergence\] and \[thm:fluct\] may be extended to the case of long range fluctuations. We refer the reader to theorem \[thm:fracBM\] in section \[sec:longrange\] below for the details. The salient features of the latter result is that the convergence properties stated in theorem \[thm:convergence\] still hold with $\beta$ replaced by $\beta-{{\mathfrak n}}$ and that the random fluctuations are now asymptotically Gaussian processes of amplitude of order ${\varepsilon}^{\frac{d-{{\mathfrak m}}-{{\mathfrak n}}}2}$. Moreover, they may conveniently be written as stochastic integrals with respect to some multiparameter fractional Brownian motion in place of the Wiener measure appearing in .
Let us also mention that all the result stated here extend to the Schrödinger equation, where $\frac{\partial}{\partial t}$ is replaced by $i\frac{\partial}{\partial t}$ in . We then verify that $-\rho$ in is replaced by $\rho$ so that the homogenized equation is given by $$\Big(i{\dfrac{\partial{}}{\partial{t}}} + P(D) +\rho
\Big) u(t,x) \,=\,0.$$ The main effect of the randomness is therefore a phase shift of the quantum waves as they propagate through the random medium. Because the semigroup associated to the free evolution of quantum waves does not damp high frequencies as efficiently as for the parabolic equation , some additional regularity assumptions on the initial condition are necessary to obtain the limiting behaviors described in theorems \[thm:convergence\] and \[thm:fluct\]. We do not consider the case of the Schrödinger equation further here.
The rest of the paper is structured as follows. Section \[sec:existence\] recasts as an infinite Duhamel series of integrals in the Fourier domain. The cross-correlations of the terms appearing in the series are analyzed by calculating moments of Gaussian variables and estimating the contributions of graphs similar to those introduced in [@Erdos-Yau2; @LS-ARMA-07]. These estimates allow us to construct a solution to in $L^2(\Omega\times\Rm^d)$ uniformly in time for sufficiently small times $t\in (0,T)$. The maximal time $T$ of validity of the theory depends on the power spectrum $\hat R(\xi)$. The estimates on the graphs are then used in section \[sec:limits\] to characterize the limit and the leading random fluctuations of the solution $u_{\varepsilon}(t,x)$. The extension of the results to long range correlations is presented in section \[sec:longrange\].
The analysis of and of similar operators has been performed for smaller potentials than those given in in e.g. [@B-CLH-08; @FOP-SIAP-82] when $u_{\varepsilon}$ converges strongly to the solution of the unperturbed equation (with $q\equiv0$). The results presented in this paper may thus be seen as generalizations to the case of sufficiently strong potentials so that the unperturbed solution is no longer a good approximation of $u_{\varepsilon}$. The analysis presented below is based on simple estimates for the Feynman diagrams corresponding to Gaussian random potentials and does not extend to other potentials such as Poisson point potentials, let alone potentials satisfying some mild mixing conditions. Extension to other potentials would require more sophisticated estimates of the graphs than those presented here or a different functional setting than the $L^2(\Omega\times\Rm^d)$ setting considered here. For related estimates on the graphs appearing in Duhamel expansion, we refer the reader to e.g. [@Chen-JSP-05; @Erdos-Yau2; @LS-ARMA-07].
Duhamel expansion and existence theory {#sec:existence}
======================================
Since $q(x)$ is a stationary mean zero Gaussian random field, it admits the following spectral representation $$\label{eq:spectq}
q(x) = \dfrac{1}{(2\pi)^d} {\displaystyle\int}_{\Rm^d} e^{i\xi\cdot x}
\hat Q(d\xi),$$ where $\hat Q(d\xi)$ is the complex spectral process such that $${\mathbb E}\Big\{ {\displaystyle\int}_{\Rm^d} f(\xi) \hat Q(d\xi)
\overline{{\displaystyle\int}_{\Rm^d} g(\xi) \hat Q(d\xi)} \Big\}
= {\displaystyle\int}_{\Rm^d} f(\xi) \bar g(\xi) (2\pi)^d\hat R(\xi) d\xi,$$ for all $f$ and $g$ in $L^2(\Rm^d;\hat R(\xi)d\xi)$ with the power spectrum and correlation function of $q$ respectively defined by $$\label{eq:RhatR}
0\leq (2\pi)^d \hat R(\xi) = {\displaystyle\int}_{\Rm^d} e^{-i\xi\cdot x} R(x) dx,\qquad
R(x) = {\mathbb E}\{q(y)q(x+y)\}.$$ In the sequel, we write $\hat Q(d\xi)\equiv\hat q(\xi) d\xi$ so that ${\mathbb E}\{\hat q(\xi)\hat q(\zeta)\}=\hat R(\xi)\delta(\xi+\zeta)$ and ${\mathbb E}\{\hat q(\xi)\overline{\hat q(\zeta)}\}=\hat
R(\xi)\delta(\xi-\zeta)$.
Duhamel expansion {#sec:duhamel}
-----------------
Let us introduce $\hat q_{\varepsilon}(\xi)={\varepsilon}^{d-\alpha}\hat q({\varepsilon}\xi)$, the Fourier transform of ${\varepsilon}^{-\alpha}q(\frac{x}{\varepsilon})$. We may now recast the parabolic equation as $$\label{eq:parabFD}
\big({\dfrac{\partial{}}{\partial{t}}} + \xi^{{\mathfrak m}}\big) \hat u_{\varepsilon}= \hat q_{\varepsilon}* \hat u_{\varepsilon},$$ with $\hat u_{\varepsilon}(0,\xi)=\hat u_0(\xi)$, where $$\hat q_{\varepsilon}* \hat u_{\varepsilon}(t,\xi) =
{\displaystyle\int}_{\Rm^d} \hat u_{\varepsilon}(t,\xi-\zeta) \hat Q_{\varepsilon}(d\zeta)
\equiv {\displaystyle\int}_{\Rm^d} \hat u_{\varepsilon}(t,\xi-\zeta) \hat q_{\varepsilon}(\zeta)d\zeta.$$ Here and below, we use the notation $\xi^{{\mathfrak m}}=|\xi|^{{\mathfrak m}}$. After integration in time, the above equation becomes $$\label{eq:intFD}
\hat u_{\varepsilon}(t,\xi) = e^{-t\xi^{{\mathfrak m}}} \hat u_0(\xi) + {\displaystyle\int}_0^t
e^{-s \xi^{{\mathfrak m}}} {\displaystyle\int}_{\Rm^d} \hat q_{\varepsilon}(\xi-\xi_1) \hat u_{\varepsilon}(t-s,\xi_1)
d\xi_1 ds.$$ This allows us to write the formal Duhamel expansion $$\begin{aligned}
\label{eq:exp}
\hat u_{\varepsilon}(t,\xi) &=& {\displaystyle\sum}_{n\in\Nm} \hat u_{n,{\varepsilon}}(t,\xi),\\
\label{eq:hatuneps}
\hat u_{n,{\varepsilon}}(t,\xi_0) &=& {\displaystyle\int}_{\Rm^{nd}} {\displaystyle\prod}_{k=0}^{n-1}
{\displaystyle\int}_0^{t_k({\mathbf s})} e^{-\xi_k^{{\mathfrak m}}s_k} e^{-(t-\sum_{k=0}^{n-1} s_k)\xi_n^{{\mathfrak m}}}
{\displaystyle\prod}_{k=0}^{n-1} \hat q_{\varepsilon}(\xi_k-\xi_{k+1})
\hat u_0(\xi_n) d{\mathbf s}d{\boldsymbol \xi}.\end{aligned}$$ Here, we have introduced the following notation: $${\mathbf s}=(s_0,\ldots,s_{n-1}),\,\, t_k({\mathbf s}) = t-s_0-\ldots - s_{k-1},\,\,
t_0({\mathbf s})=t,\,\, d{\mathbf s}=\prod_{k=0}^{n-1}ds_k,\,\, d{\boldsymbol \xi}=\prod_{k=1}^n d\xi_k.$$
We now show that for sufficiently small times, the expansion converges (uniformly for all ${\varepsilon}$ sufficiently small) in the $L^2(\Omega\times \Rm^d)$ sense. Moreover, the $L^2$ norm of $u_{\varepsilon}(t)$ is bounded by the $L^2(\Rm^d)$ norm of $\hat u_0$, which gives us an a priori estimate for the solution. The convergence results are based on the analysis of the following moments $$\label{eq:Ueps}
U^{n,m}_{\varepsilon}(t,\xi,\zeta) = {\mathbb E}\{\hat u_{{\varepsilon},n}(t,\xi)
\overline{\hat u_{{\varepsilon},m}}(t,\zeta)\},$$ which, thanks to , are given by $$\begin{array}{l}
{\displaystyle\int}_{\Rm^{d(n+m)}} \prod_{k=0}^{n-1}
{\displaystyle\int}_0^{t_k({\mathbf s})} \prod_{l=0}^{m-1} {\displaystyle\int}_0^{t_l({\boldsymbol\tau})}
e^{- s_k\xi_k^{{\mathfrak m}}}
e^{-(t-\sum_{k=0}^{n-1} s_k)\xi_n^{{\mathfrak m}}}
e^{- \tau_l\zeta_l^{{\mathfrak m}}}
e^{-(t-\sum_{l=0}^{m-1} \tau_l)\zeta_m^{{\mathfrak m}}} \\
{\mathbb E}\Big\{{\displaystyle\prod}_{k=0}^{n-1} {\displaystyle\prod}_{l=0}^{m-1}
\hat q_{\varepsilon}(\xi_k-\xi_{k+1})\bar {\hat q}_{\varepsilon}(\zeta_l-\zeta_{l+1})\Big\}
\hat u_0(\xi_n)\bar {\hat u}_0(\zeta_m) \, d{\mathbf s}d{\boldsymbol\tau}d{\boldsymbol \xi}d{\boldsymbol \zeta}.
\end{array}$$ Let us introduce the notation $s_{n}({\mathbf s})=t_n({\mathbf s})=t-\sum_{k=0}^{n-1}
s_k$ and $\tau_{m}({\boldsymbol\tau})=t_m({\boldsymbol\tau})=t-\sum_{l=0}^{m-1}\tau_l$. We also define $\xi_{n+k+1}=\zeta_{m-k}$ and $s_{n+k+1}=\tau_{m-k}$ for $0\leq k\leq m$. Since $q_{\varepsilon}$ is real-valued, we find that $$U_{\varepsilon}^{n,m}(t,\xi_0,\xi_{n+m+1}) =
{\displaystyle\int}\prod_{k=0}^{n+m+1} e^{-s_k\xi_k^{{\mathfrak m}}}
{\mathbb E}\Big\{\prod_{k=0,k\not=n}^{n+m} \hat q_{\varepsilon}(\xi_k-\xi_{k+1}) \Big\}
\hat u_0(\xi_n) \bar {\hat u}_0(\xi_{n+1}) d{\mathbf s}d{\boldsymbol \xi},$$ where the domain of integration in the $s$ and $\xi$ variables is inherited from the previous expression. Note that no integration is performed in the variables $s_n({\mathbf s})$ and $s_{n+1}({\boldsymbol\tau})$. The integral may be recast as $${\displaystyle\int}\prod_{k=0}^{n+m+1} e^{-s_k\xi_k^2}
{\mathbb E}\Big\{\prod_{k=0,k\not=n}^{n+m} \hat q_{\varepsilon}(\xi_k-\xi_{k+1}) \Big\}
\hat u_0(\xi_n) \bar {\hat u}_0(\xi_{n+1})
\delta(t-\sum_{k=0}^{n}s_k)
\delta(t-\sum_{k=n+1}^{n+m+1}s_k) d{\mathbf s}d{\boldsymbol \xi},$$ where the integrals in all the $s_k$ variables for $0\leq k\leq n+m+1$ are performed over $(0,\infty)$. The $\delta$ functions ensure that the integration is equivalent to the one presented above. The latter form is used in the proof of lemma \[lem:timeint\] below.
We need to introduce additional notation. The moments of $\hat u_{{\varepsilon},n}$ are defined as $$\label{eq:Uepsn}
U_{\varepsilon}^n(t,\xi) = {\mathbb E}\{\hat u_{{\varepsilon},n}(t,\xi) \}.$$ We also introduce the following covariance function $$\label{eq:Vepsnm}
V^{n,m}_{\varepsilon}(t,\xi,\zeta) = {\rm cov}(\hat u_{{\varepsilon},n}(t,\xi),
\hat u_{{\varepsilon},m}(t,\zeta)) =
U^{n,m}_{\varepsilon}(t,\xi,\zeta) - U_{\varepsilon}^n(t,\xi)
\overline{U_{\varepsilon}^m(t,\zeta)}.$$ These terms allow us to analyze the convergence properties of the solution $\hat u_{\varepsilon}(t,\xi)$. Let $\hat M(\xi)$ be a smooth (integrable and square integrable is sufficient) test function on $\Rm^d$. We introduce the two random variables $$\begin{aligned}
\label{eq:energy}
I_{\varepsilon}(t) &=& {\displaystyle\int}_{\Rm^d} |\hat u_{\varepsilon}(t,\xi)|^2 d\xi\\
\label{eq:Xeps}
X_{\varepsilon}(t) &=& {\displaystyle\int}_{\Rm^d} \hat u_{\varepsilon}(t,\xi) \overline{\hat M}(\xi)d\xi.\end{aligned}$$
Summation over graphs {#sec:graphs}
---------------------
We now need to estimate moments of the Gaussian process $\hat q_{\varepsilon}$. The expectation in $U_{\varepsilon}^{n,m}$ vanishes unless there is ${{\bar n}}\in\Nm$ such that $n+m=2{{\bar n}}$ is even. The expectation of a product of Gaussian variables has an explicit structure written as a sum over all possible products of pairs of indices of the form $\xi_k-\xi_{k+1}$. The moments are thus given as a sum of products of the expectation of [*pairs*]{} of terms $\hat
q_{\varepsilon}(\xi_k-\xi_{k+1})$, where the sum runs over all possible pairings. We define the [**pair**]{} $(\xi_k,\xi_l)$, $1\leq k<l$, as the contribution in the product given by $${\mathbb E}\{\hat q_{\varepsilon}(\xi_{k-1}-\xi_{k})\hat q_{\varepsilon}(\xi_{l-1}-\xi_{l})\}
= {\varepsilon}^{d-2\alpha} \hat R({\varepsilon}(\xi_k-\xi_{k-1}))
\delta(\xi_k-\xi_{k-1}+\xi_l-\xi_{l-1}).$$ We have used here the fact that $\hat R(-\xi)=\hat R(\xi)$.
The number of pairings in a product of $n+m=2{{\bar n}}$ terms (i.e., the number of allocations of the set $\{1,\ldots,2{{\bar n}}\}$ into ${{\bar n}}$ unordered pairs) is equal to $$\dfrac{(2{{\bar n}}-1)!}{2^{{{\bar n}}-1}({{\bar n}}-1)!} = \dfrac{(2{{\bar n}})!}{{{\bar n}}!2^{{\bar n}}}
=(2n-1)!!.$$ There is consequently a very large number of terms appearing in $U_{\varepsilon}^{n,m}(t,\xi_0,\xi_{n+m+1})$. In each instance of the pairings, we have ${{\bar n}}$ terms $k$ and ${{\bar n}}$ terms $l\equiv l(k)$. Note that $l(k)\geq k+1$. We denote by [**simple pairs**]{} the pairs such that $l(k)=k+1$, which thus involve a delta function of the form $\delta(\xi_{k+1}-\xi_{k-1})$.
![Graph with $n=3$ and $m=1$ corresponding to the pairs $(\xi_1,\xi_3)$ and $(\xi_2,\xi_5)$ and the delta functions $\delta(\xi_1-\xi_0+\xi_3-\xi_2)$ and $\delta(\xi_2-\xi_1+\xi_5-\xi_4)$.[]{data-label="fig:1"}](Presentation2){height="3.5cm"}
The collection of pairs $(\xi_k,\xi_{l(k)})$ for ${{\bar n}}$ values of $k$ and ${{\bar n}}$ values of $l(k)$ constitutes a graph ${\mathfrak g}\in {\mathfrak G}$ constructed as follows; see Fig.\[fig:1\] and [@Erdos-Yau2]. The upper part of the graph with $n$ bullets represents $\hat
u_{{\varepsilon},n}$ while the lower part with $m$ bullets represents $\overline{\hat u_{{\varepsilon},m}}$. The two squares on the left of the graph represent the variables $\xi_0$ and $\xi_{n+m+1}$ in $U_{\varepsilon}^{n,m}(t,\xi_0,\xi_{n+m+1})$ while the squares on the right represent $\hat u_0(\xi_n)$ and $\bar {\hat u}_0(\xi_{n+1})$. The dotted pairing lines represent the pairs of the graph ${\mathfrak g}$. Here, ${\mathfrak G}$ denotes the collection of all possible $|{\mathfrak G}|=\frac{(2{{\bar n}}-1)!}{2^{{{\bar n}}-1}({{\bar n}}-1)!}$ graphs that can be constructed for a given ${{\bar n}}$.
We denote by $A_0=A_0({\mathfrak g})$ the collection of the ${{\bar n}}$ values of $k$ and by $B_0=B_0({\mathfrak g})$ the collection of the ${{\bar n}}$ values of $l(k)$. We then find that $${\mathbb E}\Big\{\prod_{k=0,k\not=n}^{n+m} \hat q_{\varepsilon}(\xi_k-\xi_{k-1}) \Big\}
= {\displaystyle\sum}_{{\mathfrak g}\in{\mathfrak G}} {\displaystyle\prod}_{k\in A_0({\mathfrak g})}
{\varepsilon}^{d-2\alpha} \hat R({\varepsilon}(\xi_k-\xi_{k-1}))
\delta(\xi_{k}-\xi_{k-1}+\xi_{l(k)}-\xi_{l(k)-1}).$$ This provides us with an explicit expression for $U_{\varepsilon}^{n,m}(t,\xi_0,\xi_{n+m+1})$ as a summation over all possible graphs generated by moments of Gaussian random variables. We need to introduce several classes of graphs.
We say that the graph has a [**crossing**]{} if there is a $k\leq n$ such that $l(k)\geq n+2$. We denote by ${\mathfrak G}_c\subset {\mathfrak G}$ the set of graphs with at least one crossing and by ${\mathfrak G}_{nc}={\mathfrak G}\backslash{\mathfrak G}_c$ the [**non-crossing**]{} graphs. We observe that $V^{n,m}_{\varepsilon}(t,\xi_0,\xi_{n+m+1})$ is the sum over the crossing graphs and that $U_{\varepsilon}^n(t,\xi_0) \overline{U_{\varepsilon}^m(t,\xi_{n+m+1})}$ is the sum over the non-crossing graphs in $U^{n,m}_{\varepsilon}(t,\xi_0,\xi_{n+m+1})$.
The unique graph ${\mathfrak g}_s$ with only simple pairs is called the [ **simple graph**]{} and we define ${\mathfrak G}_{ns}={\mathfrak G}\backslash{\mathfrak g}_s$. We denote by ${\mathfrak G}_{cs}$ the [**crossing simple graphs**]{} with only simple pairs except for exactly one crossing. The complement of ${\mathfrak G}_{cs}$ in the crossing graphs is denoted by ${\mathfrak G}_{cns}={\mathfrak G}_c\backslash{\mathfrak G}_{cs}$.
As we shall see, only the simple graph ${\mathfrak g}_s$ contributes an $O(1)$ term in the limit ${\varepsilon}\to0$ and only the graphs in ${\mathfrak G}_{cs}$ contribute to the leading order $O({\varepsilon}^{\frac12(d-2\alpha)})$ in the fluctuations of $\hat u_{\varepsilon}$.
The graphs are defined similarly in the calculation of $U_{\varepsilon}^n(t,\xi_0)$ in for $n=2{{\bar n}}$ and $m=0$, except that crossing graphs have no meaning in such a context. A summation over $k\in A_0({\mathfrak g})$ of all the arguments $\xi_{k}-\xi_{k-1}+\xi_{l(k)}-\xi_{l(k)-1}$ of the $\delta$ functions shows that the last delta function may be replaced without modifying the integral in $U_{\varepsilon}^n(t,\xi_0)$ by $\delta(\xi_0-\xi_n)$.
This allows us to summarize the above calculations as follows: $$\label{eq:Uepsmn}
\begin{array}{l}
U_{\varepsilon}^{n,m}(t,\xi_0,\xi_{n+m+1}) =
{\displaystyle\int}\prod_{k=0}^{n+m+1} e^{-s_k\xi_k^{{\mathfrak m}}}
\hat u_0(\xi_n) \bar {\hat u}_0(\xi_{n+1}) {\displaystyle\sum}_{{\mathfrak g}\in{\mathfrak G}} \\
\qquad {\displaystyle\prod}_{k\in A_0({\mathfrak g})}
{\varepsilon}^{d-2\alpha} \hat R({\varepsilon}(\xi_k-\xi_{k-1}))
\delta(\xi_{k}-\xi_{k-1}+\xi_{l(k)}-\xi_{l(k)-1})
d{\mathbf s}d{\boldsymbol \xi}.
\end{array}$$ Similarly, $$\label{eq:Uepsm}
\begin{array}{l}
U_{\varepsilon}^{n}(t,\xi_0)
= \hat u_0(\xi_0)
{\displaystyle\int}\prod_{k=0}^{n} e^{-s_k\xi_k^{{\mathfrak m}}}{\displaystyle\sum}_{{\mathfrak g}\in{\mathfrak G}} \\\qquad
{\displaystyle\prod}_{k\in A_0({\mathfrak g})}
{\varepsilon}^{d-2\alpha} \hat R({\varepsilon}(\xi_k-\xi_{k-1}))
\delta(\xi_{k}-\xi_{k-1}+\xi_{l(k)}-\xi_{l(k)-1})
d{\mathbf s}d{\boldsymbol \xi}. \!\!\!\!\!\!\!\!\!\!\!\!
\end{array}$$
Analysis of crossing graphs
---------------------------
We now analyze the influence of the crossing graphs on $I_{\varepsilon}(t)$ and $X_{\varepsilon}(t)$ defined in and , respectively, for sufficiently small times. We obtain from and that $$\label{eq:Vepsmnint}
\begin{array}{l}
V^{n,m}_{\varepsilon}(t,\xi_0,\xi_{n+m+1})
={\displaystyle\sum}_{{\mathfrak g}\in{\mathfrak G}_c}
{\displaystyle\int}\prod_{k=0}^{n+m+1} e^{-s_k\xi_k^{{\mathfrak m}}}
\hat u_0(\xi_n) \bar {\hat u}_0(\xi_{n+1}) \\
\qquad\quad {\displaystyle\prod}_{k\in A_0({\mathfrak g})}
{\varepsilon}^{d-2\alpha} \hat R({\varepsilon}(\xi_k-\xi_{k-1}))
\delta(\xi_{k}-\xi_{k-1}+\xi_{l(k)}-\xi_{l(k)-1})
\,d{\mathbf s}\,d{\boldsymbol \xi},
\end{array}$$ involves the summation over the crossing graphs ${\mathfrak G}_c$. Let us consider a graph ${\mathfrak g}\in{\mathfrak G}_c$ with $M$ crossing pairs, $M\geq1$. Crossing pairs are defined by $k\leq n$ and $l(k)\geq n+2$. Denote by $(\xi_{q_m},\xi_{l(q_m)})$, $1\leq m\leq M$ the crossing pairs and define $Q=\max_m\{q_m\}$. By summing the arguments inside the delta functions for all $k\leq n$, we observe that the last of these delta functions may be replaced by $$\delta(\xi_0-\xi_n + {\displaystyle\sum}_{m=1}^M \xi_{q_m}-\xi_{q_m-1}).$$ Similarly, by summing over all pairs with $k\geq n+2$, we obtain that the last of these delta functions may be replaced by $$\delta(\xi_{n+1}-\xi_{n+m+1}+ {\displaystyle\sum}_{m=1}^M
\xi_{l(q_m)}-\xi_{l(q_m)-1}).$$ The product of the latter two delta functions is then equivalent to $$\delta(\xi_{n+m+1}-\xi_{n+1}+\xi_n-\xi_0)
\delta(\xi_Q-\xi_{Q-1} + \xi_0-\xi_n + {\displaystyle\sum}_{m=1}^{M-1}
\xi_{q_m}-\xi_{q_m-1}).$$ The analysis of the contributions of the crossing graphs is slightly different for the energy in and for the spatial moments in . We start with the energy.
#### Analysis of the crossing terms in $I_{\varepsilon}(t)$.
We evaluate the expression for $|V^{n,m}_{\varepsilon}(t,\xi_0,\xi_0)|$ in at $\xi_{n+m+1}=\xi_0$ and integrate in the $\xi_0$ variable over $\Rm^d$. Let us define $A'=A_0\backslash\{Q\}$. For each $k\in A'\cup\{0\}$, we perform the change of variables $\xi_k\to\frac{\xi_k}{\varepsilon}$. We then define $$\label{eq:xieps}
\xi_k^{\varepsilon}= \left\{
\begin{array}{ll}
\xi_k & k\not\in A'\cup\{0\} \\
\frac{\xi_k}{\varepsilon}& k\in A'\cup\{0\}.
\end{array}\right.$$ Note that $\xi_n=\xi_{n+1}$ since $\xi_{n+m+1}=\xi_0$. This allows us to obtain that $$\label{eq:Vepsmn2}
\begin{array}{l}
{\displaystyle\int}_{\Rm^d} |V_{\varepsilon}^{n,m}(t,\xi_0,\xi_{0})|d\xi_0 \leq
{\displaystyle\sum}_{{\mathfrak g}\in{\mathfrak G}_c}
{\displaystyle\int}e^{-(s_0+s_{n+m+1}){\varepsilon}^{-{{\mathfrak m}}}\xi_0^{{\mathfrak m}}}
\prod_{k=1}^{n+m} e^{-s_k(\xi_k^{\varepsilon})^{{\mathfrak m}}}
|\hat u_0(\xi_n)|^2 \\
\qquad {\displaystyle\prod}_{k\in A'({\mathfrak g})}
{\varepsilon}^{-2\alpha} \hat R(\xi_k-{\varepsilon}\xi^{\varepsilon}_{k-1})
\delta(\frac{\xi_{k}}{\varepsilon}-\xi^{\varepsilon}_{k-1}+\xi_{l(k)}-\xi^{\varepsilon}_{l(k)-1})\\
\qquad {\varepsilon}^{-2\alpha} \hat R(\xi_0-{\varepsilon}\xi_n +
{\displaystyle\sum}_{m=1}^M\xi_{q_m}-{\varepsilon}\xi^{\varepsilon}_{q_m-1})
\delta(\xi_{n+1}-\xi_n)
d{\mathbf s}d{\boldsymbol \xi}.
\end{array}$$ Here $d{\boldsymbol \xi}$ also includes the integration in the variable $\xi_0$. The estimates for $V^{n,m}_{\varepsilon}$ here and in subsequent sections rely on integrating selected time variables. All estimates are performed as the following lemma indicates.
\[lem:timeint\] Let $t>0$ given and consider an integral of the form $$\label{eq:intnm1}
I_{n-1} = {\displaystyle\prod}_{k=0}^{n-1}{\displaystyle\int}_0^{t_k({\mathbf s})}
\Big(\prod_{k=0}^{n-1} f_k(s_k)\Big)
\prod_{k=0}^{n-1} ds_k,$$ where $0\leq f_k({\mathbf s})\leq 1$ for $0\leq k\leq n$ and assume that $\int_0^t f_{n-1}(s_{n-1})ds_{n-1} \leq h\wedge t$. Then $$\label{eq:intnm2}
I_{n-2} \leq (h\wedge t) I_{n-1}.$$ Moreover, let ${\mathfrak s}$ be a permutation of the indices $0\leq k\leq
n-1$. Define $I^{\mathfrak s}_{n-1}$ as $I_{n-1}$ with $f_k$ replaced by $f_{{\mathfrak s}(k)}$. Then $I^{\mathfrak s}_{n-1}=I_{n-1}$.
Using the above result with the permutation leaving all indices fixed except ${\mathfrak s}(n-1)=K$ and ${\mathfrak s}(K)=n-1$ for some $0\leq K\leq n-2$ allows us to estimate $I_{n-1}$ by integrating in the $K$th variable.
The derivation of is immediate. We also calculate $$\begin{array}{rclrcl}
I_{n-1} &=& {\displaystyle\int}_{\Rm_+^{n+1}}
\Big(\prod_{k=0}^{n-1} f_k(s_k)\Big) \delta\big(t-{\displaystyle\sum}_{k=0}^n s_k\big)
\prod_{k=0}^n ds_k\\
&=& {\displaystyle\int}_{\Rm_+^{n+1}}
\Big(\prod_{k=0}^{n-1} f_{{\mathfrak s}(k)}(s_{{\mathfrak s}(k)})\Big)
\delta\big(t-{\displaystyle\sum}_{k=0}^n s_{{\mathfrak s}(k)}\big)
\prod_{k=0}^n ds_{k} \\
&=& {\displaystyle\int}_{\Rm_+^{n+1}}
\Big(\prod_{k=0}^{n-1} f_{{\mathfrak s}(k)}(s_k)\Big)
\delta\big(t-{\displaystyle\sum}_{k=0}^n s_{k}\big)
\prod_{k=0}^n ds_k &=& I^{\mathfrak s}_{n-1}.
\end{array}$$
Note that $e^{-s_{n}({\mathbf s})(\xi_{\varepsilon}^n)^{{\mathfrak m}}}$ and $e^{-s_{n+1}({\mathbf s})(\xi_{\varepsilon}^{n+1})^{{\mathfrak m}}}$ are bounded by $1$. We now estimate the integrals in the variables $s_0$, $s_{n+m+1}$, and $s_k$ for $k\in A'$ in . Note that $n+1$ cannot belong to $A'$ and that $n$ does not belong to $A'$ either since either $n=Q$ (last crossing) or $n\in B_0$ is a receiving end of the pairing line $k\to l(k)$. Each integral is bounded by: $$\label{eq:bdtime}
{\displaystyle\int}_0^{\tau\wedge t} e^{-s {\varepsilon}^{-{{\mathfrak m}}}\xi^{{\mathfrak m}}} ds \leq
\dfrac{{\varepsilon}^{{\mathfrak m}}}{\xi^{{\mathfrak m}}} \wedge t.$$ The remaining exponential terms $e^{-s_k(\xi^{\varepsilon}_k)^{{\mathfrak m}}}$ are bounded by $1$. Using lemma \[lem:timeint\], this allows us to obtain that $$\begin{array}{l}
{\displaystyle\int}_{\Rm^d} |V_{\varepsilon}^{n,m}(t,\xi_0,\xi_{0})| d\xi_0
\leq {\displaystyle\sum}_{{\mathfrak g}\in{\mathfrak G}_c} \Big({\displaystyle\int}d\tilde{\mathbf s}\Big)
{\displaystyle\int}|\hat u_0(\xi_n)|^2 \\
\qquad {\displaystyle\prod}_{k\in A'({\mathfrak g})}
{\varepsilon}^{-2\alpha} \Big(\dfrac{{\varepsilon}^{{{\mathfrak m}}}}{\xi_k^{{\mathfrak m}}} \wedge t\Big)
\hat R(\xi_k-{\varepsilon}\xi^{\varepsilon}_{k-1})
\delta(\frac{\xi_{k}}{\varepsilon}-\xi^{\varepsilon}_{k-1}+\xi_{l(k)}-\xi^{\varepsilon}_{l(k)-1})\\
\qquad {\varepsilon}^{-2\alpha} \Big(\dfrac{{\varepsilon}^{{{\mathfrak m}}}}{\xi_0^{{\mathfrak m}}} \wedge t\Big)^2
\hat R(\xi_0-{\varepsilon}\xi_n +
{\displaystyle\sum}_{m=1}^M\xi_{q_m}-{\varepsilon}\xi^{\varepsilon}_{q_m-1})
\delta(\xi_{n+1}-\xi_n) \,d{\boldsymbol \xi}.
\end{array}$$ Here, $d\tilde{\mathbf s}$ corresponds to the integration in the remaining time variables $s_k$ for $k\not\in A'\cup\{n+m+1\}$. There are $2{{\bar n}}-1-({{\bar n}}+1)={{\bar n}}-2$ such variables. Note the square on the last line, which comes from integrating in both variables $s_0$ and $s_{n+m+1}$.
The delta functions allow us to integrate in the variables $\xi_{l(k)}$ for $k\in A'({\mathfrak g})$ and the initial condition $\hat
u_0(\xi_n)$ in the variable $\xi_n$. Thanks to lemma \[lem:bound\] below, the power spectra allow us to integrate in the remaining variables in $A'\cup\{0\}$. The integrals in the variables in $A'$ are all bounded by $\rho_f$ defined in lemma \[lem:bound\] whereas the integral in $\xi_0$ results in a bound equal to ${\varepsilon}^\beta\rho_f$, where ${\varepsilon}^\beta$ is defined in . As a consequence, we have the bound $${\displaystyle\int}_{\Rm^d} |V_{\varepsilon}^{n,m}(t,\xi_0,\xi_{0})| d\xi_0
\leq {\displaystyle\sum}_{{\mathfrak g}\in{\mathfrak G}_c}
\Big({\displaystyle\int}d\tilde{\mathbf s}\Big) \rho_f^{{{\bar n}}-1} \|\hat u_0\|^2
\rho_f{\varepsilon}^\beta ={\displaystyle\sum}_{{\mathfrak g}\in{\mathfrak G}_c}
\Big({\displaystyle\int}d\tilde{\mathbf s}\Big) \rho_f^{{{\bar n}}} {\varepsilon}^\beta\|\hat u_0\|^2.$$ Using Stirling’s formula, we find that $|{\mathfrak G}_c|<\frac{(2{{\bar n}}-1)!}{2^{{{\bar n}}-1}({{\bar n}}-1)!}$ is bounded by $(\frac{2{{\bar n}}}{e})^{{{\bar n}}}$. It remains to evaluate the integrals in time. We verify that $$\label{eq:sumtime}
\prod_{k=0}^{n-1} {\displaystyle\int}_0^{t_k({\mathbf s})} ds_0\cdots ds_{n-1}
=\dfrac{t^n}{n!},\qquad
t_k({\mathbf s}) = t-s_0-\ldots - s_{k-1}.$$ Let $\bar p=\bar p({\mathfrak g})$ be the number of $s_k$ for $k\leq n$ in $\tilde{\mathbf s}$ and $\bar q=\bar q({\mathfrak g})$ be the number of $s_k$ for $k\geq n+1$ in $\tilde{\mathbf s}$, with $\bar p+\bar q={{\bar n}}-1$. Using , we thus find that $$\Big({\displaystyle\int}d\tilde{\mathbf s}\Big) = \dfrac{t^{\bar p}}{\bar p!}
\dfrac{t^{\bar q}}{\bar q!}
= \dfrac{t^{{{\bar n}}-1}}{( {{\bar n}}-1)!} {{{\bar n}}-1 \choose \bar p}
\leq t^{{{\bar n}}-1} \Big(\dfrac{{{\bar n}}-1}{2e}\Big)^{-{{\bar n}}+1}
\leq t^{{{\bar n}}-1} {{\bar n}}\Big(\dfrac{{{\bar n}}}{2e}\Big)^{-{{\bar n}}}$$ using Stirling’s formula. This shows that $$\label{eq:boundgraphs}
{\displaystyle\sum}_{{\mathfrak g}\in{\mathfrak G}_c} \Big({\displaystyle\int}d\tilde{\mathbf s}\Big) \leq
\dfrac{{{\bar n}}}{T} (4\rho_f T)^{{\bar n}},$$ uniformly for $t\in (0,T)$. We thus need to choose $T$ sufficiently small so that $4\rho_f T<1$. Then, for ${{\mathfrak r}}$ such that $4\rho_f T<{{\mathfrak r}}^2<1$, we find that $$\label{eq:bdVeps1}
{\displaystyle\int}|V_{\varepsilon}^{n,m}(t,\xi_0,\xi_{0})| d\xi_0
\leq C {{\mathfrak r}}^{n+m} {\varepsilon}^\beta \|\hat u_0\|^2,$$ for some positive constant $C$. It remains to sum over $n$ and $m$ to obtain that $$\label{eq:boundI}
\big|{\mathbb E}\{I_{\varepsilon}(t)\}-{\displaystyle\int}_{\Rm^d} {\mathbb E}\{\hat u_{\varepsilon}(t,\xi)\}^2 d\xi
\big| \leq \dfrac{C}{{{\mathfrak r}}^2} {\varepsilon}^\beta \|\hat u_0\|^2.$$ We shall analyze the non-crossing terms generating $|{\mathbb E}\{\hat
u_{\varepsilon}(t,\xi)\}|^2$ shortly. Before doing so, we analyze the influence of the crossing terms on $X_{\varepsilon}$. We can verify that the error term ${\varepsilon}^\beta$ in is optimal, for instance by looking at the contribution of the graph with $n=m=1$.
#### Analysis of the crossing terms in $X_{\varepsilon}$.
It turns out that the contribution of the crossing terms is smaller for the moment $X_{\varepsilon}$ than it is for the energy $I_{\varepsilon}$. More precisely, we show that the smallest contribution to the variance of $X_{\varepsilon}$ is of order ${\varepsilon}^{d-2\alpha}$ for graphs in ${\mathfrak G}_{cs}$ and of order ${\varepsilon}^{d-2\alpha+\beta}$ for the other crossing graphs.
We come back to and this time perform the change of variables $\xi_k\to\frac{\xi_k}{\varepsilon}$ for $k\in A'$ only. We re-define $$\label{eq:xieps2}
\xi_k^{\varepsilon}= \left\{
\begin{array}{ll}
\xi_k & k\not\in A' \\
\frac{\xi_k}{\varepsilon}& k\in A',
\end{array}\right.$$ and find that $$\label{eq:Vepsmn3}
\begin{array}{l}
V_{\varepsilon}^{n,m}(t,\xi_0,\xi_{n+m+1}) ={\displaystyle\sum}_{{\mathfrak g}\in{\mathfrak G}_c}
{\displaystyle\int}\prod_{k=0}^{n+m+1} e^{-s_k(\xi_k^{\varepsilon})^{{\mathfrak m}}}
\hat u_0(\xi_n) \bar {\hat u}_0(\xi_{n+1}) \\
\qquad {\displaystyle\prod}_{k\in A'({\mathfrak g})}
{\varepsilon}^{-2\alpha} \hat R(\xi_k-{\varepsilon}\xi^{\varepsilon}_{k-1})
\delta(\frac{\xi_{k}}{\varepsilon}-\xi^{\varepsilon}_{k-1}+\xi_{l(k)}-\xi^{\varepsilon}_{l(k)-1})\\
\qquad {\varepsilon}^{d-2\alpha} \hat R({\varepsilon}(\xi_Q-\xi^{\varepsilon}_{Q-1}))
\delta(\xi_{n+m+1}-\xi_{n+1}+\xi_n-\xi_0)
d{\mathbf s}d{\boldsymbol \xi}.
\end{array}$$ Note that neither $n$ nor $n+m+1$ belong to $A'({\mathfrak g})$. For each $k\in
A'({\mathfrak g})$, we integrate in $s_k$ and obtain using that $$\label{eq:Vepsmnbd1}
\begin{array}{l}
|V_{\varepsilon}^{n,m}(t,\xi_0,\xi_{n+m+1})| \leq {\displaystyle\sum}_{{\mathfrak g}\in{\mathfrak G}_c}
{\displaystyle\int}{\displaystyle\prod}_{k\not\in A'({\mathfrak g})} e^{-s_k \xi_k^{{\mathfrak m}}}
|\hat u_0(\xi_n) \bar {\hat u}_0(\xi_{n+1})| \\
\qquad
{\displaystyle\prod}_{k\in A'({\mathfrak g})}
{\varepsilon}^{-2\alpha} \Big(\dfrac{{\varepsilon}^{{\mathfrak m}}}{\xi_k^{{\mathfrak m}}}\wedge t\Big)
\hat R(\xi_k-{\varepsilon}\xi^{\varepsilon}_{k-1})
\delta(\frac{\xi_{k}}{\varepsilon}-\xi^{\varepsilon}_{k-1}+\xi_{l(k)}-\xi^{\varepsilon}_{l(k)-1})
\\\qquad
{\varepsilon}^{d-2\alpha} \hat R({\varepsilon}(\xi_Q-\xi^{\varepsilon}_{Q-1}))
\delta(\xi_{n+m+1}-\xi_{n+1}+\xi_n-\xi_0)
d\tilde{\mathbf s}d{\boldsymbol \xi}.
\end{array}$$ By assumption on $\hat R(\xi)$, we know the existence of a constant $\hat R_\infty$ such that $$\label{eq:unifbdhatR}
{\varepsilon}^{d-2\alpha} \hat R({\varepsilon}(\xi_Q-\xi^{\varepsilon}_{Q-1})) \leq
{\varepsilon}^{d-2\alpha} \hat R_\infty.$$ This is where the factor ${\varepsilon}^{d-2\alpha}$ arises. We need however to ensure that the integral in $\xi_Q$ is well-defined. We have two possible scenarios: either $Q=n$ or $n\in B_0$. When $Q=n$, the integration in $\xi_Q$ is an integration in $\xi_n$ for which we use $\hat u_0(\xi_n)$. When $n\in B_0$, we thus have $n=l(k_0)$ for some $k_0$ and we replace the delta function involving $\xi_n$ by a delta function involving $\xi_Q$ given equivalently by $$\label{eq:deltause}
\delta(\xi_Q-\xi_{Q-1} + \xi_0-\xi_n + {\displaystyle\sum}_{m=1}^M\xi_{q_m}-\xi_{q_m-1}).$$ In either scenario, we can integrate in the variable $\xi_Q$ without using the term $\hat R({\varepsilon}(\xi_Q-\xi_{Q-1}))$. We use the inequality $$\label{eq:ineq}
|\hat u_0(\xi_n) \bar {\hat u}_0(\xi_{n+1})|
\leq \dfrac12\Big(|\hat u_0(\xi_n) |^2 + |\hat u_0(\xi_n-\xi_0+\xi_{n+m+1})|^2
\Big),$$ to obtain the bound $$\label{eq:Vepsmnbd2}
|V_{\varepsilon}^{n,m}(t,\xi_0,\xi_{n+m+1})| \leq {\varepsilon}^{d-2\alpha} \hat R_\infty
{\displaystyle\sum}_{{\mathfrak g}\in{\mathfrak G}_c}
\Big({\displaystyle\int}d\tilde s\Big) \rho_f^{{{\bar n}}-1} \|\hat u_0\|^2.$$ The bound is uniform in $\xi_0$ and $\xi_{n+m+1}$. Using and , we obtain $$\label{eq:Vepsmnbd3}
|V_{\varepsilon}^{n,m}(t,\xi_0,\xi_{n+m+1})| \leq
{\varepsilon}^{d-2\alpha} {{\mathfrak r}}^{n+m}\|\hat u_0\|^2.$$ After summation in $n,m\in\Nm$, we thus find that $$\label{eq:Xeps2}
{\mathbb E}\{ (X_{\varepsilon}- {\mathbb E}\{X_{\varepsilon}\})^2\} \leq \dfrac{C}{{{\mathfrak r}}^2}
{\varepsilon}^{d-2\alpha} \|\hat u_0\|^2 \|\hat M\|_1^2.$$ Similarly, by setting $\xi_{n+m+1}=\xi_0$, we find that $$\label{eq:locenergy}
\Big|{\mathbb E}\Big\{ {\displaystyle\int}_{\Rm^d} |\hat u_{\varepsilon}|^2(t,\xi) \varphi(\xi) d\xi
\Big\} - {\displaystyle\int}_{\Rm^d}
|{\mathbb E}\{\hat u_{\varepsilon}(t,\xi)\}|^2 \varphi(\xi) d\xi \Big|
\leq \dfrac{C}{{{\mathfrak r}}^2}{\varepsilon}^{d-2\alpha} \|\hat u_0\|^2 \|\varphi\|_1,$$ for any test function $\varphi\in L^1(\Rm^d)$. This local energy estimate is to be compared with the global estimate obtained in .
#### Analysis of the leading crossing terms in $X_{\varepsilon}$.
The preceding estimate on $X_{\varepsilon}$ may be refined as only the crossing graphs in ${\mathfrak G}_{cs}$ have contributions of order ${\varepsilon}^{d-2\alpha}$. We return to the bound and obtain that $$\label{eq:Vepsmnbd4}
\begin{array}{l}
|V_{\varepsilon}^{n,m}(t,\xi_0,\xi_{n+m+1})| \leq
{\varepsilon}^{d-2\alpha} \hat R_\infty
{\displaystyle\sum}_{{\mathfrak g}\in{\mathfrak G}_c}
{\displaystyle\int}{\displaystyle\prod}_{k\not\in A'({\mathfrak g})} e^{-s_k(\xi_k)^{{\mathfrak m}}}
|\hat u_0(\xi_n) \bar {\hat u}_0(\xi_{n+1})| \\
\qquad
{\displaystyle\prod}_{k\in A'({\mathfrak g})}
{\varepsilon}^{-2\alpha} \Big(\dfrac{{\varepsilon}^{{\mathfrak m}}}{\xi_k^{{\mathfrak m}}}\wedge t\Big)
\hat R(\xi_k-{\varepsilon}\xi^{\varepsilon}_{k-1})
\delta(\frac{\xi_{k}}{\varepsilon}-\xi^{\varepsilon}_{k-1}+\xi_{l(k)}-\xi^{\varepsilon}_{l(k)-1})
\\\qquad
\delta(\xi_{n+m+1}-\xi_{n+1}+\xi_n-\xi_0)
d\tilde{\mathbf s}d{\boldsymbol \xi}.
\end{array}$$ The ${{\bar n}}+3$ variables in time left are $s_0$, $s_{n+1}$, $s_Q$, $s_{l(Q)}$, and the ${{\bar n}}-1$ variables $s_{l(A'({\mathfrak g}))}$.
Let ${\mathfrak g}\in{\mathfrak G}_{c}$. Let us assume that for some $k$ such that $(\xi_k,\xi_{l(k)})$ is not a crossing pair, we have $l(k)-1>k$, i.e., ${\mathfrak g}\in{\mathfrak G}_{ncs}$. The non-crossing pairs are not affected by the possible change of a delta function involving $\xi_n$ to a delta function involving $\xi_Q$. We may then integrate in the variable $s_{l(k)}$ and obtain the bound for the integral $$\begin{array}{ll}
{\varepsilon}^{d-2\alpha} \hat R_\infty
{\displaystyle\int}d\tilde{\mathbf s}d{\boldsymbol \xi}|\hat u_0(\xi_n) \bar {\hat u}_0(\xi_{n+1})|
\delta(\xi_{n+m+1}-\xi_{n+1}+\xi_n-\xi_0) \\[3mm]
{\displaystyle\prod}_{k\in A'({\mathfrak g})}
\Big(\dfrac{{\varepsilon}^{{\mathfrak m}}}{\xi_k^{{\mathfrak m}}}\wedge t\Big)
\Big( \dfrac{{\varepsilon}^{{\mathfrak m}}}{|\xi_k-{\varepsilon}\xi_{k-1}^{\varepsilon}- {\varepsilon}\xi_{l(k)-1}^{\varepsilon}|^{{\mathfrak m}}} \wedge t\Big)
\hat R(\xi_k-{\varepsilon}\xi^{\varepsilon}_{k-1})
\delta(\frac{\xi_{k}}{\varepsilon}-\xi^{\varepsilon}_{k-1}+\xi_{l(k)}-\xi^{\varepsilon}_{l(k)-1})\\
\leq {\varepsilon}^\beta {\varepsilon}^{d-2\alpha} \Big({\displaystyle\int}d\tilde{\mathbf s}\Big)
\hat R_\infty \rho_f^{{{\bar n}}-1}\|\hat u_0\|^2,
\end{array}$$ thanks to lemma \[lem:bound\] below. The summation over all graphs in ${\mathfrak G}_{ncs}$ of any quantity derived from $V_{\varepsilon}^{n,m}(t,\xi_0,\xi_{n+m+1})$ is therefore ${\varepsilon}^\beta$ smaller than the corresponding sum over all graphs in ${\mathfrak G}_c$. We thus see that any non-crossing pair has to be of the form $l(k)-1=k$, i.e., a simple pair, in order for the graph to correspond to a contribution of order ${\varepsilon}^{d-2\alpha}$.
Let us consider the graphs composed of crossings and simple pairs. We may delete the simple pairs from the graph since they contribute integrals of order $O(1)$ thanks to lemma \[lem:bound\] below and assume that the graph is composed of crossings only, thus with $n=m$ and $Q=n$ after deletion of the simple pairs. Let us consider $k<n$ with $l(k)\geq n+1$ so that the delta function $$\delta(\dfrac{\xi_{k}}{\varepsilon}- \xi_{k-1}^{\varepsilon}+ \xi_{l(k)}-\xi_{l(k)-1})$$ is present in the integral defining $V^{n,m}_{\varepsilon}$. We find for the same reason as above that the contribution of the corresponding graph is of order ${\varepsilon}^{d-2\alpha}{\varepsilon}^\beta$ by integration in the variable $s_{l(k)}$. As a consequence, the only graph composed exclusively of crossing pairs that generates a contribution of order ${\varepsilon}^{d-2\alpha}$ is the graph with $n=m=1$. This concludes our proof that the contribution of order ${\varepsilon}^{d-2\alpha}$ in $V^{n,m}_{\varepsilon}$ is given by the $nm$ graphs in ${\mathfrak G}_{cs}$ when both $n$ and $m$ are even numbers (otherwise, ${\mathfrak G}_{cs}$ is empty). All other graphs in ${\mathfrak G}_c$ provide a contribution of order ${\varepsilon}^\beta$ smaller than what we obtained in . In other words, let us define $$\label{eq:Vepsmn}
\begin{array}{l}
V_{{\varepsilon},s}^{n,m}(t,\xi_0,\xi_{n+m+1}) ={\displaystyle\sum}_{{\mathfrak g}\in{\mathfrak G}_{cs}}
{\displaystyle\int}\prod_{k=0}^{n+m+1} e^{-s_k\xi_k^{{\mathfrak m}}}
\hat u_0(\xi_n) \bar {\hat u}_0(\xi_{n+1}) \\
\qquad {\displaystyle\prod}_{k\in A_0({\mathfrak g})}
{\varepsilon}^{d-2\alpha} \hat R({\varepsilon}(\xi_k-\xi_{k-1}))
\delta(\xi_{k}-\xi_{k-1}+\xi_{l(k)}-\xi_{l(k)-1})
d{\mathbf s}d{\boldsymbol \xi}.
\end{array}$$ We have found that $$\label{eq:bounddeltaV}
|V_{{\varepsilon}}^{n,m}(t,\xi_0,\xi_{n+m+1})-V_{{\varepsilon},s}^{n,m}(t,\xi_0,\xi_{n+m+1})|
\lesssim {\varepsilon}^{d-2\alpha+\beta} {{\mathfrak r}}^{n+m}\|\hat u_0\|^2.$$
Analysis of non-crossing graphs {#sub:noncrossing}
-------------------------------
We now apply the estimates obtained in the preceding section to the analysis of the moments $U_{\varepsilon}^n(t)$ defined in and given more explicitly in . Our objective is to show that only the simple graph ${\mathfrak g}$ contributes a term of order $O(1)$ in whereas all other graphs in ${\mathfrak G}_{ns}$ contribute (summable in $n$) terms of order $O({\varepsilon}^\beta)$. Note that $n=2{{\bar n}}$, for otherwise, $U_{\varepsilon}^n(t)=0$. We recall that the simple graph is defined by $l(k)=k+1$. We thus define the simple graph contribution as $$\label{eq:Uepsns}
\begin{array}{rcl}
U_{{\varepsilon},s}^n(t,\xi_0) &=& {\mathcal U}^n_{\varepsilon}(t,\xi_0) \hat u_0(\xi_0)
\\ {\mathcal U}^n_{\varepsilon}(t,\xi_0)&=&
{\displaystyle\int}\prod_{k=0}^n e^{-s_k\xi_k^{{\mathfrak m}}} \prod_{k=0}^{{{\bar n}}-1}
{\varepsilon}^{d-2\alpha} \hat R({\varepsilon}(\xi_{2k+1}-\xi_{2k}))
\delta(\xi_{2(k+1)}-\xi_{2k}) d{\mathbf s}d{\boldsymbol \xi},
\end{array}$$ and $$\label{eq:Uepss}
U_{{\varepsilon},s}(t,\xi_0)={\displaystyle\sum}_{n\in\Nm} U_{{\varepsilon},s}^n(t,\xi_0)
:= {\mathcal U}_{\varepsilon}(t,\xi_0) \hat u_0(\xi_0).$$ For all $k\in A_0$, we perform the change of variables $\xi_k\to\frac{\xi_k}{\varepsilon}$ and (re-)define as before $$\label{eq:xieps3}
\xi_k^{\varepsilon}= \left\{
\begin{array}{ll}
\xi_k & k\not\in A_0 \\
\frac{\xi_k}{\varepsilon}& k\in A_0.
\end{array}\right.$$ This gives $$\label{eq:Uepsn2}
\begin{array}{l}
U_{\varepsilon}^n(t,\xi_0) = \hat u_0(\xi_0)
{\displaystyle\sum}_{{\mathfrak g}\in{\mathfrak G}} {\displaystyle\int}\prod_{k=0}^n e^{-s_k(\xi_k^{\varepsilon})^{{\mathfrak m}}}
\\\qquad {\displaystyle\prod}_{k\in A_0({\mathfrak g})}
{\varepsilon}^{-2\alpha} \hat R(\xi_k-{\varepsilon}\xi^{\varepsilon}_{k-1})
\delta(\dfrac{\xi_{k}}{\varepsilon}-\xi^{\varepsilon}_{k-1}+\xi_{l(k)}-\xi^{\varepsilon}_{l(k)-1})
d{\mathbf s}d{\boldsymbol \xi}.
\end{array}$$ Assuming that $l(k)-1>k$ for one of the pairings, we obtain as in the analysis leading to the following bound for the corresponding graph: $$\begin{array}{l}
|\hat u_0(\xi_0)| {\displaystyle\int}d\tilde{\mathbf s}d{\boldsymbol \xi}{\displaystyle\prod}_{k\in A_0({\mathfrak g})}
{\varepsilon}^{-2\alpha}\Big(\dfrac{{\varepsilon}^{{\mathfrak m}}}{\xi_k^{{\mathfrak m}}}\wedge t\Big)
\Big( \dfrac{{\varepsilon}^{{\mathfrak m}}}{|\xi_k-{\varepsilon}\xi_{k-1}^{\varepsilon}- {\varepsilon}\xi_{l(k)-1}^{\varepsilon}|^{{\mathfrak m}}} \wedge t\Big)\\\qquad\qquad
\hat R(\xi_k-{\varepsilon}\xi^{\varepsilon}_{k-1})
\delta(\dfrac{\xi_{k}}{\varepsilon}-\xi^{\varepsilon}_{k-1}+\xi_{l(k)}-\xi^{\varepsilon}_{l(k)-1})
\\[2mm]
\leq {\varepsilon}^{\beta} \Big({\displaystyle\int}d\tilde{\mathbf s}\Big)
\rho_f^{{{\bar n}}}|\hat u_0(\xi_0)|.
\end{array}$$ This shows that $$\label{eq:limUepsn}
|U_{\varepsilon}^n(t,\xi_0) - U_{{\varepsilon},s}^n(t,\xi_0)|
\leq |\hat u_0(\xi_0)| {\varepsilon}^{\beta} {{\mathfrak r}}^{n},$$ so that $$\label{eq:limavueps}
|{\mathbb E}\{\hat u_{\varepsilon}\}(t,\xi)- U_{{\varepsilon},s}(t,\xi)| \lesssim \dfrac{1}{{{\mathfrak r}}}
{\varepsilon}^\beta |\hat u_0(\xi)| ,$$ at least for sufficiently small times $t\in(0,T)$ such that $4\rho_fT<1$. It remains to analyze the limit of $U_{{\varepsilon},s}(t,\xi)$ to obtain the limiting behavior of $X_{\varepsilon}$ and $I_{{\varepsilon},\varphi}$. This analysis is carried out in the next section. Another application of lemma \[lem:bound\] shows that $U_{{\varepsilon},s}(t,\xi)$ is square integrable and that its $L^2(\Rm^d)$ norm is bounded by $\|\hat
u_0\|$. In other words, we have constructed a weak solution $\hat
u_{\varepsilon}(t)\in L^2(\Omega\times \Rm^d)$ to since the series converges uniformly in $L^2(\Omega\times \Rm^d)$ for sufficiently small times $t\in(0,T)$ such that $4\rho_fT<1$. Collecting the results obtained in and , we have shown that $$\begin{array}{rcl}
\|(\hat u_{\varepsilon}-U_{{\varepsilon},s})(t)\|_{L^2(\Omega\times \Rm^d)} &\lesssim&
{\varepsilon}^{\frac{\beta}2} \|\hat u_0\|_{L^2(\Rm^d)},
\end{array}$$ where $U_{{\varepsilon},s}$ is the deterministic term given in . The analysis of $U_{{\varepsilon},s}$ and that of $X_{\varepsilon}$ is postponed to section \[sec:limits\], after we state and prove lemma \[lem:bound\], which allows us to analyze the contributions of the different graphs.
\[lem:bound\] Let us assume that $\hat R$ is bounded by a smooth radially symmetric, decreasing function $f(r)$. We also assume that $f(r)\leq
\tau_f r^{-{{\mathfrak n}}}$ for some $0\leq{{\mathfrak n}}<d-{{\mathfrak m}}$ in dimension $d>{{\mathfrak m}}$ and ${{\mathfrak n}}=0$ when $d\leq{{\mathfrak m}}$. Then we obtain the following estimates.\
For $d>{{\mathfrak m}}$, we have $${\displaystyle\int}\dfrac{1}{|\xi_k|^{{\mathfrak m}}} \hat R(\xi_k-y) d\xi_k \leq \rho_f
:= c_d {\displaystyle\int}_0^\infty \dfrac{1}{|\xi|^{{\mathfrak m}}} f(|\xi|)|\xi|^{d-1}d|\xi|
\,\,\vee \,\, \tau_f,$$ uniformly in $y\in\Rm^d$, where $c_d=|S^{d-1}|$ and $a\vee
b=max(a,b)$. Moreover, $${\displaystyle\int}\dfrac{1}{|\xi_k|^{{\mathfrak m}}} \hat R(\xi_k-y)
\Big( \dfrac{{\varepsilon}^{{\mathfrak m}}}{|\xi_k-z|^{{\mathfrak m}}}\wedge t\Big) d\xi_k
\lesssim \,\rho_f \,
\left\{
\begin{array}{ll}
{\varepsilon}^{{{\mathfrak m}}-{{\mathfrak n}}} \quad &d> 2{{\mathfrak m}}-{{\mathfrak n}}\\
{\varepsilon}^{{{\mathfrak m}}-{{\mathfrak n}}}|\ln{\varepsilon}| \quad &d=2{{\mathfrak m}}-{{\mathfrak n}}\\
{\varepsilon}^{d-{{\mathfrak m}}-{{\mathfrak n}}} \quad & {{\mathfrak m}}<d<2{{\mathfrak m}}-{{\mathfrak n}}.
\end{array}
\right.$$ For $d={{\mathfrak m}}$, we define $\rho_f = c_d f(0)$ and have $${\displaystyle\int}\Big( \dfrac{{\varepsilon}^{{\mathfrak m}}}{|\xi_k-z|^{{\mathfrak m}}}\wedge t\Big)^l\hat R(\xi_k-y)
d\xi_k \lesssim \,\rho_f \,\left\{
\begin{array}{ll}
{\varepsilon}^{{\mathfrak m}}|\ln{\varepsilon}| & l=1 \\
{\varepsilon}^{{\mathfrak m}}& l=2.
\end{array}\right.$$ For $d< {{\mathfrak m}}$, we have $${\displaystyle\int}\Big( \dfrac{{\varepsilon}^{{\mathfrak m}}}{|\xi_k-z|^{{\mathfrak m}}}\wedge t\Big)^l\hat R(\xi_k-y)
d\xi_k \lesssim {\varepsilon}^d,\quad l\geq1.$$
Once $\hat R$ is bounded above by a decreasing, radially symmetric, function $f(r)$, the above integrals are maximal when $y=z=0$ thanks to lemma \[lem:decay\] below since $|\xi|^{-{{\mathfrak m}}}$ and $({\varepsilon}^{{\mathfrak m}}|\xi|^{-{{\mathfrak m}}}\wedge t)$ are radially symmetric and decreasing. The first bound is then obvious and defines $\rho_f$. The second bound is obvious in dimension $d>2{{\mathfrak m}}$ since $|\xi_k|^{-2{{\mathfrak m}}}$ is integrable.
All the bounds in the lemma are thus obtained from a bound for $${\displaystyle\int}_0^\infty \Big( \dfrac{{\varepsilon}^{{\mathfrak m}}}{r^{{\mathfrak m}}}\wedge t\Big)^l
r^{d-1} f(r) dr.$$ We obtain that the above integral restricted to $r\in (1,\infty)$ is bounded by a constant times ${\varepsilon}^{{{\mathfrak m}}l} \rho_f$ for $d\geq{{\mathfrak m}}$ and by a constant times ${\varepsilon}^{{{\mathfrak m}}l}$ for $d<{{\mathfrak m}}$. It thus remains to bound the integral on $r\in(0,1)$, which is equal to $${\displaystyle\int}_0^{{\varepsilon}t^{-\frac1{{{\mathfrak m}}}}} t^l r^{d-1} f(r) dr
+ {\displaystyle\int}_{{\varepsilon}t^{-\frac1{{{\mathfrak m}}}}}^1 \dfrac{{\varepsilon}^{l{{\mathfrak m}}}}{r^{l{{\mathfrak m}}}}r^{d-1} f(r)dr.$$ Replacing $f(r)$ by $\tau_f r^{-{{\mathfrak n}}}$, we find that the first integral is bounded by a constant times ${\varepsilon}^{d-{{\mathfrak n}}}$ and the second integral by a constant times ${\varepsilon}^{d-{{\mathfrak n}}} \vee {\varepsilon}^{l{{\mathfrak m}}}$ when $d-{{\mathfrak n}}-l{{\mathfrak m}}\not=0$ and ${\varepsilon}^{2{{\mathfrak m}}} |\ln{\varepsilon}|$ when $d=2{{\mathfrak m}}-{{\mathfrak n}}$. It remains to divide through by ${\varepsilon}^{{\mathfrak m}}$ when $l=2$ to obtain the desired results.
\[lem:decay\] Let $f$, $g$, and $h$ be non negative, bounded, integrable, and radially symmetric functions on $\Rm^d$ that are decreasing as a function of radius. Then the integral $$\label{eq:intab}
I_{\zeta,\tau} = {\displaystyle\int}_{\Rm^d} f(\xi-\zeta)g(\xi-\tau)h(\xi) d\xi,$$ which is well defined, is maximal at $\zeta=\tau=0$.
In a first step, we rotate $\zeta$ to align it with $\tau$. The first claim is that the integral cannot increase while doing so. Then we send $\zeta$ and $\tau$ to $0$. The second claim is that the integral again does not increase.
We assume that the functions $f$, $g$, and $h$ are smooth and obtain the result in the general case by density. We choose a system of coordinates so that $\tau=|\tau|e_1$, where $(e_1,\ldots,e_d)$ is an orthonormal basis of $\Rm^d$, and $\zeta=|\zeta|\hat\theta$ with $\hat \theta=(\cos\theta,\sin\theta,0,\ldots,0)$. Without loss of generality, we may assume that $\theta\in(0,\pi)$. Then $I_{\zeta,\tau}$ may be recast as $I_\theta$ and we find that $$I_\theta = {\displaystyle\int}_0^\infty |\xi|^{d-1} h(|\xi|) J_\theta(|\xi|) d|\xi|,$$ where we denote $h(|\xi|)\equiv h(\xi)$ with the same convention for $f$ and $g$ and define $$J_\theta(|\xi|) = {\displaystyle\int}_{S^{d-1}} f(|\xi|\psi-\zeta)g(|\xi|\psi-\tau)
d\psi.$$ It is sufficient to show that $\partial_\theta J_\theta \leq0$. We find $$\partial_\theta J_\theta = {\displaystyle\int}_{S^{d-1}} -\hat\theta^\perp\cdot
\nabla f(|\xi|\psi-\zeta) g(|\xi|\psi-\tau) d\psi,$$ with $\hat\theta^\perp=(-\sin\theta,\cos\theta,0,\ldots,0)$. We decompose the sphere as $\psi=(\psi\cdot\hat\theta,\tilde\psi)$ and find, for some positive weight $w(\mu)$ that $$\begin{array}{rcl}
\partial_\theta J_\theta &=& {\displaystyle\int}_{-1}^1 d(\psi\cdot\hat\theta)
(-f')(||\xi|\psi-\zeta|) w(\psi\cdot\hat\theta)
{\displaystyle\int}_{S^{d-2}} (\hat\theta^\perp\cdot\tilde\psi) g(|\xi|\psi-\tau)
d\tilde\psi.
\end{array}$$ We now observe that $$\begin{array}{rcl}
&& {\displaystyle\int}_{S^{d-2}} (\hat\theta^\perp\cdot\tilde\psi) g(|\xi|\psi-\tau)
d\tilde\psi\\[3mm]
&=&{\displaystyle\int}_{\hat\theta^\perp\cdot\tilde\psi>0}
(\hat\theta^\perp\cdot \tilde\psi )\big(
g(||\xi|(\hat\theta\cdot\psi\hat\theta+\tilde\psi)-\tau|)
- g(||\xi|(\hat\theta\cdot\psi\hat\theta-\tilde\psi)-\tau|) \big)
d\tilde\psi\,\leq\,0,
\end{array}$$ as $||\xi|(\hat\theta\cdot\psi\hat\theta+\tilde\psi)-\tau|\leq||\xi|(\hat\theta\cdot\psi\hat\theta-\tilde\psi)-\tau|$ by construction. Indeed, we find that $||\xi|(\hat\theta\cdot\psi\hat\theta\pm\tilde\psi)-\tau|^2-|\xi|^2-|\tau|^2+2|\tau||\xi|\hat\theta\cdot\psi\hat\theta\cdot\tau=\pm2|\tau||\xi|\tilde\psi\cdot\tau=\pm2|\tau||\xi|\hat\theta^\perp\cdot\tau$ whereas $\hat\theta^\perp\cdot\tau=-\sin\theta|\tau|<0$ by construction. This shows that $|\xi|(\hat\theta\cdot\psi\hat\theta+\tilde\psi)$ is closer to $\tau$ than $|\xi|(\hat\theta\cdot\psi\hat\theta-\tilde\psi)$ is, and since $g(r)$ is decreasing, that $\partial_\theta
J_\theta\leq0$. This concludes the proof of the first claim.
If $\beta=0$ or $\tau=0$, we set $b=0$ below. Otherwise, we may assume without loss of generality that $\tau=-b\zeta$ for some $b>1$. We still define $\zeta=|\zeta|\hat\theta$. We now define the integral $ I_a = I_{a\zeta,b\zeta},\,0\leq a\leq1,
$ and compute $$\partial_a I_a = {\displaystyle\int}_{\Rm^d} -\zeta\cdot\nabla f(\xi-a\zeta)
g(\xi+b\zeta)h(\xi) d\xi=
{\displaystyle\int}_{\Rm^d} -\zeta\cdot\nabla f(\xi) g(\xi+(b-a)\zeta)h(\xi+a\zeta)
d\xi.$$ Define $l(\xi,\zeta)=g(\xi+(b-a)\zeta)h(\xi+a\zeta)$. Then because $f$ is radially symmetric, we have $$\partial_a I_a = {\displaystyle\int}_0^\infty m(|\xi|) |\xi|^{d-1} d |\xi|,\,\,\quad
m(|\xi|) = -f'(|\xi|) {\displaystyle\int}_{S^{d-1}} \hat\theta\cdot\psi\,
l(|\xi|\psi,\zeta) d\psi.$$ We recast $$m(|\xi|) = -f'(|\xi|) {\displaystyle\int}_{\hat\theta\cdot\psi>0} (\hat\theta\cdot\psi)
\big( l(|\xi|\psi,\zeta) - l (-|\xi|\psi,\zeta)\big) d\psi
\leq0,$$ since $\big||\xi|\psi+\gamma\zeta\big|\geq
\big|-|\xi|\psi+\gamma\zeta\big|$ by construction for all $\gamma>0$ and thus for $\gamma=a$ and $\gamma=b-a$. This shows that $\partial_\alpha I_\alpha\leq0$ and concludes the proof of the second claim.
Homogenized limit and Gaussian fluctuations {#sec:limits}
===========================================
In this section, we conclude the proof of theorems \[thm:convergence\] and \[thm:fluct\].
Homogenization theory for $u_{\varepsilon}$ {#sec:convergence}
-------------------------------------------
We come back to the analysis of $U_{{\varepsilon},s}(t,\xi)$ defined in . Since only the simple graph is retained in the definition of mean field solution $U_{{\varepsilon},s}(t,\xi)$, the equation it satisfies may be obtained from that for $\hat u_{\varepsilon}$ by simply assuming the mean field approximation ${\mathbb E}\{\hat q_{\varepsilon}\hat q_{\varepsilon}\hat u_{\varepsilon}\} \sim {\mathbb E}\{\hat q_{\varepsilon}\hat q_{\varepsilon}\}{\mathbb E}\{\hat u_{\varepsilon}\}$ since the Duhamel expansions then agree. As a consequence, we find that ${{U_{{\varepsilon},s}}}$ is the solution to the following integral equation $$\label{eq:intUes}
\begin{array}{l}
{{U_{{\varepsilon},s}}}(t,\xi)= e^{-t\xi^{{\mathfrak m}}} \hat u_0(\xi) \\\quad + {\displaystyle\int}_0^te^{-\xi^{{\mathfrak m}}s}
{\displaystyle\int}_0^{t-s} e^{-\xi_1^{{\mathfrak m}}s_1} {\displaystyle\int}{\varepsilon}^{d-2\alpha}\hat R({\varepsilon}(\xi_1-\xi))
{{U_{{\varepsilon},s}}}(t-s-s_1,\xi) d\xi_1 dsds_1\\
= e^{-t\xi^{{\mathfrak m}}} \hat u_0(\xi) + {\displaystyle\int}_0^t {\displaystyle\int}_0^v
e^{-\xi^{{\mathfrak m}}(v-s_1)} e^{-\xi_1^{{\mathfrak m}}s_1}
{\varepsilon}^{d-2\alpha}{\displaystyle\int}\hat R({\varepsilon}(\xi_1-\xi))
{{U_{{\varepsilon},s}}}(t-v,\xi) d\xi_1 ds_1 dv \\
= e^{-t\xi^{{\mathfrak m}}} \hat u_0(\xi) + {\varepsilon}^{{{\mathfrak m}}-2\alpha}
{\displaystyle\int}_0^t {\displaystyle\int}_0^{\frac v{{\varepsilon}^{{\mathfrak m}}}} e^{-\xi^{{\mathfrak m}}(v-{\varepsilon}^{{\mathfrak m}}s_1)}
e^{-\xi_1^{{\mathfrak m}}s_1} {\displaystyle\int}\hat R(\xi_1-{\varepsilon}\xi) d\xi_1 ds_1
{{U_{{\varepsilon},s}}}(t-v,\xi) dv \\
:= e^{-t\xi^{{\mathfrak m}}} \hat u_0(\xi) + A_{\varepsilon}{{U_{{\varepsilon},s}}}(t,\xi).
\end{array}$$ The last integral results from the change of variables ${\varepsilon}\xi_1\to\xi_1$ and $s_1{\varepsilon}^{-{{\mathfrak m}}}\to s_1$. It remains to analyze the convergence properties of the solution to the latter integral equation. Note that $\xi$ acts as a parameter in that equation. Let us decompose $$\label{eq:decompAeps}
A_{\varepsilon}U(t,\xi) = \rho_{\varepsilon}{\displaystyle\int}_0^t e^{-\xi^{{\mathfrak m}}v} U(t-v,\xi) dv
+ E_{\varepsilon}U(t,\xi),$$ with $\rho_{\varepsilon}=\int_{\Rm^d} \frac{\hat R(\xi_1-{\varepsilon}\xi)}{\xi_1^{{\mathfrak m}}}
d\xi_1$ when $d>{{\mathfrak m}}$ and $\rho_{\varepsilon}=c_d\hat R({\varepsilon}\xi)$ when $d={{\mathfrak m}}$. Then we have
\[lem:Eeps\] Let $\xi\in\Rm^d$ and $f(r)$ as in lemma \[lem:bound\]. Then the operator $E_{\varepsilon}$ defined above in is bounded in the Banach space of continuous functions on $(0,T)$. Moreover, we have $$\label{eq:boundCeps}
\|E_{\varepsilon}\|_{\mathcal L(\mathcal C(0,T))} \lesssim {\varepsilon}^{\beta-{{\mathfrak n}}}.$$
We start with the case $d>{{\mathfrak m}}$ so that and ${\varepsilon}^{{{\mathfrak m}}-2\alpha}=1$. Note that ${{\mathfrak n}}$ in lemma \[lem:bound\] is defined such that $d>{{\mathfrak m}}-{{\mathfrak n}}$ as well. With $B_{\varepsilon}=A_{\varepsilon}-E_{\varepsilon}$ in , we find that $$B_{\varepsilon}{{U_{{\varepsilon},s}}}(t,\xi) = {\displaystyle\int}_0^t e^{-\xi^{{\mathfrak m}}v} {\displaystyle\int}_0^\infty
{\displaystyle\int}e^{-\xi_1^{{\mathfrak m}}s_1}
\hat R(\xi_1-{\varepsilon}\xi) d\xi_1ds_1 {{U_{{\varepsilon},s}}}(t-v,\xi) dv.$$ The remainder $E_{\varepsilon}$ is then given by $$\begin{array}{ll}
E_{\varepsilon}{{U_{{\varepsilon},s}}}(t,\xi) &= {\displaystyle\int}_0^t {\displaystyle\int}_0^{\frac v{{\varepsilon}^{{\mathfrak m}}}} {\displaystyle\int}e^{-\xi^{{\mathfrak m}}v} (e^{{\varepsilon}^{{\mathfrak m}}\xi^{{\mathfrak m}}s_1}-1)
e^{-\xi_1^{{\mathfrak m}}s_1} \hat R(\xi_1-{\varepsilon}\xi) d\xi_1 ds_1
{{U_{{\varepsilon},s}}}(t-v,\xi) dv\\
& - {\displaystyle\int}_0^t {\displaystyle\int}_{\frac v{{\varepsilon}^{{\mathfrak m}}}}^\infty {\displaystyle\int}e^{-\xi^{{\mathfrak m}}v}
e^{-\xi_1^{{\mathfrak m}}s_1} \hat R(\xi_1-{\varepsilon}\xi) d\xi_1 ds_1
{{U_{{\varepsilon},s}}}(t-v,\xi) dv.
\end{array}$$ The continuity of $E_{\varepsilon}{{U_{{\varepsilon},s}}}(t,\xi)$ in time is clear when ${{U_{{\varepsilon},s}}}(t,\xi)$ is continuous in time. Without loss of generality, we assume that ${{U_{{\varepsilon},s}}}(\cdot,\xi)$ is bounded by $1$ in the uniform norm. We decompose the integral in the $s_1$ variable in the first term of the definition of $E_{\varepsilon}$ into two integrals on $0\leq s_1\leq
\frac{v}{2{\varepsilon}^{{\mathfrak m}}}$ and $\frac{v}{2{\varepsilon}^{{\mathfrak m}}}\leq s_1\leq \frac
v{{\varepsilon}^{{\mathfrak m}}}$. Because $e^{-\xi^{{\mathfrak m}}v} (e^{{\varepsilon}^{{\mathfrak m}}\xi^{{\mathfrak m}}s_1}-1)\leq1$, the second integral is estimated as $$\begin{array}{l}
{\displaystyle\int}_0^t {\displaystyle\int}_{\frac v{2{\varepsilon}^{{\mathfrak m}}}}^{\frac v{{\varepsilon}^{{\mathfrak m}}}} {\displaystyle\int}e^{-\xi^{{\mathfrak m}}v} (e^{{\varepsilon}^{{\mathfrak m}}\xi^{{\mathfrak m}}s_1}-1)
e^{-\xi_1^{{\mathfrak m}}s_1} \hat R(\xi_1-{\varepsilon}\xi) d\xi_1 ds_1 dv \\[4mm]
\leq {\displaystyle\int}_0^t {\displaystyle\int}\dfrac{1}{\xi_1^{{\mathfrak m}}}e^{-\xi_1^{{\mathfrak m}}\frac{v}{2{\varepsilon}^{{\mathfrak m}}}}
\hat R(\xi_1-{\varepsilon}\xi) d\xi_1 dv
\leq {\displaystyle\int}\dfrac{2}{\xi_1^{{\mathfrak m}}}\Big(\dfrac{{\varepsilon}^{{\mathfrak m}}}{\xi_1^{{\mathfrak m}}}\wedge t\Big)
\hat R(\xi_1-{\varepsilon}\xi) d\xi_1 \lesssim {\varepsilon}^{\beta-{{\mathfrak n}}} \rho_f,
\end{array}$$ thanks to lemma \[lem:bound\]. The above bound is uniform in $\xi$. The last integral defining $E_{\varepsilon}$ on the interval $s_1\geq
\frac{v}{{\varepsilon}^{{\mathfrak m}}}$ is treated in the exact same way and also provides a contribution of order $O({\varepsilon}^{\beta-{{\mathfrak n}}})$.
The final contribution involves the integration over the interval $0\leq s_1\leq\frac v{2{\varepsilon}^{{\mathfrak m}}}$. Using $e^{-\xi^{{\mathfrak m}}v}
(e^{{\varepsilon}^{{\mathfrak m}}\xi^{{\mathfrak m}}s_1}-1)\leq {\varepsilon}^{{\mathfrak m}}\xi^{{\mathfrak m}}s_1 e^{-\frac{\xi^{{\mathfrak m}}v}2}$ on that interval, it is bounded by $$\begin{array}{l}
I_3 :={\displaystyle\int}_0^t {\displaystyle\int}_0^{\frac{v}{2{\varepsilon}^{{\mathfrak m}}}} {\displaystyle\int}_{\Rm^d}
{\varepsilon}^{{\mathfrak m}}\xi^{{\mathfrak m}}s_1 e^{-\frac{\xi^{{\mathfrak m}}v}2}
e^{-\xi_1^{{\mathfrak m}}s_1} \hat R(\xi_1-{\varepsilon}\xi) d\xi_1 ds_1 dv \\
\leq \dfrac{2{\varepsilon}^{{\mathfrak m}}\xi^{{\mathfrak m}}}{\xi^{{\mathfrak m}}}\big(1-e^{-\frac{\xi^{{\mathfrak m}}t}2}\big)
{\displaystyle\int}_0^{\frac{t}{2{\varepsilon}^{{\mathfrak m}}}} s_1{\displaystyle\int}_{\Rm^d} e^{-\xi_1^{{\mathfrak m}}s_1}
\hat R(\xi_1-{\varepsilon}\xi) d\xi_1 ds_1,
\end{array}$$ by switching the variables $0\leq s\leq \frac{v}{2{\varepsilon}^{{\mathfrak m}}} \leq
\frac{t}{2{\varepsilon}^{{\mathfrak m}}}$. Using lemma \[lem:decay\], we may replace $\hat R(\xi_1-{\varepsilon}\xi)$ by $\hat R(\xi_1)$ in the above expression. This shows that $$I_3 \leq 2{\varepsilon}^{{\mathfrak m}}{\displaystyle\int}_{\Rm^d} {\displaystyle\int}_0^{\frac{t}{2{\varepsilon}^{{\mathfrak m}}}}s_1
e^{-\xi_1^{{\mathfrak m}}s_1}ds_1 \hat R(\xi_1) d\xi_1.$$ We observe that $${\displaystyle\int}_0^{\tau} s_1
e^{-\xi_1^{{\mathfrak m}}s_1}ds_1 \lesssim \dfrac{1}{\xi_1^{2{{\mathfrak m}}}} \wedge \tau^2,$$ so that $$I_3 \lesssim {\varepsilon}^{{\mathfrak m}}{\displaystyle\int}_0^\infty f(r) r^{d-1}
\big(r^{-2{{\mathfrak m}}}\wedge \tau^2\big)dr,\qquad \tau = \dfrac{t}{2{\varepsilon}^{{\mathfrak m}}}\vee 1.$$ The integral over $(1,\infty)$ is bounded by ${\varepsilon}^{{\mathfrak m}}\rho_f$. Using the assumption that $f(r)\lesssim r^{-{{\mathfrak n}}}$, we obtain that the integral over $(0,1)$ is bounded by a constant times $${\displaystyle\int}_0^{\tau^{-\frac 1{{\mathfrak m}}}} r^{d-1-{{\mathfrak n}}} dr
+ {\displaystyle\int}_{\tau^{-\frac 1{{\mathfrak m}}}}^1 r^{d-1-{{\mathfrak n}}-2{{\mathfrak m}}} dr \lesssim
\tau^{2-\frac{d-{{\mathfrak n}}}{{{\mathfrak m}}}} \vee 1,$$ when $d-{{\mathfrak n}}-2{{\mathfrak m}}\not=0$ and $|\ln\tau|$ when $d={{\mathfrak n}}+2{{\mathfrak m}}$. Since $\tau$ is bounded by a constant times ${\varepsilon}^{-{{\mathfrak m}}}$, this shows that $I_3$ is bounded by ${\varepsilon}^{d-{{\mathfrak m}}-{{\mathfrak n}}}$ when $d-{{\mathfrak n}}-2{{\mathfrak m}}\not=0$ and ${\varepsilon}^d|\ln{\varepsilon}|$ when $d={{\mathfrak n}}+2{{\mathfrak m}}$. This concludes the proof when $d>{{\mathfrak m}}-{{\mathfrak n}}$.
We now consider the proof when $d={{\mathfrak m}}$ with ${{\mathfrak n}}=0$. Then, ${\varepsilon}^{{{\mathfrak m}}-2\alpha}=\frac{1}{|\ln{\varepsilon}|}$. The leading term is given by $U_{{\varepsilon},s}$, which solves the integral equation: $$\label{eq:intUes2s}
\begin{array}{l}
{{U_{{\varepsilon},s}}}(t,\xi)= e^{-t\xi^{{\mathfrak m}}} \hat u_0(\xi) \\\quad + {\displaystyle\int}_0^te^{-\xi^{{\mathfrak m}}s}
{\displaystyle\int}_0^{t-s} e^{-\xi_1^{{\mathfrak m}}s_1} {\displaystyle\int}\dfrac{1}{|\ln{\varepsilon}|}
\hat R({\varepsilon}(\xi_1-\xi))
{{U_{{\varepsilon},s}}}(t-s-s_1,\xi) d\xi_1 dsds_1\\
= e^{-t\xi^{{\mathfrak m}}} \hat u_0(\xi) + \dfrac{1}{|\ln{\varepsilon}|}
{\displaystyle\int}_0^t {\displaystyle\int}_0^{\frac v{{\varepsilon}^{{\mathfrak m}}}}
e^{-\xi^{{\mathfrak m}}(v-{\varepsilon}^{{\mathfrak m}}s_1)}
e^{-\xi_1^{{\mathfrak m}}s_1} \hat R(\xi_1-{\varepsilon}\xi) d\xi_1 ds_1
{{U_{{\varepsilon},s}}}(t-v,\xi) dv \\
= e^{-t\xi^{{\mathfrak m}}} \hat u_0(\xi) + A_{\varepsilon}{{U_{{\varepsilon},s}}}(t,\xi),\qquad
A_{\varepsilon}= B_{\varepsilon}+E_{\varepsilon}.
\end{array}$$ Here we have defined $$B_{\varepsilon}U (t,\xi) = \rho_{\varepsilon}{\displaystyle\int}_0^t e^{-\xi^{{\mathfrak m}}v} U(t-s,\xi)ds,\quad
\rho_{\varepsilon}= c_d \hat R({\varepsilon}\xi),$$ and $E_{\varepsilon}$ is the remainder. As in the case $d>{{\mathfrak m}}$, a contribution to $|\ln{\varepsilon}|E_{\varepsilon}$ comes from $${\displaystyle\int}_0^t {\displaystyle\int}_0^{\frac v{{\varepsilon}^{{\mathfrak m}}}} {\displaystyle\int}e^{-\xi^{{\mathfrak m}}v} (e^{{\varepsilon}^{{\mathfrak m}}\xi^{{\mathfrak m}}s_1}-1)
e^{-\xi_1^{{\mathfrak m}}s_1} \hat R(\xi_1-{\varepsilon}\xi) d\xi_1 ds_1
{{U_{{\varepsilon},s}}}(t-v,\xi) dv.$$ We again decompose the integral in $s_1$ into $0\leq s_1\leq
\frac{v}{2{\varepsilon}^{{\mathfrak m}}}$ and $\frac{v}{2{\varepsilon}^{{\mathfrak m}}}\leq s_1\leq \frac
v{{\varepsilon}^{{\mathfrak m}}}$. We have $$\begin{array}{l}
{\displaystyle\int}_0^t {\displaystyle\int}_{\frac v{2{\varepsilon}^{{\mathfrak m}}}}^{\frac v{{\varepsilon}^{{\mathfrak m}}}} {\displaystyle\int}e^{-\xi^{{\mathfrak m}}v} (e^{{\varepsilon}^{{\mathfrak m}}\xi^{{\mathfrak m}}s_1}-1)
e^{-\xi_1^{{\mathfrak m}}s_1} \hat R(\xi_1-{\varepsilon}\xi) d\xi_1 ds_1 dv \\[4mm]
\leq {\displaystyle\int}\dfrac{2}{{\varepsilon}^{{\mathfrak m}}}\Big(\dfrac{{\varepsilon}^{{\mathfrak m}}}{\xi_1^{{\mathfrak m}}}\wedge t\Big)^2
\hat R(\xi_1-{\varepsilon}\xi) d\xi_1 \lesssim \rho_f,
\end{array}$$ according to lemma \[lem:bound\]. Also, $$\begin{array}{l}
{\displaystyle\int}_0^t {\displaystyle\int}_0^{\frac v{2{\varepsilon}^{{\mathfrak m}}}}{\displaystyle\int}e^{-\xi^{{\mathfrak m}}v} (e^{{\varepsilon}^{{\mathfrak m}}\xi^{{\mathfrak m}}s_1}-1)
e^{-\xi_1^{{\mathfrak m}}s_1} \hat R(\xi_1-{\varepsilon}\xi) d\xi_1 ds_1 dv \lesssim {\varepsilon}^{{\mathfrak m}}\big(\dfrac{t}{2{\varepsilon}^{{\mathfrak m}}} \vee 1\big)
\end{array}$$ according to the calculations performed above on $I_3$, which is uniformly bounded, and thus provides a $|\ln{\varepsilon}|^{-1}$ contribution to $E_{\varepsilon}$.
We are thus left with the analysis of $$U(t,\xi)\mapsto {\displaystyle\int}_0^t
e^{-\xi^{{\mathfrak m}}v} \Big( \dfrac{1}{|\ln{\varepsilon}|}
{\displaystyle\int}\dfrac{1-e^{-\frac{\xi_1^{{\mathfrak m}}v}{{\varepsilon}^{{\mathfrak m}}}}}{\xi_1^{{\mathfrak m}}}
\hat R(\xi_1-{\varepsilon}\xi) d\xi_1 -\rho_{\varepsilon}\Big) U(t-v,\xi) dv,$$ as an operator in $\mathcal L(\mathcal C(0,T))$ for $\xi$ fixed. Define $\hat R_{\varepsilon}(\xi_1)=\hat R(\xi_1-{\varepsilon}\xi)$. The integral in $\xi_1$ may be recast as $${\displaystyle\int}_0^\infty \dfrac{1-e^{-\frac{r^{{\mathfrak m}}v}{{\varepsilon}^{{\mathfrak m}}}}}{r}
\Big({\displaystyle\int}_{S^{d-1}}\hat R_{\varepsilon}(r\theta)d\mu(\theta)\Big) dr.$$ We observe that the integral on $(1,\infty)$ is bounded by $\|\hat
R\|_1$. Assuming that $\hat R$ is of class $\mathcal
C^{0,\gamma}(\Rm^d)$ for $\gamma>0$, we write $\hat
R_{\varepsilon}(\xi_1)=\hat R_{\varepsilon}(0)+(\hat R_{\varepsilon}(\xi_1)-\hat R_{\varepsilon}(0))$. The second contribution generates a term proportional to $r^\gamma$ in the integral and thus is bounded independent of ${\varepsilon}$. It remains to estimate $$c_d \hat R_{\varepsilon}(0){\displaystyle\int}_0^1 \dfrac{1-e^{-\frac{r^{{\mathfrak m}}v}{{\varepsilon}^{{\mathfrak m}}}}}{r} dr
=c_d\hat R_{\varepsilon}(0) {\displaystyle\int}_0^{\frac{v^{\frac1{{\mathfrak m}}}}{{\varepsilon}}}
\dfrac{1-e^{-r^{{\mathfrak m}}}}{r} dr.$$ The latter integral restricted to $(0,1)$ is bounded. On $r\geq1$, $e^{-r^{{\mathfrak m}}}/r$ is uniformly integrable so that $$c_d\hat R_{\varepsilon}(0){\displaystyle\int}_0^1 \dfrac{1-e^{-\frac{r^{{\mathfrak m}}v}{{\varepsilon}^{{\mathfrak m}}}}}{r} dr=
c_d\hat R({\varepsilon}\xi) |\ln{\varepsilon}| + O(1).$$ This shows that $E_{\varepsilon}$ is of order $\frac{1}{|\ln{\varepsilon}|}={\varepsilon}^\beta$ as an operator on $\mathcal C(0,T)$ and concludes the proof of the lemma.
Note that $A_{\varepsilon}$ may be written as $$A_{\varepsilon}U(t,\xi) = {\displaystyle\int}_0^t \varphi_{\varepsilon}(s,\xi) U(t-s\xi) ds,$$ where $\varphi_{\varepsilon}(s,\xi)$ is uniformly bounded in $s$, $\xi$, and ${\varepsilon}$ by a constant $\varphi_\infty$. The equation $$(I-A_{\varepsilon}) U(t,\xi) = S(t,\xi),$$ admits a unique (by Gronwall’s lemma) solution given by the Duhamel expansion and bounded by $$|U(t,\xi)| \leq \|S\|_\infty e^{t\varphi_\infty}.$$
As in the proof of lemma \[lem:Eeps\], let us define $B_{\varepsilon}=A_{\varepsilon}-E_{\varepsilon}$. We verify that ${{\mathfrak U}}_{{\varepsilon}}(t,\xi)$, the solution to $$(I-B_{\varepsilon}){{\mathfrak U}}_{{\varepsilon}} = e^{-t\xi^{{\mathfrak m}}}\hat u_{\varepsilon}(\xi),$$ is given by $$\label{eq:tildeUeps}
{{\mathfrak U}}_{{\varepsilon}}(t,\xi) = e^{-t(\xi^{{\mathfrak m}}-\rho_{\varepsilon}(\xi))} \hat u_0(\xi).$$ The solution may thus grow exponentially in time for low frequencies. The error $V_{\varepsilon}(t,\xi)=(U_{{\varepsilon},s}(t,\xi)-{{\mathfrak U}}_{{\varepsilon}}(t,\xi))$ is a solution to $$(I-A_{\varepsilon}) V_{\varepsilon}= E_{\varepsilon}{{\mathfrak U}}_{{\varepsilon}}(t,\xi),$$ so that over bounded intervals in time (with a constant growing exponentially with time but independent of $\xi$), we find that $$\label{eq:bdVeps}
|V_{\varepsilon}(t,\xi)| \lesssim {\varepsilon}^\beta.$$
Up to an order $O({\varepsilon}^\beta|\hat u_0(\xi)|)$, we have thus obtained that ${\mathbb E}\{\hat u_{\varepsilon}(t,\xi)\}$ is given by $$e^{-t(\xi^{{\mathfrak m}}-\rho_{\varepsilon}(\xi))} \hat u_0(\xi),$$ which in the physical domain gives rise to a possibly non-local equation. It remains to analyze the limit of the above term, and thus the error $\rho_{\varepsilon}(\xi)-\rho$, which depends on the regularity of $\hat R(\xi)$. For $\hat R(\xi)$ of class $\mathcal C^2(\Rm^d)$, we find that $$\big|e^{-t(\xi^{{\mathfrak m}}-\rho_{\varepsilon}(\xi))}-e^{-t(\xi^{{\mathfrak m}}-\rho)}\big|
\leq te^{Ct} e^{-\xi^{{\mathfrak m}}t} \big|\rho_{\varepsilon}(\xi)-\rho\big|
\lesssim e^{Ct} e^{-\xi^{{\mathfrak m}}t} {\varepsilon}^2 t\xi^2.$$ The reason for the second order accuracy is that $\hat R(-\xi)=\hat
R(\xi)$ and $\nabla \hat R(0)=0$ so that first-order terms in the Taylor expansion vanish. For $\hat R(\xi)$ of class $\mathcal
C^\gamma(\Rm^d)$ with $0<\gamma<2$, we obtain by interpolation that $$\big|e^{-t(\xi^{{\mathfrak m}}-\rho_{\varepsilon}(\xi))}-e^{-t(\xi^{{\mathfrak m}}-\rho)}\big|
\lesssim e^{Ct} e^{-\xi^{{\mathfrak m}}t} {\varepsilon}^\gamma t\xi^\gamma.$$ When ${{\mathfrak m}}\geq\gamma$, the above term is bounded by $O({\varepsilon}^\gamma)$ uniformly in $\xi$ and uniformly in time on bounded intervals. When ${{\mathfrak m}}\leq\gamma$, the above term is bounded by $O({\varepsilon}^{{\mathfrak m}})$ uniformly in $\xi$ and uniformly in time on bounded intervals. This concludes the proof of theorem \[thm:convergence\]. In terms of the propagators defined in , we may recast the above result as $$\label{eq:errorprop}
\big|{\mathcal U}_{\varepsilon}(t,\xi)-{\mathcal U}(t,\xi)\big| \lesssim
{\varepsilon}^{\gamma\wedge \beta},
\qquad
{\mathcal U}(t,\xi) = e^{-(\xi^{{\mathfrak m}}-\rho)t},$$ where the bound is uniform in time for $t\in (0,T)$ and uniform in $\xi\in\Rm^d$.
Fluctuation theory for $u_{\varepsilon}$
----------------------------------------
We now address the proof of theorem \[thm:fluct\]. The first term in the decomposition of $\hat u_{n,{\varepsilon}}$ defined in is its mean ${\mathbb E}\{\hat u_{n,{\varepsilon}}\}$, which was analyzed in the preceding section. The second contribution corresponds to the graphs ${\mathfrak G}_{cs}$ in the analysis of the correlation function and is constructed as follows. Let $n=2p+1$, $p\in\Nm$. We introduce the corrector $\hat u_{n,{\varepsilon}}^c$ given by $$\label{eq:hatunepsc}
\begin{array}{l}
\hat u_{n,{\varepsilon}}^c(t,\xi_0) = {\displaystyle\int}{\displaystyle\prod}_{k=0}^n e^{-s_k \xi_k^{{\mathfrak m}}}
{\displaystyle\sum}_{q=0}^p \Big[{\displaystyle\prod}_{r=1}^{q}
{\mathbb E}\{\hat q_{\varepsilon}(\xi_{2(r-1)}-\xi_{2r-1})\hat q_{\varepsilon}(\xi_{2r-1}-\xi_{2r})\}
\Big] \\\qquad \qquad\hat q_{\varepsilon}(\xi_{2q}-\xi_{2q+1})
\Big[{\displaystyle\prod}_{r=q+1}^{p}
{\mathbb E}\{\hat q_{\varepsilon}(\xi_{2r-1}-\xi_{2r})\hat q_{\varepsilon}(\xi_{2r}-\xi_{2r+1})\}
\Big] \hat u_0(\xi_n) d{\mathbf s}d{\boldsymbol \xi}.
\end{array}$$ In other words, all the random terms are averaged as simple pairs except for one term. There are $p+1$ such graphs. We define $$\label{eq:hatuepsc}
\hat u_{\varepsilon}^c(t,\xi) = {\displaystyle\sum}_{n\geq1} \hat u_{n,{\varepsilon}}^c(t,\xi).$$ We verify that $$V_{{\varepsilon},s}^{n,m}(t,\xi_0,\xi_{n+m+1}) :=
{\mathbb E}\{\hat u_{n,{\varepsilon}}^c(t,\xi_0) \bar{\hat u}_{n,{\varepsilon}}^c(t,\xi_{n+m+1})\}$$ is equal to the sum in $V^{n,m}_{\varepsilon}(t,\xi_0,\xi_{n+m+1})$ only over the graphs in ${\mathfrak G}_{cs}$. Indeed, the above correlation involves all the graphs composed of simple pairs with a single crossing.
Now let us define the variable $$\label{eq:Yeps}
Y_{\varepsilon}= (\hat u_{\varepsilon}-\hat u_{\varepsilon}^c-{\mathbb E}\{\hat u_{\varepsilon}\},\hat M).$$ Summing over $n,m\in\Nm$ the inequality in as we did to obtain , we have demonstrated that $$\label{eq:bdYeps}
{\mathbb E}\{Y_{\varepsilon}^2\} \lesssim {\varepsilon}^{d-2\alpha+\beta} \|\hat u_0\|^2
\|\hat M\|_1^2,$$ for sufficiently small times. The leading term in the random fluctuations of $u_{\varepsilon}$ is thus given by $u_{\varepsilon}^c$. It remains to analyze the convergence properties of $$\label{eq:Zeps}
Z_{\varepsilon}(t) = \dfrac{1}{{\varepsilon}^{\frac{d-2\alpha}2}} (\hat u_{\varepsilon}^c,\hat M).$$
We thus come back to the analysis of $\hat u_{\varepsilon}^c$ and observe that for $n=2p+1$, $$\begin{array}{l}
\hat u_{n,{\varepsilon}}^c(t,\xi_0) = {\displaystyle\int}{\displaystyle\prod}_{k=0}^n e^{-s_k \xi_k^{{\mathfrak m}}}
{\displaystyle\sum}_{q=0}^p
\Big[{\displaystyle\prod}_{r=1}^{q}
{\varepsilon}^{d-2\alpha} \hat R({\varepsilon}(\xi_{2r-1}-\xi_0))
\delta(\xi_{2r}-\xi_0)
\Big] \\\qquad \qquad\hat q_{\varepsilon}(\xi_{0}-\xi_{n})
\Big[{\displaystyle\prod}_{r=q+1}^{p}
{\varepsilon}^{d-2\alpha} \hat R({\varepsilon}(\xi_{2r}-\xi_n))
\delta(\xi_{2r-1}-\xi_n)
\Big] \hat u_0(\xi_n) d{\mathbf s}d{\boldsymbol \xi}.
\end{array}$$ Using the propagator defined in , we verify that $$\begin{array}{l}
\hat u_{n,{\varepsilon}}^c(t,\xi_0) = {\displaystyle\sum}_{q=0}^p
{\displaystyle\int}{\displaystyle\prod}_{k=0}^{2q+1} e^{-s_k \xi_k^{{\mathfrak m}}}
\Big[{\displaystyle\prod}_{r=1}^{q}
{\varepsilon}^{d-2\alpha} \hat R({\varepsilon}(\xi_{2r-1}-\xi_0))
\delta(\xi_{2r}-\xi_0)
\Big] \\\qquad \qquad\hat q_{\varepsilon}(\xi_{0}-\xi_{n})
{\mathcal U}^{n-2q}_{\varepsilon}(t_{2q+1},\xi_n) \hat u_0(\xi_n) d\tilde{\mathbf s}d\tilde{\boldsymbol \xi}\\
= {\displaystyle\sum}_{q=0}^p {\displaystyle\int}_0^t {\mathcal U}^{2q}_{\varepsilon}(t-t_{2q+1},\xi_0)
\hat q_{\varepsilon}(\xi_{0}-\xi_{n})
{\mathcal U}^{n-2q}_{\varepsilon}(t_{2q+1},\xi_n) \hat u_0(\xi_n)
dt_{2q+1} d\xi_n\\
= {\displaystyle\sum}_{q=0}^p {\displaystyle\int}_0^t {\mathcal U}^{2q}_{\varepsilon}(t-s,\xi_0)
\hat q_{\varepsilon}(\xi_{0}-\xi_{1})
{\mathcal U}^{n-2q}_{\varepsilon}(s,\xi_1) \hat u_0(\xi_1) ds d\xi_1.
\end{array}$$ Upon summing over $n$, we obtain $$\label{eq:uc}
\hat u_{\varepsilon}^c(t,\xi) = {\displaystyle\int}_0^t {\mathcal U}_{\varepsilon}(t-s,\xi)
\hat q_{\varepsilon}(\xi-\xi_1) {\mathcal U}_{\varepsilon}(s,\xi_1)
\hat u_0(\xi_1)ds d\xi_1.$$ We can use the error on the propagator obtained in to show that the leading order of $\hat u_{\varepsilon}^c$ is not modified by replacing ${\mathcal U}_{\varepsilon}$ by ${\mathcal U}$. In other words, replacing ${\mathcal U}_{\varepsilon}$ by ${\mathcal U}$ modifies $Z_{\varepsilon}$ in by a term of order $O({\varepsilon}^{\frac12(\beta\wedge \gamma)})$ in $L^2(\Omega\times \Rm^d)$, which thus goes to $0$ in law.
Note that $\hat u_{\varepsilon}^c(t,\xi)$ is a mean zero Gaussian random variable. It is therefore sufficient to analyze the convergence of its variance in order to capture the convergent random variable for each $t$ and $\xi$. The same is true for the random variable $Z_{\varepsilon}$. Up to a lower-order term, which does not modify the final convergence, we thus have that $$(\hat u_{\varepsilon}^c,\hat M) ={\displaystyle\int}{\displaystyle\int}_0^t
{\bar {\mathcal U}}_{\hat M}(t-s,\xi)
\hat q_{\varepsilon}(\xi_1) {\mathcal U}_{\hat u_0}(s,\xi-\xi_1) ds d\xi d\xi_1.$$ We have defined ${\mathcal U}_f(t,\xi)={\mathcal U}(t,\xi) f(\xi)$ for a function $f(\xi)$. As a consequence, we find that, still up a vanishing contribution, $$\begin{array}{rcl}
{\mathbb E}\{|Z_{\varepsilon}|^2\}& =&
{\displaystyle\int}{\displaystyle\int}_0^t{\displaystyle\int}_0^t {\bar{\mathcal U}}_{\hat M}(t-s,\xi)
{\mathcal U}_{\hat M}(t-\tau,\zeta) \hat R({\varepsilon}\xi_1)
\delta(\xi_1-\zeta_1) \\ && \qquad \qquad \times\,\,
{\mathcal U}_{\hat u_0}(s,\xi-\xi_1)
\bar{\mathcal U}_{\hat u_0}(\tau,\zeta-\zeta_1)
d[s\tau\zeta\zeta_1\xi\xi_1].
\end{array}$$ Here and below, we use the notation $d[x_1\ldots x_n]\equiv dx_1\ldots
dx_n$. By the dominated Lebesgue convergence theorem, we obtain in the limit $${\mathbb E}\{|Z|^2\} :=
\hat R(0) {\displaystyle\int}\Big| {\displaystyle\int}{\displaystyle\int}_0^t{\mathcal U}_{\hat M}(t-s,\xi)
{\mathcal U}_{\hat u_0}(s,\xi-\xi_1) d\xi_1 ds\Big|^2 d\xi.$$ Here, $Z$ is defined as a mean zero Gaussian random variable with the above variance. Let us define $ {\mathcal G}^\rho_t f(x),$ the solution at time $t$ of with $f(x)$ as initial conditions, which is also the inverse Fourier transform of ${\mathcal U}_{\hat f}(t,\xi)$. We then recognize in $ \int \int_0^t{\mathcal U}_{\hat M}(t-s,\xi)
{\mathcal U}_{\hat u_0}(s,\xi-\xi_1) d\xi_1 ds$ the Fourier transform of $\mathcal M_t(x)$ defined in so that by an application of the Plancherel identity, we find that $$\label{eq:varZ}
{\mathbb E}\{Z^2\} = (2\pi)^d \hat R(0) \!\!{\displaystyle\int}_{\Rm^d}
\Big({\displaystyle\int}_0^t {\mathcal G}_{t-s}^\rho M(x) {\mathcal G}_s^\rho u_0(x) ds
\Big)^2 dx = (2\pi)^d \hat R(0) \!\!{\displaystyle\int}_{\Rm^d} \!\!\mathcal M_t^2(x)dx.$$ This shows that $Z(t)$ is indeed the Gaussian random variable written on the right hand side in by an application of the Itô isometry formula. This concludes the proof of theorem \[thm:fluct\].
Long range correlations and correctors {#sec:longrange}
--------------------------------------
Let us now assume that $$\label{eq:decR}
\hat R(\xi) = h(\xi) S(\xi),\qquad 0<h(\lambda\xi)=|\lambda|^{-{{\mathfrak n}}}h(\xi),$$ where $h(\xi)$ is thus a positive function homogeneous of degree $-{{\mathfrak n}}$ and $\hat S(\xi)$ is bounded on $B(0,1)$. We assume that $\hat R(\xi)$ is still bounded on $\Rm^d\backslash B(0,1)$. We also assume that ${{\mathfrak m}}+{{\mathfrak n}}<d$ and that $\rho$ in is still defined. We denote by $\varphi(x)$ the inverse Fourier transform of $h(\xi)$. Then we have the following result.
\[thm:fracBM\] Let us assume that $h(\xi)=|\xi|^{-{{\mathfrak n}}}$ for ${{\mathfrak n}}>0$ and ${{\mathfrak m}}+{{\mathfrak n}}<d$. We also impose the following regularity on $\hat u_0$: $$\label{eq:regu0}
{\displaystyle\int}_{B(0,1)} |\hat u_0(\xi+\tau)|^2 h(\xi) d\xi \leq C,
\quad \mbox{ for all }\, \tau\in\Rm^d.$$ Then theorem \[thm:convergence\] holds with $\beta$ replaced by $\beta-{{\mathfrak n}}$.
Let us define the random corrector $$\label{eq:u1eps2}
u_{1,{\varepsilon}}(t,x) = \dfrac{1}{{\varepsilon}^{\frac{d-{{\mathfrak m}}-{{\mathfrak n}}}2}}
\big(u_{\varepsilon}-{\mathbb E}\{u_{\varepsilon}\}\big)(t,x).$$ Then its spatial moments $(u_{1,{\varepsilon}}(t,x),M(x))$ converge in law to centered Gaussian random variables $\mathcal N(0,\Sigma_M(t))$ with variance given by $$\label{eq:sigmaM}
\Sigma_M(t) = (2\pi)^{d}\hat S(0)
{\displaystyle\int}_{\Rm^{2d}} \mathcal M_t(x) \varphi(x-y) \mathcal M_t(y) dx dy.$$
The proof of theorem \[thm:convergence\] relies on three estimates: those of lemma \[lem:bound\] and lemma \[lem:Eeps\] and the uniform bound in for $\hat R$. Lemmas \[lem:bound\] and \[lem:Eeps\] were written to account for power spectra bounded by $|\xi|^{-{{\mathfrak n}}}$ in the vicinity of the origin. It thus remains to replace by $${\varepsilon}^{d-{{\mathfrak m}}} \hat R({\varepsilon}(\xi_Q-\xi_{Q-1}^{\varepsilon}))
\leq {\varepsilon}^{d-{{\mathfrak m}}-{{\mathfrak n}}} h(\xi_Q-\xi_{Q-1}^{\varepsilon}) \hat S_\infty,$$ when $|\xi_Q-\xi_{Q-1}^{\varepsilon}|\leq1$ while we still use otherwise. We have defined $\hat S_\infty$ as the supremum of $\hat S(\xi)$ in $B(0,1)$. It now remains to show that the integration with respect to $\xi_Q$ in is still well-defined. Note that either $Q=n$ or $\xi_Q-\xi_{Q-1}^{\varepsilon}$ may be written as $\xi_n-\zeta$ for some $\zeta\in\Rm^d$ thanks to . Upon using , we thus observe that in all cases, the integration with respect to $\xi_Q$ in is well-defined and bounded uniformly provided that is satisfied uniformly in $\tau$. Using the Hölder inequality, we verify that holds e.g. when $\hat u_0(\cdot-\tau)\in
L^q(B(0,1))$ uniformly in $\tau$ for $q>\frac{2d}{d-{{\mathfrak n}}}$. This concludes the proof of the first part of the theorem.
Let us now define $$\tilde Z_{\varepsilon}(t) =
\dfrac{1}{{\varepsilon}^{\frac{d-{{\mathfrak m}}-{{\mathfrak n}}}2}} (\hat u_{\varepsilon}^c,\hat M)
= {\varepsilon}^{\frac{{{\mathfrak n}}}2} Z_{\varepsilon}(t).$$ We verify as for the derivation of ${\mathbb E}\{Z_{\varepsilon}^2\}$ that $$\begin{array}{rcl}
{\mathbb E}\{\tilde Z_{\varepsilon}^2\}& =&
{\displaystyle\int}{\displaystyle\int}_0^t{\displaystyle\int}_0^t {\mathcal U}_{\hat M}(t-s,\xi)
\hat{\mathcal U}_{\hat M}(t-\tau,\zeta)\hat S({\varepsilon}\xi_1)h(\xi_1)
\delta(\xi_1-\zeta_1) \\ && \qquad \qquad \times\,\,
{\mathcal U}_{\hat u_0}(s,\xi-\xi_1)
\hat{\mathcal U}_{\hat u_0}(\tau,\zeta-\zeta_1)
d[s\tau\zeta\zeta_1\xi\xi_1].
\end{array}$$ The dominated Lebesgue convergence theorem yields in the limit ${\varepsilon}\to0$ $$\begin{array}{rcl}
{\mathbb E}\{\tilde Z^2\} &:=&
\hat S(0) {\displaystyle\int}\Big|{\displaystyle\int}_0^t {\displaystyle\int}{\mathcal U}_{\hat M}(t-s,\xi)
{\mathcal U}_{\hat u_0}(s,\xi-\xi_1) h^{\frac12}(\xi_1)
d\xi_1 ds\Big|^2 d\xi \\
&=& \hat S(0) {\displaystyle\int}|\hat{\mathcal M}_t (\xi)|^2 h(\xi) d\xi,
\end{array}$$ where ${\mathcal M}_t$ is defined in . An application of the inverse Fourier transform yields .
Note that generalizes , where $\varphi(x)=\delta(x)$, to functions $\mathcal M_t(x)\in
L^2_\varphi(\Rm^d)$ with inner product $$\label{eq:innerprod}
(f,g)_\varphi = {\displaystyle\int}_{\Rm^{2d}} f(x)g(y)\varphi(x-y) dxdy.$$ For $h(\xi)=|\xi|^{-{{\mathfrak n}}}$, we find that $\varphi(x)=c_{{{\mathfrak n}}}|x|^{{{\mathfrak n}}-d}$, with $c_{{{\mathfrak n}}}=\Gamma(\frac{d-{{\mathfrak n}}}2)/(2^{{\mathfrak n}}\pi^{\frac
d2}\Gamma(\frac{{{\mathfrak n}}}2))$ a normalizing constant. Following e.g. [@HOUZ-Birk-96; @L-BLM-93], we may then define a stochastic integral with fractional Brownian $$\label{eq:Z}
Z = {\displaystyle\int}_{\Rm^d} {\mathcal M}_t (x) dB^H(x),$$ where $B^H$ is fractional Brownian motion such that $${\mathbb E}\{ B^H(x)B^H(y) \} = \frac12\big( |x|^{2H}+|y|^{2H}-|x-y|^{2H}\big),
\qquad 2H = 1+\frac{{{\mathfrak n}}}d.$$ We then verify that ${\mathbb E}\{Z^2\}=\Sigma_M$ so that the random variable $Z$ is indeed given by the above formula . When ${{\mathfrak n}}=0$, we retrieve the value for the Hurst parameter $H=\frac 12$ so that $B^H=W$. The above isotropic fractional Brownian motion is often replaced in the analysis of stochastic equations by a more Cartesian friendly fractional Brownian motion defined by $$\varphi_H(x) = {\displaystyle\prod}_{i=1}^d H_i (2H_i-1)|x_i|^{2H_i-2}.$$ The above is then defined as the Fourier transform of $$h_H(\xi) = {\displaystyle\prod}_{i=1}^d |\xi_i|^{-{{\mathfrak n}}_i}, \qquad
{\displaystyle\sum}_{i=1}^d {{\mathfrak n}}_i = {{\mathfrak n}}, \qquad 2H_i = 1 + \dfrac{{{\mathfrak n}}_i}d.$$ The results of theorem \[thm:convergence\] and \[thm:fracBM\] may also be extended to this framework by modifying the proofs in lemmas \[lem:bound\] and \[lem:Eeps\]. We then obtain that holds for a multiparameter anisotropic fractional Brownian motion $B^H$, $H=(H_1,\ldots,H_d)$, with covariance $${\mathbb E}\{ B^H(x)B^H(y) \}= \frac{1}{2^d}\prod_{i=1}^d
\big(|x_i|^{2H_i}+|y_i|^{2H_i}-|x_i-y_i|^{2H_i}\big).$$ Note that homogenization theory is valid as soon as $d>{{\mathfrak m}}+{{\mathfrak n}}$. As in the case ${{\mathfrak n}}=0$, we expect that when $d<{{\mathfrak m}}+{{\mathfrak n}}$ (rather than $d<{{\mathfrak m}}$), the limit for $u_{\varepsilon}$ will be the solutions in $L^2(\Omega\times
\Rm^d)$ to a stochastic differential equation of the form with white noise replaced by some fractional Brownian motion; see also [@Hu-AMO-01].
The stochastic representation in is not necessary since $\Sigma_M(t)$ fully characterizes the random variable $Z$. However, the representation emphasizes the following conclusion. Let $Z_1^H$ and $Z_2^H$ be the limiting random variables corresponding to two moments with weights $M_1(x)$ and $M_2(x)$ and a given Hurst parameter $H$. When $H=\frac12$, we deduce directly from that ${\mathbb E}\{Z_1^{\frac12}Z_2^{\frac12}\}=0$ when $M_1(x)M_2(x)=0$, i.e., when the supports of the moments are disjoint. This is not the case when $H\not=\frac12$ as fractional Brownian motion does not have independent increments. Rather, we find that ${\mathbb E}\{Z_1^{H}Z_2^{H}\}$ is given by $(\mathcal M_{t,1},\mathcal M_{t,2})_{\varphi}$, where the inner product is defined in and $\mathcal
M_{t,k}$ is defined in with $M$ replaced by $M_k$, $k=1,2$. Similar results were obtained in the context of the one-dimensional homogenization with long-range diffusion coefficients [@BGMP-AA-08].
Acknowledgment {#acknowledgment .unnumbered}
==============
This work was supported in part by NSF Grants DMS-0239097 and DMS-0804696.
[^1]: Department of Applied Physics and Applied Mathematics, Columbia University, New York NY, 10027; [email protected]
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Extracting texts of various size and shape from images containing multiple objects is an important problem in many contexts, especially, in connection to e-commerce, augmented reality assistance system in natural scene etc. The existing works (based on only CNN) often perform sub-optimally when the image contains regions of high entropy having multiple objects. This paper presents an end-to-end text detection strategy combining a segmentation algorithm and an ensemble of multiple text detectors of different types to detect text in every individual image segments independently. The proposed strategy involves a super-pixel based image segmenter which splits an image into multiple regions. A convolutional deep neural architecture is developed which works on each of the segments and detects texts of multiple shapes, sizes and structures. It outperforms the competing methods in terms of coverage in detecting texts in images especially the ones where the text of various types and sizes are compacted in a small region along with various other objects. Furthermore, the proposed text detection method along with a text recognizer outperforms the existing state-of-the-art approaches in extracting text from high entropy images. We validate the results on a dataset consisting of product images on an e-commerce website.'
bibliography:
- 'bmvc\_final.bib'
title: 'Semi-Bagging Based Deep Neural Architecture to Extract Text from High Entropy Images'
---
Introduction
============
Text detection from an image having multiple texts and other objects has been a very important problem in recent times due to the recent advancement and increasing application of computer vision techniques in different domains, for example, e-commerce, automatic driving assistance system, security camera surveillance etc. Detecting regions of texts and extracting that information from natural images is a challenging problem due to the presence of multiple texts of various shapes and sizes and many other objects together. One important problem in this context is extracting texts from product images, particularly in e-commerce, especially, when the intent is to extract brand, product type, various attributes from the product label. Large e-commerce companies sell billions of products through their websites. All these products are associated with one or more product images containing various textual information about them. Extracting this information are important not only to enhance the quality of the product catalogue but also to examine that product information with respect to various compliance policies of the organization. One of the primary requisites to extract information from product images is to extract text from those images with high accuracy and coverage.
A vast amount of literature is already available regarding the detection and extraction of text from an image. The problem of text detection in an image has been extensively studied from different perspectives. Neuman [@Neuman2012] proposed a technique to detect individual characters in an image and then group them into words depending on various attributes of the detected characters. Jaderberg [@Jaderberg2016] proposed a technique to directly detect the words in an image. Zhang [@Zhang2015] proposed a method to detect a text line in a scene image directly and then splitting that line into words. TextBoxes [@TB2016] is a word-based method to detect text in a scene image. Unlike Jaderberg [@Jaderberg2016], which comprises three detection steps and each further includes more than one algorithm, the algorithm of [*Textbox*]{} is much simpler. It requires only one network to be trained end-to-end. [*Textbox++*]{} [@TBPP2018] is an improvement over [*Textbox*]{} and it offers the capability to detect oriented text as well. However, both [*Textbox++*]{} and [*Textbox*]{} focus on detecting text in scene images. Our proposed method to detect text in high entropy images is primarily inspired by both [*Textbox++*]{} and [*Textbox*]{}. [*TextBox*]{} and [*TextBox++*]{} [@TB2016; @TBPP2018] exhibit reasonably accurate performance in detecting text from scene images.
However, none of the existing techniques performs with high accuracy when it comes to detecting text from images where multiple texts are enclosed within regions along with possibly various other objects. This essentially means that the image has a high variability of heterogeneous information within compact regions. We relate this with the [*entropy*]{} of the image, defined as the average information in an image and can be determined approximately from the histogram of multichannel and colour-space features of the image (see [@ENT1984]). In this work, we propose to detect text from such images by first segmenting it into multiple regions. The segmented regions are evaluated by first computing super-pixel level colour-space features and texture features from a dilated version of the image and hence, combing multiple super-pixels to form various segmented regions based on the super-pixel similarities. We propose an ensemble modelling approach for feature extraction by combining multiple Convolutional Neural Network (CNN) based models using selective non-max suppression. We use the ensemble model algorithm for detecting text of varying scales in each of the segments.
Our key contributions in this paper are two-fold: First, we designed a segmentation algorithm which is tailor-fitted to segment out the regions of the image containing text. Secondly, we propose an ensemble of multiple neural networks which extract features from segments of the image in order to detect text of varied sizes in the form of compact bounding boxes. The proposed architecture is highly parallelizable and the results show comparable and in cases, better accuracies in comparison to the competitors.
The rest of the paper is arranged as follows. In [**Section 1**]{} we describe the framework of the image segmentation and the text detection in the form of a pipeline process. [**Section 2**]{} is dedicated to describing the super-pixel based image segmentation methodology. The text detection method based on selective non-max suppression of the CNN ensemble is described in [**Section 3.**]{} The implementation of the proposed framework along with the accuracy of the proposed method on various image datasets is discussed in [**Section 4.**]{} We conclude the paper with a brief discussion in [**Section 5.**]{}
Framework
=========
The framework consists of two main components - (i) Image segmentation and (ii) Text detection. Image segmentation component splits the high entropy regions of the image into multiple segments so that it is easier for the subsequent text detection module to extract features from them more efficiently. The text detection component embodies a semi-bagging styled model. It is a pipeline process where the output of the image segmentation component is fed to the text detection component. The above components are elaborated more in the following sections. Fig.\[fig:arch\] details the architecture of the complete framework end-to-end.
![(a) Input image, (b) Segmented image, (c) and (d) TextBoxes[@TB2016] and TextBoxes++[@TBPP2018] CNN architectures trained on ICDAR2013, (e) Selective non-maximal suppression, (f) Final output text boxes.[]{data-label="fig:arch"}](images/Arch.png){width="0.7\linewidth"}
Segmentation
============
The proposed segmentation method consists of a number of intermediate steps resulting in spatially connected segments of homogeneous pixels. The goal of the segmentation module primarily is to make sure that the complete image is segmented into a number of regions such that the different text objects are enclosed in the individual regions. To ensure spatial continuity in the segments, we first perform dilation of the image to make sure the holes and gaps inside objects are nullified and small intrusions at object boundaries are somewhat smoothened. We consider super-pixels of fixed size in the dilated image to calculate various features summarising super-pixel level information. Based on the super-pixel level feature information on the dilated image, we fit a Gaussian Mixture model to identify the class of super-pixels in an unsupervised manner. The details of the steps are described below.
Dilation
--------
The image dilation is performed by convolving the image with a suitable kernel, in our case, the Gaussian Kernel. The anchor point of the kernel is chosen to be the centre of the kernel. As we scan the chosen kernel over the image, the pixel value at the anchor point is replaced by the maximum pixel value of the image region overlapping the kernel. This results in the interesting regions of the image to grow and the holes and gaps within the object to get nullified. As a result, the segments having the objects are over-compensated making sure there is less chance of a segment to truncate objects inside its true boundaries and split at holes and gaps within the object.
Super-Pixel Features
--------------------
For each super-pixel of fixed size $s\in \mathcal S$, we calculate a set of features as $\bf x_s.$ For each super-pixel $s$ and each of the colour channels $c\in \mathcal C$, we calculate mean, standard deviation and energy, denoted by $\bf x^{(1)}_{s,c}=[\mu_{s,c}, \sigma_{s,c}, e_{s,c}].$ To summarize the texture features, we consider the Leung-Malik filterbank (see [@Leung]) at multiple scales and orientations. In total, we consider, first and second derivatives at 6 orientations, 8 Laplacians of Gaussian filters and 4 Gaussians and hence take the convolution with the pixel at different channels. To make sure there is orientation invariance, we take the maximum response over all orientations at each pixel. We calculate the mean, standard deviation and energy for all pixels within a super-pixel for all the filter convolution to get features $\bf x^{(2)}_{s,c}=[\mu_{s,c, j}, \sigma_{s,c, j}, e_{s,c, j}]_{j\in \mathcal J},$ for all colour channels $c\in \mathcal C$ and super-pixel $s\in \mathcal S$. The combined feature set for a given super-pixel is given by $\bf x_s=[\bf x^{(1)}_{s,c},x^{(2)}_{s,c}]_{c\in \mathcal C }.$
![The results of the proposed strategy of segmentation are shown for three different images in the three columns (a), (b) and (c). The actual image, the dilated image and the segmented image is shown for each of the examples in row 1, row 2 and row 3 respectively. []{data-label="fig:seg"}](images/segmentation.png){width="0.6\linewidth"}
Super-Pixel Similarity
----------------------
Following the strategy adopted by [@main], we incorporate the similarity of neighbourhood super-pixels based on a function $w(s,s')$, for all $s,s'\in\mathcal S$. We combine information available over the entire set of features and spatial distance to calculate the similarity function w(.,.) between neighbourhood super-pixels. Denote the Euclidean distance between features of two super-pixels $s,s'$ by $d(\bf x_s, \bf x_{s'})$ and the standard deviation across all super-pixels by $\sigma_x$. The spatial Euclidean distance between a pair of super-pixels is given by $d(s, s')$ and the average distance across all super-pixels by $\overline d(\mathcal S).$ Combining the feature level information and spatial distance between super-pixels, the similarity function is given by $w(s, s')=exp(-\frac{d(\bf x_s, \bf x_{s'})}{2\sigma_x^2})\big(\frac{d(s,s')}{\overline d(\mathcal S)}\big)^{-1}.$
Segment Classification
----------------------
Based on the computed features and weight function we classify the super-pixels into a number of classes in an unsupervised manner. Let us denote the unknown classes of the super-pixels by $Y=\{y_s, s\in \mathcal S\}$. If there are $K$ segments present in a given image, denoting $K$ classes, we have $y_s\in \{1,2,\ldots, K\}$ for $s\in \mathcal S.$ The class information given by the joint class probability function is factorized as $p(Y)=\prod_{s\in \mathcal S}\pi(y_s)\prod_{s,s'\in \mathcal S}R(y_s,y_s')$ where the class prior probabilities are given by $\pi(y_s).$ The mutual information between a pair of neighbourhood super-pixels are given by $R(y_s,y_{s'})=\beta w(s,s')B(y_s,y_{s'}),$ $\beta>0$ being a tuning parameter controlling the spatial regularization. Here $B(y_s,y_{s'})$ is a spatial regularisation function indicating the chance of two neighbouring super-pixels to belong to the same class. We chose a diagonal structure of the matrix $[B(y_s,y_{s'}), s, s'\in\mathcal S]$ making all the diagonal elements to be identical to $1$.
Given a fixed class $k$, the features are assumed to have a Gaussian distribution with fixed mean $\mu_k$ and variance-covariance matrix $\Sigma_k$ given by $p(\bf {x}_s|y_s=k)=N_k(\mu_k, \Sigma_k).$ Hence the super-pixel class is predicted by estimating the model parameters using the Expectation-Maximization algorithm and hence evaluating $$\begin{split}
(\hat{y}_s, s\in \mathcal S)=ArgMax_{{y}_s, s\in \mathcal S} \prod_{s\in \mathcal S}p(\bf {x}_s|y_s)\pi(y_s)\\\prod_{s,s'\in \mathcal S}R(y_s,y_{s'}).
\end{split}$$ The estimated class information of the super-pixels is used to merge super-pixels of the same class level to get different segments. For three selected examples of images, the results of the proposed strategy after the dilation and segmentation is shown in Fig.\[fig:seg\]. In the next section, we describe the text detection strategy where the ensemble of the CNN models is used to probe in each of the detected segments to extract texts of various sizes.
Text Detection
==============
The task of text detection in an image is very similar to object detection, where the text can be treated as an object. Hence, all object detection models can be used by making them binary classifiers - text (word level) and non-text. But all these object classifiers have their own limitations.
- Sometimes the image has a large amount of text compacted in a region forming a text cluster. Detecting these words separately becomes hard for conventional object detection techniques as they are trained to recognize a few numbers of separable objects in an image.
- Text in a single image can vary in both font-sizes and font-styles in a single image. Though it is claimed that most object detection methods are scale invariant, the results say otherwise as shown in Singh [@Singh18].
As shown in Liao [@TB2016; @TBPP2018] having wide kernels ($3 \times 5$) rather than the generic square kernels help in better coverage in detecting text. Text in most cases, unlike objects, has a rectangular aspect ratio. Wide kernels will capture information about the aspect ratio of text objects better than square kernels. An ensemble of multiple CNN based models ensures a different level of information will be captured by different kind of models resulting in better coverage in information gathered from image.
The models are then stitched together using selective non-max suppression algorithm. Non-Maximal Suppression removes multiple overlapping boxes detected for the same text location and keeps the one with the highest probability. Selective non-maximal suppression does the same but also takes into account the accuracy of the model from which the bounding box has been generated, giving it higher preference. Predictions from models which have a higher accuracy are preferred over others even if the individual probability might be slightly smaller.
Non-Maximal Suppression (NMS)
-----------------------------
Let us assume that there are $n$ models and the number of bounding boxes predicted by $j^{th}$ model be $n_j$. Let $K$ be the list of all bounding boxes such that $k_{ij}$ is the $i^{th}$ bounding box predicted by model $j$ with $p_{k_{ij}}$ being the probability of that bounding box containing text. Let $\ell$ represent a sorted ordering of all these bounding boxes. That implies $\lvert \ell \rvert = \lvert K \rvert = n_1 + n_2 + n_3 + ... + n_n$.
$\ell$ = sort($k$, $p$, $desc$) $\ell.pop(j)$
Selective Non-Maximal Suppression (sNMS)
----------------------------------------
Let $M_q$ denote the model with the highest accuracy $a_q$ among all other models $M_i$ where $i \in \{1, 2, ..., n$}. Let $P_t$ be the threshold probability - the probability that the bounding box is considered a true text box predicted by a model. $P_t$ is kept high for $M_q$, say $P_{t_h}$ while $P_t$ is kept slightly lower for the other $n-1$ models, say $P_{t_l}$. The bounding boxes predicted by each of the models are first filtered using this. After that, the probability of all the $n_q$ predicted boxes of model $M_q$ is assigned to $1$, while the probability of other boxes is left untouched. Post this reassignment of probabilities, NMS is performed on all the predicted boxes from all $n$ models.
$q = max(a)$ remove boxes where $p_q < P_{t_h}$ $p_q = 1$ remove boxes where $p_r < P_{t_l}$ $bbs = $NMSalgorithm($k$, $p$, $nmsThreshold$)
#### Notes and Comments.
Selective NMS ensures that the text boxes predicted with high probability by a model with the highest accuracy will always have priority over similar text boxes predicted with high probability by other models.
Implementation and Experiments
==============================
Our proposed methodology employs multiple models, to detect text from an image, which are stitched together using the selective-NMS algorithm. Multiple pre-trained models, as developed in Liao [@TB2016; @TBPP2018] were used to detect text boxes from images. For selective-NMS, corresponding to the model with the highest accuracy of $0.9$ was set as a probability threshold above which a bounding box was considered a true text box with high confidence. The same parameter was set to 0.8 for the other models. NMS Threshold, the ratio of intersection area of two text boxes to union area of them (IOU), was set to $95\%$ i.e. with IOU above $95\%$ between two text boxes, they are considered to contain the same text. For text recognition, we synthesised 9 million text images using **SynthText** for various size, style and background of text for training. The full training set was run in a computer with standard K80 GPU and average execution time for detecting text in a single image is recorded to be around $0.15s$.
The ICDAR2013 dataset consists of images where the user is explicitly directing the focus of the camera on the text content of interest in a real scene. Product image dataset, on the other hand, consists of images of items taken from a high-resolution camera and have no background (white). By converting the image to grey-scale we calculate the entropy of the images in areas where the text is present. The average entropy of a sample of images from ICDAR2013 [@ICDAR2013] dataset was around $7.0$ while that of images from Walmart dataset [@WALMART2019] was around $6.0$ with $6.5$ marking a demarcation boundary for separating the two datasets.
Experimental Results
--------------------
ICDAR2013 [@ICDAR2013] contains high-resolution real-world images. The models have been trained on the ICDAR2013 train set and then tested on the ICDAR2013 validation set. The results from all the models are then passed through selective NMS and the final bounding boxes are used for computing the metrics for precision, recall and f-score. Table.\[tab:comp\] summarizes and compares our results with other methods. A sample of the results are shown in Fig.\[fig:icdres\]. The predicted text regions are placed on top of the true text regions, indicated by blue and red boxes respectively.
![Sample of results for text localization on ICDAR2013 dataset (top row) and high entropy image dataset (bottom row). Blue boxes represent the predicted bounding box while the red represent the ground truth bounding box.[]{data-label="fig:icdres"}](images/icdar_result.png){width="0.8\linewidth"}
**Datasets** **Time/s**
------------------------ ---------- ------------ ---------- ----------
**Methods** **P** **R** **F**
0.86 0.70 0.77 0.75
0.85 0.76 0.80 1.4
0.88 0.78 0.83 2.1
0.80 0.60 0.68 0.1
TextBoxes[@TB2016] 0.86 0.74 0.80 0.09
TextBoxes++[@TBPP2018] 0.86 0.74 0.80 0.10
Ensemble **0.83** **0.77** **0.80** **0.15**
: Text localization on ICDAR2013. P, R and F refer to precision, recall and F-measure respectively.[]{data-label="tab:comp"}
The ensemble model is also tested on a dataset containing publicly available product images on the Walmart website [@WALMART2019]. These are high resolution and high entropy images of the front face of processed food items used on a daily basis by consumers. The predicted text region bounding boxes enclose regions containing texts of multiple sizes and mixed font types in the same image. This is particularly important for product images as the product labels often contain texts of multiple fonts. The proposed text detection strategy also successfully detects text regions when the text is moderately rotated or curved due to the shape of the product package, eg- a can or a bottle (see the 2nd, 3rd and 4th images in the bottom row in Fig.\[fig:icdres\] ). The use of wide kernels turn out to be useful in detecting horizontal text boxes and on top of it, the image segmentation and CNN ensemble network consider image convolution filters at multiple scales and rotation angles. This contributes to ensuring that the text box detection accuracy is invariant at least under limited distortion and rotation of the horizontal orientation of the texts.
The models trained on ICDAR2013 train set are used on 50 images from this dataset where the ground truth boxes are known. The main difference between the images in this dataset and the other publicly available datasets is that the images have no background noise that is usually present in scene text. However multiple texts are usually present in a small region of the image along with various other objects resulting in high local entropy. Most of the models currently available perform poorly on detecting text in such regions in the image. In such cases, the model described in this paper performs better than the existing ones in terms of precision, recall as well as f-score. Fig.\[fig:icdres\] gives the results achieved on the datasets [@WALMART2019; @ICDAR2013] by the model. In the case of ICDAR2013 dataset, the model has performed at par with the existing models currently available, but this improves drastically in the case of the dataset containing high entropy images [@WALMART2019]. The precision is at least $6\%$ higher than the existing methods while recall is higher by around $15\%$. Table.\[tab:wal\] compares the results achieved on the Walmart High Entropy Images.
**Datasets**
-------------- ----------- ----------- -----------
**Methods** **P** **R** **F**
TextBoxes 0.867 0.264 0.405
TextBoxes++ 0.831 0.311 0.453
Ensemble **0.920** **0.467** **0.619**
: Text localization results on High entropy image dataset. P, R and F refer to precision, recall and F-measure respectively.[]{data-label="tab:wal"}
Conclusion
==========
We have presented an algorithm which employs an ensemble of multiple fully convolutional networks preceded by an image segmenter for text detection. It is highly stable and parallelizable. It can detect words of varied sizes in an image which is very high on entropy. Comprehensive evaluations and comparisons on benchmark datasets clearly validate the advantages of our method in three related tasks including text detection, word spotting and end-to-end recognition. It even exhibits better performance than [*TextBox*]{} and [*TextBox++*]{} [@TB2016; @TBPP2018] in detecting graphical text in an image. The ICDAR2013 dataset images have real-world contents and background noise surrounding the true text regions, unlike the Walmart High Entropy Images, where the challenge is largely the presence of multiple textual elements within small regions resulting higher entropy. The proposed methodology is particularly targeted to work on such high entropy text regions and hence performs very well on Walmart High Entropy Images. However, a more targeted background removal strategy, image segmentation and text candidate pre-filtering using text region specific key point identification and feature descriptions such as Stroke width descriptors, Maximally Stable Extremal Region descriptors should enhance the performance of the CNN ensemble model more. We keep this development for future communication.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
We compute the two-loop corrections of ${\cal O}(\alpha_s^2)$ to the Yukawa couplings in the framework of the Minimal Supersymmetric Standard Model (MSSM). The calculation is performed using the effective Lagrangian approach under the approximation of neglecting the Higgs boson mass with respect to the top quark, gluino and all squark flavour masses. As an application we derive the ${\cal O}(\alpha_s^2)$ corrections to the partial decay width of the lightest Higgs boson to a bottom quark pair. We find that the two-loop corrections are sizable for large values of $\tan\beta$ and low CP-odd Higgs boson mass. With our calculation of the ${\cal O}(\alpha_s^2)$ corrections the remaining theoretical uncertainties reduce below a few percent.
PACS numbers: 11.30.Pb, 12.38.-t, 12.38.Bx, 12.10.Kt
author:
- |
L. Mihaila, C. Rei[ß]{}er\
[*Institut f[ü]{}r Theoretische Teilchenphysik,*]{}\
[*Karlsruhe Institute of Technology (KIT)*]{}\
[*76128 Karlsruhe, Germany*]{}\
title: |
-3cm
SFB/CPP-09-120\
TTP/09-45
1.5cm ${\cal O}(\alpha_s^2)$ corrections to fermionic Higgs decays\
in the MSSM
---
Introduction
============
One of the main purposes of the CERN Large Hadron Collider (LHC) is the search for Higgs bosons. The discovery of a light Higgs boson is a decisive test for all models predicting supersymmetric (SUSY) particles at the TeV scale and in particular of the MSSM [@Nilles:1983ge]. A remarkable feature of the MSSM is the restricted Higgs sector. This allows Higgs searches without any assumption about the mechanism of SUSY breaking, but only constraints from the Higgs sector [@Carena:2002qg]. For Higgs boson searches at the hadron colliders, two new complementary benchmark scenarios, the “small $\alpha_{\rm eff}$” and the “gluophobic” scenarios, have been proposed in addition to those used at the CERN Large Electron-Positron Collider (LEP) for the MSSM Higgs searches at hadron colliders.\
More precisely, in the “gluophobic” scenario the gluon fusion process is strongly suppressed due to cancellation between top quark and squark loop contributions. Nevertheless, the channel $t\bar{t}\to t\bar{t} h\to
t\bar{t} b\bar{b}$ is enhanced as compared to the Standard Model (SM) case, so that this becomes the most promising detection mode. In the “small $\alpha_{\rm eff}$” scenario the decay width for $h\to b\bar{b}$ is much smaller than its SM value. In this case the complementary channel $h\to\gamma\gamma$ is enhanced as compared to the SM and it becomes the preferred detection mode.\
For a light Higgs boson ($m_h\le 130$ GeV) the decay $h\to b\bar{b}$ is the dominant mode, but its detection at hadron colliders is difficult due to large QCD backgrounds for the $b$ jets. However, at lepton colliders the Higgs boson search relies on $b$ tagging that can be performed with high efficiency.
For the discovery of the Higgs bosons, the cross section of the main production channels, decay widths and branching ratios are necessary to be known with high accuracy. Within the SM the radiative corrections to the fermionic Higgs decay were intensively studied in the literature. The QCD and EW corrections are known up to the three-loop order: ${\cal O}(\alpha_s^3)$ originating from the light degrees of freedom were first derived in Ref. [@Chetyrkin:1996sr] and the top-induced ${\cal O}(\alpha_s^3)$ corrections in Ref. [@Chetyrkin:1997vj]; the QCD-EW interference contributions of ${\cal O}(\alpha_s^2 x_t)$[^1] can be found in Ref. [@Chetyrkin:1996ke]. In this paper we concentrate on the fermionic Higgs decay in the MSSM, for moderate Higgs boson masses $M_h\le 130$ GeV. The processes $h\to b\bar{b}$ and $h \to \tau^+\tau-$ (the second most important decay mode) are affected by large radiative corrections for scenarios with large values of $\tan\beta$ and moderate values of the neutral CP-odd Higgs boson mass $M_A$ ($\tan\beta\ge 20\,, M_A\le 250$ GeV) [@Hall:1993gn]. Apart from pure QCD and EW corrections mentioned above, there are Higgs boson propagator corrections and vertex corrections due to SUSY particles. The first class of radiative corrections, can be taken into account by introducing the effective mixing angle $\alpha_{\rm eff}$ that diagonalizes the neutral Higgs boson mass matrix [@Heinemeyer:2000fa]. Such type of corrections are known analytically up to two-loop order in supersymmetric QCD (SQCD) and the supersymmetric electroweak theory (SEW), see [@FeynHiggs] and references therein. The second class of radiative corrections are especially important for large values of $\tan\beta$. Usually, they are derived using the effective Lagrangian approach [@Carena:1999py]. The one-loop contributions are known since long [@Hall:1993gn; @Guasch:2003cv], while the two-loop corrections have been computed very recently[@Noth:2010jy].\
It is the aim of this paper to present the complete two-loop SQCD corrections to the decay width $h\to b\bar{b}$, taking into account the exact dependence on the supersymmetric particle masses and working in the full theory.
The paper is organized as follows: in the next section we introduce our framework and the quantities required for the computation of the decay width $\Gamma(h\to q\bar{q})$ through two loops in the MSSM. In Section 3 we describe the actual two-loop calculation, pointing out the connection between the radiative corrections to the vertex $hq\bar{q}$ and those to the quark propagator through the low-energy theorem. The formalism discussed in this section is valid for a general quark flavour. However, we specify our calculation for the case of bottom quark which generates the dominant decay mode of a light Higgs boson. In Section 4 we perform a numerical analysis and discuss the phenomenological implications of the two-loop SQCD vertex corrections to $\Gamma(h\to b\bar{b})$. We present the conclusions in Section 5.
\[sec::framework\] Notation and theoretical framework
=====================================================
The part of the MSSM Lagrangian describing the fermionic Higgs decay can be written in the following form $$\begin{aligned}
{\cal L}&=&{\cal L}_{\rm QCD}+ {\cal L}_{\rm SQCD}
+\sum_{i=1,2}{\cal L}_{{\rm q}\phi_i}
+ \sum_{i=1,2}{\cal L}_{{\rm \tilde{q}}\phi_i}\,\end{aligned}$$ where $$\begin{aligned}
{\cal L}_{{\rm q}\phi_i}=-\sum_{q=1}^6 \frac{m_q}{v} g_q^{\phi_i}
\bar{q} q \phi_i \quad \mbox{and} \quad
{\cal L}_{{\rm \tilde{q}}\phi_i}=-\sum_{q=1}^6\sum_{r,k=1,2}
\frac{m_q}{v} g_{\tilde{q};kr}^{\phi_i}\tilde{q}^\star_{k}\tilde{q}_r \phi_i\,.\end{aligned}$$ ${\cal L}_{\rm QCD}+ {\cal L}_{\rm SQCD}$ denotes the supersymmetric extension of the full QCD Lagrangian with six quark flavours. The couplings $g_q^{\phi_i}$ and $g_{\tilde{q};kr}^{\phi_i}$ are defined in Table \[tab::yukawa\_coeff\], $m_q$ denotes the mass of quark $q$, $v=\sqrt{v_1^2+v_2^2}$ with $v_i\,,i=1,2$, the vacuum expectation values of the two Higgs doublets of the MSSM. The fields $\tilde{q}_i\,,i=1,2$, denote the squark mass eigenstates, while $\theta_q$ stands for the mixing angle defined through: $$\begin{aligned}
\sin 2\theta_q=\frac{2 m_q X_q}{m_{\tilde{q}_1}^2-m_{\tilde{q}_2}^2}\,,\quad
X_q = A_q-\mu_{\rm SUSY} \bigg\{\begin{array}{ll}
\tan\beta\,, &\mbox{ for down-type quarks}\\
\cot\beta\,, &\mbox{ for up-type quarks}
\end{array} \,,
\label{eq::mixangle}\end{aligned}$$ where $A_q$ is the trilinear coupling and $\mu_{\rm SUSY} $ the Higgs-Higgsino bilinear coupling. The fields $\phi_i\,, i=1,2$, denote the neutral CP-even components of the MSSM Higgs doublets and they are related to the Higgs mass eigenstates through the orthogonal transformation $$\begin{aligned}
\left(
\begin{array}{c}
H\\h
\end{array}
\right) =
\left(
\begin{array}{cc}
\cos\alpha&\sin\alpha\\
-\sin\alpha & \cos\alpha
\end{array}
\right)
\left(
\begin{array}{c}
\phi_1\\
\phi_2
\end{array}
\right)\,.
\label{eq::hmix}\end{aligned}$$ As usual, $h$ stands for the lightest Higgs boson. The mixing angle $\alpha$ is determined at the leading order through $$\begin{aligned}
\tan 2\alpha= \tan 2 \beta \frac{M_A^2+M_Z^2}{M_A^2-M_Z^2}\,;
\quad -\frac{\pi}{2}<\alpha<0\, ,
\label{eq::h_alpha}\end{aligned}$$ where $M_Z$ is the mass of the Z boson and $\tan \beta=v_2/v_1$.
------------------------------------------------------------------------------------------------------------------------------------------
f $g_q^{\phi_1}$ $g_{\tilde{q};11}^{\phi_1}$ $g_{\tilde{q};12}^{\phi_1}= $g_{\tilde{q};22}^{\phi_1}$
g_{\tilde{q};21}^{\phi_1}$
------ ---------------- ----------------------------- ------------------------------------------------------ -----------------------------
up $0$ $-\mu S_q/S_\beta$ $-\mu C_q/S_\beta$ $\mu S_q/S_\beta$
down $1/C_\beta$ $(2 m_q+A_q S_q)/C_\beta$ $ $(2 m_q-A_q S_q)/C_\beta$
A_q C_q/C_\beta$
f $g_q^{\phi_2}$ $g_{\tilde{q};11}^{\phi_2}$ $ $
g_{\tilde{q};12}^{\phi_2}=g_{\tilde{q};21}^{\phi_2}$ g_{\tilde{q};22}^{\phi_2}$
up $1/S_\beta$ $(2 m_q+A_q S_q)/S_\beta$ $A_q C_q/S_\beta$ $
(2 m_q-A_q S_q)/S_\beta$
down $0$ $-\mu S_q/C_\beta$ $-\mu C_q/C_\beta$ $\mu S_q/C_\beta$
------------------------------------------------------------------------------------------------------------------------------------------
: \[tab::yukawa\_coeff\]Yukawa coupling coefficients for up and down type quark and squark, where $S_q=\sin 2\theta_q$ and $C_q=\cos 2 \theta_q$, and $S_\beta=\sin \beta$ and $C_\beta=\cos \beta$.
In the following, we assume the mass of the lightest Higgs boson $h$ to be much smaller than the mass of the top-quark and of the SUSY particles, as well as all the other Higgs bosons. In this case, the physical phenomena at low energies can be described with an effective theory containing five quark flavours and the light Higgs. At leading order in the heavy masses, the effective Lagrangian ${\cal L}_Y^{\rm eff}$ can be written as a linear combination of three physical operators [@Spiridonov:1984; @Chetyrkin:1997un] constructed from the light degrees of freedom $$\begin{aligned}
{\cal L}\longrightarrow {\cal L}_Y^{\rm eff} + {\cal L}_{\rm QCD}^{(5)}\,;\quad
{\cal L}_Y^{\rm eff} = -\frac{h^{(0)}}{v^{(0)}}\left[
C_1^0{\cal O}_1^0 + \sum_{q}\left( C_{2q}^0{\cal O}_{2q}^0
+ C_{3q}^0{\cal O}_{3q}^0\right)
\right]\,,
\label{eq::eft}\end{aligned}$$ where $ {\cal L}_{\rm QCD}^{(5)}$ denotes the Lagrangian of QCD with five active flavours and the coefficient functions $C_i\,, i=1,2q,3q$, parametrize the effects of the heavy particles on the low-energy phenomena. The superscript $0$ labels bare quantities. The three operators are defined as
$$\begin{aligned}
{\cal O}_1^0 &=& (G_{\mu,\nu}^{0,\prime,a})^2\,,\nonumber\\
{\cal O}_{2q}^0 &=& m_q^{0,\prime}\bar{q}^{0,\prime}
q^{0,\prime}\,,\nonumber\\
{\cal O}_{3q}^0 &=& \bar{q}^{0,\prime}(i\,/\!\!\!\! D^{0,\prime}
-m_q^{0,\prime})q^{0,\prime}\,,
\label{eq::ops}\end{aligned}$$
where $G_{\mu,\nu}^{0,\prime,a}$ and $ D_{\mu}^{0,\prime}$ are the gluon field strength tensor and the covariant derivative, respectively, and the primes label the quantities in the effective theory. The operator ${\cal O}_{3q}$ vanishes by the fermionic equation of motion and it will not contribute to physical observables. So, the last term in Eq. (\[eq::eft\]) might be omitted, once the coefficients $C_1^0, C_{2q}^0$ are determined. The coefficient functions contain information about the heavy particles that were integrated out in the construction of the effective theory. On the contrary, as can be understood from Eq. (\[eq::ops\]), the operators encounter only the effects of the light degrees of freedom.
The relations between the parameters and fields in the full and effective theories are given by $$\begin{aligned}
G_{\mu,\nu}^{0,\prime,a}&=&(\zeta_3^{(0)})^{1/2}
G_{\mu,\nu}^{0,a}\,,\nonumber\\
q^{0,\prime}&=& (\zeta_2^{(0)})^{1/2} q^{0}\,,\nonumber\\
g_s^{0,\prime}&=& \zeta_g^{(0)} g_s^{0}\,,\nonumber\\
m_q^{0,\prime}&=&\zeta_m^{(0)}m_q^{0}\,,
\label{eq::dec}\end{aligned}$$ where $g_s=\sqrt{4\pi\alpha_s}$ is the strong coupling. The coefficients $\zeta_3^{(0)}\,,\zeta_2^{(0)}\,,\zeta_g^{(0)}\,,\zeta_m^{(0)}$ are the bare decoupling coefficients. They may be computed from the transverse part of the gluon polarization function and the vector and scalar part of the quark self-energy via [@Chetyrkin:1997un] $$\begin{aligned}
\zeta_3^{(0)}=1+\Pi^{0,h}(0)\,,\nonumber\\
\zeta_2^{(0)}=1+\Sigma_v^{0,h}(0)\,,\nonumber\\
\zeta_m^{(0)}=\frac{1-\Sigma_s^{0,h}(0)}{1+\Sigma_v^{0,h}(0)}\,.
\label{eq::letse}\end{aligned}$$ For the derivation of the coefficient $\zeta_g^{(0)}$ one has to consider in addition one vertex involving the strong coupling, for example $\bar{q}qg$ or $\bar{c}cg$, where $c$ denotes the Faddeev-Popov ghost. The decoupling coefficients are independent of the momentum transfer, so that they can be evaluated at vanishing external momenta. The superscript $h$ indicates that in the framework of Dimensional Regularization (DREG) or Dimensional Reduction (DRED) only diagrams containing at least one heavy particle inside the loops contribute and that only the hard regions in the asymptotic expansion of the diagrams are taken into account. They have been computed in QCD including corrections up to the four-loop order for the strong coupling [@Schroder:2005hy] and three-loop order for quark masses [@Chetyrkin:1997un]. In the MSSM the two-loop SQCD [@Harlander:2005wm; @Bednyakov:2007vm; @Bauer:2008bj] and SEW [@Bednyakov:2009wt] expressions are known. Similar to the case of SM, the decoupling coefficients derived within the MSSM can be connected through the Low Energy Theorem (LET) [@Ellis:1975ap] with the coefficients $C_1^0, C_{2q}^0$. We discuss in more detail the relation between the coefficients $C_{2q}^0, C_{3q}^0$ and $\zeta_m^{(0)}, \zeta_2^{(0)}$ in Subsection \[sec::letsusy\].\
The renormalization procedure of the dimension four operators in the Minimal Subtraction Scheme within DREG ([$\overline{\mbox{{\scalefont{.9}}MS}}$]{}) [@Bardeen:1978yd] is known since long time [@Spiridonov:1984]. The main aspect is that different operators in general mix under renormalization. For the convenience of the reader we reproduce the results for the renormalization constants of the operators ${\cal O}_1^0$ and ${\cal O}_{2q}^0$ that are of interest for the fermionic Higgs decays $$\begin{aligned}
&&{\cal O}_1= Z_{11} {\cal O}_1^0 +Z_{12} {\cal O}_{2q}^0\,,\qquad
{\cal O}_2= Z_{22}{\cal O}_{2q}^0\,,\quad \mbox{where} \nonumber\\
&&Z_{11}=
\left(1-\frac{\pi}{\alpha_s^\prime}\frac{\beta(\alpha_s^\prime)}
{\epsilon}\right)^{-1}
, \,\,\, Z_{12}=-\frac{4\gamma_m(\alpha_s^\prime)}{\epsilon}
\left(1-\frac{\pi}{\alpha_s^\prime}\frac{\beta(\alpha_s^\prime)}
{\epsilon}\right)^{-1},
\, Z_{22}=1\,,\\
&&C_1= Z_{11}^{-1} C_1^0 \,,\qquad\qquad\qquad C_{2q}=C_{2q}^0-
\frac{Z_{12}}{Z_{11}} C_1^0 \,.\end{aligned}$$ The explicit expressions for the $\beta$-function and quark mass anomalous dimension $\gamma_m$ of QCD with $n_l=5$ active flavours at the one-loop order that are needed in the present paper, are given by $$\begin{aligned}
\beta(\alpha_s^\prime)&=&-\left(\frac{\alpha_s^\prime}{\pi}\right)^2\beta_0
+{\cal O}((\alpha_s^\prime)^3)\,,\quad \beta_0 =\frac{11}{4}
-\frac{n_l}{6}\,,\nonumber\\
\gamma_m(\alpha_s^\prime) &=& -\frac{\alpha_s^\prime}{\pi} \gamma_0 +{\cal
O}((\alpha_s^\prime)^2)\,,\qquad\quad \gamma_0= \frac{3}{4} C_F\,.
\label{eq::andim}\end{aligned}$$ The bare coefficient functions $C_i^0\,, i=1,2q$, must be computed diagrammatically. For the calculation of the ${\cal O}(\alpha_s^2)$ corrections to the process $h\to q \bar{q}$, the knowledge of the coefficient functions $C_1^0$ and $C_{2q}^0$ is required at the one- and two-loop order, respectively.
The renormalized coefficient functions and operators are finite but not renormalization group (RG) invariant. In Ref. [@Chetyrkin:1996ke], a redefinition of the coefficient functions and operators was introduced so that they are separately RG invariant. This procedure allows us to choose independent renormalization scales for coefficient functions and operators. In practice, one makes a separation of scales: one chooses $\mu\approx M_h$ for the renormalization scale of the operators and $\mu\approx \tilde{M}$ (where $\tilde{M}$ denotes an averaged mass for the heavy supersymmetric particles) for the coefficient functions. The new coefficient functions read [@Chetyrkin:1996ke] $$\begin{aligned}
{\cal C}_1(\tilde{M}, M_h)&=&
\frac{\alpha_s^\prime(\tilde{M})\beta^{(5)}(\alpha_s^\prime(M_h))}
{\alpha_s^\prime(M_h)\beta^{(5)}(\alpha_s^\prime(\tilde{M}))}
C_1(\tilde{M})\,,\nonumber\\
{\cal C}_2(\tilde{M}, M_h)&=& \frac{4 \alpha_s^\prime(\tilde{M})}{\pi
\beta^{(5)}(\alpha_s^\prime(\tilde M))}[\gamma_m^{(5)}(\alpha_s^\prime(\tilde{M}))
-\gamma_m^{(5)}(\alpha_s^\prime(M_h))]C_1(\tilde{M})+
C_{2q}(\tilde{M})\,,\end{aligned}$$ where the superscript $(5)$ marks that $n_l=5$ in Eq. (\[eq::andim\]). We employ this approach for the evaluation of the decay width $\Gamma(h\to \bar{q}q)$. More details about the practical calculation are discussed in the next section.
\[sec::hdecay\] Higgs decay width
---------------------------------
Once the renormalized coefficient functions ${\cal C}_1, {\cal C}_2$ are known, the decay width for the process $h\to \bar{q}q$ can be predicted. From Eqs. (\[eq::eft\]) and (\[eq::ops\]) one can derive a general formula for the inclusive $h\to \bar{q}q$ decay width [@Chetyrkin:1996ke] $$\begin{aligned}
\Gamma(h\to \bar{q}q) = \Gamma^{(0)}(1+\bar{\delta}_u)^2\bigg[
(1+\Delta_q^{\rm QCD}) {\cal C}_2^2+ \Xi_q^{\rm QCD}{\cal C}_1{\cal C}_2
\bigg]\,,
\label{eq::gamma_gen}\end{aligned}$$ where $\Gamma^{(0)}$ represents the complete leading order (LO) result given by $$\begin{aligned}
\Gamma^{(0)} &=& \frac{N_c G_F M_h m_q^2}{4\pi\sqrt{2}}
\left(1-\frac{4m_q^2}{M_h^2}\right)^{3/2}\,.
\label{eq::born}\end{aligned}$$ As is well known, the large logarithms of the type $\ln(M_h^2/m_q^2)$ can be resummed by taking $m_q$ in Eq. (\[eq::born\]) to be the [$\overline{\mbox{{\scalefont{.9}}MS}}$]{} mass $m_q^{{\overline{\rm{\scalefont{.9}}MS}}}(\mu)$ evaluated at the scale $\mu=M_h$. The QCD correction $\Delta_q^{\rm QCD}$ is known since long time [@Gorishnii:1990zu], $$\begin{aligned}
\Delta_q^{\rm QCD} &=&
\frac{\alpha_s^{\prime}(\mu)}{\pi}\left(\frac{17}{3} + 2
\ln\frac{\mu^2}{M_h^2}\right)\nonumber\\
&+&
\left(\frac{\alpha_s^{\prime}(\mu)}{\pi}\right)^2\bigg[
\frac{8851}{144}-\frac{47}{6}\zeta(2)-\frac{97}{6}
\zeta(3)+\frac{263}{9}\ln\frac{\mu^2}{M_h^2}
+ \frac{47}{12}\ln^2\frac{\mu^2}{M_h^2}
\bigg]\,,\end{aligned}$$ with $\zeta(x)$ being the Riemann’s zeta function. The additional QCD correction generated through double-triangle topologies $\Xi_q^{\rm QCD}$ was first computed in Ref. [@Chetyrkin:1996ke], $$\begin{aligned}
\Xi_q^{\rm QCD} &=& \frac{\alpha_s^{\prime}(\mu)}{\pi} {C_F} \left(-19
+6\zeta(2)
-\ln^2\frac{m_q^2}{M_h^2} -6 \ln\frac{\mu^2}{M_h^2}\right)\,.\end{aligned}$$ The universal corrections $\bar{\delta}_u$ of ${\cal
O}(\alpha_s^n x_t)$, where $x_t=(\alpha_t/4\pi)^2=G_F M_t^2/(8\pi^2\sqrt{2})$, with $\alpha_t$ the top-Yukawa coupling, contain the contributions from the renormalization of the Higgs wave function and the vacuum expectation value [@Kwiatkowski:1994cu], $$\begin{aligned}
\bar{\delta}_u &=& x_t\left[\frac{7}{2} +
\frac{\alpha_s^{\prime}(\mu)}{\pi}\left(\frac{19}{3} -2\zeta(2)
+7\ln\frac{\mu^2}{M_t^2} \right) +{\cal O}(\alpha_s^2)
\right]\,.\end{aligned}$$ The coefficient functions $C_1\,, C_{2q}$ (and implicitly ${\cal C}_1, {\cal C}_2$) are known within SQCD at the one-loop order since quite some time [@Spira:1993bb; @Guasch:2003cv]. For completeness, we display them here providing also ${\cal O}(\epsilon)$ terms that are necessary for the two-loop calculation. $$\begin{aligned}
C_1&=&-\frac{\alpha_s^\prime(\mu)}{12\pi}\Bigg\{-\frac{\sin\alpha}{\cos\beta}
\Bigg[\frac{M_t^2\mu_{\rm SUSY} X_t}{4{m_{\tilde{t}_1}}^2{m_{\tilde{t}_2}}^2\tan\beta}
-\epsilon \frac{M_t\mu_{\rm SUSY}\sin
2\theta_t}{8\tan\beta}\left(\frac{{L_{\tilde{t}_1}}}{{m_{\tilde{t}_1}}^2} -
\frac{{L_{\tilde{t}_2}}}{{m_{\tilde{t}_2}}^2}\right) \Bigg]
\nonumber\\
&&+\frac{\cos\alpha}{\sin\beta}\Bigg[
\frac{4{m_{\tilde{t}_1}}^2{m_{\tilde{t}_2}}^2 + {m_{\tilde{t}_1}}^2 M_t^2 + {m_{\tilde{t}_2}}^2 M_t^2 - A_t M_t^2
X_t}{4{m_{\tilde{t}_1}}^2{m_{\tilde{t}_2}}^2}\nonumber\\
&&
+\epsilon \frac{A_t M_t
\sin 2\theta_t}{8}\left(\frac{{L_{\tilde{t}_1}}}{{m_{\tilde{t}_1}}^2} -
\frac{{L_{\tilde{t}_2}}}{{m_{\tilde{t}_2}}^2}\right) +
\epsilon\frac{M_t^2}{4}\left(\frac{4 {L_t}}{M_t^2}
+ \frac{{L_{\tilde{t}_1}}}{{m_{\tilde{t}_1}}^2} + \frac{{L_{\tilde{t}_2}}}{{m_{\tilde{t}_2}}^2}\right)
\Bigg]
\Bigg\}\,,
\\
C_{2q}&=&-\frac{\sin\alpha}{\cos\beta}\frac{1+\frac{\alpha_s^\prime(\mu)}{2\pi} C_F
A_b {m_{\tilde g}}\bigg[F_1({m_{\tilde{b}_1}}^2,{m_{\tilde{b}_2}}^2,{m_{\tilde g}}^2)+\epsilon
F_2({m_{\tilde{b}_1}}^2,{m_{\tilde{b}_2}}^2,{m_{\tilde g}}^2)\bigg]}{1+\frac{\alpha_s^\prime(\mu)}{2\pi} C_F
X_b {m_{\tilde g}}\bigg[F_1({m_{\tilde{b}_1}}^2,{m_{\tilde{b}_2}}^2,{m_{\tilde g}}^2)+\epsilon
F_2({m_{\tilde{b}_1}}^2,{m_{\tilde{b}_2}}^2,{m_{\tilde g}}^2)\bigg]}
\nonumber\\
&&
+\frac{\cos\alpha}{\sin\beta}
\frac{\frac{\alpha_s^\prime(\mu)}{2\pi} C_F
(-\mu_{\rm SUSY}\tan\beta) {m_{\tilde g}}\bigg[F_1({m_{\tilde{b}_1}}^2,{m_{\tilde{b}_2}}^2,{m_{\tilde g}}^2)+\epsilon
F_2({m_{\tilde{b}_1}}^2,{m_{\tilde{b}_2}}^2,{m_{\tilde g}}^2)\bigg]}{1+\frac{\alpha_s^\prime(\mu)}{2\pi} C_F
X_b {m_{\tilde g}}\bigg[F_1({m_{\tilde{b}_1}}^2,{m_{\tilde{b}_2}}^2,{m_{\tilde g}}^2)+\epsilon
F_2({m_{\tilde{b}_1}}^2,{m_{\tilde{b}_2}}^2,{m_{\tilde g}}^2)\bigg]}
\,,\end{aligned}$$ where $C_F=4/3$, $L_x=\ln (\mu^2/m_x^2)$ and the functions $F_1$ and $F_2$ are defined through $$\begin{aligned}
F_1(x,y,z)&=&-\frac{x y
\ln\frac{y}{x}+yz\ln\frac{z}{y}+zx\ln\frac{x}{z}}{(x-y)(y-z)(z-x)}\,,
\nonumber\\
F_2(x,y,z)&=&-\frac{1}{(x-y)(y-z)(z-x)}
\Bigg[x y
\ln\frac{y}{x}(1+\ln\frac{\mu^2}{\sqrt{xy}})
\nonumber\\
&&
+yz\ln\frac{z}{y}(1+\ln\frac{\mu^2}{\sqrt{yz}})
+zx\ln\frac{x}{z}(1+\ln\frac{\mu^2}{\sqrt{xz}})\Bigg]\,.\end{aligned}$$ In the above formulas, $\alpha_s^\prime(\mu)$ denotes the strong coupling constant computed in the [$\overline{\mbox{{\scalefont{.9}}MS}}$]{} scheme and taking into account $n_l=5$ active quark flavours.\
The computation of the coefficient function $C_{2q}$ through two loops in SQCD is discussed in some detail in the next section.
\[sec::nnlo\]Calculation of the coefficient function $C_{2q}$ at next-to-next-to-leading order (NNLO)
======================================================================================================
For the derivation of the coefficient functions one has to compute Green functions in the full and effective theory and make use of the decoupling relations Eq. (\[eq::dec\]) to connect them [@Chetyrkin:1997un]. In general, one Green function contains several coefficient functions. For example, the amputated Green function involving the $q \bar{q}$ pair and the zero-momentum insertion of the operator ${\cal O}_h$ which mediates the couplings to the light Higgs boson $h$ contains both coefficient functions $C_{2q}$ and $C_{3q}$. Similarly, one possibility to compute the coefficient function $C_1$ involves the Green function formed by the coupling of the operators ${\cal O}_h$ to two gluons.
In the following, we restrict the discussion to the computation of the coefficient function $C_{2q}$. Considering the appropriate one-particle-irreducible (1PI) Green function, we get $$\begin{aligned}
\Gamma_{\bar{q}q{\cal O}_h}^{0}(p,-p)=i^2\int {\rm d}x {\rm d}y e^{i
p (x-y)}\langle T q^{0}(x) \bar{q}^0(y){\cal
O}_h(0)\rangle^{\rm 1PI}\,,\end{aligned}$$ where $p$ is the outgoing momentum of $q$. In a next step we express the operator ${\cal O}_h$ with the help of Eqs. (\[eq::eft\]) and (\[eq::ops\]) and make use of the decoupling relations Eq. (\[eq::dec\]). One can easily see that the above Green function will get contributions only from the operators ${\cal O}_{2q}$ and ${\cal O}_{3q}$ $$\begin{aligned}
\Gamma_{\bar{q}q{\cal O}_h}^{0,h}(p,-p)&=&-\zeta_2^{(0)}\int {\rm d}x {\rm d}y
e^{i p (x-y)}\langle T q^{\prime,0}(x)
\bar{q}^{\prime,0}(y)( C_{2q}{\cal O}_{2q} +C_{3q}{\cal
O}_{3q})\rangle^{\rm 1PI}\,,\nonumber\\
&=&\zeta_2^{(0)}\zeta_m^{(0)}(C_{2q}^{0}-C_{3q}^{0})m_b^{0}
+\zeta_2^{(0)}C_{3q}^{0}/\!\!\!p\,.
\label{eq::gf}\end{aligned}$$ In the last step we have used the Feynman rules for the scalar dimension four operators that can be found in Ref. [@Surguladze:1990sp] and the fact that $\Gamma_{\bar{q}q{\cal O}_h}^{0,h}(p,-p)$ denotes an amputated Green function. Exploiting the fact that the coefficient functions do not depend on the momentum transfer, one can set also $p=0$. In this case, on the l.h.s. of Eq. (\[eq::gf\]) only the hard parts of the Green function survive, as the massless tadpoles are set to zero in DRED and DREG. As before, the superscript $h$ stands for hard contributions.
The validity of the approximation $m_h^2=p_h^2\approx 0$ was extensively studied in the context of the SM and reconfirmed for the case of gluon fusion at two-loop order in SQCD in Ref. [@Degrassi:2008zj]. We expect that this approximation holds also in the case of fermionic Higgs decays, due to the heavy supersymmetric mass spectrum.
There are two possibilities currently used in the literature for the derivation of $\Gamma_{\bar{q}q{\cal O}_h}^{0,h}(0,0)$. The first one is the direct computation of the scalar and vector components of the vertex function making use of the appropriate projectors. The second one uses the LET, which relates the vertex corrections to the $hq\bar{q}$ coupling to the quark self-energy corrections via appropriate derivatives [@Ellis:1975ap; @Degrassi:2008zj].
Direct calculation of the coefficient function $C_{2q}^0$ at NNLO
------------------------------------------------------------------
Decomposing the Green function $\Gamma_{\bar{q}q{\cal O}_h}^{0,h}$ into its scalar and vector components $\Gamma_{\bar{q}q{\cal O}_h; s}^{0,h}$, $\Gamma_{\bar{q}q{\cal O}_h; v}^{0,h}$ one derives from Eq. (\[eq::gf\]) two linearly independent relations for the coefficients $C_{2q}^{0}$ and $C_{3q}^{0}$ $$\begin{aligned}
\Gamma_{\bar{q}q{\cal O}_h; s}^{0,h}(0,0)&=&\zeta_2^{(0)}
\zeta_m^{(0)}(C_{2q}^{0}-C_{3q}^{0})\,,
\nonumber\\
\Gamma_{\bar{q}q{\cal O}_h; v}^{0,h}(0,0)&=&\zeta_2^{(0)}C_{3q}^{0}\,.\end{aligned}$$ In SQCD at the two-loop order there is also an axial contribution to $\Gamma_{\bar{q}q{\cal O}_h}^{0,h}(0,0)$, which arises from diagrams where a top quark-squark pair is exchanged from a gluino propagator. However, it generates only contributions ${\cal O}(\alpha_s^4)$ to $\Gamma(h\to q\bar{q})$ which are beyond the precision we are interested in this paper.
Finally, using Eqs. (\[eq::letse\]) the expression for $C_{2q}^{0}$ through two loops reads $$\begin{aligned}
C_{2q}^{0}&=&\frac{\Gamma_{\bar{q}q{\cal
O}_h; s}^{0,h}(0,0)}{1-\Sigma_s^{0,h}(0)}
+\frac{\Gamma_{\bar{q}q{\cal O}_h;
v}^{0,h}(0,0)}{1+\Sigma_v^{0,h}(0)}\,.
\label{eq::c2mssm}\end{aligned}$$ One can either work in the $(\phi_1,\phi_2)$ basis, [*i.e.*]{} generating two sets of vertex corrections for each Higgs field, or one derives the Feynman rules for the couplings of quarks and squarks to the light Higgs using the relation Eq. (\[eq::hmix\]).
For the computation of the vertex corrections up to two loops we have implemented two independent setups. In one of them, the Feynman diagrams are generate with the help of the program [ FeynArts]{} [@Hahn:2000kx], then its output is handled in a self-written [Mathematica]{} code which includes the two-loop tensor reduction and the mapping of the vertex topologies to the two-loop tadpole ones [@Davydychev:1992mt]. The last step is possible due to the fact, that we neglect the mass of the light Higgs boson and of the external light quarks. In the second setup, the Feynman diagrams are generated with the program [QGRAF]{} [@Nogueira:1991ex], and further processed with [q2e]{} and [ exp]{} [@Harlander:1997zb; @Seidensticker:1999bb]. The reduction of various vacuum integrals to the master integral was performed by a self-written [FORM]{} [@Vermaseren:2000nd] routine.
\[sec::letsusy\] LET derivation of the coefficient function $C_{2q}^0$ at NNLO
-------------------------------------------------------------------------------
The connection between the coefficient functions $C_1^0, C_{2q}^0$ and the decoupling coefficients $\zeta_s^0, \zeta_m^0 $ or equivalently $\Pi^{0,h}(0), \Sigma_s^{0,h}(0)$ and $\Sigma_v^{0,h}(0)$ was extensively studied in the context of the SM. The validity of the LET was verified up to three-loop order in QCD [@Chetyrkin:1997un]. In the framework of the MSSM, however, the derivation of LET is much more involved due to the presence of the two Higgs doublets and of many massive particles and mixing angles. The applicability of LET in SQCD at two-loop order was verified only very recently. Namely, the relationship between the coefficient function $C_1$ and the hard part of the transverse gluon polarization function $\Pi^{0,h}(0)$ has been established in Ref. [@Degrassi:2008zj]. Furthermore, the leading two-loop contributions to the effective bottom Yukawa couplings have been derived from the scalar part of the bottom quark self-energy in Ref. [@Noth:2010jy]. It is one of the aims of this paper to verify the relationship between the coefficient function $C_{2q}$ and the hard part of the scalar and vector contributions to the quark self-energy.
For our calculation it is very convenient to work in $(\phi_1,\phi_2)$ basis, which means that we have to decompose the Green functions according to Eq. (\[eq::hmix\]) $$\begin{aligned}
\Gamma_{\bar{q}q{\cal O}_h; a}^{0,h}(0,0)=-\sin\alpha\,
\Gamma_{\bar{q}q{\cal O}_{\phi_1}; a}^{0,h}(0,0)
+ \cos\alpha\, \Gamma_{\bar{q}q{\cal O}_{\phi_2}; a}^{0,h}(0,0)\,,\quad a=s,v\,.\end{aligned}$$ Similarly, the coefficient function can be written as follows $$\begin{aligned}
C_{2q}^{0}&=& -\sin\alpha \, C_{2q,\phi_1}^{0}+\cos\alpha \,
C_{2q,\phi_2}^{0}\,.
\label{eq::c2_12}\end{aligned}$$ Applying the LET[^2] to the individual components $\Gamma_{\bar{q}q{\cal O}_{\phi_i}; a}$ we get $$\begin{aligned}
\Gamma_{\bar{q}q{\cal O}_{\phi_i};\, s}^{0,h}(0,0)&=&\frac{1}{m_b}\frac{\partial
}{\partial\phi_i}\bigg[m_b(1-\Sigma_s^{0,h}(0))\bigg]\bigg|_{\phi_i=v_i}\equiv\frac{1}{m_b} \hat{D}_{q,\phi_i} \bigg[m_b(1-\Sigma_s^{0,h}(0))\bigg]
\,,
\nonumber\\
\Gamma_{\bar{q}q{\cal O}_{\phi_i};\, v}^{0,h}(0,0)&=&\frac{\partial
}{\partial\phi_i}\bigg[-\Sigma_v^{0,h}(0)\bigg]\bigg|_{\phi_i=v_i}\equiv \hat{D}_{q,\phi_i} \bigg[-\Sigma_v^{0,h}(0)\bigg]
\,,
\label{eq::let}\end{aligned}$$ with $ i=1,2$ and $a=v,s$. As we are considering only the hard parts of the above Green functions no complication related to the occurrence of infrared divergences is encountered. In practice, it is convenient to express the operators $\hat{D}_{q,\phi_i}$ introduced in Eq. (\[eq::let\]) in terms of derivatives w.r.t. masses and mixing angles. This can be achieved using the field-dependent definition of the parameters, in our case quark and squark masses and squark mixing angles [@Brignole:1991pq].
The formulas derived up to now are valid for a generic light quark flavour $q$. However, for phenomenological applications the decay channel $h\to b\bar{b}$ is the most important one. The explicit expressions for the operators $\hat{D}_{b,\phi_i}$ can be easily derived from the Eqs. (11) and (12) in Ref. [@Degrassi:2008zj]. We quote them here for completeness and to fix our normalization: $$\begin{aligned}
\hat{D}_{b,\phi_1}&=&\frac{1}{\cos\beta}( m_b A_b {\cal F}_b +m_b {\cal G}_b)
-\frac{1}{\sin\beta} m_t
\mu_{\rm SUSY}\sin 2 \theta_t {\cal F}_t\,,\nonumber\\
\hat{D}_{b,\phi_2}&=& \frac{1}{\cos\beta} (-m_b\mu_{\rm SUSY} {\cal F}_b)
+ \frac{1}{\sin\beta} (m_t A_t \sin 2 \theta_t {\cal F}_t
+2 m_t^2 {\cal G}_t)\,, \quad \mbox{with}\nonumber\\
{\cal F}_b&=& \frac{2}{{m_{\tilde{b}_1}}^2-{m_{\tilde{b}_2}}^2} (1-{\sin^2 2 \theta_b})
\frac{\partial}{\partial {\sin 2 \theta_b}}\,,\quad {\cal G}_b = \frac{\partial}{\partial m_b}\,,\nonumber\\
{\cal F}_t&=& \frac{\partial}{\partial {m_{\tilde{t}_1}}^2}-
\frac{\partial}{\partial {m_{\tilde{t}_2}}^2}
+\frac{2}{{m_{\tilde{t}_1}}^2-{m_{\tilde{t}_2}}^2} \frac{ (1-{\sin^2 2 \theta_t})}{{\sin 2 \theta_t}}
\frac{\partial}{\partial {\sin 2 \theta_t}}\,,\nonumber\\
{\cal G}_t&=&\frac{\partial}{\partial {m_{\tilde{t}_1}}^2}
+\frac{\partial}{\partial {m_{\tilde{t}_2}}^2}
+\frac{\partial}{\partial m_t^2}\,.
\label{eq::derivop}\end{aligned}$$ In the above formulas we keep only the terms that do not vanish in the limit $m_b\to 0$.
As is well known, in Eqs. (\[eq::let\]) one has first to apply the derivative operators $\hat{D}_{q,\phi_i}$ and afterwards perform the renormalization. For simplicity of the notation we suppress the superscript $(0)$, labeling bare quantities. We checked explicitly at the diagram level that Eqs. (\[eq::let\]) hold through two loops. The computation of the two-loop diagrams contributing to $\Sigma_a^{0,h}(0)$ goes along the same line as that for $\Gamma_{\bar{q}q{\cal O}_{\phi_i};\, a}^{0,h}(0,0)$. The exact results together with few expansions for special mass hierarchies can be found in Ref. [@Bednyakov:2007vm; @Bauer:2008bj].
Let us mention at this point that for large values of $\tan\beta$ the dominant contribution to the coefficient function $C_{2q}$ is contained in the first term in Eq. (\[eq::c2mssm\]). The $\mu_{\rm SUSY}\tan\beta$-enhanced contributions are implicitly resummed in Eq. (\[eq::c2mssm\]), through the presence of the denominator $1-\Sigma_s^{0,h}(0)$, which contains contribution of the form $\alpha_s^n\mu_{\rm SUSY}\tan\beta$. In the framework of LET the $\mu_{\rm SUSY}\tan\beta$-enhanced contributions to $\Gamma_{\bar{q}q{\cal O}_h;s}$ are generated through the term proportional to the derivative ${\cal F}_b$ in $\hat D_{b,\phi_2}$. Taking into account the parametric dependence of $\Sigma_s^{0,h}$ on masses and mixing angles, one can easily derive these contributions from the terms proportional to $\sin 2\theta_b$ in $\Sigma_s^{0,h}$.[^3] Such contributions have also been derived in Ref. [@Noth:2010jy] using the effective Lagrangian approach. Indeed, after discarding the additional pieces comprised in our computation of the coefficient $C_{2q}$, namely the vector part and the rest of the derivative operators in Eq. (\[eq::derivop\]), we get good numerical agreement with the results of Ref. [@Noth:2010jy].
\[sec::ren\] Regularization and renormalization scheme
------------------------------------------------------
It is well known that the appropriate regularization scheme for the computation of radiative corrections in supersymmetric theories is DRED. However, the most convenient regularization scheme for the handling of the dimension four operators at higher orders is DREG as discussed in Section \[sec::framework\]. For the renormalization we employed two different approaches. In one of them we computed the radiative corrections to the coefficient functions directly in DREG. This implies that some of the supersymmetric relations between couplings of quarks and squarks do not hold anymore. More precisely, one has to distinguish between the gluino-quark-squark coupling $\hat{g}_s$ and the gauge coupling $g_s$, and between the Yukawa couplings of Higgs bosons to quarks $g_q^{\phi_i}$ and squarks $g_{\tilde{q}; kr}^{\phi_i}$. The relationships between the different couplings are necessary only at the one-loop order and they are well known since long time [@Martin:1993yx].\
In the second approach, we performed the two-loop computation of the coefficient functions in the DRED scheme and afterwards converted the results into the DREG scheme using the two-loop translation relations for the quark masses and strong couplings defined in the full [@Mihaila:2009bn] and effective theory [@Harlander:2006rj], and Eqs. (\[eq::let\]). As a consistency check, we explicitly verified that the results obtained with the two methods agree.
For the renormalization of the divergent parameters we used the on-shell scheme for the gluino and bottom squark masses and mixing angle and the minimal subtraction scheme [$\overline{\mbox{{\scalefont{.9}}MS}}$]{} or [$\overline{\mbox{{\scalefont{.9}}DR}}$]{} for the strong coupling and the bottom quark mass. The renormalization of the trilinear coupling $A_b$ was performed implicitly through the use of relation Eq. (\[eq::mixangle\]). The explicit formulas for the one-loop counterterms are well-known in the literature (see for example Ref. [@Pierce:1996zz]).
The complete two-loop results for the SQCD corrections to the coefficient function $C_{2q}$ discussed in this section are too lengthy to be given here. They are available in [MATHEMATICA]{} format upon request from the authors. For further applications, we also provide results for the case where all parameters are renormalized minimally.\
In principle, the results obtained in this section can be easily generalized to the heavy Higgs decays taking into account the necessary changes in the Yukawa couplings as given in Table \[tab::yukawa\_coeff\]. However, the application of effective theory formalism introduced in Section \[sec::nnlo\] is not justified in this case, [*i.e. *]{} the condition $M_H\ll M_t\,,M_{\rm SUSY}$ does not hold anymore.
Numerical results
=================
In this section we study the phenomenological implications of the two-loop corrections to the coefficient function $C_{2q}$ on the Higgs decay width $\Gamma(h\to b\bar{b})$. The SM input parameters are the strong coupling constant at the Z-boson mass scale $\alpha_s(M_Z)=0.1184$ [@Bethke:2009jm], the top quark pole mass $M_t=173.1$ GeV [@:2009ec] and the running bottom quark mass in the [$\overline{\mbox{{\scalefont{.9}}MS}}$]{} scheme $m_b(m_b)=4.163$ GeV [@Chetyrkin:2009fv]. For the supersymmetric mass spectrum we adopted the corresponding values of the “small $\alpha_{\rm eff}$” and “gluophobic” scenarios as defined in Ref. [@Carena:2002qg].[^4]\
For the running of the strong coupling constant within QCD we use the [ Mathematica]{} package [RunDec]{} [@Chetyrkin:2000yt]. For the evaluation of the strong coupling constant within the six-flavour SQCD and the [$\overline{\mbox{{\scalefont{.9}}DR}}$]{} scheme we follow Ref. [@Bauer:2008bj].[^5]\
An important ingredient for the computation of the decay width $\Gamma(h\to b \bar{b})$ is the effective mixing angle of the neutral Higgs sector $\alpha_{\rm
eff}$ [@Heinemeyer:2000fa] that takes into account the radiative corrections to the Higgs propagator. In practical applications one replaces the tree-level mixing angle $\alpha$ defined in Eq. (\[eq::h\_alpha\]) with $\alpha_{\rm eff}$. This can be computed in perturbation theory from the knowledge of the radiative corrections to the self-energy matrix of the neutral Higgs doublet $\hat{\Sigma}_{\phi_1},\,\hat{\Sigma}_{\phi_2},\,\hat{\Sigma}_{\phi_1\phi_2} $, $$\begin{aligned}
\tan\alpha_{\rm
eff}=\frac{-(M_A^2+M_Z^2)\sin\beta\cos\beta-
\hat{\Sigma}_{\phi_1\phi_2}}{M_Z^2\cos^2\beta+M_A^2
\sin^2\beta-M_h^2-\hat{\Sigma}_{\phi_1}}\,,
\label{eq::h_alpha_eff}\end{aligned}$$ where $M_h$ stands for the on-shell mass of the light Higgs boson. For the numerical analyses we implemented the exact two-loop results from Ref. [@Degrassi:2001yf].[^6]
[c]{}\
(a)\
\
(b)
In Fig. \[fig::muren\_alp\] we show separately the renormalization scale dependence of the coefficient functions ${\cal C}_{2q,\phi_1}$ and ${\cal C}_{2q,\phi_2}$, which describe the effective Yukawa couplings of the neutral Higgs fields $\phi_1$ and $\phi_2$. We choose the “small $\alpha_{\rm
eff}$” scenario with $\tan\beta=50$, $M_A=300$ GeV and evaluate $\alpha_{\rm eff}$ with the tree-level formula in order to avoid additional scale dependence. The dashed and solid lines correspond to the one- and two-loop results, respectively. As can be read from the figure, at the one-loop order the scale dependence amounts to about $50$% when the renormalization scale is varied around the average value $\mu_0=({m_{\tilde{b}_1}}+{m_{\tilde{b}_2}}+{m_{\tilde g}})/3\simeq 658$ GeV by a factor $10$. At the two-loop order the variation with the renormalization scale is significantly improved. The remaining scale dependence is below $6$%. An interesting aspect is the opposite evolution of the two coefficient functions with the renormalization scale. This feature is reflected in a milder scale dependence of the full coefficient function ${\cal C}_{2q}$ of about $4$% and $0.5$% at one- and two-loop order, respectively. However, for low A-boson masses $M_A\le 120$ GeV this numbers change to $27$% and $5$%, respectively. As usual, we can interpret the last number as an estimation of the theoretical uncertainties due to unknown higher order corrections. So, the two-loop SQCD corrections are essential for the accurate prediction of the decay width $\Gamma(h\to b \bar{b})$.
[c]{}\
(a)\
\
(b)
In Fig. \[fig::muren\_gl\] we display the renormalization scale dependence for the “gluophobic” scenario, where we fixed the value of $\tan\beta$ to $20$, $M_A=300$ GeV and maintain the same convention for the lines. A similar behaviour as for the “small $\alpha_{\rm eff}$” scenario is observed. However, in this case the coefficient function ${\cal C}_{2q,\phi_2}$ has a much stronger scale dependence at the one-loop order than the coefficient ${\cal C}_{2q,\phi_1}$. The scale variation of the full coefficient function ${\cal C}_{2q}$ sums up to $1.5$% and $0.2$% for $M_A=300$ GeV at the one- and two-loop order, respectively. However for $M_A\le 120$ GeV the scale variation amounts to $8$% and $1.5$% at one- and two-loop order, respectively, which shows the importance of the two-loop SQCD corrections for this region of the parameter space.\
We can conclude that for phenomenological analyses the choice of the renormalization scale around $\mu_0$ ensures small radiative corrections and a good convergence of the perturbative scale. In the following, we set $\mu\simeq \mu_0$ for the computation of the SQCD corrections to the coefficient functions ${\cal C}_{2q}$ and the decay width $\Gamma(h\to b\bar{b})$.
[c]{}\
(a)\
\
(b)
In Fig. \[fig::ma\_mh\] we depict the dependence of the full coefficient function $C_{2q}$ on the A boson mass $M_A$ (a) and the light Higgs boson mass $M_h$ (b) for the “small $\alpha_{\rm eff}$” scenario. We set $\tan\beta=50$ and vary the mass of the A boson between $100$ GeV$\le M_A\le 200$ GeV. For the evaluation of the Higgs boson mass and the effective mixing angle $\alpha_{\rm eff}$ we employed the two-loop SQCD results [@Degrassi:2001yf]. As can be seen from Fig. \[fig::ma\_mh\](a) for low $M_A$ values there are large one-loop corrections of about $60$% of the tree-level values. They originate from the large corrections to the scalar part of the quark self-energy, that are actually resummed through the use of the formula given in the Eq. (\[eq::c2mssm\]). This feature is also reflected by the relatively small two-loop corrections of about $2$% of the tree-level values. For $M_A$ values larger than $130$ GeV one observes a steep increase of the coefficient function $C_{2q}$ (decrease of the absolute value) due to cancellation of Eq. (\[eq::h\_alpha\_eff\]), that implies $\alpha_{\rm eff}\to 0$. In this case, $C_{2q}$ reaches its minimal absolute value. From panel (b) one notices a similar steep increase of the coefficient function $C_{2q}$ when $M_h$ reaches its maximal value of about $125$ GeV. A similar behaviour is also observed for the “gluophobic” scenario. So, for both scenarios we expect a strong suppression of the decay width $\Gamma(h\to b\bar{b})$ for large A-boson masses or equivalently for $M_h$ close to its maximal value for which $\alpha_{\rm eff} \to 0$.
[c]{}\
(a)\
\
(b)
In Fig. \[fig::tbeta\] the dependence of the coefficient function $C_{2q}$ on $\tan\beta$ is shown for the “small $\alpha_{\rm eff}$” (a) and “gluophobic” (b) scenarios, where the A-boson mass was fixed to $M_A=130$ GeV. As expected, we observe a significant increase of the magnitude of the radiative corrections with the increase of the $\tan\beta$ value. At the one-loop order the radiative corrections amount to about 70% (a) and to 33% (b) from the tree level values. At the two-loop order they sum up to 3% and 5%, respectively.
[c]{}\
(a)\
\
(b)
In Fig. \[fig::gamma\] we display the decay width for $h\to b\bar{b}$ as a function of the Higgs boson mass $M_h$, considering the “small $\alpha_{\rm eff}$”(a) and “gluophobic”(b) scenarios. We chose $\tan\beta=50$ and $\tan\beta=20$ for the case (a) and (b), respectively. The two-loop genuine QCD and EW corrections to the process $h\to b\bar{b}$, as well as the two-loop SQCD corrections to the Higgs boson propagator are depicted by the dotted lines. More precisely, they are derived from Eq. (\[eq::gamma\_gen\]), where the coefficient functions ${\cal C}_1 \, \mbox{and}\, {\cal C}_2$ are set to their tree-level values. The additional SQCD vertex corrections parametrized through the coefficient functions ${\cal C}_1 \, \mbox{and}\, {\cal C}_2$ are represented at the one- and two-loop order by the dashed and solid lines, respectively. We also take into account the one-loop SEW corrections to the coefficient function ${\cal C}_2$ and fix their renormalization scale at $\mu_{\rm SEW}=({m_{\tilde{t}_1}}+{m_{\tilde{t}_2}}+\mu_{\rm SUSY})/15$, for which the two-loop SEW corrections become negligible [@Noth:2010jy].\
For a relatively light Higgs boson mass $M_h$, the large one-loop radiative corrections of about 70% (a) and 50% (b) are still amplified by mild two-loop corrections that can reach as much as about $8$% from the decay width including QCD corrections even for the selected choice of the renormalization scale of SQCD corrections. The large SQCD radiative corrections to $\Gamma(h\to b\bar{b})$ have only a relatively small impact on the branching ratio $BR(h\to b\bar{b})$ but they can have a large impact on $BR(h\to \tau^+\tau-)$. For sufficiently large $\tan\beta$ and $\mu_{\rm SUSY}$, the measurement of $BR(h\to \tau^+\tau-)$ can provide information about the distinction between the SM and MSSM predictions.\
For a large Higgs boson mass for which $\alpha_{\rm eff}\to 0$, the partial decay widths for $h\to b\bar{b}$ and $h\to \tau^+\tau^-$ are significantly suppressed. In this case the radiative corrections (in particular the corrections to the Higgs boson propagator in Eq. (\[eq::h\_alpha\_eff\])) are essential for an accurate prediction of $\Gamma(h\to b\bar{b})$ and $\Gamma(h\to \tau^+\tau^-)$. Furthermore, the $BR(h\to \gamma\gamma)$ will be strongly enhanced, improving the LHC prospects of finding a light Higgs.
\[sec::concl\]Conclusions
=========================
The knowledge of the Higgs boson couplings is essential for its searches at the present hadron colliders. In this paper we calculate the two-loop corrections of ${\cal O}(\alpha_s^2)$ to the Yukawa couplings within the MSSM. We employed the effective Lagrangian approach under the assumption of large top quark and supersymmetric particle masses. We calculate analytically the two-loop corrections to the coefficient function $C_{2q}$ , taking into account the complete mass dependence. For large values of $\tan\beta$ the radiative corrections need to be resummed, which in our approach is performed through the use of the formula given in Eq. (\[eq::c2mssm\]). Furthermore, we verified at the diagram level the applicability of the low-energy theorem for Higgs interactions in the framework of the MSSM as stated in Ref. [@Degrassi:2008zj].
From the phenomenological point of view, the two-loop corrections presented here reduce significantly the theoretical uncertainties, estimated through the variation with the renormalization scale, at the percent level. The two-loop SQCD corrections become sizable for $\tan\beta\ge 20$ and $M_A\le 130$ GeV.
[**Acknowledgements**]{}\
We are grateful to Matthias Steinhauser and Konstantin Chetyrkin for enlightening conversations and many valuable comments. We thank Michael Spira for providing us with results necessary for the numerical comparison with Ref. [@Noth:2010jy] and Pietro Slavich for providing us with the analytic result for the two-loop SQCD corrections to the light Higgs boson mass. This work was supported by the DFG through SFB/TR 9 “Computational Particle Physics”.
[99]{}
H. P. Nilles, Phys. Rept. [**110**]{} (1984) 1.\
H. E. Haber and G. L. Kane, Phys. Rept. [**117**]{} (1985) 75. M. S. Carena, S. Heinemeyer, C. E. M. Wagner and G. Weiglein, Eur. Phys. J. C [**26**]{} (2003) 601
K. G. Chetyrkin, Phys. Lett. B [**390**]{} (1997) 309
K. G. Chetyrkin and M. Steinhauser, Phys. Lett. B [**408**]{} (1997) 320
K. G. Chetyrkin, B. A. Kniehl and M. Steinhauser, Phys. Rev. Lett. [**78**]{} (1997) 594; Nucl. Phys. B [**490**]{} (1997) 19
L. J. Hall, R. Rattazzi and U. Sarid, Phys. Rev. D [**50**]{} (1994) 7048;\
R. Hempfling, Phys. Rev. D [**49**]{}, 6168 (1994);\
M. S. Carena, M. Olechowski, S. Pokorski and C. E. M. Wagner, Nucl. Phys. B [**426**]{} (1994) 269
S. Heinemeyer, W. Hollik and G. Weiglein, Eur. Phys. J. C [**16**]{} (2000) 139
M. S. Carena, D. Garcia, U. Nierste and C. E. M. Wagner, Nucl. Phys. B [**577**]{} (2000) 88
J. Guasch, P. Hafliger and M. Spira, Phys. Rev. D [**68**]{} (2003) 115001
D. Noth and M. Spira, \[arXiv:1001.1935 \[hep-ph\]\].\
D. Noth and M. Spira, Phys. Rev. Lett. [**101**]{} (2008) 181801
V. P. Spiridonov, Report No. INR P-0378, Moscow, 1984.
K. G. Chetyrkin, B. A. Kniehl and M. Steinhauser, Nucl. Phys. B [**510**]{} (1998) 61
Y. Schroder and M. Steinhauser, JHEP [**0601**]{} (2006) 051 K. G. Chetyrkin, J. H. Kuhn and C. Sturm, Nucl. Phys. B [**744**]{} (2006) 121 R. Harlander, L. Mihaila and M. Steinhauser, Phys. Rev. D [**72**]{} (2005) 095009
A. V. Bednyakov, Int. J. Mod. Phys. A [**22**]{} (2007) 5245
A. Bauer, L. Mihaila and J. Salomon, JHEP [**0902**]{} (2009) 037
A. V. Bednyakov, Int. J. Mod. Phys. A [**25**]{} (2010) 2437
J. R. Ellis, M. K. Gaillard and D. V. Nanopoulos, Nucl. Phys. B [**106**]{} (1976) 292;\
M. A. Shifman, A. I. Vainshtein and V. I. Zakharov, Phys. Lett. B [**78**]{} (1978) 443;\
M. A. Shifman, A. I. Vainshtein, M. B. Voloshin and V. I. Zakharov, Sov. J. Nucl. Phys. [**30**]{} (1979) 711 \[Yad. Fiz. [**30**]{} (1979) 1368\];\
A. I. Vainshtein, V. I. Zakharov and M. A. Shifman, Sov. Phys. Usp. [**23**]{} (1980) 429 \[Usp. Fiz. Nauk [**131**]{} (1980) 537\];\
B. A. Kniehl and M. Spira, Z. Phys. C [**69**]{} (1995) 77;\
W. Kilian, Z. Phys. C [**69**]{} (1995) 89;\
M. Spira, A. Djouadi, D. Graudenz and P. M. Zerwas, Nucl. Phys. B [**453**]{} (1995) 17;
W. A. Bardeen, A. J. Buras, D. W. Duke and T. Muta, Phys. Rev. D [**18**]{} (1978) 3998
S. G. Gorishnii, A. L. Kataev, S. A. Larin and L. R. Surguladze, Mod. Phys. Lett. A [**5**]{} (1990) 2703; Phys. Rev. D [**43**]{} (1991) 1633.
A. Kwiatkowski and M. Steinhauser, Phys. Lett. B [**338**]{} (1994) 66 \[Erratum-ibid. B [**342**]{} (1995) 455\]
M. Spira, A. Djouadi, D. Graudenz and P. M. Zerwas, Phys. Lett. B [**318**]{} (1993) 347.
L. R. Surguladze and F. V. Tkachov, Nucl. Phys. B [**331**]{} (1990) 35
G. Degrassi and P. Slavich, Nucl. Phys. B [**805**]{} (2008) 267
T. Hahn, Comput. Phys. Commun. [**140**]{} (2001) 418\
T. Hahn and C. Schappacher, Comput. Phys. Commun. [**143**]{} (2002) 54
A. I. Davydychev and J. B. Tausk, Nucl. Phys. B [**397**]{}, 123 (1993)
P. Nogueira, J. Comput. Phys. [**105**]{} (1993) 279
R. Harlander, T. Seidensticker and M. Steinhauser, Phys. Lett. B [**426**]{} (1998) 125, T. Seidensticker, hep-ph/9905298
J. A. M. Vermaseren, arXiv:math-ph/0010025
A. Pilaftsis, Phys. Lett. B [**422**]{} (1998) 201
A. Brignole, J. R. Ellis, G. Ridolfi and F. Zwirner, Phys. Lett. B [**271**]{} (1991) 123. J. R. Ellis, G. Ridolfi and F. Zwirner, Phys. Lett. B [**262**]{} (1991) 477.
S. P. Martin and M. T. Vaughn, Phys. Lett. B [**318**]{} (1993) 331
L. Mihaila, Phys. Lett. B [**681**]{} (2009) 52
R. Harlander, P. Kant, L. Mihaila and M. Steinhauser, JHEP [**0609**]{} (2006) 053
D. M. Pierce, J. A. Bagger, K. T. Matchev and R. j. Zhang, Nucl. Phys. B [**491**]{} (1997) 3 S. Bethke, Eur. Phys. J. C [**64**]{} (2009) 689
\[Tevatron Electroweak Working Group and CDF Collaboration and D0 Collab\], arXiv:0903.2503 \[hep-ex\]. K. G. Chetyrkin, J. H. Kuhn, A. Maier, P. Maierhofer, P. Marquard, M. Steinhauser and C. Sturm, Phys. Rev. D [**80**]{} (2009) 074010
K. G. Chetyrkin, J. H. Kuhn and M. Steinhauser, Comput. Phys. Commun. [**133**]{} (2000) 43.
G. Degrassi, P. Slavich and F. Zwirner, Nucl. Phys. B [**611**]{} (2001) 403
P. Kant, R. V. Harlander, L. Mihaila and M. Steinhauser, arXiv:1005.5709 \[hep-ph\].
[^1]: For the definition of the parameter $x_t$ see Section \[sec::hdecay\].
[^2]: The gauge-fixing condition for SQCD is independent of the vacuum expectation values $v_{1,2}$, so that LET holds in its trivial form [@Pilaftsis:1997fe].
[^3]: A similar observation was made in Ref. [@Noth:2010jy], too.
[^4]: We used the tree-level formulas to derive the mass eigenvalues for squark fields.
[^5]: For a more detailed discussion see Ref. [@Bauer:2008bj] and the references therein.
[^6]: Very recently the three-loop SQCD corrections to $M_h$ have been computed [@Kant:2010tf]. However, they are valid only for specific mass hierarchies.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We formulate the interaction between the communicating nodes and an adversary within a game-theoretic context. We show that earlier information-theoretic capacity results for a jammed channel correspond to a pure Nash Equilibrium (NE). However, when both players are allowed to randomize their actions (i.e., coding rate and jamming power) new mixed Nash equilibria appear with surprising properties. We show the existence of a threshold ($J_{TH}$) such that if the jammer average power exceeds $J_{TH}$, the channel capacity at the NE is the same as if the jammer was using its maximum allowable power, $J_{Max}$, all the time. This indicates that randomization significantly advantages powerful jammers. We also show how the NE strategies can be derived, and we provide very simple (e.g., semi-uniform) approximations to the optimal communication and jamming strategies. Such strategies are very simple to implement in current hardware and software.'
author:
- |
Koorosh Firouzbakht\
\
\
\
\
\
\
Guevara Noubir\
\
\
\
\
\
Masoud Salehi\
\
\
\
\
\
subtitle: 'TO APPEAR IN Fifth ACM Conference on Security and Privacy in Wireless and Mobile Networks (WiSec’2012)'
title: 'On the Capacity of Rate-Adaptive Packetized Wireless Communication Links under Jamming '
---
Introduction {#Sec:Introduction}
============
Over the last decades, wireless communication proved to be an enabling technology to an increasingly large number of applications. The convenience of wireless and its support of mobility has revolutionized the way we access data, information services, and interact with the physical world. Beyond enabling mobile devices to access information and data services ubiquitously, wireless technology is widely used in cyber-physical systems such as air-traffic control, power plants synchronization, transportation systems, and human body implantable devices. This pervasiveness elevated wireless communication systems to the level of critical infrastructure. Radio-Frequency wireless communications occur over a broadcast medium, that is not only shared between the communicating nodes but is also exposed to adversaries. Jamming is one of the most prominent security threats as it not only can lead to denial of service attacks, but can also be the prelude to spoofing attacks.
Anti - jamming has been an active area of research for decades. Various techniques for combating jamming have been developed at the physical layer [@SimonOSL01] which include directional antennas, spread spectrum communication, power / modulation / coding control. At the time, most of the wireless communication were not packetized nor networked. Reliable communication in the presence of adversaries regained significant interest in the last few years, as new jamming attacks and the need for more complex applications and deployment environments have emerged. Several specifically crafted attacks and counter-attacks were proposed for packetized wireless data networks [@negi+p:jam; @LinN04; @LiKP07; @WilhelmMSL11], multiple access resolution [@BenderFHKL05; @GilbertRC06; @BayraktarogluKLNRT08; @AwerbuchRS08], multi-hop networks [@XuMTZ06; @TagueSNP08; @LiKP07], broadcast and control communication [@KBKV06; @ChiangH07; @ChanLNT07; @TagueLP07; @LazosLK09; @LiuPDL10; @LiuLK11], cross-layer resiliency [@LinN05], wireless sensor networks [@XuKWY06; @XuTZ07; @XuTZ08], spread-spectrum without shared secrets [@StrasserCSM08; @SlaterPRB09; @StrasserCS09; @JinNT09], and navigation information broadcast systems [@RasmussenCC07].
Nevertheless, very little work has been done on protecting rate adaptation algorithms against adversarial attacks. Rate adaptation plays an important role in widely used wireless communication systems such as IEEE802.11 standard as the link quality in a WLAN is often highly dynamic. In recent years, a number of algorithms for rate adaptation have been proposed in literature [@HollandVB01; @JuddXP08; @VutukuruHK09; @RahulFDC09; @RamachandranKZG08; @CampE08; @KimSSD06; @WongHSV06], and some are widely deployed [@Bicket05; @LacageMT04]. Recently, rate adaptation for the widely used IEEE 802.11 protocol was investigated in [@PelechrinisBKG09; @BroustisKDSL09; @NoubirRST11]. Experimental and theoretical analysis of optimal jamming strategies against currently deployed rate adaptation algorithms indicate that IEEE 802.11 can be significantly degraded with very few interfering pulses. The commoditization of software radios makes these attacks very practical and calls for investigation of the capacity of packetized communication under adaptive jamming.
In this work, we focus on the problem of determining the optimal rate control and adaptation mechanisms for a channel subject to a power constrained jammer. We consider a setup where a pair of nodes (transmitter and receiver) communicate using data packets. An adversary (jammer) can interfere with the communication but is constrained by an instantaneous maximum power per packet ($J_{Max}$) and a long-run average power ($J_{Ave}$). Appropriately coded packets can overcome interference and are lost otherwise. Over-coding (coding at low rates) reduces the throughput, while under-coding (coding at high rates) increases the chances of loosing a packet. An important question is to understand the interaction between the communicating nodes and the adversary, determine the long-term achievable maximum throughput and the optimal strategy to achieve it, as well as the optimal strategy for the adversary. While, the capacity of a channel under a fixed-power jammer, and the optimal strategies for communication and jamming, derive from fundamental information theoretic results (See Section \[Sec:Game\_Analysis\]), these questions are still open for a packetized communication system.
Our contribution can be summarized as follows:
- We formulate the interaction between the communicating nodes and an adversary within a game-theoretic context. We show the existence of the Nash Equilibrium for this non-typical game. We also show that the Nash Equilibrium strategies can be computed using Linear Programming.
- We show that earlier information-theoretic capacity results for a jammed channel correspond to a pure Nash Equilibrium (NE).
- We further characterize the game by showing that, when both players are allowed to randomize their actions (i.e., coding rate and jamming power) new mixed Nash equilibria appear with surprising properties. We show the existence of a threshold ($J_{TH}$) such that if the jammer average power exceeds $J_{TH}$, the channel capacity at the NE is the same as if the jammer was using $J_{Max}$ all the time.
- We also show that the optimal NE strategies can be approximated by very simple (e.g., semi-uniform) distributions. Such strategies are very simple to implement in current hardware and software.
The rest of the paper is structured as follows. In Section \[Sec:System\_Model\], we present our model for the communication link, communicating nodes and the adversary. In Section \[Sec:Game\_Model\], we introduce the players, the *transmitter* and the *jammer*, and their respective strategies and payoffs. We discuss how additional constraint on jammer’s mixed strategy space makes our game model different from a typical zero-sum game. In Section \[Sec:4\], we show that the Nash equilibrium indeed exists. We also prove the existence of a threshold, $J_{TH}$, for the jammer and its effect on the game outcome. In Section \[Sec:Game\_Analysis\], we study two particular cases. The case of a *powerful jammer*, when jammer’s average power is greater than the threshold, and the case of a *weak jammer*, when jammer’s average power is less than the threshold. We will also provide transmitter’s optimal strategies in these two cases. In Section \[Sec:Continuous\], we study the case where players have infinite number of pure strategies (the continuous zero-sum game) and finally, we conclude the paper in Section \[Sec:Conclusion\].
\[Table:Parameters\]
Parameter Description
------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------
$P_T$ Transmitter’s power
$N$ Noise power spectral density
$J_{Max}$ Jammer’s maximum power per packet
$J_{Ave}$ Jammer’s average power
$J_{TH}$ Jamming power threshold
$J$ Variable denoting jammer’s power
$J_T$ Jamming power corresponding to the transmitter’s rate
$\boldsymbol{J}^T = \begin{bmatrix} J_0 & \dots & J_j & \dots & J_{N_J} \end{bmatrix}_{1\times(N_J+1)}$ Jamming power vector
$J_j = \frac{j}{N_J}J_{Max}$
$\boldsymbol{R}^T = \begin{bmatrix} R_0 & \dots & R_i & \dots & R_{N_T} \end{bmatrix}_{1\times(N_T+1)}$ Vector corresponding to transmitter’s rates
$R_i = \frac{1}{2}\log \left(1 + \frac{P_T}{N + \frac{i}{N_T}J_{Max}}\right)$
$ \boldsymbol{x}^T = \begin{bmatrix} x_0 & \dots & x_i & \dots & x_{N_T} \end{bmatrix}_{1\times(N_T+1)} \in \mathbb{X}$ Transmitter’s mixed-strategy vector
$ \boldsymbol{y}^T = \begin{bmatrix} y_0 & \dots & y_j & \dots & y_{N_J} \end{bmatrix}_{1\times(N_J+1)} \in \mathbb{Y}$ Jammer’s mixed strategy vector
$\mathbb{X,Y}$ Mixed-strategy space, transmitter’s and jammer’s respectively
$ C_{(N_T+1)\times(N_J+1)}\ \text{or}\ C(\boldsymbol{x},\boldsymbol{y})\ \text{or}\ C(J_{Ave})$ Game matrix and expected game payoffs
System Model {#Sec:System_Model}
============
In this section we introduce and define our system model. The overall system model is shown in Figure \[Fig:Fig1\]. The communication link between the transmitter and the receiver is an AWGN channel with a fixed noise variance. Beside the channel noise, transmitted packets are being disrupted by an additive jammer. Jammer’s peak and average power are assumed to be limited to produce a more realistic model.
Channel Model {#SubSec:Channel_Model}
-------------
The overall system model is shown in Figure \[Fig:Fig1\]. The communication link between the transmitter and the receiver is assumed to be a single-hop, additive white Gaussian noise (AWGN) channel with a fixed and known noise variance, $N$, referred to the receiver’s front end. Furthermore, the communication link is being disrupted by an additive adversary, the *jammer*. The jammer transmits radio signals to degrade the capacity between the transmitter and the receiver. We assume transmissions are *packet-based*, i.e., transmissions take place in disjoint time intervals during which transmitter’s and jammer’s state (parameters) remain unchanged. We assume packets are long enough that channel capacity theorem could be applied to each packet being transmitted, this is justified by today’s Internet protocols that use packet sizes of up to $1,500$ bytes[^1].
In section \[Sec:Game\_Model\] we introduce and study a two-player zero-sum game in which transmitter-receiver goal is to achieve highest possible rate while jammer tries to minimize the achievable rate.
Jammer Model {#SubSec:Jammer_Model}
------------
Radio jamming or simply *jamming* is deliberate transmission of radio signals with the intention of degrading a communication link. The effect of jammer on the communication link is reduction of the effective signal to noise ratio (SNR) at the receiver and hence decreasing the channel capacity. As long as reduction in effective signal to noise ratio is concerned, the jammer can use arbitrary random signals for transmission but, it can be shown [@CoverJ06] that in the AWGN channel with a fixed and known noise variance, a Gaussian jammer with a flat power spectral density is the most effective in minimizing the the capacity between the transmitter and the receiver. In other words, in the communication game described above, the optimal strategy for the transmitter is to use a zero-mean white Gaussian input with variance equal to $P$, the transmitter power, and the best strategy for the jammer is is to use a similar distribution with variance $J$, the jammer power.
A fairly large number of jamming models have been proposed in the literature [@Peterson]. The most benign jammer is the *barrage noise jammer*. The barrage noise jammer transmits bandlimited white Gaussian noise with power spectral density (psd) of $J$. It is usually assumed that the barrage noise jammer power spectrum covers exactly the same frequency range as the communicating system. This kind of jammer simply increases the Gaussian noise level from $N$ to $(N+J)$ at the receiver’s front end. Another frequently used jamming model is the *pulse-noise jammer*. The pulse noise jammer transmits pulses of bandlimited white Gaussian noise having total average power of $J_{Ave}$ referred to the receiver’s front end. It is usually assumed that the jammer chooses the center frequency and bandwidth of the noise to be the same as the transmitter’s center frequency and bandwidth. The jammer chooses its pulse duty factor to cause maximum degradation to the communication link while maintaining the average jamming power $J_{Ave}$. For a more realistic model, the pulse-noise jammer could be subject to a maximum peak power constraint. Other jamming models, to name a few, are the *partial-band jammer* and *single/multiple-tune jammer*.
However, we study a more sophisticated jamming model. The jammer in study is a reactive and additive jammer, i.e., he is only active when a packet is being transmitted and silent otherwise. We assume that the jammer has a set of discrete jamming power levels uniformly distributed between $J=0$ and $J=J_{Max}$. The jammer can choose any jamming power level given that he maintains an overall average jamming power, $J_{Ave}$. The jammer uses his available power levels according to a distribution (his strategy), he chooses an optimal distribution to minimize the achievable capacity of the communication link while maintaining his maximum and average power constraints, i.e., $J_{Max}$ and $J_{Ave}$, respectively.
For reasons given in section \[SubSec:Transmitter\_Model\], burst jamming (transmitting a burst of white noise to disrupt a few bits in a packet) is not an optimal jamming scheme. Hence, we assume the jammer remains active during the entire packet transmission, i.e., the jammer transmits a continuous Gaussian noise with a fixed variance $J \in \big[ 0, J_{Max} \big]$ for each transmitting packet.
Transmitter Model {#SubSec:Transmitter_Model}
-----------------
Transmitter has a rate adaptation block which enables him to transmit at different rates. Popular techniques to increase or decrease the rate of a code are puncturing or extending. Puncturing and extending increase the flexibility of the system without significantly increasing its complexity. Considering jammer’s activity, the transmitter changes his rate according to a distribution (his strategy). Changing the rate can be accomplished using techniques like rate-compatible puncturing. The transmitter chooses an optimal distribution to achieve the best possible average rate (payoff). Same as before, we assume transmissions are *packet-based*, i.e., transmissions are taken place in disjoint time intervals during which, transmitter’s rate remain unchanged. Transmitter’s model is shown in Figure \[Fig:Transmitter\].
The interleaver block in transmitter’s model is a countermeasure to burst errors and burst jamming. Interleaving is frequently used in digital communications and storage devices to improve the burst error correcting capabilities of a code. Burst errors are specially troublesome in short length codes as they have very limited error correcting capabilities. In such codes, a few number of errors could result in a decoding failure or an incorrect decoding. A few incorrectly decoded codewords within a larger frame could make the entire frame corrupted.
Fortunately, combining effective interleaving schemes such as cryptographic interleaving and capacity-achieving codes such as turbo codes and LDPC codes results in transmission schemes that have good burst error correcting properties (see [@LinN04]) which make burst jamming ineffective. Therefore, in our study we do not consider burst jamming and instead assume that the jammer remains active during the entire packet transmission.
Game Model {#Sec:Game_Model}
==========
In this section we discuss the game setup in detail; we introduce and define the players, their respective strategies and the constraints in the game. We present the game model and define and formulate the payoff function in a game theoretic frame work. As discussed in section \[SubSec:Jammer\_Model\], in the AWGN channel, the additive white Gaussian jammer is the optimal jammer, in the sense that the white Gaussian jammer minimizes the channel capacity. Henceforward, we will only consider the additive Gaussian jammer.
We present the jammer’s strategy set and introduce the jammer’s average power constraint and its impact on the mixed strategy space. The additional constraint makes our game model different from a typical two-player zero-sum game. We also introduce transmitter’s strategy set and define the game utility function and the payoff matrix.
We begin by introducing a discrete version of the game to prove basic concepts and conclusions. Generalization to the continuous case is given in Section \[Sec:Continuous\]
Strategy Set {#SubSec:Jammer}
-------------
The jammer has the option to select discrete values of jamming power, uniformly distributed over $
\big[ 0, J_{Max} \big]$. We assume there are $(N_J + 1)$ pure strategies available to the jammer. Hence, the jammer’s strategy set (set of jamming powers), $\mathcal{J}$, is given by $$\label{Eq:Jammer_Mixed_Set}
\mathcal{J} = \Big\{J_j, 0 \leq j \leq N_J \Big\}$$ where $$\label{Eq:Power_Levels}
J_j = \frac{j}{N_J}J_{Max}$$ We can write the possible jammer power levels in vector form, hence the jammer’s pure strategies vector, $\boldsymbol{J}$, is $$\label{Eq:Jammer_Vector}
\boldsymbol{J}^T = \begin{bmatrix} J_0 & \dots & J_j & \dots & J_{N_J} \end{bmatrix}_{1\times(N_J+1)}$$ where $^T$ indicates transposition and $J_j$ is defined in . Unlike typical zero-sum games in which there are no other constraints on the mixed-strategies, in our model, the jammer’s mixed-strategy must satisfy the additional average power constraint, $J_{Ave} \leq
J_{Max}$. Hence, in this model, not all mixed-strategies (and not even the pure strategies that are greater than $J_{Ave}$) are feasible strategies [@Owen Sec. III.7]. If we let $\boldsymbol{y}$ be the jammer’s mixed-strategy vector and $\mathbb{Y}$ be the $(N_J + 1)$-simplex, we have the following relations: $$\begin{aligned}
\boldsymbol{y}^T& = \begin{bmatrix} y_0 & \dots & y_j & \dots & y_{N_J} \end{bmatrix}_{1\times(N_J+1)} \in \mathbb{Y}\\
\sum_{j=0}^{N_J} y_j &= 1; \quad y_j \geq 0, \quad 0\leq j \leq N_J \nonumber\end{aligned}$$ By using the jammer’s pure strategy vector we define the constrained mixed strategy space $\mathbb{Y}_{\text{E}}$ as $$\label{Eq:Jammer_Mixed_Set_1}
\mathbb{Y}_{\text{E}} = \{ \boldsymbol{y} \in \mathbb{Y} | \; \boldsymbol{y}^T \cdot
\boldsymbol{J} = J_{Ave} \}$$ which is a subset of the $(N_J+1)$-simplex that satisfies the average power constraint. By substituting the equality constraint in with the less than or equal sign, we define a new mixed strategy space which consists of all mixed strategies that result in an average power less than or equal to $J_{Ave}$. The new mixed-strategy space, $\mathbb{Y}_{\text{LE}}$, is $$\label{Eq:Jammer_Mixed_Set_2}
\mathbb{Y}_{\text{LE}} = \{ \boldsymbol{y} \in \mathbb{Y} | \; \boldsymbol{y}^T \cdot
\boldsymbol{J} \leq J_{Ave} \}$$ It is obvious that $$\mathbb{Y}_{\text{E}} \subset \mathbb{Y}_{\text{LE}} \subset \mathbb{Y}$$
A typical mixed strategy space with equality constraint, as defined in , is shown in Figure \[Fig:Jammer\_Mixed\_Space\] where $N_J = N_T =
3$. In this case jammer’s mixed and pure strategy vectors are $\begin{bmatrix} y_0 & y_1 & y_2 &
y_3 \end{bmatrix}_{1\times 4}$ and $\begin{bmatrix} 0 & \frac{1}{3}J_{Max} & \frac{2}{3}J_{Max} &
J_{Max} \end{bmatrix}_{1\times 4}$.
Since by introducing the new mixed strategy spaces of and we are eliminating some mixed strategies that could have been otherwise selected, the existence of the Nash equilibrium for this case must be first established. This is unlike a typical zero-sum game with a finite number of pure strategies in which the existence of the Nash Equilibrium is assured. In section \[SubSec:Nash\_Existence\], we provide an outline of the proof of the existence of the Nash Equilibrium in our game where the jammer’s mixed strategy space is limited to $\mathbb{Y}_{\text{E}}$ or $\mathbb{Y}_{\text{LE}}$.
Strategy Set {#SubSec:Transmitter}
-------------
The transmitter strategy set is a set of discrete transmission rates corresponding to different assumed jamming power levels, i.e, the transmitter chooses his rate, $R$, from the set $$\mathcal{R} = \Big\{ R_i, 0 \leq i \leq N_T \Big\}$$ where $$\label{Eq:Rate}
R_i = \frac{1}{2}\log \left( 1 + \frac{P_T}{N+\frac{i}{N_T}J_{Max}} \right)$$ and $\frac{i}{N_T}J_{Max}$ denotes the jammer’s power level assumed by the transmitter. If the actual jammer’s power level is less than or equal to the assumed value of $\frac{i}{N_T}J_{Max}$, then transmission at rate $R_i$ is possible, otherwise reliable transmission is not possible, the packet is lost, and the actual transmission rate drops to zero. Same as the case with the jammer, we define the vector of mixed-strategies for the transmitter, $\boldsymbol{x}$, as $$\boldsymbol{x}^T = \begin{bmatrix} x_0 & \dots & x_i & \dots & x_{N_T}
\end{bmatrix}_{1\times(N_T+1)} \in \mathbb{X}$$ where $\mathbb{X}$ is the $(N_T+1)$-simplex with no additional constraints.
The Payoff Function {#SubSec:Payoff}
-------------------
The *payoff* to the transmitter is defined assuming transmissions at the channel capacity. Defining the payoff based on channel capacity (or other variations of channel capacity) is a common practice in the games involving a transmitter-receiver pair and an adversary [@Giannakis08; @Koorosh11; @Altman07].
Because transmissions occur in the presence of an adversary, recovery of the transmitted information at the receiver is not always guaranteed. The information can only be recovered when the actual jamming power, $J$, is less than or equal to the jamming power level assumed by the transmitter, $J_T$, i.e., if and only if $J \leq J_T$. If $J_T < J$, the corresponding transmission rate would exceed the channel capacity and the information would be lost. Therefore, the transmitter’s payoff function is given by $$\label{Eq:Payoff}
C(J_T,J) =
\begin{cases}
R(J_T) = \frac{1}{2}\log \left( 1 + \frac{P_T}{N+J_T} \right) & J_T \geq J\\
0 & J_T<J\\
\end{cases}$$ Since the game in study is a zero-sum game, the payoff to the jammer is the negative of the transmitter’s payoff. We can formulate the payoffs in a payoff matrix where the transmitter and the jammer would be the row and column players respectively. The resulting payoff matrix, $C$, is $$\label{Eq:Game_Matrix}
C =
\begin{bmatrix}
R_0 & 0 & 0 & \dots & 0 \\
\vdots & \ddots & 0 & 0 & \vdots \\
R_i & \dots & R_i & 0 & \vdots \\
\vdots & & & \ddots & \vdots \\
R_{N_T} & R_{N_T} & \dots & R_{N_T} & R_{N_T}\\
\end{bmatrix}_{(N_T+1)\times(N_T+1)}$$ where $R_i$ is defined in . The expected payoff (or the game value) of the game is $$\label{Eq:Expected_Payoff}
C(\boldsymbol{x},\boldsymbol{y}) = \boldsymbol{x}^T \cdot C \cdot \boldsymbol{y}, \qquad
\boldsymbol{y} \in \mathbb{Y}_{\text{E}}\ \text{or}\ \mathbb{Y}_{\text{LE}}$$
In defining we have assumed $N_J=N_T$. As discussed below, without loss of generality, we can always assume that $N_T = N_J$.
\[Lem:N\_T\] Let $C$ be the payoff matrix in the two-player zero-sum game defined by the utility function . The payoff matrix resulted by removing the dominated strategies is a square lower triangular matrix with size less than or equal to $\min \big[ N_T, N_J \big]$. Furthermore, if the power levels were uniformly distributed over $\big[ 0, J_{Max} \big]$, the size of the non-dominated payoff matrix would be the minimum of $N_T$ and $N_J$.
Assume the jammer’s power levels are arbitrary distributed over some range, $\big[ 0, J_{Max}
\big]$, and $N_T < N_J$. A typical case where $N_T < N_J$ is depicted in Figure \[Fig:Dominate\_Strategies\] (top). In Figure \[Fig:Dominate\_Strategies\], the transmitter’s pure strategies are mapped to the jammer’s power levels for better visualization. Between some of the transmitter’s pure strategies there might be a pure strategy of the jammer but since $N_T < N_J$, according to the *Pigeonhole* principle, between at least two of the transmitter’s pure strategies (not necessarily any two pure strategy as sketched) there must be more than one jamming power level (shown as dashed or solid lines ending in squares). Any of these jamming power levels (or pure strategies) could be used to terminate the information transmitted by the rate corresponding to the power level immediately to the left of them (shown as solid line ending in circles). From these pure strategies, a rational jammer would choose the one with the lowest power level (the solid line) and hence, it would dominate the rest (dashed lines). therefore, the number of non-dominated pure strategies for the jammer is at most equal to the the number of the transmitter’s pure strategies (first part of the lemma).
If the pure strategies were uniformly distributed over $\big[ 0,$ $J_{Max} \big]$, as sketched, for every transmitter’s pure strategy there would be exactly one non-dominated strategy for the jammer and hence, there would be no intention for the jammer to use more pure strategies than the transmitter. The same discussion can be given for the number of pure strategies a rational transmitter should use for the case $N_T > N_J$ (see Figure \[Fig:Dominate\_Strategies\]( bottom)). Henceforward, without loss of generality, we assume $N_T = N_J$.
As a consequence of Lemma \[Lem:N\_T\], in our study, we need to consider only square matrices which simplifies further studies and assumptions. In the section that follows, we will study the outcome of the game when jammer’s average power assumes different values.
Game Characterization {#Sec:4}
=====================
In this section, we study the basic properties of the game. We will show that although we have put an additional constraint on the jammer’s mixed strategy space, the existence of the Nash equilibrium is still guaranteed.
Furthermore, we will show that by randomizing his strategy, the jammer can force the transmitter to operate at his lowest rate, given that he uses an average jamming power, $J_{Ave}$, that is more than a certain threshold, $J_{TH} < J_{Max}$. We also provide an upper bound for $J_{TH}$ in this section.
Existence of the Nash Equilibrium {#SubSec:Nash_Existence}
---------------------------------
We begin this section by the following lemma that shows existence of the Nash equilibrium under the additional average power constraint is guaranteed.
For the two-player zero-sum game defined by the utility function $C \left( J_T,J \right)$, given in and the payoff matrix $C$, given by and the transmitter’s mixed strategy, $ \boldsymbol{x} \in \mathbb{X} $, and the jammer’s mixed strategy, $\boldsymbol{y} \in \mathbb{Y}_{\text{E}}\ \text{or}\ \mathbb{Y}_{\text{LE}}$ (defined in and , respectively), at least one Nash equilibrium exists.
Nash in his 1951 seminal paper, “Non Cooperative Games” [@Nash51], proved that for any game with finite set of pure strategies, there exists at least one (pure or mixed) equilibrium such that no player can do better by unilaterally deviating from his strategy. In the proof of the existence of the Nash equilibrium, no additional constraints were assumed on the mixed strategy spaces. But, in our game model, we are assuming an additional constraint on the jammer’s mixed strategy space; the jammer must maintain a fixed or maximum average jamming power (corresponding to and , respectively). These additional assumptions change the jammer’s mixed strategy space from the n-simplex to a subset of it. Therefore, the Nash equilibrium theorem cannot be applied to our model directly and the existence of the Nash equilibrium must be established.
The proof of the existence of Nash equilibrium hinges on the *Sperner’s lemma* and *Brouwer’s fixed point theorem* and a corollary of this theorem on simplotopes [^2]. Sperner’s lemma applies to simplicially subdivided n-simplexes. It can easily be shown that by using a *radial projection*, the mixed strategy space in our model, which is a result of additional constraint of maintaining an average jamming power (or maintaining a maximum average power), can be projected to an appropriate lower dimension m-simplex where $m <
n$. A similar argument can be used to generalize the Brouwer’s fixed point theorem to any arbitrary convex and compact set. Since the additional average power constraint does not effect the convexity or compactness of the mixed strategy space, we can conclude that all the conditions and requirements assumed by the Sperner’s lemma and the Brouwer’s fixed point theorem are satisfied [@Basar99] and the existence of the Nash equilibrium for our problem is guaranteed.
Existence of Jamming Power Threshold {#SubSec:Power_Threshold}
------------------------------------
The following theorem proves the existence of a threshold jammer power that plays an important role in our further development.
\[The:Threshold\] For the two-player zero-sum game defined with the utility function $C \left( J_T,J \right)$, given in , and the payoff matrix $C$, given in , and the transmitter’s mixed strategy, $ \boldsymbol{x} \in \mathbb{X} $, and the jammer’s mixed strategy $\boldsymbol{y} \in \mathbb{Y}_{\text{LE}}$, given in and for all $P_T,N,J_{Max}~>~0 $ $$\exists J_{TH};\quad 0 < J_{TH} < J_{Max}$$ such that, if $J_{Ave} \geq J_{TH}$ then, $\exists \boldsymbol{y}^* \in \mathbb{Y}_{LE}$ for which we have $$\begin{aligned}
& \boldsymbol{x}^{*^T} = \begin{bmatrix} \boldsymbol{0}_{1\times N_T} & 1 \\ \end{bmatrix}_{1\times(N_T+1)}\\
& C(\boldsymbol{x}^*,\boldsymbol{y}^*) = R_{N_T}
\end{aligned}$$ where $\boldsymbol{x}^*,\boldsymbol{y}^*$ are transmitter’s and jammer’s optimal mixed-strategies, respectively and $ C(\boldsymbol{x}^*,\boldsymbol{y}^*) $ represents the value of the game.
Theorem \[The:Threshold\] states that there exists a jamming threshold ($J_{TH}$) such that if the jammer’s average power exceeds $J_{TH}$ then the transmitter’s optimal mixed-strategy is to use the lowest rate.
Assume the jammer is using a mixed strategy with the pmf given in Figure \[Fig:Uniform\] (*semi-uniform*)[^3] which is not necessarily an optimal mixed strategy. The parameters of this pmf are $$\begin{aligned}
y_0 & = 1 - \frac{2N_T}{N_T+1}\cdot \frac{J_{Ave}}{J_{Max}}\\
y & = \frac{2}{N_T+1}\cdot \frac{J_{Ave}}{J_{Max}}\\
\end{aligned}$$ It can be easily verified that the semi-uniform pmf satisfies the average power constraint $$\begin{split}
& \sum^{N_T}_{j=0} J\cdot \text{Pr} \big[ J \big]\\
& = \sum^{N_T}_{j=0} \left( \frac{j}{N_T} J_{Max} \right) \cdot \text{Pr} \bigg[J = \left(
\frac{j}{N_T} J_{Max} \right) \bigg]\\
& = J_{Ave}
\end{split}$$ We assume the transmitter is using an arbitrary mixed strategy in which rates $R_{N_T}$ (the lowest rate corresponding to $J_T = J_{Max}$) and $R_i$ (an arbitrary rate corresponding to $J_T = $ $\frac{i}{N_T} J_{Max},$ $ 0\leq i < N_T$) have probabilities $x_{N_T}$ and $x_i$ respectively. Define $C$ to be the expected payoff for the jammer’s semi-uniform mixed strategy against the transmitter’s arbitrary mixed-strategy: $$\label{Eq:C_Payoff}
\begin{split}
C & = C_{-i,N_T} + R_{N_T} x_{N_T} \times 1 + R_i x_i \times \text{Pr} \Big[ J \leq J_T = J_i \Big]\\
& = C_{-i,N_T} R_{N_T}x_{N_T} + R_i x_i \left( y_0 + iy \right)\\
\end{split}$$ where $C_{-i,N_T}$ is the partial expected payoff resulting from all pure strategies except for the $i$’th and $N_T$’th strategies. In order to improve his payoff, the transmitter, deviates from his current strategy to $x'_{N_T} = x_{N_T} + \delta$ and $x'_j = x_j - \delta$ where $\delta >0$. Defining $C'$ to be the expected payoff for the new strategy, we have $$\label{Eq:C_Prime_Payoff}
\begin{split}
C' &= C_{-i,N_T} + R_{N_T} \left( x_{N_T} + \delta \right)\\
& \phantom{=} + R_i \left( x_i - \delta \right) \times \text{Pr} \Big[ J \leq J_T = J_i \Big]\\
&= C + \delta \Big[ R_{N_T} - R_i \left( y_0 + iy \right) \Big]\\
\end{split}$$ Let $\Delta C $ be the difference in the expected payoff caused by deviating to the new strategy $$\label{Eq:Delta_C_InEq}
\begin{split}
\Delta C & = C' - C\\
& = \delta \left( R_{N_T} - 2 R_i \frac{N_T - i}{N_T + 1} \cdot \frac{J_{Ave}}{J_{Max}} \right)\\
\end{split}$$ where $\delta>0$ and $0\leq i < N_T$. We show that there exists a jammer power threshold, denoted by $J_{TH}$, such that if $J_{Ave} \geq J_{TH}$, then for all $\delta > 0$ and for all $i \in [0,
J_{Max})$, we have $$\label{Eq:Delta_C_Pos}
\Delta C >0$$ Assuming (for now) that $\Delta C > 0$ we can rewrite as $$\label{Eq:Z}
\begin{split}
J_{Ave} &\geq \frac{1}{2}J_{Max} \frac{N_T + 1}{N_T - i} \cdot \left( 1 - \frac{R_{N_T}}{R_i} \right)\\
& = Z_i, \qquad 0\leq i < N_T\\
\end{split}$$ where $Z_i$’s, for $ i = 0, \dots ,N_T - 1$, are a set of $N_T$ finite values. Let us define $J_{TH} = \max Z_i$, then for $$\label{Eq:J_TH_Upper_Bound}
J_{Ave} \geq J_{TH}$$ and for all $\delta > 0$ and $i\in [0, N_T)$ the inequalities in and are satisfied.
We showed that for $J_{Ave} \geq J_{TH}$, the transmitter can improve his expected payoff by dropping probability from any arbitrary rate (except for the lowest rate) and adding this probability to the lowest rate. We can continue this process until all other probabilities are added to the lowest rate probability and no further improvement to the expected payoff is possible. This shows that the low rate is indeed an optimal strategy for the transmitter against the jammer’s semi uniform mixed strategy.
By using the semi-uniform pmf and $J_{Ave} \geq J_{TH}$, the jammer can force the transmitter to operate at the lowest rate and given that the expected payoff is bounded between the transmitter’s lowest and highest rates, we can conclude that the semi-uniform distribution is indeed an optimal mixed strategy for the jammer when is the mixed strategy space[^4].
It is interesting to note that the packetized transmission model employed here and the transmitter’s lack of knowledge of the actual jammer power level benefits the jammer. In fact, the jammer uses a power level less than $J_{Max}$ but forces the transmitter to transmit at a rate corresponding to $J_{Max}$. This is similar to the situation in fading channels where although the ergodic capacity can be large, the outage capacity is considerably lower.
It is shown in Appendix \[App:I\] that $Z_i$ in is maximized for $i=0$. Therefore an upper bound for $J_{TH}$ is $$\label{Eq:J_TH_Upper_Bound2}
J_{TH,U} = \frac{1}{2} \frac{ N_T + 1 }{N_T} \left( 1 - \frac{R_{N_T}}{R_0} \right) J_{Max}$$
In section \[SubSec:Powerful\_Jammer\] we show that by using an optimal mixed strategy, the jammer can achieve a lower threshold than .
Game Analysis {#Sec:Game_Analysis}
=============
In this section we study the optimal mixed strategies for the jammer and the transmitter. We provide analytic and computer simulated results and a comparison between power thresholds resulted from computer simulation and the upper bound derived in section \[Sec:4\].
Based on relative values of $J_{Ave}$ and $J_{TH}$, we study two cases, the *powerful jammer* where $J_{Ave} \geq J_{TH}$ and the *weak Jammer* where$J_{Ave} < J_{TH}$.
Powerful Jammer {#SubSec:Powerful_Jammer}
---------------
As a result of the Theorem \[The:Threshold\], there exists a jamming threshold ($J_{TH}$), such that if the jammer’s average power exceeds $J_{TH}$, then the transmitter’s optimal mixed strategy (or more accurately, the optimal pure strategy in this case) is to use the lowest rate. We formulate this fact in the following theorem.
\[The:Powerful\_Jammer\] There exists a threshold $J_{TH}$ such that if $J_{Ave} \geq J_{TH}$, the expected payoff of the game is $$\begin{split}
C\Big( J_{Ave} \Big) & = R_{N_T}=\frac{1}{2} \log \left( 1 + \frac{P_T}{N+J_{Max}} \right)
\end{split}$$ The value of $J_{TH}$ is given by $$\label{Eq:Power_Threshold}
J_{TH} = \left( 1 - \frac{1}{N_T}\alpha^{-1} R_{N_T} \right) J_{Max}$$ where $R_i$ is defined in and $$\alpha^{-1} = \sum^{N_T - 1}_{i = 0} \left( R_i \right) ^{-1}$$
In other words, if the average jamming power exceeds $J_{TH}$ given in , by randomizing his strategy, the jammer forces the transmitter to operate at his lowest rate as if the jammer was using $J_{Max}$ all the time (Barrage noise jammer). If we define the effective jamming power, $J_{\text{Eff}}$, to be the jamming power a Barrage noise jammer needs to force the transmitter to operate at the same rate ($R_{N_T}$ in this case) then, for the powerful jammer the effective jamming power becomes $$J_{\text{Eff}} = J_{Max}$$
Typical optimal mixed strategies for the transmitter and the jammer in a powerful jammer case are given in Figure \[Fig:Powerful\_Jammer\]. Proof of Theorem \[The:Powerful\_Jammer\] is similar to the proof of Theorem \[The:Threshold\]. Details of deriving relation are given in Section \[SubSec:Weak\_Jammer\].
Unfortunately, jammer’s optimal mixed strategy cannot be formulated in a closed form relation and the optimal distribution has to be calculated numerically. As we showed in section \[SubSec:Power\_Threshold\], the simple semi-uniform pmf, shown in Figure \[Fig:Uniform\], could be used to derive an upper bound for the jamming threshold and as an approximation to the jammers optimal mixed strategy (see Figure \[Fig:Powerful\_Jammer\] (right)). The price paid by deviating from the optimal mixed strategy to the simple semi-uniform distribution is that the jammer has to use more average power to force the transmitter to operate at the lowest rate. A comparison between the jammer’s average power threshold given in and the upper derived in is given in Figure \[Fig:Comparison\].
Weak Jammer {#SubSec:Weak_Jammer}
-----------
A weak jammer has an average jamming power less than the threshold, $J_{Ave} < J_{TH}$. Typical optimal mixed strategies for the weak jammer case are given in Figure \[Fig:Weak\_Jammer\]. In this case the expected payoff, $C\Big( J_{Ave} \Big) \in \big( R_{N_T}, R_0 \ \big]$. Although a useful closed form relation between the expected payoff and the jammer’s average power where $J_{Ave} \in \big[ 0,$ $ J_{TH} \big)$ cannot be derived, for specific values of the average jamming power the relation reduces to a simple form. For these specific values, the expected payoff of the game, $C\Big( J_{Ave} \Big)$, corresponds to one of the transmitter’s rates $R_i$, $i =
0,\dots,N_T-1$. We present this fact in the following theorem without providing the full proof.
\[The:Weak\_Jammer\] Assuming $ J_{Ave} < J_{TH}$
1. The expected payoff of the game is $$\begin{split}
C\Big( J_{Ave} \Big) = & R_{m+1}\\
=& \frac{1}{2} \log \left ( 1 + \frac{P_T}{N + \frac{m + 1}{N_T}J_{Max}} \right)\\
\end{split}$$ where $m$ is the solution of $$\label{Eq:J_AVE_M}
J_{Ave} = \left( m + 1 - \alpha^{-1}R_{m+1} \right) \frac{J_{Max}}{N_T}$$
2. The transmitter’s optimal mixed strategy is $$\boldsymbol{x}^{*^T}_m = \begin{bmatrix} x_0 & x_1 & \dots & x_m & 0 & \dots & 0
\end{bmatrix}_{1\times(N_T+1)}$$ where $$\label{Eq:Trans_Optimal_Weak_Jammer}
x_i =\text{Pr} \left[ J_T = \left( \frac{i}{N_T} \right) J_{Max} \right]= \alpha_m R_i^{-1},\quad 0\le i\leq m$$ and $$\alpha^{-1}_m = \sum^{m}_{i = 0} \left( R_i \right) ^{-1}$$
The optimal mixed strategies for a typical zero-sum two-player game could be calculated by *linear programming*. Our game model differs from a typical zero-sum game however, linear programming could still be used to calculate the optimal mixed strategies by making the proper modifications [@Owen] and even though we do not provide the full proof for the transmitter’s optimal mixed-strategy, the consistency of can be verified by computer simulation. Numerical calculations verify that results achieved by using as the transmitter’s optimal mixed strategies are accurate to the order of $10^{-15}$.
In order to prove , we first introduce the following lemma without a proof.
\[Lem:Semi\_Uniform\_Approx\] The semi-uniform distribution and the jammer’s optimal mixed strategy (see Figure \[Fig:Weak\_Jammer\] (left)) result in the same expected payoff against the transmitter mixed strategy given in , if they have the same support and average jamming power.
The outline of the proof for will be given next.
Assume $J_{Ave}$ is such that the transmitter is using $(m+1)$ of his pure strategies, i.e., $$\boldsymbol{x}^{*^T}_m = \begin{bmatrix} x_0 & x_1 & \dots & x_m & 0 & \dots & 0
\end{bmatrix}_{1\times(N_T+1)}$$ where $\boldsymbol{x}^{*^T}_m$ is given in . Using Lemma \[Lem:N\_T\], the jammer only needs to use the strategies $J_j$ where $j = 0,\dots,$ $(m+1)$ and the expected payoff of the game would be at least $R_{m+1}$ (otherwise the jammer had to use more strategies). Lemma \[Lem:Semi\_Uniform\_Approx\] suggests that the following semi-uniform distribution which has the same support and average power as the jammer’s optimal mixed strategy could be used instead to compute the expected payoff of the game. $$\boldsymbol{y}^T_{\text{SU}} = \begin{bmatrix} y_0 & y_1 & \dots & y_{m+1} & 0 & \dots & 0
\end{bmatrix}_{1\times(N_T+1)}$$ $$\label{Eq:Semi_Uniform_M}
y_j =
\begin{cases}
1 - \frac{2N_T}{(m+2)}\cdot\frac{J_{Ave}}{J_{Max}} &\quad j = 0\\
\frac{2N_T}{(m+1)(m+2)}\cdot\frac{J_{Ave}}{J_{Max}} &\quad j = 1, \dots,m + 1
\end{cases}$$ If we let the expected payoff of the transmitter be exactly $R_{m+1}$, then $$\label{Eq:R_M_Plus_One}
\boldsymbol{x}^{*^T}_m C \boldsymbol{y}_{\text{SU}} = R_{m+1}$$ Substituting and in and solving for $J_{Ave}$ results in .
Finally, letting $R_m = R_{N_T}$ or equivalently letting $m = (N_T-1)$ in we obtain the desired relation in .
For a weak jammer, the effective jamming power, $J_{\text{Eff}}$ is $$J_{\text{Eff}} = \left( \frac{m+1}{N_T} \right) J_{Max}$$ If we define the effectiveness factor $E$ to be the ratio of the effective jamming power to the actual average jamming power, we have $$\begin{aligned}
E^{-1} & = \frac{J_{Ave}}{J_{\text{Eff}}}\\
& = \frac{\left( m + 1 - \alpha^{-1}_m R_{m+1} \right) \cdot \left( \frac{J_{Max}}{N_T}
\right)}{\left( \frac{m+1}{N_T} \right) J_{Max}}\\
& = 1 - \frac{1}{m+1} \alpha^{-1}_m R_{m+1} < 1\\
\end{aligned}$$ Similar to the case of the powerful jammer, the weak jammer can cause more damage to the communication link than a Barrage noise jammer with an average power $J_{Ave}$.
Continuous Case {#Sec:Continuous}
===============
In this section we study the case where the jammer and the transmitter have infinite pure strategies. In this case, instead of finite number of pure strategies, the transmitter and the jammer have a continuum of pure strategies that could be represented as points in intervals $ R \in
\Big[ R(J_{Max}), R(0) \Big]$ and $J \in \big[ 0, J_{Max} \big] $ respectively.
By letting $N_T \rightarrow \infty$ in , we can find the jamming power threshold for the continuous case to be $$\begin{aligned}
J_{TH,\text{Lim}} = & \lim_{N_T\rightarrow \infty} J_{TH}\\
= & J_{Max} - \frac{1}{2} \log \left( 1 + \frac{P_T}{N + J_{Max}} \right) \\
& \times \int^{J_{Max}}_0 \left[ \frac{1}{2} \log \left( 1 + \frac{P_T}{N + J} \right) \right]^{-1} \cdot dJ\\
\end{aligned}$$
Similar to the discrete case, we can use a continuous semi-uniform distribution to approximate the jammer’s optimal mixed strategy and find an upper bound for $J_{TH,\text{Lim}}$.
$$J_{TH,\text{Lim,UB}} = \frac{1}{2} \left[ 1 - \frac{R(J_{Max})}{R(0)} \right]J_{Max}$$
Conclusions {#Sec:Conclusion}
===========
We formulated the interaction between rate-adaptive communicating nodes and a smart power-limited jammer in a game-theoretic context. We show that packetization and adaptivity advantage the jammer. While, previous stationary information-theoretic capacity results correspond to a pure Nash-Equilibrium, packetized adaptive communication leads to lower game values. We show the existence of a mixed Nash Equilibrium and how to compute it. More importantly and surprisingly, we show the existence of a threshold on the average power of the jammer, above which the transmitter is forced to use a rate that corresponds to the maximum power of the jammer (and not the average power). We finally show how the optimal strategies can be computed and also derive a very simple (semi-uniform) jamming strategies that forces the transmitter to operate at the lowest rate (as if the jammer was continuously using its maximum power and not its average power).
[10]{}
E. Altman, K. Avrachenkov, and A. Garnaev. A jamming game in wireless networks with transmission cost. , 4465:1–12, 2007.
B. Awerbuch, A. Richa, and C. Scheideler. A jamming-resistant mac protocol for single-hop wireless networks. In [*ACM PODC*]{}, 2008.
T. Basar and G. J. Olsder. . Academic Press, 1999.
E. Bayraktaroglu, C. King, X. Liu, G. Noubir, R. Rajaraman, and B. Thapa. On the performance of ieee 802.11 under jamming. In [*Proceedings of IEEE INFOCOM*]{}, 2008.
M. A. Bender, M. Farach-Colton, S. He, B. C. Kuszmaul, and C. E. Leiserson. Adversarial contention resolution for simple channels. In [*SPAA*]{}, 2005.
J. Bicket. Bit-rate selection in wireless networks. , 2005.
I. Broustis, K. Pelechrinis, D. Syrivelis, S. V. Krishnamurthy, and L. Tassiulas. Fiji: Fighting implicit jamming in 802.11 wlans. , 2009.
J. Camp and E. Knightly. Modulation rate adaptation in urban and vehicular environments: Cross-layer implementation and experimental evaluation. , 2008.
A. Chan, X. Liu, G. Noubir, and B. Thapa. Control channel jamming: Resilience and identification of traitors. In [*IEEE ISIT*]{}, 2007.
J. Chiang and Y.-C. Hu. Cross-layer jamming detection and mitigation in wireless broadcast networks. In [*MobiCom*]{}, 2007.
T. M. Cover and J. A. Thomas. . Wiley, 2006.
K. Firouzbakht, G. Noubir, and M. Salehi. Superposition coding in an adversarial environment. , May 2011.
S. Gilbert, R. Guerraoui, and C. Newport. Of malicious motes and suspicious sensors: On the efficiency of malicious interference in wireless networks. In [*OPODIS*]{}, 2006.
G. Holland, N. Vaidya, and V. Bahl. A rate-adaptive mac protocol for multihop wireless networks. , 2001.
T. Jin, G. Noubir, and B. Thapa. Zero pre-shared secret key establishment in the presence of jammers. In [*Proceedings of the tenth ACM international symposium on Mobile ad hoc networking and computing, MobiHoc’09*]{}, pages 219–228, New York, NY, USA, 2009. ACM.
G. Judd, X. Wang, and P. Steenkiste. Efficient channel-aware rate adaptation in dynamic environments. , 2008.
J. Kim, S. Kim, S. Choi, and D. Qiao. Cara: Collision-aware rate adaptation for ieee 802.11 wlans. , 2006.
C. Koo, V. Bhandari, J. Katz, and N. Vaidya. Reliable broadcast in radio networks: The bounded collision case. In [*ACM PODC*]{}, 2006.
M. Lacage, M. H. Manshaei, and T. Turletti. Ieee 802.11 rate adaptation: A practical approach. , 2004.
L. Lazos, S. Liu, and M. Krunz. Mitigating control-channel jamming attacks in multi-channel ad hoc networks. In [*Proceedings of the second ACM conference on Wireless network security*]{}, WiSec ’09, pages 169–180, New York, NY, USA, 2009. ACM.
M. Li, I. Koutsopoulos, and R. Poovendran. Optimal jamming attacks and network defense policies in wireless sensor networks. In [*INFOCOM*]{}, 2007.
G. Lin and G. Noubir. On link layer denial of service in data wireless lans. , 5, 2004.
G. Lin and G. Noubir. On link layer denial of service in data wireless lans. , 5(3):273–284, 2005.
S. Liu, L. Lazos, and M. Krunz. Thwarting inside jamming attacks on wireless broadcast communications. In [*Proceedings of the fourth ACM conference on Wireless network security*]{}, WiSec ’11, pages 29–40, New York, NY, USA, 2011. ACM.
Y. Liu, P. Ning, H. Dai, and A. Liu. Randomized differential dsss: jamming-resistant wireless broadcast communication. In [*Proceedings of the 29th conference on Information communications*]{}, INFOCOM’10, pages 695–703, Piscataway, NJ, USA, 2010. IEEE Press.
J. Nash. Non-cooperative games. , pages 286–295, 1951.
R. Negi and A. Perrig. Jamming analysis of [MAC]{} protocols. Technical report, Carnegie Mellon University, 2003.
G. Noubir, R. Rajaraman, B. Sheng, and B. Thapa. On the robustness of ieee 802.11 rate adaptation algorithms against smart jamming. In [*Proceedings of the fourth ACM conference on Wireless network security*]{}, WiSec ’11, pages 97–108, New York, NY, USA, 2011. ACM.
G. Owen. . Academic Press, 1995.
K. Pelechrinis, I. Broustis, S. V. Krishnamurthy, and C. Gkantsidis. Ares: an anti-jamming reinforcement system for 802.11 networks. In [*Proceedings of the 5th international conference on Emerging networking experiments and technologies*]{}, CoNEXT ’09, pages 181–192, New York, NY, USA, 2009. ACM.
R. K. Peterson, R. E. Ziemer, and D. E. Borth. . Prentice-Hall, 1995.
H. Rahul, F. Edalat, D. Katabi, and C. Sodini. Frequency-aware rate adaptation and mac protocols. , 2009.
K. Ramachandran, R. Kokku, H. Zhang, and M. Gruteser. Symphony: Synchronous two-phase rate power control in 802.11 wlans. , 2008.
K. B. Rasmussen, S. Capkun, and M. Cagalj. Secnav: secure broadcast localization and time synchronization in wireless networks. In [*MobiCom*]{}, 2007.
M. K. Simon, J. K. Omura, R. A. Scholtz, and B. K. Levitt. . McGraw-Hill, 2001.
D. Slater, P. Tague, R. Poovendran, and B. Matt. A coding-theoretic approach for efficient message verification over insecure channels. In [*2nd ACM Conference on Wireless Network Security (WiSec)*]{}, 2009.
M. Strasser, C. Popper, and S. Capkun. Efficient uncoordinated fhss anti-jamming communication. In [*MobiHoc*]{}, 2009.
M. Strasser, C. Popper, S. Capkun, and M. Cagalj. Jamming-resistant key establishment using uncoordinated frequency hopping. In [*ISSP*]{}, 2008.
P. Tague, M. Li, and R. Poovendran. Probabilistic mitigation of control channel jamming via random key distribution. In [*Proceedings of International Symposium on Personal, Indoor and Mobile Radio Communications*]{}, 2007.
P. Tague, D. Slater, G. Noubir, and R. Poovendran. Linear programming models for jamming attacks on network traffic flows. In [*WiOpt*]{}, 2008.
M. Vutukuru, H. Balakrishnan, and K. Jamieson. Cross-layer wireless bit rate adaptation. , 2009.
T. Wang and G. B. Giannakis. Mutual information jammer-relay games. , 3(2):290–303, June 2008.
M. Wilhelm, I. Martinovic, J. B. Schmitt, and V. Lenders. Short paper: reactive jamming in wireless networks: how realistic is the threat? In [*Proceedings of the fourth ACM conference on Wireless network security*]{}, WiSec ’11, pages 47–52, New York, NY, USA, 2011. ACM.
S. H. Wong, H. Yang, S. Lu, and V. Bharghavan. Robust rate adaptation for 802.11 wireless networks. , 2006.
W. Xu, K. Ma, W. Trappe, and Y. Zhang. Jamming sensor networks: attack and defense strategies. , 2006.
W. Xu, K. Ma, W. Trappe, and Y. Zhang. Jamming sensor networks: attack and defense strategies. , 20(3):41–47, 2006.
W. Xu, W. Trappe, and Y. Zhang. Channel surfing: defending wireless sensor networks from interference. In [*Proceedings of the 6th international conference on Information processing in sensor networks*]{}, IPSN ’07, pages 499–508, New York, NY, USA, 2007. ACM.
W. Xu, W. Trappe, and Y. Zhang. Defending wireless sensor networks from radio interference through channel adaptation. , 4:18:1–18:34, September 2008.
Upper Bound for Jamming\
Power Threshold {#App:I}
========================
In Section \[SubSec:Power\_Threshold\] we showed that the $$\label{Eq:Zi}
J_{Ave} \geq Z_i = \frac{1}{2} J_{Max} \frac{N_T + 1}{N_T - i} \left(1 - \frac{R_{N_T}}{R_i} \right)$$ where $$R_i = \frac{1}{2} \log \left( 1 + \frac{P_T}{N + \left( \frac{i}{N_T} \right) J_{Max}} \right)$$ We stated that $Z_i$ given in is strictly decreasing function of $i$ hence, $\max_i Z_I = Z_0 $ and an upper bound for the average power threshold is $$\label{Eq:JTHU}
J_{TH,U} = \frac{1}{2} J_{Max} \frac{N_T + 1}{N_T} \left(1 - \frac{R_{N_T}}{R_0} \right)$$
To prove first, we rewrite $Z_i$ as $$\label{Eq:Zi2}
\begin{aligned}
Z_i &= \frac{1}{2} J_{Max} \frac{N_T + 1}{N_T} \frac{1}{1-\frac{i}{N_T}} \left(1 - \frac{R_{N_T}}{R_i} \right) ;\quad 0 \leq i < N_T\\
&= \left( \frac{1}{2} J_{Max} \frac{N_T + 1}{N_T} \right) \frac{J_{Max}}{J_{Max} - (\frac{i}{N_T} J_{Max})} \left(1 - \frac{R_{N_T}}{R_i} \right)\\
\end{aligned}$$ define $J$ and $R(J)$ as $$\label{Eq:J_and_RJ}
\begin{aligned}
& J = \left( \frac{i}{N_T} J_{Max} \right) & 0 & \leq i < N_T \\
& R(J) = \frac{1}{2} \log \left( 1 + \frac{P_T}{N + J} \right) & 0 & \leq J < J_{Max} \\
\end{aligned}$$ substituting in and we have $$\label{Eq:ZJ}
\begin{aligned}
Z(J) &= \left( \frac{1}{2} J_{Max} \frac{N_T + 1}{N_T} \right) \frac{J_{Max}}{J_{Max} - J} \left[1 - \frac{R(J_{Max})}{R(J)} \right]\\
&= a \times \frac{J_{Max}}{J_{Max} - J} \left[1 - \frac{R(J_{Max})}{R(J)} \right]\\
&= a\times F(J); \quad \text{where} \quad a > 0\ \text{and}\ 0\leq J < J_{Max}\\
\end{aligned}$$ If $F(J)$ in were a decreasing function of $J$ then, $Z_i$ and $Z(J)$ would also be decreasing functions of $i$ and $J$ respectively. Now let $$\label{Eq:FJ}
F(J) = f(J)g(J) \quad \text{where} \quad
\begin{aligned}
&f(J) = \frac{J_{Max}}{J_{Max} - J}\\
&g(J) = 1 - \frac{R(J_{Max})}{R(J)}\\
\end{aligned}$$ For decreasing $F(J)$ we have $$\label{Eq:Decreasing_FJ}
\begin{aligned}
\frac{\partial}{\partial J}F & = g \frac{\partial}{\partial J}f + f \frac{\partial}{\partial J}g < 0\\
& f,g>0 \quad \text{for} \quad 0\leq J < J_{Max} \Rightarrow \frac{\frac{\partial}{\partial J}f}{f} < - \frac{\frac{\partial}{\partial J}g}{g}\\
\end{aligned}$$ but from we have $$\label{Eq:dg_and_df}
\begin{aligned}
- \frac{\frac{\partial}{\partial J}g}{g} =&
\frac{1}{N+J} \times \left( \frac{x}{1 + x} \right) \left( \frac{\log (1 + x_m)}{\log (1 +x)} \right) \\
& \quad \times \left( \frac{1}{\log (1 + x) - \log (1 + x_m) } \right) \\
\frac{\frac{\partial}{\partial J}f}{f} =& \frac{P_T^{-1} x_m x}{x - x_m}\\
\end{aligned}$$ where $$\label{Eq:x}
x = \frac{P_T}{N + J} \quad \text{and} \quad x_m = \frac{P_T}{N + J_{Max}}$$ obviously $$0 < x_m < x$$ if we plug and in and simplify inequality we have $$\begin{aligned}\label{Eq:MathcalZ}
\mathcal{Z} =& \frac{x_m^{-1}x}{1+x} \frac{x-x_m}{\log(1+x) - \log(1+x_m)} \frac{\log(1+x_m)}{\log(1+x)} \\
& > 1 \qquad \text{for} \quad 0 < x_m < x \\
\end{aligned}$$ We need to show that holds for all $ 0 < x_m < x $ but first, we notice that $$\label{Eq:MathcalZ_Limit}
\lim _{x \rightarrow x_m^{+}} \mathcal{Z} \sim \frac{x_m^{-1} \log(1+x_m)}{x^{-1} \log(1+x)} \rightarrow 1^+ \quad \forall \ 0 < x_m < x$$ since we have $$\frac{\partial}{\partial z} z^{-1} \log(1+z) < 0 \quad \forall \ 0 < z$$ where we used the following natural logarithm property $$\label{Eq:Log_Inequality}
\frac{z}{1+z} < \log(1+z) \leq z \quad \text{for all} \quad z>0$$ For simplicity we rewrite inequality in as $$\label{Eq:MathcalZ2}
\begin{aligned}
\mathcal{Z}_2 =& \left[ x(x-x_m) \log(1+x_m) \right] \\
-& \left[ x_m (1+x) \log(1+x) \left( \log(1+x) - \log(1+x_m) \right) \right] \\
& > 0
\end{aligned}$$ As a result of we have $ \lim_{x \rightarrow x_m^{+}} \mathcal{Z}_2 \rightarrow 0^+ $ for all $0<x_m<x$. Since holds for $x\rightarrow x_m^+ $, if $\mathcal{Z}_2$ was strictly increasing function of $x$ for all $x>x_m$, and would also hold as a corollary.
To show that $\mathcal{Z}_2$ is strictly increasing, we first verify that $$\label{Eq:MathcalZ2_Derivative}
\frac{\partial \mathcal{Z}_2}{\partial x} \bigg( x = x_m \bigg) = 0$$ given that is true, an alternative way to proceed is to show that $\frac{\partial \mathcal{Z}_2}{\partial x}$ is itself strictly increasing function of $x$ (to show that the second partial derivate is strictly positive). Define $\mathcal{Z}_3$ $$\label{Eq:MathcalZ3}
\begin{aligned}
\mathcal{Z}_3 =& \frac{\partial^2 \mathcal{Z}_2}{\partial x^2} \times (1+x)\\
=& 2\log(1+x_m) - 2x_m + 2x \log(1+x_m)\\
& - 2x_m \log(1+x) + x_m \log(1+x_m)\\
\end{aligned}$$ It can be verified that for all $x>x_m$ and $x_m > 0 $ we have $ \lim_{x \rightarrow x_m^{+}} \mathcal{Z}_3 > 0 $. Taking the partial derivate of $\mathcal{Z}_3$ with respect to $x$ we have $$\frac{\partial \mathcal{Z}_3}{\partial x} = 2 \left[ \log(1+x_m) - \frac{x_m}{1+x} \right]$$ but from we have $$\begin{aligned}
&\log(1+x_m) > \frac{x_m}{1+x_m}>\frac{x_m}{1+x} \quad \text{for all} \quad x>x_m\\
& \Rightarrow \quad 2 \left[ \log(1+x_m) - \frac{x_m}{1+x} \right] >0 \quad \forall x>x_m>0\\
\end{aligned}$$ and hence we conclude that $$\frac{\partial \mathcal{Z}_3}{\partial x} > 0 \qquad \text{for all} \quad x>x_m>0$$ and $\mathcal{Z}_2$ is indeed an increasing function of $x$ for all $0<x_m<x$. Taking the reverse steps that resulted in and we conclude that $Z_i$ in is indeed a decreasing function and hence $J_{TH,U}$ given in is an upper bound for $J_{TH}$ in .
Optimal Mixed-Strategies {#App:II}
========================
Assume $0 \geq J_{Ave} < J_{TH} $ is such that jammer’s optimal mixed strategy is to use $(m+1)$ of his pure strategies. It is easy to show that in such a case, the transmitter’s optimal mixed strategy includes, at most, $(m+1)$ of his pure strategies. For now, we assume the transmitter is using $(m)$ of his pure strategies, i.e.; $$\label{Eq:Mixed_Strategy_Vectors}
\begin{aligned}
\boldsymbol{x}^T &= \begin{bmatrix}
x_0 & x_1 & \dots & x_m & 0 & \dots & 0
\end{bmatrix}_{1\times (1+N_T)}\\
\boldsymbol{y}^T &= \begin{bmatrix}
y_0 & y_1 & \dots & y_m & y_{m+1} & 0 & \dots & 0
\end{bmatrix}_{1\times (1+N_T)}\\
& \qquad \text{where} \quad 0 \leq m < N_T\\
\end{aligned}$$ The expected payoff of the game for the mixed-strategy pair $(\boldsymbol{x},\boldsymbol{y}) $ given in is $$\begin{aligned}
C(\boldsymbol{x},\boldsymbol{y}) &= \boldsymbol{x}^T C \boldsymbol{y}\\
&= \begin{bmatrix}
x_0 & x_1 & \dots & x_m & 0 & \dots & 0
\end{bmatrix}\\
& \times \begin{bmatrix}
R_0 & 0 & 0 & \dots & 0 \\
\vdots & \ddots & 0 & 0 & \vdots \\
R_i & \dots & R_i & 0 & \vdots \\
\vdots & & & \ddots & \vdots \\
R_{N_T} & R_{N_T} & \dots & R_{N_T} & R_{N_T}\\
\end{bmatrix}
\begin{bmatrix}
y_0 \\ y_1 \\ \dots \\ y_m \\ y_{m+1} \\ 0 \\ \dots \\ 0
\end{bmatrix}\\
\end{aligned}$$ $$\begin{aligned}
& = \Bigg[ \sum^{m}_{i=0} x_iR_i \quad \sum^{m}_{i=1} x_iR_i \quad \dots \\
& \quad \sum^{m}_{i=j} x_iR_i \quad \dots \quad \sum^{m}_{i=m} x_iR_i \quad 0 \quad \dots \Bigg]
\begin{bmatrix}
y_0 \\ y_1 \\ \dots \\ y_m \\ y_{m+1} \\ 0 \\ \dots \\ 0
\end{bmatrix}\\
\end{aligned}$$ $$\label{Eq:General_Expected_Payoff}
\begin{aligned}
=& y_0 \sum^{m}_{i=0} x_iR_i + y_1 \sum^{m}_{i=1} x_iR_i + \dots \\
& \qquad + y_j \sum^{m}_{i=j} x_iR_i + \dots + y_mx_mR_m\\
\end{aligned}$$ We can rewrite in terms of $R_i$’s; $$\begin{aligned}
C(\boldsymbol{x},\boldsymbol{y}) = & x_0 R_0 (y_0) + x_1 R_1 (y_0 + y_1) + \dots \\
& \quad + x_i R_i \sum^{i}_{j=0} y_j + \dots x_m R_m \sum^{m}_{j=0} y_j
\end{aligned}$$ Assume the transmitter is using the mixed-strategy $\widehat{\boldsymbol{x}}$, defined below, which not necessary an optimal mixed-strategy. $$\label{Eq:x_Hat_Vector}
\widehat{\boldsymbol{x}} = \alpha^{-1}_m \begin{bmatrix} R_0^{-1}&R_1^{-1}&\dots&R_m^{-1}&0&\dots&0 \end{bmatrix} _{1\times (N_T + 1)}$$ where $\alpha_m$ is defined as $$\label{Eq:Alpha_m}
\alpha_m^{-1} = \sum^{m}_{i=0} R_i^{-1} \qquad 0<m<N_T$$ The expected payoff of the game for the mixed-strategy pair $(\widehat{\boldsymbol{x}},\boldsymbol{y})$ is $$\label{Eq:Expected_Payoff_for_x_Hat}
\begin{aligned}
C \big(\widehat{\boldsymbol{x}},\boldsymbol{y} \big) =& \widehat{\boldsymbol{x}}^T C \boldsymbol{y}\\
=& \Bigg[ (y_0) + (y_0 + y_1) + \dots\\
& \quad + \left( \sum^{i}_{j=0} y_j \right) + \dots + \left( \sum^{i}_{j=m} y_j \right) \Bigg] \alpha_m\\
=& \Bigg[ \left( 1 - \sum^{m+1}_{j=1} y_j \right) + \left( \sum^{m+1}_{j=2} y_j \right) + \dots \\
& \quad + \left( 1- \sum^{m+1}_{j=i+1} y_j \right) + \dots + \left( 1 - y_m \right) \Bigg] \alpha_m\\
\end{aligned}$$ By expanding the sums in we can rewrite the expected payoff of the game for $(\widehat{\boldsymbol{x}},\boldsymbol{y})$ in a more compact form $$\label{Eq:Expected_Payoff_for_x_Hat_Compact}
\begin{aligned}
&C \big(\widehat{\boldsymbol{x}},\boldsymbol{y} \big) = \alpha_m \times\\
&
\begin{pmatrix}
1 - & ( & y_1 & + y_2 & + & \dots & + y_{i+1} & \dots & + y_{m+1} )+\\
1 - & ( & &\ \ y_2 & + & \dots & + y_{i+1} & \dots & + y_{m+1} )+\\
\vdots & \vdots & & & & \ddots & \vdots & & \vdots \\
1 - & ( & & & & & + y_{i+1} & \dots & + y_{m+1} )+\\
\vdots & \vdots & & & & & & \ddots& \vdots \\
1 - & ( & & & & & & \dots & + y_{m+1} )\ \ \\
\end{pmatrix} \\
&= \left[ \big( m+1 \big) - \left( \sum^{m+1}_{j=0} jy_j \right) \right] \alpha_m
\end{aligned}$$ but from and for all $\boldsymbol{y} \in \mathbb{Y}_{\text{E}} $ we have $$\label{Eq:Average_Power_Constraint}
\begin{aligned}
\boldsymbol{J}^T \boldsymbol{y} = J_{Ave} &= \sum^{m+1}_{j=0} \left( \frac{j}{N_T} J_{Max} \right) y_j \\
&= \left( \sum^{m+1}_{j=0} j y_j \right) \frac{J_{Max}}{N_T}\\
\Rightarrow \qquad & \sum^{m+1}_{j=0} j y_j = N_T \frac{J_{Ave}}{J_{Max}}
\end{aligned}$$ substituting in and the expected payoff of the game for the the mixed-strategy pair $(\widehat{\boldsymbol{x}}, \boldsymbol{y})$ becomes $$\label{Eq:Expected_Payoff_for_x_Hat_Final}
C \big(\widehat{\boldsymbol{x}}, \boldsymbol{y}\big) = \left( m+1 - N_T \frac{J_{Ave}}{J_{Max}} \right) \alpha_m$$ Hence, by using $\boldsymbol{\widehat{x}}$ (which is not necessary an optimal mixed strategy) against jammer’s arbitrary mixed-strategy with average power $J_{Ave}$, the transmitter can achieve the expected payoff given in . Therefore, the expected payoff of the game at equilibrium must at least be equal to , i.e., $$\label{Eq:Payoff_for_Trans_Optimal}
C \big(\boldsymbol{x}^*, \boldsymbol{y}^* \big) \geq C \big(\widehat{\boldsymbol{x}}, \boldsymbol{y}^* \big) = \left( m+1 - N_T \frac{J_{Ave}}{J_{Max}} \right) \alpha_m$$ In the same way, it can be shown that if the transmitter and the jammer were using the same number of pure strategies, $(m+1)$, the mixed-strategy given in results in the same expected payoff give that $m$ is replaced by $(m+1)$.
Now, assume $J_{Ave}$ is such that the jammer is using $ (m+1)$ of his pure strategies. Define the the following mixed-strategy for the jammer which is not necessary an optimal mixed-strategy. $$\begin{gathered}
\label{Eq:y_Hat}
\widehat{\boldsymbol{y}}^T = \begin{bmatrix}
y_0 & y_1 & \dots & y_m & y_{m+1} & 0 & \dots & 0
\end{bmatrix}_{1\times (1+N_T)}\\
y_j = \begin{cases}
R_0^{-1}R_{m+1}& j=0\\
\Big( R_j^{-1} - R_{j-1}^{-1} \Big)R_{m+1} & j=1,\dots, m+1\\
0 & m+1<j\leq N_T\\
\end{cases} \\\end{gathered}$$ it can be verified that $\widehat{\boldsymbol{y}}$ is indeed a mixed-strategy vector; $$\begin{aligned}
\sum^{m+1}_{j=0}y_j & = R_{m+1}R_0^{-1} + R_{m+1} \sum^{m+1}_{j=1} \Big( R_j^{-1} - R_{j-1}^{-1} \Big)\\
& = R_{m+1} \left( R_0^{-1} + \sum^{m+1}_{j=1} R_j^{-1} - \sum^{m+1}_{j=1} R_{j-1}^{-1} \right) \\
& = R_{m+1} \left( R_{0}^{-1} + R_{m+1}^{-1} - R_{0}^{-1} \right) = 1\\
\end{aligned}$$ Furthermore, we have $$\begin{aligned}
\sum^{m+1}_{j=0} jy_j &= \sum^{m+1}_{j=1} jy_j = R_{m+1} \left( \sum^{m+1}_{j=1} j \Big( R_j^{-1} - R_{j-1}^{-1} \Big) \right)\\
& = R_{m+1} \left( \sum^{m+1}_{j=1} jR_j^{-1} - \sum^{m+1}_{j=1} \big(j-1+1\big)R_{j-1}^{-1} \right) \\
&= R_{m+1} \Bigg( \sum^{m+1}_{j=1}jR_j^{-1} - \sum^{m+1}_{j=1} (j-1)R_{j-1}^{-1} \\
& \qquad \qquad - \sum^{m}_{j=0} R_{j}^{-1} \Bigg)\\
&= R_{m+1} \left( \big(m+1\big) R_{M=1}^{-1} - \sum^{m}_{j=0} R_j^{-1} \right)\\
&= \big(m+1\big) - \alpha_m^{-1} R_{m+1}\\
\end{aligned}$$ hence the average power of $\widehat{\boldsymbol{y}}$ becomes; $$\label{Eq:y_Hat_Average_Power}
\boldsymbol{J}^T \widehat{\boldsymbol{y}} = J_{Ave,m} = \Big( m+1 - \alpha_m^{-1} R_{m+1} \Big) \frac{J_{Max}}{N_T}$$ Assuming the jammer is using $\widehat{\boldsymbol{y}}$ against transmitter’s arbitrary mixed-strategy, the expected payoff of the game is $$\begin{aligned}
& C \big( \boldsymbol{x}, \widehat{\boldsymbol{y}} \big) = \boldsymbol{x}^T C \widehat{\boldsymbol{y}}
= R_{m+1} \boldsymbol{x}^T \times \\
&\begin{bmatrix}
R_0 & 0 & 0 & \dots & 0 \\
\vdots & \ddots & 0 & 0 & \vdots \\
R_i & \dots & R_i & 0 & \vdots \\
\vdots & & & \ddots & \vdots \\
R_{N_T} & R_{N_T} & \dots & R_{N_T} & R_{N_T}\\
\end{bmatrix}
\begin{bmatrix}
R_0^{-1}\\
R_1^{-1} - R_2^{-1}\\
\vdots\\
R_{j}^{-1} - R_{j-1}^{-1}\\
\vdots\\
R_{m+1}^{-1} - R_m^{-1}\\
0\\
\vdots\\
0
\end{bmatrix}\\
&\qquad = R_{m+1} \boldsymbol{x}^T \begin{bmatrix}
\boldsymbol{1}_{(m+2)\times 1}\\
0\\
\vdots\\
0
\end{bmatrix} = R_{m+1}
\end{aligned}$$ since $ \boldsymbol{x}$ is a mixed-strategy and has at most $(m+1)$ non-zero elements (see ).
Therefore, by using $\widehat{\boldsymbol{y}}$ given in against transmitter’s arbitrary mixed-strategy, the jammer guarantees not to lose more than $R_{m+1}$ given that his average jamming power is $J_{Ave,m}$ given in . Since $\widehat{\boldsymbol{y}}$ is not necessary an optimal mixed-strategy for the jammer, the optimal mixed-strategy would at most be less than $R_{m=1}$, i.e., $$\label{Eq:Game_Upper_Bound}
C \big( \boldsymbol{x}^*, \boldsymbol{y}^* \big) \leq C \big( \boldsymbol{x}^*, \widehat{\boldsymbol{y}} \big) = R_{m+1}$$ It can be shown (by induction) that for specific values of average jamming power given by and for $ m = 0,\dots,N_T-1 $, optimal mixed-strategy for the transmitter is to use $(m+1)$ of his pure strategies. From and by letting $m \rightarrow (m+1) $ we have $$\label{Eq:Game_Lower_Bound}
\begin{aligned}
C\big(\boldsymbol{x}^*,\boldsymbol{y}^*\big) \geq & \left(m+2 - N_T \frac{J_{Ave,m}}{J_{Max}} \right) \alpha_{m+1}\\
& \quad = \Big(1 + \alpha_m^{-1} R_{m+1} \Big) \alpha_{m+1}\\
\alpha_{m+1}^{-1} = \alpha_m^{-1} + R_{m+1}^{-1} \Rightarrow & \quad = \Big( 1 + \alpha_{m+1}^{-1} R_{m+1} -1 \Big) \alpha_{m+1}\\
& \quad = R_{m+1}\\
\end{aligned}$$ Therefore from and we have $$\begin{gathered}
R_{m+1} = C\big(\widehat{\boldsymbol{x}},\boldsymbol{y}^*\big) \leq C\big(\boldsymbol{x}^*,\boldsymbol{y}^* \big) \leq C\big(\boldsymbol{x}^*,\widehat{\boldsymbol{y}}\big) = R_{m+1}\\
\Rightarrow C\big(\boldsymbol{x}^*,\boldsymbol{y}^* \big) = R_{m+1} \quad \text{for} \quad J_{Ave} = J_{Ave,m}\end{gathered}$$ and hence $\widehat{\boldsymbol{x}}$ and $\widehat{\boldsymbol{y}}$ defined in and are indeed optimal mixed-strategies for the transmitter and the jammer respectively[^5].
[^1]: IEEE 802.3 and IEEE 802.11x protocols allow MAC frame sizes of up to 1,642 and 2304 bytes respectively.
[^2]: There are alternative proofs for the existence of the Nash equilibrium, i.e., using *Kakutani fixed point theorem* [@Owen].
[^3]: We will refer to this class of pmf/pdf as the *semi-uniform*
[^4]: The $J_{TH}$ given by is not necessarily the lowest possible threshold since we have limited jammer’s strategies to semi-uniform distributions. However, it is an upper bound for the lowest $J_{TH}$.
[^5]: Not necessary unique though.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We show that the fundamental group of the $3$-manifold obtained by $\frac{p}{q}$-surgery along the $(n-2)$-twisted $(3,3m+2)$-torus knot, with $n,m \ge 1$, is not left-orderable if $\frac{p}{q} \ge 2n + 6m-3$ and is left-orderable if $\frac{p}{q}$ is sufficiently close to $0$.'
address: 'Department of Mathematical Sciences, The University of Texas at Dallas, Richardson, TX 75080, USA'
author:
- 'Anh T. Tran'
title: 'Left-orderability for surgeries on twisted torus knots'
---
[^1]
[^2]
Introduction
============
The motivation of this paper is the L-space conjecture of Boyer, Gordon and Watson [@BGW] which states that an irreducible rational homology 3-sphere is an L-space if and only if its fundamental group is not left-orderable. Here a rational homology 3-sphere $Y$ is an $L$-space if its Heegaard Floer homology $\widehat{HY}(Y)$ has rank equal to the order of $H_1(Y; \BZ)$, and a nontrivial group $G$ is left-orderable if it admits a total ordering $<$ such that $g<h$ implies $fg < fh$ for all elements $f,g,h$ in $G$.
Many hyperbolic $L$-spaces can be obtained via Dehn surgery. A knot $K$ in $S^3$ is called an L-space knot if it admits a positive Dehn surgery yielding an L-space. For an L-space knot $K$, Ozsvath and Szabo [@OS] proved that the $\frac{p}{q}$-surgery of $K$ is an $L$-space if and only if $\frac{p}{q}\ge 2g(K)-1$, where $g(K)$ is the genus of $K$. In view of the L-space conjecture, one would expect that the fundamental group of the $\frac{p}{q}$-surgery of an L-space knot $K$ is not left-orderable if and only if $\frac{p}{q} \ge 2g(K)-1$.
By [@BM] among all Montesinos knots, the $(-2,3,2n+1)$-pretzel knots, with $n \ge 3$, and their mirror images are the only hyperbolic L-space knots. Nie [@Nie] has recently proved that the fundamental group of the $3$-manifold obtained by $\frac{p}{q}$-surgery along the $(-2,3,2n+1)$-pretzel knot, with $n \ge 3$, is not left-orderable if $\frac{p}{q} \ge 2n + 3$ and is left-orderable if $\frac{p}{q}$ is sufficiently close to $0$. This result extends previous ones by Jun [@Jun], Nakae [@Nakae], and Clay and Watson [@CW]. Note that the genus of the $(-2,3,2n+1)$-pretzel knot, with $n \ge 3$, is equal to $n+2$.
In this paper, we study the left-orderability for surgeries on the twisted torus knots. Some results about non left-orderable surgeries of twisted torus knots were obtained by Clay and Watson [@CW], Ichihara and Temma [@IT2015; @IT2018], and Christianson, Goluboff, Hamann, and Varadaraj [@C]. We will focus our study to the $(n-2)$-twisted $(3,3m+2)$-torus knots, which are the knots obtained from the $(3,3m+2)$-torus knot by adding $(n-2)$ full twists along an adjacent pair of strands. For $n,m \ge 1$, these knots are known to be L-space knots, see [@Va]. Moreover, the $(n-2)$-twisted $(3,5)$-torus knots are exactly the $(-2,3,2n+1)$-pretzel knots. Note that the genus of the $(n-2)$-twisted $(3,3m+2)$-torus knot, with $n,m \ge 1$, is equal to $n+3m-1$.
The following result generalizes the one in [@Nie].
\[main\] Suppose $n, m \ge 1$. Then the fundamental group of the $3$-manifold obtained by $\frac{p}{q}$-surgery along the $(n-2)$-twisted $(3,3m+2)$-torus knot is
\(i) not left-orderable if $\frac{p}{q} \ge 2n + 6m-3$,
\(ii) left-orderable if $\frac{p}{q}$ is sufficiently close to $0$.
The rest of this paper is devoted to the proof of Theorem \[main\]. In Section \[nonLO\] we prove part (i). To do so, we follow the method of Jun [@Jun], Nakae [@Nakae] and Nie [@Nie] which was developed for studying the non left-orderable surgeries of the $(-2,3,2n+1)$-pretzel knots. In Section \[LO\] we prove part (ii). To this end, we apply a criterion for the existence of left-orderable surgeries of knots which was first developed by Culler and Dunfield [@CD], and then improved by Herald and Zhang [@HZ].
Non left-orderable surgeries {#nonLO}
============================
Let $K_{n,m}$ denote the $(n-2)$-twisted $(3,3m+2)$-torus knot. By [@IT2018] (see also [@IT2015], [@CW]), the knot group of $K_{n,m}$ has a presentation $$\pi_1(S^3 \setminus K_{n,m}) = {\langle}a, w \mid w^n (aw)^m a^{-1} (aw)^{-m} = (wa)^{-m} a (wa)^m w^{n-1} {\rangle},$$ where $a$ is a meridian. Moreover, the preferred longitude corresponding to $\mu = a$ is $$\label{long}
\lambda = a^{-(4n+9m-2)} [(wa)^m w^{n}] (aw)^{m-1} a [w^{n} (aw)^m].$$ Note that the first homology class of $w$ is twice that of the meridian $a$.
\(i) It is known that $K_{n,1}$ is the $(-2,3,2n+1)$-pretzel knot. The above presentation of the knot group of $K_{n,1}$ was first derived in [@LT], [@Nakae].
\(ii) The formula for the longtitude of $K_{n,m}$ in [@IT2015], [@IT2018] contains a small error: $a^{-(4n+9m-2)}$ was written as $a^{-(2n+9m+2)}$.
Let $M_{\frac{p}{q}}$ be the 3-manifold obtained by $\frac{p}{q}$-surgery along the $(n-2)$-twisted $(3,3m+2)$-torus knot $K_{n,m}$. Then $$\pi_1(M_{\frac{p}{q}}) = {\langle}a, w \mid w^n (aw)^m a^{-1} (aw)^{-m} = (wa)^{-m} a (wa)^m w^{n-1}, a^p \lambda^q =1 {\rangle}.$$
Since $a^p \lambda^q =1$ in $\pi_1(M)$ and $a\lambda = \lambda a$, there exists an element $k \in \pi_1(M)$ such that $a = k^q$ and $\lambda = k^{-p}$, see e.g. [@Nakae Lemma 3.1].
Suppose $m, n \ge 1$. Assume $\pi_1(M_{\frac{p}{q}})$ is left-orderable for some $\frac{p}{q} \ge 2n + 6m-3$, where $q>0$. Then there exists a monomorphism $\rho: \pi_1(M_\frac{p}{q}) \to \text{Homeo}^+(\BR)$ such that there is no $x \in \BR$ satisfying $\rho(g)(x) = x$ for all $g \in \pi_1(M)$, see e.g. [@CR Problem 2.25].
From now on we write $gx$ for $\rho(g)(x)$.
We have $kx \not=x$ for any $x \in \BR$.
Assume $kx=x$ for some $x \in \BR$. Then $x=k^q x=ax$. If $x=wx$ then $gx=x$ for all $g \in \pi_1(M)$, a contradiction. Otherwise, without loss of generality, we assume that $x < wx$. Then we have $$\begin{aligned}
x &=& a^{(4n+9m-2)} k^{-p} x = a^{(4n+9m-2)} \lambda x = [(wa)^m w^{n}] (aw)^{m-1} a [w^{n} (aw)^m]x > x,\end{aligned}$$ which is also a contradiction.
Since $kx \not=x$ for any $x \in \BR$ and $kx$ is a continuous function of $x$, without loss of generality, we may assume $x < kx$ for any $x \in \BR$. Then $x < k^q x = ax$.
\[base\] We have $(aw)^m a x < w(aw)^m x$ for any $x \in \BR$.
Since $w^n (aw)^m a^{-1} (aw)^{-m} = (wa)^{-m} a (wa)^m w^{n-1}$ in $\pi_1(M_{\frac{p}{q}})$, we have $$\begin{aligned}
w(aw)^m x &=& [(aw)^m a (aw)^{-m} w^{-n} (wa)^{-m} a (wa)^m w^{n-1}] w(aw)^m x \\
&=& (aw)^m a [(wa)^m w^{n}(aw)^m]^{-1} a [(wa)^m w^{n}(aw)^m] x.\end{aligned}$$
Writing $g$ for $(wa)^m w^{n}(aw)^m$, we then obtain $$w(aw)^m x = (aw)^m a g^{-1} a g x > (aw)^m a x,$$ since $g^{-1} a g x > g^{-1} g x = x$.
Lemma \[base\] implies that $(aw)^m x < (aw)^m a x < w(aw)^m x$. Hence $x < wx$ for any $x \in \BR$.
\[induction\] For any $x \in \BR$ and $k \ge 1$ we have $$(aw)^m a^k x < w^k(aw)^m x \qquad \text{and} \qquad a^k (wa)^m x < (wa)^m w^k x.$$
We prove the lemma by induction on $k \ge 1$. The base case ($k=1$) is Lemma \[base\]. Assume $(aw)^m a^k x < w^k(aw)^m x$ for any $x \in \BR$. Then $$\begin{aligned}
(aw)^m a^{k+1} x &=& (aw)^m a^{k} (ax) \\
&<& w^k(aw)^m ax \\
&<& w^k (wa)^m w x = w^{k+1}(aw)^m x.\end{aligned}$$ Similarly, assuming $a^k (wa)^m x < (wa)^m w^k x$ for any $x \in \BR$ then $$\begin{aligned}
a^{k+1} (wa)^m x < a (wa)^m w^k x = (aw)^m a w^k x< w (aw)^m w^{k} x = (wa)^m w^{k+1} x.\end{aligned}$$ This completes the proof of Lemma \[induction\].
\[key\] With $\frac{p}{q} \ge 2n + 6m-3$ we have $wx< ax$ for any $x \in \BR$.
With $\frac{p}{q} \ge 2n + 6m-3$ and $q>0$, we have $-p + (2n + 6m-3)q \le 0$. Since $a = k^q$, $\lambda = k^{-p}$ and $x < kx$ for any $x \in \BR$, we have $$\begin{aligned}
ax &\ge& k^{-p+(2n + 6m-3)q} ax = a^{2n + 6m-2}\lambda x \\
&=& a^{-n} [(wa)^m w^{n}] (aw)^{m-1}a [w^{n} (aw)^m] a^{-(n+3m)} x.\end{aligned}$$
Then, by Lemma \[induction\], we obtain $$\begin{aligned}
ax
&>& a^{-n} [a^{n}
(wa)^{m}] (aw)^{m-1}a [(aw)^m a^n] a^{-(n+3m)} x
\\
&=&
w(aw)^{m-1} a (aw)^{m-1} a (aw)^m a^{-3m} x \\
&>& w a^{m-1} a a^{m-1}a a^{m} a^{-3m} x = wx.\end{aligned}$$ Here, in the last inequality, we use the fact that $x < wx$ for any $x \in \BR$.
With $\frac{p}{q} \ge 2n + 6m-3$, by Lemmas \[induction\] and \[key\] we have $$\begin{aligned}
(aw)^m x &=& [(aw)^m a] a^{-1} x \\
&<& [w(aw)^m ] a^{-1} x = (wa)^m w (a^{-1} x) \\
&<& (wa)^m a (a^{-1} x) = a^{-1} [(aw)^m a] x \\
&<& a^{-1} [w(aw)^m ] x = a^{-1} w[(aw)^m x]\\
&<& a^{-1} a[(aw)^m x] = (aw)^m x,\end{aligned}$$ a contradiction. This proves Theorem \[main\](i).
Left-orderable surgeries {#LO}
========================
To prove Theorem \[main\](ii) we apply the following result. It was first stated and proved by Culler and Dunfield [@CD] under an additional condition on $K$.
\[HZ\] [@HZ] For a knot $K$ in $S^3$, if its Alexander polynomial $\Delta_K(t)$ has a simple root on the unit circle, then the fundamental group of the manifold obtained by $\frac{p}{q}$-surgery along $K$ is left-orderable if $\frac{p}{q}$ is sufficiently close to $0$.
In view of Theorem \[HZ\], to prove Theorem \[main\](ii) it suffices to show that the Alexander polynomial of the twisted torus knot $K_{n,m}$ has a simple root on the unit circle. The rest of the paper is devoted to the proof of this fact. We start with a formula for the Alexander polynomial of a knot via Fox’s free calculus.
The Alexander polynomial
------------------------
Let $K$ be a knot in $S^3$ and $E_K = S^3 \setminus K$ its complement. We choose a deficiency one presentation for the knot group of $K$: $$\pi_1(E_K)=
\langle a_1,\ldots,a_l~|~r_1,\ldots,r_{l-1}\rangle.$$ Note that this does not need to be a Wirtinger presentation. Consider the abelianization homomorphism $\alpha:\pi_1(E_K)\to H_1(E_K;\BZ)
\cong {\BZ}
=\langle t
\rangle$.
The map $\alpha$ naturally induces a ring homomorphism $\tilde{\alpha}:{\BZ}[\pi_1(E_K)]\rightarrow {\BZ}[t^{\pm1}]$, where ${\BZ}[\pi_1(E_K)]$ is the group ring of $\pi_1(E_K)$. Consider the $(l-1)\times l$ matrix $A$ whose $(i,j)$-entry is $\tilde{\alpha}(\frac{\partial r_i}{\partial a_j})
\in {\BZ}[t^{\pm1}]$, where $\frac{\partial}{\partial a}$ denotes the Fox’s free differential. For $1\leq j\leq l$, denote by $A_j$ the $(l-1)\times(l-1)$ matrix obtained from $A$ by removing the $j$th column. Then it is known that the rational function $$\frac{\det A_j}{\det\tilde{\alpha}(a_j-1)}$$ is an invariant of $K$, see e.g. [@Wa]. It is well-defined up to a factor $\pm t^{k}~(k\in{\BZ})$ and is related to the Alexander polynomial $\Delta_K(t)$ of $K$ by the following formula $$\frac{\det A_j}{\det\tilde{\alpha}(a_j-1)} = \frac{\Delta_K(t)}{t-1}.$$
Proof of Theorem \[main\](2)
----------------------------
Let $r_1 = w^n (aw)^m a^{-1}(aw)^{-m}$ and $r_2 = (wa)^{-m} a (wa)^m w^{n-1}$. Then we can write $\pi_1(E_{K_{n,m}}) = {\langle}a, w \mid r_1 r_2^{-1} = 1 {\rangle}$. In $\pi_1(E_{K_{n,m}})$ we have $$\frac{\partial r_1 r_2^{-1}}{\partial a} = \frac{\partial r_1}{\partial a} + r_1 \frac{\partial r_2^{-1}}{\partial a} = \frac{\partial r_1}{\partial a} - r_1r_2^{-1} \frac{\partial r_2}{\partial a} = \frac{\partial r_1}{\partial a} -\frac{\partial r_2}{\partial a}.$$
Let $\delta_k(g) = 1 + g + \cdots + g^k$. Then $$\begin{aligned}
\frac{\partial r_1 r_2^{-1}}{\partial a}
&=& w^n \big[ \delta_{m-1}(aw) - (aw)^m a^{-1}(aw)^{-m} \big( \delta_{m-1}(aw) + (aw)^m \big) \big] \\
&& - \, \big[ -(wa)^{-m} \delta_{m-1}(wa) w + (wa)^{-m} \big( 1 + a \delta_{m-1}(wa) w \big) \big] \\
&=& - w^n (aw)^m a^{-1} \big[ 1 - (a-1) (aw)^{-m} \delta_{m-1}(aw) \big] \\
&& - \, (wa)^{-m} \big[ 1 + (a-1)w \delta_{m-1}(aw) \big].\end{aligned}$$
The Alexander polynomial $\Delta_{K_{n,m}}(t)$ of $K_{n,m}$ satisfies $$\frac{\Delta_{K_{n,m}}(t)}{t-1} = \frac{\tilde{\alpha}(\frac{\partial r_1 r_2^{-1}}{\partial a})}{\tilde{\alpha}(w)-1}.$$ Hence, since $\tilde{\alpha}(a)=t$ and $\tilde{\alpha}(w) = t^2$, we have $$\begin{aligned}
-(t+1)\Delta_{K_{n,m}}(t) &=& t^{2n+3m-1} \big[ 1 - (t-1) t^{-3m} \delta_{m-1}(t^3)\big] + t^{-3m} \big[ 1+(t-1)t^2 \delta_{m-1}(t^3)\big] \\
&=& t^{2n+3m-1} + t^{-3m} - (t^{2n-1} - t^{2-3m}) (t-1)\delta_{m-1}(t^3) \\
&=& t^{2n+3m-1} + t^{-3m} - (t^{2n-1} - t^{2-3m}) (t^{3m}-1)/(t^2+t+1) \\
&=& t^{-3m}(1+t+t^{3m+2}+t^{2n+3m-1}+t^{2n+6m}+t^{2n+6m+1})/(t^2+t+1).\end{aligned}$$
Let $f(t) = t^{n+3m+1/2} + t^{-(n+3m+1/2)}+t^{n+3m-1/2} + t^{-(n+3m-1/2)} + t^{n-3/2} + t^{-(n-3/2)}$. Then $$\Delta_{K_{n,m}}(t) = - \frac{t^{n-1}f(t)}{(t^{1/2}+t^{-1/2})(t+t^{-1}+1)}.$$ Hence $
\Delta_{K_{n,m}}(e^{i\theta}) = - \frac{e^{i(n-1)\theta}f(e^{i\theta})}{2\cos(\theta/2)(2\cos\theta+1)}.
$
Let $g(\theta) = f(e^{i\theta})/2$. To show that $\Delta_{K_{n,m}}(t)$ has a simple root on the unit circle, it suffices to show that $g(\theta)$ has a simple root on $(0, 2\pi/3)$. We have $$\begin{aligned}
g(\theta) &=& \cos (n+3m+1/2)\theta + \cos (n+3m-1/2)\theta + \cos (n-3/2) \theta \\
&=& 2 \cos (\theta/2) \cos (n+3m)\theta + \cos (n-3/2) \theta.\end{aligned}$$
If $n=1$ then $g(\theta) = \cos (\theta/2) (2\cos (n+3m)\theta + 1)$. It is clear that $\theta = \frac{2\pi/3}{n+3m}$ is a simple root of $g(\theta)$ on $(0,\pi/6]$.
Suppose $n \ge 2$. We claim that $g(\theta)$ has a simple root on $(\theta_0, \theta_1)$ where $\theta_0 = \frac{\pi/2}{n+3m}$ and $\theta_1 = \frac{\pi/2}{n+3m/2-3/4}$. Note that $0 < \theta_0 < \theta_1 \le \frac{\pi/2}{7/4}= \frac{2\pi}{7}$. We have $$g(\theta_0) = \cos (n-3/2) \theta_0 = \cos \left(\frac{\pi}{2} \cdot \frac{n-3/2}{n+3m} \right) > 0,$$ since $0 < \frac{\pi}{2} \cdot \frac{n-3/2}{n+3m} < \frac{\pi}{2}$. At $\theta = \theta_1 = \frac{\pi/2}{n+3m/2-3/4}$ we have $$\cos (n+3m)\theta + \cos (n-3/2) \theta = 2 \cos (n+3m/2-3/4)\theta \cos (3m/2+3/4)\theta=0.$$ Hence $$\begin{aligned}
g(\theta_1) &=& (1-2\cos(\theta_1/2)) \cos (n-3/2) \theta_1 \\
&=& (1-2\cos(\theta_1/2)) \cos \left(\frac{\pi}{2} \cdot \frac{n-3/2}{n+3m/2-3/4} \right) < 0,\end{aligned}$$ since $1-2\cos(\theta_1/2) < 0 < \cos \left(\frac{\pi}{2} \cdot \frac{n-3/2}{n+3m/2-3/4} \right)$.
We show that $g(\theta)$ is a strictly decreasing function on $(\theta_0, \theta_1)$. Indeed, we have $$\begin{aligned}
-g'(\theta) &=& \sin (\theta/2) \cos (n+3m)\theta + 2 (n+3m)\cos (\theta/2) \sin (n+3m)\theta \\
&& + \, (n-3/2) \sin (n-3/2) \theta\end{aligned}$$
Since $$0 < (n-3/2) \theta < \frac{\pi}{2} \cdot \frac{n-3/2}{n+3m/2-3/4} < \frac{\pi}{2},$$ we have $(n-3/2) \sin (n-3/2) \theta >0$. Since $\frac{\pi/2}{n+3m} < \theta <\frac{\pi/2}{n+3m/2-3/4}$ we have $${\pi/2} < (n+3m)\theta <\frac{\pi}{2} \cdot \frac{n+3m}{n+3m/2-3/4} < \pi,$$ which implies that $\cos (n+3m)\theta < 0 <\sin (n+3m)\theta$. Hence $$-g'(\theta) > \sin (\theta/2) \cos (n+3m)\theta + \cos (\theta/2) \sin (n+3m)\theta = \sin (n+3m+1/2)\theta.$$ Since $0< (n+3m+1/2)\theta < \frac{\pi}{2} \cdot \frac{n+3m+1/2}{n+3m/2-3/4} \le \pi$, we have $\sin (n+3m+1/2)\theta \ge 0$. Hence $-g'(\theta) > 0$ on $(\theta_0, \theta_1)$. This, together with $g(\theta_0) > 0 > g(\theta_1)$, implies that $g(\theta)$ has a simple root on $(\theta_0, \theta_1)$. The proof of Theorem \[main\](ii) is complete.
Acknowledgements {#acknowledgements .unnumbered}
================
The author has been partially supported by a grant from the Simons Foundation (\#354595 to AT). He would like to thank K. Ichihara for helpful discussions.
[99999]{}
S. Boyer, C. Gordon and L. Watson, [*On L-spaces and left-orderable fundamental groups*]{}, Math. Ann. **356** (2013), no. 4, 1213–1245.
K. Baker and A. Moore, [*Montesinos knots, Hopf plumbings, and L-space surgeries*]{}, J. Math. Soc. Japan **70** (2018), no. 1, 95–110.
N. Dunfield and M. Culler, [*Orderability and Dehn filling*]{}, Geom. Topol. **22** (2018), 1405–1457.
K. Christianson, J. Goluboff, L. Hamann, and S. Varadaraj, [*Non-left-orderable surgeries on twisted torus knots*]{}, Proc. Amer. Math. Soc. **144** (2016), no. 6, 2683–2696.
A. Clay and D. Rolfsen, Ordered groups and topology, Graduate Studies in Mathematics, 176. American Mathematical Society, Providence, RI, 2016. x+154 pp.
A. Clay and L. Watson, [*Left-orderable fundamental groups and Dehn surgery*]{}, Int. Math. Res. Not. IMRN 2013, no. 12, 2862–2890.
C. Herald and X. Zhang, [*A note on orderability and Dehn filling*]{}, arXiv:1807.00742.
K. Ichihara and Y. Temma, [*Non-left-orderable surgeries and generalized Baumslag-Solitar relators*]{}, J. Knot Theory Ramifications **24** (2015), no. 1, 1550003, 8 pp.
K. Ichihara and Y. Temma, [*Non-left-orderable surgeries on negatively twisted torus knots*]{}, Proc. Japan Acad. Ser. A Math. Sci. **94** (2018), no. 5, 49–52.
J. Jun, [*$(-2,3,7)$-pretzel knot and Reebless foliation*]{}, Topol. Appl. **145**(1-3) (2004) 209–232.
Y. Nakae, [*A good presentation of $(-2,3,2s+1)$-type pretzel knot group and $\BR$-covered foliation*]{}, J. Knot Theory Ramifications **22** (2013), no. 1, 1250143, 23 pp.
Z. Nie, [*Left-orderablity for surgeries on $(-2,3,2s+1)$-pretzel knots*]{}, arXiv:1803.00076.
T. Le and A. Tran, [*On the AJ conjecture for knots*]{}, with an appendix written jointly with V. Huynh, Indiana Univ. Math. J. **64** (2015), no. 4, 1103–1151.
P. Ozsvath and Z. Szabo, [*Knot Floer homology and rational surgeries*]{}. Algebr. Geom. Topol. **11** (2011), no. 1, 1–68.
F. Vafaee, [*On the knot Floer homology of twisted torus knots*]{}, Int. Math. Res. Not. IMRN 2015, no. 15, 6516–6537.
M. Wada, *Twisted Alexander polynomial for finitely presentable groups*, Topology [**33**]{} (1994), 241–256.
[^1]: 2000 [*Mathematics Subject Classification*]{}. Primary 57M27, Secondary 57M05, 57M25.
[^2]: [*Key words and phrases.*]{} Dehn surgery, left-orderable, L-space, twisted torus knot.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- |
Teppo Niinimäki$^{1}$, Mikko Heikkilä$^{2}$, Antti Honkela$^{2,3,4,}$[^1] , and Samuel Kaski$^{1,*}$\
$^{1}$Helsinki Institute for Information Technology HIIT, Department of Computer Science, Aalto University, Finland,\
$^{2}$Helsinki Institute for Information Technology HIIT, Department of Mathematics and Statistics, University of Helsinki, Finland,\
$^{3}$Department of Public Health, University of Helsinki, Finland and\
$^{4}$Helsinki Institute for Information Technology HIIT, Department of Computer Science, University of Helsinki, Finland.
bibliography:
- 'paper.bib'
title: Representation Transfer for Differentially Private Drug Sensitivity Prediction
---
Introduction
============
Privacy-preserving machine learning has the potential to enable the research use of many sensitive datasets, that would otherwise be out of reach for the community. This is especially the case for medical data, which almost always contain sensitive information traceable back to the data subjects. As an example, it has been shown that individuals can be identified from genomic data [@gymrek2013] including mixtures from several individuals [@homer2008]. It is likely that functional genomics data such as gene expression data are also identifiable. While different anonymisation strategies [@sweeney2002; @machanavajjhala2007; @li2007] can protect the privacy of the data subjects to some degree, they do not have formal guarantees and can fail to provide sufficient protection in practice [@ganta2008].
*Differential privacy* (DP) [@dwork2006; @dwork2014] is a framework that guarantees strict bounds for the amount of leaked private information, even in the presence of arbitrary side information. The guarantees are obtained by adding specific forms of randomisation to the computation process. In a machine learning context this usually means adding noise either directly to the input of the algorithm (*input perturbation*), to the output (*output perturbation*), or modifying the algorithm itself, for instance, by perturbing the optimisation objective (*objective perturbation*).
The privacy guarantee is controlled by a “privacy budget” parameter, usually denoted by $\epsilon > 0$; smaller $\epsilon$ means stricter guarantees, and can be achieved by increasing the amount of noise. Formally, a randomised mechanism ${\mathcal{M}}$ is said to be *$\epsilon$-differentially private*, if for all pairs of *neighbouring* datasets $X, X'$ differing[^2] on a single sample and all measurable subsets $S$ of possible outputs, $$\Pr({\mathcal{M}}(X) \in S) \le e^\epsilon \Pr({\mathcal{M}}(X') \in S).$$ Intuitively, this means that changing one sample in the dataset can change the output distribution only by a factor $e^\epsilon$.
As an extension, ${\mathcal{M}}$ is said to be *($\epsilon$, $\delta$)-differentially private*, if $$\Pr({\mathcal{M}}(X) \in S) \le e^\epsilon \Pr({\mathcal{M}}(X') \in S) + \delta,$$ for all measurable $S$ and all neighbouring datasets $X, X'$. The condition with nonzero $\delta > 0$ is often easier to achieve than pure $\epsilon$-DP.
In this article we are interested in DP learning for drug sensitivity prediction using gene expression data. First proposed by @Staunton2001, the drug sensitivity prediction problem has attracted significant attention recently, including from a DREAM challenge in 2012 [@Costello2014] that provided standardised evaluation metrics. The scale of the cytotoxicity assays needed has kept the sizes of the available data sets relatively small from a machine learning perspective. @honkela2018 were the first to apply DP learning to this problem. They needed to specifically limit the sensitivity of the learning and the dimensionality of the input data to make the learning feasible.
In abstract terms, our goal in this problem is DP learning of predictive models with high-dimensional input data, where both input and output variables need DP protection. This is a case where DP methods tend to run into trouble with moderate dataset sizes: the amount of noise that needs to be added usually increases quickly with the dimensionality, leading to output that is dominated by the noise. This warrants the use of dimensionality reducing methods with the aim of finding a good low-dimensional representation of the original data. However, unless one uses a random projection or some other “dummy” method that does not depend on the data, finding a good representation can also leak private information. For this reason, the dimension reduction method itself would also need to be made differentially private, which can completely invalidate the noise magnitude savings obtained in any downstream task like prediction.
Different solutions have been proposed for various special cases: @kifer2012 solve sparse linear regression problems by using an $\epsilon$-DP feature selection algorithm. @honkela2018 utilise external knowledge to select a relevant subset of features. @kasiviswanathan2016 show theoretical results on using random projections to improve DP learning on high-dimensional problems. Differentially private versions of methods such as PCA [@chaudhuri2012; @dwork2014pca] or deep learning [@abadi2016; @acs2018] exist and could be used to learn a representation, but the noise cost can be impractically large for small but high-dimensional datasets.
We study a straightforward solution based on feature representation transfer, similar to self-taught learning of @Raina2007. By using an additional non-sensitive dataset to learn the representation, we can apply more advanced representation learning methods. This approach has many advantages: we do not need labels for the additional data set, although in our case we make use of labels for a different task; and only the main learning algorithm needs to be differentially private, while the representation can be learned using any non-DP method. Additionally, the public data can also be used for optimising any hyperparameters for the representation learning. In this article, we consider PCA and variational autoencoders.
Differentially private transfer learning was recently considered by @Wang2019 in a hypothesis transfer setting, where models trained on several related source domains are used to improve learning in the desired target domain. While related, this approach is not directly applicable to our problem because the setup is different.
Another related approach was considered by @papernot2017, who propose differentially private semi-supervised knowledge transfer that uses an ensemble of “teacher” models trained on private data to label unlabeled public data, which is then used to train a “student” model that will be released. The method is flexible in a sense that it can use any “black-box” model as teachers and student. However, it is limited to classification tasks. In addition, it seems to require a rather large private dataset in addition to a small public dataset—a setting that is somewhat different from what we are interested in.
Yet another strategy is to learn a differentially private unsupervised generative model for the data (including the target variable for the prediction task of interest), use it to generate a synthetic version of the data, and use a non-DP algorithm for the actual learning task of interest. Several methods have been proposed for differentially private data sharing [@zhang2017; @xie2018; @acs2018] that could be used for generative model learning and data generation. For the problem we are considering, however, this approach is problematic as it requires solving a more general and difficult learning task, good solution of which would typically require orders of magnitude more private data than a direct solution of the original prediction task.
The rest of this article is organised as follows: In Section \[sec:approach\] we formalise the problem setting and give an overview of our proposed approach. Section \[sec:methods\] gives more details on the implementation of different parts of the proposed approach. And finally, in Section \[sec:results\] we conduct experiments with the approach on two different prediction tasks on genomic data.
Approach {#sec:approach}
========
We assume a setting where we have a private dataset containing a high-dimensional $n \times d$ feature matrix ${X_{\text{priv}}}$ and an $n \times 1$ target vector ${Y_{\text{priv}}}$, where $n$ is the number of samples and $d$ is the number of features. The goal is to learn a differentially private predictor from ${X_{\text{priv}}}$ to ${Y_{\text{priv}}}$. As learning to predict ${Y_{\text{priv}}}$ from high-dimensional ${X_{\text{priv}}}$ directly is typically not feasible, with moderate sample size and a reasonable privacy budget, we opt for using public data to learn a low-dimensional representation for ${X_{\text{priv}}}$. Therefore, we also assume a publicly available dataset of an $m \times d$ feature matrix ${X_{\text{pub}}}$ and an $m \times 1$ auxiliary target vector ${Y_{\text{pub}}'}$ for a related auxiliary prediction task. While a representation can be learned with ${X_{\text{pub}}}$ only, the availability of ${Y_{\text{pub}}'}$ is useful for selecting the size of the representation and any other hyperparameters.
We make the following informal assumptions about the relation of the public and the private data: (1) ${X_{\text{priv}}}$ and ${X_{\text{pub}}}$ contain the same set of features and are either draws from the same distribution or otherwise distributed similarly enough that using the same mapping to compute a representation is reasonable. (2) ${Y_{\text{pub}}'}$ may or may not be of the same type as ${Y_{\text{priv}}}$, but the prediction tasks should resemble each other enough that the prediction of ${Y_{\text{pub}}'}$ can be used for optimising the hyperparameters for the main task of predicting ${Y_{\text{priv}}}$.
![The process of learning $f$ and $g$. Since the learning of $g$ is DP, the leakage of information outside of the “privacy wall” is controlled.[]{data-label="fig:learning_flow"}](fig/learning_process.pdf){width="\linewidth"}
We propose the following procedure:
1. Use the public data to learn a dimension-reducing representation mapping $f: {\mathbb{R}}^d \rightarrow {\mathbb{R}}^r$, where $r \ll d$, such that $f^{-1}(f({X_{\text{pub}}})) \approx {X_{\text{pub}}}$.
2. Obtain a low-dimensional representation ${Z_{\text{priv}}}$ of the private feature data by applying $f$ to ${X_{\text{priv}}}$.
3. Learn a differentially private predictor $g$ such that $g({Z_{\text{priv}}}) \approx {Y_{\text{priv}}}$.
4. Publish $g \circ f$.
An overview of the learning process is shown in Figure \[fig:learning\_flow\].
It is easy to see that the proposed process has the same DP-guarantees as the learning algorithm of step 3:
If step 3 is $(\epsilon, \delta)$-DP w.r.t. ${Z_{\text{priv}}}$ and ${Y_{\text{priv}}}$, then the whole process is also $(\epsilon, \delta)$-DP w.r.t ${X_{\text{priv}}}$ and ${Y_{\text{priv}}}$.
As the learning of $f$ does not use private data, it does not leak any private information. Since each row of ${Z_{\text{priv}}}$ depends only on the corresponding row of ${X_{\text{priv}}}$, $(\epsilon, \delta)$-guarantees w.r.t. ${Z_{\text{priv}}}$ translate directly to guarantees w.r.t. ${X_{\text{priv}}}$.
In the following section we give some methods that can be used to implement the DP predictor $g$ and the representation mapping $f$. In addition, we describe a procedure for tuning the hyperparameters of $f$.
Methods {#sec:methods}
=======
Differentially private prediction
---------------------------------
Later in Section \[sec:results\] we will consider prediction tasks that are either real-valued regression or binary classification tasks. Linear regression will be applied to the former and logistic regression to the latter. For now, denote by $X$ the feature matrix and by $y$ the prediction target vector (either real-valued or binary $\{-1, 1\}$)
Logistic regression can be made differentially private with objective perturbation. The usual non-DP version of the problem can be solved by minimising the regularised negative log-likelihood $n^{-1} \sum_{i=1}^n \log(1 + e^{y_i w^T x_i}) + \lambda w^T w$ with respect to the weight vector $w$, where $x_i$ and $y_i$ denote the $i$th sample in $X$ and $y$ respectively and $\lambda$ controls the strength of $L_2$ regularisation. In a method presented by @chaudhuri2009, $\epsilon$-DP privacy is obtained by adding a random bias term $b^T w / n$ (where $b$ is a random vector drawn from a distribution with density proportional to $e^{-\epsilon ||b|| / 2}$) to the optimisation objective. The method requires that the samples in the input feature data are bounded into a 1-sphere.
Like DP logistic regression, also a DP linear regression algorithm can be obtained with an analogous objective perturbation method [@kifer2012]. However, since the underlying model belongs to the exponential family, there is also an alternative output-perturbation based $\epsilon$-DP method that does not require iterative optimisation: Compute the sufficient statistics ($X^T X$, $X^T y$ and $y^T y$) and add noise to them via the Laplace-mechanism [@foulds2016]. We use Bayesian linear regression with sufficient statistic perturbation and data clipping as described by @honkela2018.
Representation learning
-----------------------
*Random projection* [see e.g. @bingham2001] projects the $d$-dimensional data to an $r$-dimensional subspace by multiplying it with a random $d \times r$ projection matrix. This transformation has been shown to preserve approximately the distances between data points [@johnson1984], which is often a desired property for dimensionality reduction methods.
*Principal component analysis* (PCA) finds an orthogonal linear transformation that converts the data to coordinates that are uncorrelated and whose variance decreases from first to last coordinate. When used for dimensionality reduction, only the first $r$ coordinates are kept—these correspond to the $r$ orthogonal directions in which the variance of the original data is the highest.
*Variational autoencoder* (VAE) [@kingma2014] learns a generative decoder model $p_\theta(x | z)$, where $z$ is a latent representation of $x$, and an encoder model $q_\xi(z | x)$ that approximates the posterior distribution $p_\theta(z | x)$. Both $p_\theta$ and $q_\xi$ are implemented as neural networks (typically MLPs) and optimised concurrently with variational inference.
We fix $z$ to be low-dimensional, in which case the learned encoder $q_\xi$ can be used for dimensionality reduction by setting $f(x) = {\mathrm{E}}_{z \sim q_\xi(\cdot | x)}[z]$. (As usual, define $q_\xi(\cdot | x)$ as a multivariate Gaussian distribution parametrised by mean $\mu_\xi(x)$ and diagonal covariance $\Sigma_\xi(x)$, in which case $f(x) = \mu_\xi(x)$.)
Optimisation of hyperparameters
-------------------------------
![The process of hyperparameter optimisation. In this example the auxiliary prediction task is assumed to be classification with multiple classes (denoted by a, b, $\ldots$, n), which are partitioned into subsets that consist of two classes each. These are then used in a cross-validation-like hyperparameter optimisation procedure.[]{data-label="fig:param_opt"}](fig/param_opt_process.pdf){width="\linewidth"}
For selecting the dimension of the representation and any other hyperparameters of the representation-learning algorithm, we propose a combination of any parameter optimisation approach (such as Bayesian optimisation, random search or grid search) and a cross-validation-like procedure for optimising an auxiliary task of predicting ${Y_{\text{pub}}'}$ from ${X_{\text{pub}}}$. As no private data are used, the parameter optimisation phase does not consume any of the available privacy budget. In addition, if the auxiliary prediction task uses the same method as the main prediction task, then the hyperparameters could be optimised at the same time.
First the (public) data are divided into $k$ disjoint subsets. Instead of using one of the subsets as “validation” data and the rest as “training” data as in cross-validation, we use one of the subsets to simulate the private data and the rest to simulate the public data. From now on, these are referred to as *pseudo-private* and *pseudo-public* sets. The proposed framework (from Section \[sec:approach\]) is then applied to these, that is, a representation mapping $f$ is learned from the pseudo-public data, $f$ is applied to the features of pseudo-private data, a predictor $g$ is learned for the pseudo-private target variable and its accuracy is measured. As in $k$-fold cross-validation, this is repeated for all $k$ possible selections of the pseudo-private subset. For measuring the accuracy of $g$, (actual) cross-validation can be used, i.e., the pseudo-private data can be further divided into different learning and validation sets.
To mimic the case in which the public and private data do not have exactly the same distribution, we also want the pseudo-public and pseudo-private data to be sufficiently different. This guides the optimiser towards selecting conservative hyperparameters that are more likely to work well on a wide range of different private datasets. If the auxiliary prediction task is classification and ${Y_{\text{pub}}'}$ has multiple classes, the subset division can be based on the classes: Form each subset by selecting the samples from two (or more) classes. This strategy is based on the assumption that samples belonging to different classes have different distributions. Otherwise, for instance clustering (based on either ${X_{\text{pub}}}$, ${Y_{\text{pub}}'}$ or both) could be used for finding a good subset division. An overview of the proposed hyperparameter optimisation method is shown in Figure \[fig:param\_opt\].
Results {#sec:results}
=======
We conducted experiments with two prediction tasks using cancer cell line gene expression data: cancer type classification and drug sensitivity prediction.
Representation learning for DP cancer type classification
---------------------------------------------------------
We first demonstrate the method by classifying TCGA pan-cancer samples according to the annotated cancer type (e.g. lung squamous cell carcinoma) using RNA-seq gene expression data. In this task we use the data from The Cancer Genome Atlas (TCGA) project [@tcga] as both the private and public datasets. We use this example because it can be performed within the large TCGA dataset. Because most cancer type pairs are quite easy to identify, we focus on a number of most difficult pairs.
We used preprocessed TCGA pan-cancer RNA-seq data available at <https://xenabrowser.net/datapages/>. After further preprocessing (filtering out low-expression genes, applying RLE normalisation) the dataset contains 10534 samples, 14796 genes and 33 distinct cancer types. We pick two cancer types as private data and the remaining cancer types form the public dataset. The main and auxiliary prediction tasks are therefore both cancer type classification tasks, but for distinct classes. For prediction we use the differentially private logistic regression algorithm by @chaudhuri2009.
Case First cancer type Second cancer type
------ ----------------------------------- ---------------------------------------
1 lung squamous cell carcinoma head & neck squamous cell carcinoma
2 bladder urothelial carcinoma cervical & endocervical cancer
3 colon adenocarcinoma rectum adenocarcinoma
4 stomach adenocarcinoma esophageal carcinoma
5 kidney clear cell carcinoma kidney papillary cell carcinoma
6 glioblastoma multiforme sarcoma
- adrenocortical cancer uveal melanoma
- testicular germ cell tumor uterine carcinosarcoma
- lung adenocarcinoma pancreatic adenocarcinoma
7 ovarian serous cystadenocarcinoma uterine corpus endometrioid carcinoma
- brain lower grade glioma pheochromocytoma & paraganglioma
- skin cutaneous melanoma mesothelioma
- liver hepatocellular carcinoma kidney chromophobe
8 breast invasive carcinoma prostate adenocarcinoma
- acute myeloid leukemia diffuse large B-cell lymphoma
- thyroid carcinoma cholangiocarcinoma
: The list of cancer type pairs ordered in descending order by the difficulty of classification. The pairs selected to be tested are numbered.
\[tbl:selected\_pairs\]
![Logistic regression prediction accuracy (the fraction of correctly classified samples) with $\epsilon = 1.0$ in 8 cancer type classification tasks (see Table \[tbl:selected\_pairs\]). Data: TCGA. Error bars show the standard deviation of the mean accuracy over 9 independent runs of the testing phase.[]{data-label="fig:cancertype_acc"}](fig/pred_acc.pdf)
\[tbl:hyperparam\]
While the split to private and public data could be done in multiple ways, the prediction task would be quite easy in many of those. Hence, we use the following procedure to produce several of these splits: (1) Consider all $\binom{33}{2}$ possible splits and run a non-DP version of the pipeline (as in Figure \[fig:learning\_flow\]) with PCA-based reduction to eight-dimensional space. (2) Build a sequence of cancer-type pairs by picking the pair that was the hardest to predict (i.e. has lowest classification accuracy), then from the remaining cancer types again the pair that was hardest, and so on. The result is a sequence of 16 pairs ordered by the prediction difficulty (see Table \[tbl:selected\_pairs\]). (3) Of these pairs, select the 6 hardest, as well as those 2 of the remaining pairs that had at least 200 samples in both classes.
The full testing pipeline, including the hyperparameter optimisation phase, was then run separately for each of the 8 selected pairs as a private dataset. In each case, the remaining 15 pairs form the $k = 15$ subsets that were used for optimising the hyperparameters.
#### Methods {#methods}
We compare three different representation learning methods: random projection (RP), principal component analysis (PCA) and variational autoencoder (VAE) [@kingma2014]. VAE was implemented with PyTorch [@paszke2017] and uses 1–3 hidden layers with ReLU activation functions for both the encoder and the decoder. The learning phase uses the Adam optimiser [@kingma2015] and is given one hour of GPU time with early stopping. The size of the representation (for RP, PCA and VAE) and other hyperparameters for VAE (the number of layers, layer sizes, learning rate) are optimised with GPyOpt [@gpyopt2016]. We also experimented with optimising a much larger set of hyperparameters, 12 in total, but GPyOpt had difficulties in obtaining similar levels of performance.
For each of the 8 test cases we ran the hyperparameter optimisation phase once, giving it 5 days of time. Then with the best found hyperparameters we ran the final testing 9 times with different RNG seeds, and report the mean prediction accuracy as well as the standard deviation of the mean. In measuring the prediction accuracy (both for hyperparameter optimisation and for final testing) we use 10-fold cross-validation.
![Logistic regression classification accuracy in cancer type classification as a function of $\epsilon$ for case 1. The error bars denote the standard error of the mean when repeating the DP learning but do not cover the uncertainty from hyperparameter selection.[]{data-label="fig:cancertype_acc_eps"}](fig/pred_acc_eps.pdf)
\[tbl:hyperparam\_eps\]
#### Results {#results}
Figure \[fig:cancertype\_acc\] shows the final prediction accuracy in the selected 8 cases for $\epsilon = 1$. While none of the methods fully dominates the others, VAE seems to get some edge, being clearly the best in about half of the cases and doing decent job in the rest of the cases too. The selected hyperparameters are listed in Table \[tbl:hyperparam\]. Interestingly, VAE seems to always end up with lower dimensionality of the representation than the other two methods. This could be due to the fact that VAE allows nonlinear transformations which can help to compress the relevant information in the data into a smaller number of dimensions. On the other hand, it is not clear why RP also always chooses lower dimension than PCA.
The prediction accuracy as a function of $\epsilon$ in the case 1 is shown in Figure \[fig:cancertype\_acc\_eps\] and the corresponding hyperparameters are shown in Table \[tbl:hyperparam\_eps\]. As expected, larger $\epsilon$ results in better accuracy. There is some variability compared to case 1 in Figure \[fig:cancertype\_acc\], which is mostly likely due to the results having been computed with different hyperparameters. Due to the high computational cost, variability due to hyperparameter adaptation is not included in the error bars.
![The accuracy of drug sensitivity prediction (Spearman’s rank correlation coefficient between the measured ranking of the cell lines and the ranking predicted by the models) with differentially private linear regression ($\epsilon = 1.0$) on the GDSC data.[]{data-label="fig:drugsens_acc"}](fig/avg-corr-norm2-fixed.pdf)
Representation learning for DP drug sensitivity prediction
----------------------------------------------------------
Our main learning task is to predict the sensitivities of cancer cell lines to certain drugs. In this task we use data from the Genomics of Drug Sensitivity in Cancer (GDSC) project [@yang2013] as private data. After preprocessing the data contains 985 samples, 11714 genes and 265 drugs. The data are sparse in the sense that not all drugs have been tested on all samples. For prediction we use the differentially private Bayesian linear regression algorithm by @honkela2018. The DP linear regression is applied for each drug separately, using the full $\epsilon$ budget as if it was the only drug we are interested in. We then measure and report the average prediction accuracy over all drugs.
As public data we use the gene expression measurements from the TCGA data with cancer type classification as the auxiliary prediction task. The private and public datasets are unified by removing genes not appearing in both datasets. In addition, since the TCGA gene expression data are RNA-seq-based while GDSC data are based on microarrays, we apply quantile normalisation to each gene in the TCGA data to make it match the distribution of the gene in the GDSC data. (While this operation theoretically breaks the privacy guarantees, in practice we can avoid the issue by assuming that the expression distributions obtained with the microarray technology are public knowledge.)
#### Methods {#methods-1}
In addition to RP, PCA and VAE, we also compare to DP feature selection by Sample and Aggregate framework (SAF) as presented by @kifer2012, as well as to using a set of 10 preselected genes that were used by @honkela2018 in the same prediction task. In the case of SAF half of the privacy budget is reserved for feature selection.
RP, PCA and VAE learning was performed in a similar manner as in the cancer type classification task. For selecting the size of the representation of SAF, we simply ran it with all possible sizes and select the best result (which is obviously unfair for the other methods and would yield a weaker privacy guarantee).
#### Results {#results-1}
Figure \[fig:drugsens\_acc\] shows the average prediction performance, measured by Spearman’s rank correlation. Here PCA and VAE are the best by some margin, both improving significantly over the previous state-of-the-art with preselected genes. On the other hand, SAF is clearly the worst as the DP feature selection is essentially random due to small privacy budget, and since it leaves only half of the privacy budget for the main prediction task.
Discussion
==========
Our results clearly demonstrate that representation learning with public data can significantly improve the accuracy of differentially private learning, compared to using a set of preselected dimensions or doing differentially private feature selection. Whether it is beneficial to use more advanced representation learning methods such as variational autoencoders instead of simple methods such as PCA or random projections, depends on the task. On some tasks that certainly seems to be the case.
In our current approach, the representation is learned in an unsupervised manner and the auxiliary supervised task is only used for hyperparameter selection. A natural question, that we leave for further work, is whether representation learning would also benefit from having an integral auxiliary prediction task that would be learned concurrently with the representation. The optimisation target would in that case be a combination of unsupervised reconstruction error and supervised prediction error. This approach would require an auxiliary target variable, as is the case in this work with hyperparameter optimisation.
In general, we believe DP learning can be important in opening genomic and other biomedical datasets to broader use. This can significantly advance open science and open data, and lead to more accurate models for precision medicine. So far, the accuracy of DP learning in most practical applications is not comparable to realistic non-private alternatives. Our present work makes an important contribution toward making DP learning practical.
In the present work the representation learning was not performed under DP. This is a clear limitation if the other data set also needs privacy protection. This can in theory be addressed easily, by simply training the representation model under DP, but this will likely have an impact on the accuracy of the final model. Ultimately we believe that a clever combination of private and non-private data such as in our paper can lead to the best results.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank the reviewers of an earlier workshop version of this article for helpful comments.
Funding {#funding .unnumbered}
=======
This work has been supported by the Academy of Finland \[Finnish Center for Artificial Intelligence FCAI and grants 292334, 294238, 303815, 303816, 313124\].
[^1]: These authors jointly supervised the work.
[^2]: There are two slightly different definitions of neighbouring datasets. In *bounded* case, the value of one sample is allowed to change. In *unbounded* case, the addition or removal of one sample is allowed. Unbounded $\epsilon$-DP guarantee implies bounded $2\epsilon$-DP guarantee. This article uses the bounded case.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Ice flow velocities close to the terminus of major outlet glaciers of the Greenland Ice Sheet can vary on the time scale of years to hours. Such flow speed variations can be explained as the reaction to changes in terminus geometry with help of a 3D full-Stokes ice flow model. Starting from an initial steady state geometry, parts of an initially 7 km long floating terminus are removed. Flow velocity increases everywhere up to 4 km upstream of the grounding line, and complete removal of the floating terminus leads to a doubling of flow speed. The model results conclusively show that the observed velocity variations of outlet glaciers is dominated by the terminus geometry. Since terminus geometry is mainly controlled by calving processes and melting under the floating portion, changing ocean conditions most probably have triggered the recent geometry and velocity variations of Greenland outlet glaciers.'
---
[**Terminus Geometry as Main Control on\
Outlet Glacier Velocity**]{} Martin P. Lüthi\
[Versuchsanstalt für Wasserbau, Hydrologie und Glaziologie (VAW), ETH Zürich, 8092 Zürich, Switzerland]{}
now at
[University of Zurich, Dept. of Geography, 8057 Zürich, Switzerland\
Email: [email protected]]{}
Introduction
============
Flow velocities close to the terminus of major outlet Glaciers of the Greenland Ice Sheet can vary substantially on the time scale of years , and to a lesser degree seasonally , and episodically during calving events . Several major Greenland Ice Sheet outlet glaciers accelerated within the last decade , sometimes to double their pre-acceleration velocity within three years, occasionally followed by a slowdown in the following years . The dynamic changes of these outlet glaciers control the short term evolution of the ice sheet geometry, and their changing calving flux impacts the ice sheet mass budget, and therefore sea level.
The close timing of the acceleration of Helheim and Kangerdlussuaq Glaciers on the east coast and Jakobshavn Isbræ on the west coast hints to an external forcing not related to internal dynamic instabilities of the outlet glacier system. The obvious possible causes are atmospheric forcing through high meltwater production that affects basal motion , or the influence of ocean temperature on terminus melt rate, and therefore the geometry of the calving front [@Holland2008]. Increased meltwater production can supply important amounts of water to the ice-bedrock interface by hydro-fracturing , temporarily increasing the already very high water pressure under the ice sheet in vicinity of the ice stream . Due to stress transfer from the ice stream trunk to its surroundings , stream velocities are susceptible to changes in basal motion in the surrounding ice sheet, and could be affected by higher sliding velocities there. If on the other hand increased heat flux from the ocean is the driver, thinning of the floating terminus and higher calving rates would be expected, as were indeed observed at Jakobshavn Isbræ .
In this contribution we use a three-dimensional full-Stokes ice flow model to investigate the relation between terminus geometry and ice flow velocity. The model results show that the velocity at the grounding line is controlled by the length of a floating terminus, even in absence of friction at pinning points. Removing a floating terminus part by part leads to step-wise increases in flow velocity, similar to what has been observed at Jakobshavn Isbræ and Helheim Glacier .
Flow model {#sec:methods}
==========
The FISMO Full Ice Stream Model was used to investigate the effect of changing geometry on flow velocities. The finite element code FISMO solves the Stokes equations for slow, incompressible flow with variable viscosity, expressed in terms of the field variables velocity $\mathbf{v}$ and pressure $p$
\[eq:stokes\] $$\begin{aligned}
- \nabla p + \eta\, \nabla^2 \mathbf{v} +
2\,\mathbf{D} \cdot \nabla \eta + \rho \mathbf{g} &= \mathbf{0}\,,\\
\mathrm{tr}\mathbf{D} = \nabla \mathbf{v} &= 0\,,
\end{aligned}$$
where $\mathbf{D}=\frac{1}{2}\left( \nabla\mathbf{v}+(\nabla\mathbf{v})^T
\right)$ is the Cauchy strain rate tensor. The viscosity $\eta$ of glacier ice is strain rate dependent and is calculated according to Glen’s flow law (a power-law rheology) as $$\label{eq:glen}
\mathbf{D} = A\, \tau^{n-1}\, {\mbox{\boldmath$\sigma$\unboldmath}}^d \qquad \Longleftrightarrow \qquad
\eta = \frac{1}{2}\,A^{-\frac{1}{n}}\, II_{\mathbf{D}}{}^{\frac{1-n}{n}}\,.$$ where ${\mbox{\boldmath$\sigma$\unboldmath}}^d$ is the deviatoric stress tensor, $II_{{\mbox{\boldmath$\sigma$\unboldmath}}^d}= \left(
\frac{1}{2} \sigma^d_{ij}\sigma^d_{ij} \right)^{\frac{1}{2}}$ and $II_{\mathbf{D}} = \left( \frac{1}{2} D_{ij}D_{ij} \right)^{\frac{1}{2}}$ the the second invariants of ${\mbox{\boldmath$\sigma$\unboldmath}}^d$ and $\mathbf{D}$, $A=215
\,\mathrm{MPa}^{-3}\,\mathrm{a}^{-1}$ is the rate factor commonly assumed for temperate ice [@Paterson1999], and $n=3$. To avoid inifinte viscosity at vanishing strain rates, a small constant $\epsilon=1\cdot
10^{-6}\,\mathrm{a}^{-1}$ was added to $II_{\mathbf{D}}$.
Velocity was prescribed as Dirichlet boundary condition on the parts of the domain representing bedrock. On the parts of the boundary in contact with air a zero stress boundary condition was applied (which in the Galerkin finite element method employed requires no effort). On the faces in contact with the ocean, normal stress was set equal to the hydrostatic water pressure $\sigma_n
= -\rho_\mathrm{water} g z$ (ocean level is assumed at $z=0$). To limit the geometrical extent of the model, a stress boundary condition was prescribed on the boundaries to the inland ice, with the normal stress equal to ice overburden pressure $\sigma_n = -\rho_\mathrm{ice} g (z_s-z)$ (the minus sign indicates a compressive force). The latter boundary condition has been tested to work well for an infinite inclined slab of ice, and is useful since it does not force ice flow into the computational domain.
Equations (\[eq:stokes\]) and (\[eq:glen\]) together with the boundary conditions were solved numerically with the FISMO finite element (FE) code, which builds on the Libmesh FE-library (Kirk, 2007) that uses the PETSc parallel solver library. To obtain a numerically stable and divergence-free velocity solution, Q2Q1 isoparametric Taylor-Hood Elements on 27-node hexahedra were used. In Libmesh all boundary conditions are enforced with a penalty method.
The nonlinear equation system arising from Equation (\[eq:glen\]) was solved with a fixed-point iteration. An ALE (arbitrary Lagrange-Euler) formulation was used in an explicit time stepping scheme for the evolution of the model geometry. Given the velocity $\mathbf{v}$ from the previous time step, vertical mesh node positions $z$ at the free boundaries (surface and floating terminus) are updated during the time step $\Delta t$ according to $$\label{eq:6}
\Delta z = \left( \left(\mathbf{v}\cdot \mathbf{n}\right)_z + b \right) \Delta t$$ where $\mathbf{n}$ is the face normal, index $z$ indicates the vertical component, and $b$ is the annual net balance, i.e. the amount of ice added or removed during a year at the surface or under the floating terminus. The time step $\Delta t$ was chosen so that the maximum vertical displacement was $1\,\mathrm{m}$. After each time step, all mesh nodes were adjusted to their initial relative positions between surface and bedrock.
Model setup {#sec:assumptions}
===========
The bedrock is parametrized as a fjord geometry that resembles Jakobshavn Isbræ. The grounding line location is fixed at $x=0$, ice is grounded in the domain $x>0$ and floating for $x<0$. Bedrock elevation $z_b(x,y)$ is assumed to be $z_b=0$ outside of a straight channel geometry that is parametrized in the domain $x>0$ by $$\begin{aligned}
\label{eq:channel}
z_c(x) &=
\begin{cases}
1 + \beta_r(x_{r}-x)\,, & x < x_r \\
1 \,, & x_r \le x < x_t \\
e^{-(\beta_{t}(x_t-x))^2}, & x \ge x_t \\
\end{cases}\notag\\
z_b(x,y) &= - z_c(x) \, e^{-(\mu y )^2}\,H_{m} \end{aligned}$$ where $z_c(x)$ is the unit centerline depth of the channel and $\mu$ determines the cross-sectional shape of the channel. The position $x_t$ determines where the channel reaches is maximum depth and $x_r$ is the position from where on the channel starts raising towards the calving front (parameters are given in Table \[tab:params\]). The initial ice surface was parametrized with $$\label{eq:surface}
z_s(x) =
\begin{cases}
50 + 4.5 \sqrt{x}\,, & x \ge 0\,,\\
50 \,, & x < 0\,.
\end{cases}$$ For the initial bottom elevation of the floating terminus the bedrock geometry given in Equation (\[eq:channel\]) was extended. The geometry of the ice surface and the floating terminus was evolved until it became stationary. This resulted in surface draw-down in the main ice stream channel where longitudinal extension rates are highest.
The usually complicated polythermal structure of polar ice streams is neglected and all ice is assumed to be at the melting point. This assumption includes the important effect on ice dynamics of a thick layer of temperate ice close to the base, the thickness of which is at least $300\,\mathrm{m}$ at Jakobshavn Isbræ and might even amount to $700\,\mathrm{m}$ . Such an approach underestimates the stress transfer to the surrounding ice sheet through the kilometer-thick layer of very cold ice close to the surface .
Velocity at the ice-bedrock contact is set to zero and basal motion is therefore ignored. This assumption is not realistic, but reasonable for the task of investigating the importance of geometry change on flow velocities. Moreover, the flow of Jakobshavn Isbræ, for example, seems not to be dominated by basal motion, and high flow velocities are thought to be largely due to a thick layer of temperate ice and the steep surface slope.
The geometry was discretized with hexahedra elements (the central part of the mesh is shown in Figure \[fig:modelvelo\]). Since the domain is symmetric about the $x$-$z$-plane, only one half of the geometry has to be meshed, and the boundary condition on the $x$-$z$-plane is $v_y=0$. The computational mesh of the grounded ice consists of 25x20 elements in the horizontal, and 10 elements in the vertical. The floating terminus was discretized with up to 30x11x10 elements, depending on terminus length. Element sizes are reduced in $x$-direction around the calving front where velocity gradients are biggest.
The annual net mass balance at the surface is assumed to be elevation dependent with $b(z)= \gamma (z - z_\mathrm{ELA})$ with a mass balance gradient $\gamma = 0.005\,\mathrm{a}^{-1}$ and an equilibrium line altitude $z_\mathrm{ELA}=1100\,\mathrm{m}$, which is the order of measured values close to the ice sheet margin at Jakobshavn Isbræ. Net mass balance under the floating terminus was set to $-50\,\mathrm{m}\,\mathrm{a}^{-1}$.
The model experiment was started with an initial geometry as described above, and with a floating terminus of $7\,\mathrm{km}$ length. The elevation of the surface and the bottom of the floating terminus were allowed to evolve until they reached a stationary state.
Model results {#sec:results}
=============
To investigate the influence of the length of the floating terminus on the stress state and the flow velocities, parts from the front of the floating terminus were removed. New velocity solutions were calculated without any further evolution of the geometry. The geometries with terminus lengths of 7, 2, 1 and $0\,\mathrm{km}$ are plotted in Figure \[fig:modelvelo\] next to the corresponding centerline velocities. Flow velocities increase with shorter terminus length, and nearly double at the grounding line when the floating terminus is completely removed.
The characteristic local maximum of flow velocity around the grounding line (black and blue curves in Fig. \[fig:modelvelo\]) is due to changing surface slopes in a small over-deepening plotted in the top panel of Figure \[fig:modelvelo\]. The ice is up to $5\,\mathrm{m}$ below hydrostatic equilibrium at the grounding line (top plot of Fig. \[fig:modelvelo\]) and reaches equilibrium at about four ice thicknesses along the floating terminus. A similar surface depression close to the calving front has been found in other modeling studies [e.g. @Lestringant1994].
The removal of the floating terminus induces changes in the stress state at the grounding line, as shown in Figures \[fig:modelgeometry\] and \[fig:sigmaxx\]. Longitudinal deviatoric stress, which determines the rate of longitudinal extension, changes from negative (compressive) to positive (extensive) at the grounding line when the floating terminus is removed. Also the mean stress (plotted in Fig. \[fig:sigmaxx\]b as the deviation from overburden pressure) shows a marked change with reversed slope and a kink at the water line (red lines in Fig. \[fig:sigmaxx\]).
Discussion {#sec:discussion}
==========
The length and shape of the floating terminus controls the size and distribution of glacier flow velocity in vicinity of the grounding line, as is illustrated in Figure \[fig:modelvelo\]. There is a striking similarity of the modeled velocities along the centerline of our prototype outlet glacier with the measured evolution of terminus velocities at Jakobshavn Isbræ. Characteristic features like the maximum of flow velocity at the grounding line are also visible in the modeled velocities. The major difference between model and reality is the measured increase in flow velocity up to $40\,\mathrm{km}$ upstream of the grounding line, while our model experiment only shows a speedup in the last $4\,\mathrm{km}$ upstream of the grounding line. The difference is due to the model experiment setup that did not evolve the surfce after removal of the floating terminus, which corresponds to rapid disintegration of the floating ice. The flow acceleration would migrate inland together with the surface drawdown, as was exemplified in model studies of Pine Island Glacier (Antarctica) . In these studies the removal of the terminus immediately influenced ice flow far inland due to very weak coupling to the bed over large parts of the glacier. In contrast, we assumed full coupling, which affects flow velocity only within 5 ice thicknesses from the grounding line.
Ice shelf buttressing is usually assumed to result from frictional forces acting on the floating terminus or ice shelf at pinning points, or at its sides . In the absence of friction on the terminus the total resistive horizontal force in flow direction is that of water pressure acting as a normal stress on the interface between ice and ocean. The integral of this normal stress over the interface area, weighted by the face normal, is constant and independent of the length and shape, presence or absence of the terminus. In the presence of a floating terminus the horizontal compressive force is evenly distributed over the cross section at the grounding line (Fig. \[fig:sigmaxx\]). In the absence of the terminus the water pressure acts only below the water line, which leads to extensive deviatoric stresses with a maximum at the water line, and corresponding high extensional ice deformation rates there. The geometry of compact floating ice in contact with grounded ice has a major influence on the stress distribution in the grounding line area. The model results presented above show that ice shelf buttressing is mainly an effect of the geometry of floating portion.
Since the flow velocity is sensitive to terminus geometry, a growing floating terminus leads to slower flow velocities. This effect might explain the observation that Jakobshavn Isbræ slows in winter when a compact floating portion forms in the terminus area .
The cause for big changes of terminus geometry, and therefore flow velocity, is most likely the influence of increased heat flux from the ocean causing increased melt under the floating portion of the terminus. For Jakobshavn Isbræ a strong increase of ocean bottom temperature has been measured, the influence of which coincides with thinning of the floating terminus and flow acceleration [@Holland2008]. Alternatively, changes of the calving process by ponding meltwater in crevasses could explain the disintegration of floating termini , although observations of increase of such ponding water have not been made for the Greenland outlet glaciers discussed above.
Conclusions {#sec:conclusions}
===========
The geometry of the terminus area is the dominant control on the velocity field close to the grounding line of an outlet glacier. Short term geometry changes, such as the disintegration of a floating terminus, calving, or the creation of embayments in grounded ice, greatly affects the flow field close to the terminus.
A typical calving event at big outlet glaciers such as Jakobshavn Isbræremoves up to $400\,\mathrm{m}$ of ice from the glacier terminus. Measurements have shown that the glacier velocity reacts immediately (within a 15 minute measurement interval), but no extra movement could be observed during calving events . Such almost step-wise increase in flow velocity can be reproduced with a 3D flow model, when parts of a floating terminus are removed. Short term flow velocity variations thus are mainly an effect of stress redistribution, which in turn is controlled by changes of the terminus geometry.
Acknowledgments {#acknowledgments .unnumbered}
===============
Thoughtful comments by Martin Truffer, Martin Funk and two anonymous referees have helped improve the clarity of presentation. Funding was provided by the Swiss National Science Foundation (200021-113503/1) and by NASA’s Cryospheric Sciences Program (NNG06GB49G).
Amundson, J., Truffer, M., L[ü]{}thi, M. P., Fahnestock, M., Motyka, R. J., and West, M. (2008). Glacier, fjord, and seismic response to recent large calving events, [J]{}akobshavn [I]{}sbr[æ]{}, [G]{}reenland. , 35:L22501. doi:10.1029/2008GL035281.
Benn, D. I., Warren, C. R., and Mottram, R. H. (2007). Calving processes and the dynamics of calving glaciers. , 82:143–179.
Das, S. B., Joughin, I., Behn, M. D., Howat, I. M., King, M. A., Lizarralde, D., and Bhatia, M. P. (2008). Fracture propagation to the base of the [G]{}reenland [I]{}ce [S]{}heet during supraglacial lake drainage. , 320. doi: 10.1126/science.1153360.
Dupont, T. and Alley, R. (2005). Assessment of the importance of ice-shelf buttressing to ice-sheet flow. , 32:L04503.
Holland, D. M., Thomas, R. H., de Young, B., Ribergaard, M. H., and Lyberth, B. (2008). Acceleration of [J]{}akobshavn [I]{}sbr[æ]{} triggered by warm subsurface ocean waters. .
Howat, I. M., Joughin, I., and Scambos, T. A. (2007). Rapid changes in ice discharge from [G]{}reenland outlet glaciers. , 315:1559–1561.
Iken, A., Echelmeyer, K., Harrison, W. D., and Funk, M. (1993). Mechanisms of fast flow in [J]{}akobshavns [I]{}sbrae, [G]{}reenland, [P]{}art [I]{}: [M]{}easurements of temperature and water level in deep boreholes. , 39(131):15–25.
Joughin, I., Howat, I. M., Fahnestock, M., Smith, B., Krabill, W., Alley, R. B., Stern, H., , and Truffer, M. (2008). Continued evolution of [J]{}akobshavn [I]{}sbrae following its rapid speedup. , 113:F04006. doi:10.1029/2008JF001023.
Joughin, I. R., Abdalati, W., and Fahnestock, M. A. (2004). Large fluctuations in speed on [G]{}reenland’s [J]{}akobshavn [I]{}sbr[æ]{} glacier. , 432:608–610.
Lestringant, R. (1994). A two-dimensional finite-element study of flow in teh transition zone between an ice sheet and an ice shelf. , 20:67–72.
L[ü]{}thi, M., Funk, M., and Iken, A. (2003). Indication of active overthrust faulting along the [H]{}olocene-[W]{}isconsin transition in the marginal zone of [J]{}akobshavn [I]{}sbr[æ]{}. , 108(B11). doi:10.1029/2003JB002505.
L[ü]{}thi, M., Funk, M., Iken, A., Gogineni, S., and Truffer, M. (2002). Mechanisms of fast flow in [J]{}akobshavns [I]{}sbrae, [G]{}reenland; [P]{}art [III]{}: measurements of ice deformation, temperature and cross-borehole conductivity in boreholes to the bedrock. , 48(162):369–385.
L[ü]{}thi, M. P., Fahnestock, M., and Truffer, M. (2008). Observational evidence for a thick temperate layer at the base of [J]{}akobshavn [I]{}sbr[æ]{}. . (submitted).
Nettles, M., Larsen, T. B., Elesegui, P., Hamilton, G. S., Stearns, L. A., Ahlstr[o]{}m, A. P., Davis, J. L., Andersen, M. L., de Juan, J., Khan, S. A., Stenseng, L., Ekstr[o]{}m, G., and Forsberg, R. (2008). Step-wise changes in glacier flow speed coincide with calving and glacial earthquakes at [H]{}elheim [G]{}lacier, [G]{}reenland. .
Paterson, W. S. B. (1999). . Butterworth-Heinemann, third edition.
Payne, A. J., Vieli, A., Shepherd, A. P., Wingham, D. J., and Rignot, E. (2004). Recent dramatic thinning of largest [W]{}est [A]{}ntarctic ice stream triggered by oceans. , 31:L23401. doi:10.1029/2004GL021284.
Rignot, E. and Kanagaratnam, P. (2006). Changes in the velocity structure of the [G]{}reenland [I]{}ce [S]{}heet. , 311:986–990. doi:10.1126/science.1121381.
Schmeltz, M., Rignot, E., Dupont, T. K., and MacAyeal, D. R. (2002). Sensitivity of pine island glacier, west antarctica, to changes in ice-shelf and basal conditions: a model study. , 48(163):552–558.
Thomas, R. H., Abdalati, W., Frederick, E., Krabill, W. B., Manizade, S., and Steffen, K. (2003). Investigation of surface melting and dynamic thinning on jakobshavn isbrae, greenland. , 49(165):231–239.
Truffer, M. and Echelmeyer, K. (2003). Of [I]{}sbr[æ]{} and [I]{}ce [S]{}treams. , 36:66–72.
, C. J. (1998). Fracture mechanics approach to penetration of surface crevasses on glaciers. , 27(1):31–47.
Zwally, H. J., Abdalati, W., Herring, T., Larson, K., Saba, J., and Steffen, K. (2002). Surface melt-induced acceleration of [G]{}reenland ice-sheet flow. , 297(5579):218–222.
----------------------------------- ------------- ------ ------------
maximum channel depth $H_m$ 1600 m b.s.l.
location of deepest point $x_t$ 15 km
location of bed inflexion $x_r$ 7 km
steepness of trough on centerline $\beta_{t}$ 2 $ 10^{-4}$
centerline slope $\beta_r$ -8 $10^{-5}$
steepness of sides $\mu$ 7 $10^{-4}$
----------------------------------- ------------- ------ ------------
: Parameters for the ice stream geometry[]{data-label="tab:params"}
![Modeled centerline flow velocity at the surface for the terminus geometries shown next to the curves. Decreasing the length of the floating terminus affects the value and spatial distribution of flow velocities. Grounding line position is indicated with a vertical dotted line. Top panel: surface elevation (blue, displayed $40\,\mathrm{m}$ lower) and height above buoyancy (black).[]{data-label="fig:modelvelo"}](fig1.png){width="30pc"}
![Modeled deviatoric stress $\sigma_{xx}^d$ in flow direction at the grounding line. The model geometry is sliced along the centerline, and at the grounding line (i.e. the terminus is not shown). Terminus lengths are 7, 2, 1 km, and no terminus.[]{data-label="fig:modelgeometry"}](fig2.pdf){width="20pc"}
![(a) Modeled deviatoric component of longitudinal stress $\sigma_{xx}^d$ in a vertical profile $20\,\mathrm{m}$ upstream of the grounding line. (b) Deviation of the mean stress $\sigma_m$ with respect to the hydrostatic overburden stress $\sigma_o$.[]{data-label="fig:sigmaxx"}](fig3.pdf){width="30pc"}
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We consider a monitoring application where sensors periodically report data to a common receiver in a time division multiplex fashion. The sensors are constrained by the limited and unpredictable energy availability provided by Energy Harvesting (EH), and by the channel impairments. To maximize the quality of the reported data, the packets transmitted contain newly generated data blocks together with up to $r - 1$ previously unsuccessfully delivered ones, where $r$ is a design parameter; such blocks are compressed, concatenated and encoded with a channel code. The scheme applies lossy compression, such that the fidelity of the individual blocks is traded with the reliability provided by the channel code. We show that the proposed strategy outperforms the one in which retransmissions are not allowed. We also investigate the tradeoff between the value of $r$, the compression and coding rates, under the constraints of the energy availability, and, once $r$ has been decided, use a Markov Decision Process (MDP) to optimize the compression/coding rates. Finally, we implement a reinforcement learning algorithm, through which devices can learn the optimal transmission policy without knowing a priori the statistics of the EH process, and show that it indeed reaches the performance obtained via MDP.'
author:
-
bibliography:
- 'bibb.bib'
title: 'Joint Retransmission, Compression and Channel Coding for Data Fidelity under Energy Constraints'
---
Introduction {#sec:intro}
============
the last few years, Wireless Sensor Networks (WSNs) and their evolution into the Internet of Things (IoT) have spurred some significant research efforts [@Zorzi-2010]. A major challenge in such IoT applications is to ensure uninterrupted service with minimal device maintenance, driven by the need to minimize the operational expenses. A typical periodically-reporting device is battery-driven, but expected to operate for long periods without human intervention: in particular, the industry aims to achieve a minimum of 10 years of battery lifetime [@nokia; @ericsson; @huawei]. Energy harvesting is a promising technique through which sensors can scavenge energy from the environment and replenish their batteries, thereby fostering self-sustainability of the devices and, thus, of the IoT application. Ideally, EH could guarantee infinite lifetime; however, its intermittent nature requires the use of flexible protocols able to adapt to a stochastic energy availability. A general overview of recent advances in wireless communications with EH is presented in [@Uetal2015], while [@Gunduz-2014] discusses the challenges of designing an intelligent EH communication system and presents a general mathematical model that can be adapted to some specific contexts.
In this work, we study how to combine an efficient energy utilization with Quality of Service (QoS) requirements, in terms of quality of the reported data. We consider a monitoring system where multiple sensor nodes report their readings to a common receiver in a single-hop access network. Given the predictability of the reporting patterns, the devices access the channel according to a Time-Division Multiple Access (TDMA) scheme. Each device aims at using its available energy in such a way to optimize the reconstruction fidelity, which depends on the distortion introduced by lossy compression and on the channel impairments. To this end, we develop a joint source-channel coding mechanism, which comprises transmission of new data, possible retransmission of previously unsuccessfully received data, and lossy compression. Specifically, each transmitted packet contains a newly generated data block and up to $r - 1$ previous data blocks that were not successfully received, all of them compressed so as to fit the fixed frame size; the retransmission mechanism will be thoroughly explained in Section \[sec:retx\]. The compression is lossy and affects data quality, but on the other hand reduces the volume of information bits to send over the channel, allowing the use of more redundancy to combat the channel impairments. The selection of the optimal operating strategy in terms of the tradeoff between selection of compression rate and channel coding rate is constrained by the energy availability and statistics of the EH process, and, at the same time, driven by the target minimization of the distortion of the reported data. In the paper, we investigate the impact of the value of $r$ on the distortion, and once $r$ has been decided, use a MDP to study the tradeoff between the compression and the channel coding rates. MDPs are often employed to derived energy management policies, as they represent an appealing solution to optimize some long-term utilities in the presence of stochastic EH [@tutorial]. Moreover, in the paper we implement a Reinforcement Learning (RL) algorithm, namely R-learning [@schwartz-1993], through which a device can learn the optimal transmission policy through a trial-and-error discovery process and, thus, overcome the need to know the statistics of the network and the energy dynamics in advance. In particular, RL algorithms are commonly used to solve problems that are modeled as MDPs [@Sutton-98] and several works in the literature use them in energy management problems in the presence of EH, cf. [@Blasco-2013; @Ortiz-2016]. We also note that more sophisticated approaches, like deep learning algorithms, cannot be pursued because they have too demanding resource and computation requirements that the simple IoT devices could not bear.
The key contributions of this work are summarized as follows:
- We set up a transmission problem where energy consumption and data quality are balanced. We consider both lossy compression and the impairments due to transmission over a fading channel, and propose an exhaustive parametric model of the energy dynamics of a device.
- We jointly consider source and channel coding, accounting for realistic rate-distortion curves that match those of practical data compression algorithms, and characterize the outage probability that affects the transmission of short packets over a fading channel.
- Driven by the convexity of the distortion function, we introduce a hybrid retransmission mechanism where previously lost packets are sent along with the new one. Although this implies choosing a larger compression ratio so that all packets can fit in the same slot, the long-term average distortion is improved. Further, by fixing the maximum number of retransmissions allowed, latency is limited.
- We implemented an RL algorithm that learns from experience by trial-and-error and gradually converges to the optimal policy determined through the MDP.
- We evaluate numerically the improvement brought by the retransmission scheme, as well as the role that various system variables play in determining the performance.
A preliminary version of the work presented in this paper, focused on the MDP solution to the problem of joint selection of compression and channel-coding ratio, appeared in [@Pielli-2017]. This paper presents its substantial extension, where the retransmission mechanism and the RL algorithm constitute the two main novel contributions.
The rest of the paper is structured as follows. We conclude this section with an account of the related work. Section \[sec:comm\_model\] describes the communication model, i.e., how data is processed and transmitted, while Section \[sec:en\_model\] focuses on the energy consumption and EH profiles. In Section \[sec:retx\] we present our energy- and QoS-aware retransmission scheme. The optimization problem is defined in Section \[sec:problem\] and optimally solved in Section \[sec:solution\]; Section \[sec:results\] shows the numerical evaluation. In Section \[sec:learning\] we discuss how nodes can learn the optimal policy. Finally, the conclusions are drawn in Section \[sec:conclu\].
Related work {#related-work .unnumbered}
------------
Investigating the effects of packet losses on data distortion is not a new topic. However, many works limit their studies to Gaussian data sources or neglect the energy limitations, cf. [@Zachariadis; @Aguerri-2011]. A common approach is to use the distortion exponent as the performance metric [@Laneman; @Zhao-2016], but this is meaningful only for the high Signal-to-Noise Ratio (SNR) regime. Since the SNR is generally low in IoT scenarios, and especially in EH-powered nodes, we do not make use of the distortion exponent, but rather consider a distortion metric that accounts for compression and channel outages and that is affected also by the energy availability. Another beaten path is that of layered transmission schemes, where the source is coded in superimposed layers, like in [@etemadi]. Each layer successively refines the data description and is transmitted with a larger coding rate, thus the transmission is less robust to failures. This practice is often used in multimedia applications [@ng-2014], but is not very meaningful in other contexts. Several works focus on the minimization of data distortion in the presence of energy limitations and/or EH. For example, [@bhat] analyses the tradeoff between the energy required by quantization and transmission in the presence of EH, but neglects the effect of packet losses on the data quality at the receiver. A similar problem is treated in [@Castiglione], where the sensor nodes jointly perform source-channel coding and balance energy between processing and transmission in such a way to guarantee a minimum average distortion and, at the same time, maintain the data queue stable. Also [@motlagh] proposes a joint source-channel coding scheme for Gaussian and binary sources. The authors derive lower bounds on the distortion achievable when the energy buffer may have some leakage. In [@bui], the sensor nodes can tune their duty cycle and information generation rate with the objective of guaranteeing energy self-sufficiency to a multi-hop network.
Two works that bear similarities with ours are [@zordan] and [@Fullana-2017]. In [@zordan], the goal is to maximize the long-term average quality of the transmitted packets by adapting the degree of lossy compression and the transmission power. Power control is used to maintain the packet error probability below a chosen threshold, depending on the state of the channel. Like in our work, the optimal transmission strategy is determined through a MDP. In [@Fullana-2017], the reconstruction of time-correlated sources in a point-to-point communication is formulated as a convex optimization problem and solved with an iterative algorithm. The node has to decide on the transmission power and rate and is subject to EH constraints.
It is difficult to find works that deal with retransmissions in conjunction with data processing outside of the multimedia networks. Retransmissions in IoT constrained networks are considered, e.g., in [@aprem-2013], which addresses the problem of transmit power control in the presence of EH when a sensor node makes use of an Automatic Repeat Request (ARQ) protocol and retransmits the lost packets. Differently from us, they assume that some Channel State Information (CSI) is available at the nodes, and, further, they do not consider data processing. In [@Costa-2013], sensor nodes implement an energy-aware hop-by-hop retransmission mechanism where only packets carrying critical information are ensured to be retransmitted. This differs from our scheme, where packets are not differentiated according to priorities but can be compressed at the source.
The retransmissions are also often investigated in terms of cooperative retransmissions, where devices collaborate with each other, cf. [@Hu-2013; @Lee-2012]. However, in this work we do not to consider cooperation among nodes, as we are interested in scenarios where the devices are unaware of the presence of the others and of the resources they have available, and communicate solely with the gateway, e.g., as in a LoRa network [@lorawan_specs].
Communication model {#sec:comm_model}
===================
{width=".8\textwidth"}
We assume a , star topology network, comprising a multitude of IoT devices that periodically monitor some phenomena of interest and report data to a common receiver. This data collector is assumed to be connected to the energy grid, whereas the sensor nodes are battery-powered, but also endowed with energy harvesting capabilities.
The predictability of the traffic pattern makes TDMA a good choice for transmission scheduling among the devices. Since in this case the devices do not interfere with each other, we focus just on a single device, with the aim to find its optimal transmission strategy. The corresponding model is represented in Fig. \[fig:node\]. A device is endowed with a circuitry to scavenge energy from the environment; this energy is stored in a buffer of finite size and used by the node to generate, process, and send data to the common receiver through a Rayleigh-fading channel. In this section, we describe the joint source-channel coding scheme, while in Section \[sec:en\_model\] we discuss the energy harvesting aspects. For the sake of simplicity, we focus on the case when no retransmissions are allowed, while the extension that takes into account retransmissions is discussed in Section \[sec:retx\].
Compression at the source
-------------------------
A device is capable of compressing in a lossy fashion the time series generated through sensing the environment, as described next. Between two consecutive reporting events, a sensor node collects a block of readings[^1], where each block has a constant size of $L_0$ bits and is independent of the previous ones. The device compresses the generated block by selecting a compression level $k$, where $k\in\{0,\dots,m\}$, and produces a compressed block of size $L$ bits. When $k=0$, the size of the compressed block is $L=0$, i.e., no packet is transmitted, and when $k=m$, the size of the compressed block is $L = L_0$, i.e., no compression takes place. We define the *compression ratio* as the ratio between the size of the compressed block and that of the original one: $L/L_0\!=\!k/m\!\in\![0,1]$. Lossy compression makes it possible to trade some accuracy in the data representation using a lower compression ratio with additional error-correction redundancy, and consequently have an increased robustness to channel impairments, as will be explained in Section \[ssec:data\_tx\]. The distortion metric we employ is defined as the maximum absolute error between the original and the compressed signal normalized to the amplitude range of the signal in the considered time window; it has been used on real data series in, e.g., [@Zordan-2017].
The curve is signal- and algorithm-dependent. We consider an automatic sensor profiling approach where a device dynamically decides upon the compression algorithm to use depending on the type of signal it generates and the corresponding distortion. We use the following expression for the distortion introduced by lossy compression, that was derived in our previous work [@eccentric]: $$\label{eq:dist}
D(k) = \begin{cases}
b \left( \left(\frac{k}{m} \right)^{-a} - 1\right) \, &\text{if } k\ge 1 \\
D_\fl \, &\text{if } k = 0
\end{cases}$$ where $b\!>\!0$, $0\!<a\!<\!1$. The choice $k=0$ entails that the packet is discarded (due to energy restrictions) and the corresponding distortion is equal to the maximum value $D_\fl \triangleq 1$. Eq. shows that the distortion is a convex and decreasing function of the compression ratio $k/m$.
Data transmission {#ssec:data_tx}
-----------------
Each IoT node transmits during its dedicated time slot, whose duration $T$ defines the maximum number of bits that can be sent $S=T/T_{\rm b}$, with $T_{\rm b}$ being the (fixed) bit duration. It is reasonable to assume that $L_0\le S$, i.e., a device may avoid compressing a packet. After compression, there are $L \le L_0$ information bits and the corresponding coding rate is $R\!=\!L/S$. Depending on $R$ and the actual channel conditions, the packet may not be correctly received with an outage probability $\pr_\out(R)$.
We envisage the presence of an acknowledgment mechanism, so that the receiver sends feedback to the transmitter; however, no channel state estimation is performed, and consequently the transmission power $P_\ttx$ is kept constant. The communication channel is affected by block Rayleigh fading; when the channel is in a deep fade the packet is lost and an outage occurs.
Because the packets sent by IoT devices are likely to be short, we exploit the recent results of finite-length information theory that adapt the classical concepts of channel capacity to the case of short data packets [@polyansky]. In particular, the results of [@yang] legitimate the use of the quantity $\log_2(1+\gamma)$ to represent the maximum rate even in the finite-length regime, where $\gamma$ is the SNR at the receiver, given by: $$\gamma = \dfrac{|H|^2 P_\ttx}{A^2\, (d/d_0)^{\eta} N} \triangleq |H|^2 \, \bar{\gamma}.$$ $H$ is the channel gain coefficient that represents fading (assumed to be constant over the packet duration in the quasi-static scenario), and $\bar{\gamma}$ is the expected SNR at the receiver, that depends on the transmission power $P_\ttx$, the noise power $N$, and the term $A^2\, (d/d_0)^{\eta}$ that accounts for path loss. The latter depends on the path-loss exponent $\eta$, the distance between transmitter and receiver $d$, and a path-loss coefficient $A=4\pi d_0 f_0/c$, where $f_0$ is the transmission frequency, $c$ the speed of light, and $d_0$ a reference distance for the antenna far field [@Rappaport-1996]. Thus, the outage probability can be approximated as: $$\pr_\out(R) = \prr\left(\,\log_2\left(1+\gamma\right) < R\right). \label{eq:out_gen}$$
As we consider Rayleigh fading, $H$ follows a complex Gaussian distribution with zero mean and unit variance. In this case, the outage probability becomes: $$\pr_\out(R) = 1 - e^{-(2^{R}-1)/\bar{\gamma}}, \label{eq:outage}$$ which is non-decreasing in $R$ (and thus in $k$, since $L=k/m \,L_0$), and initially convex and then concave. Clearly, the farther the device from the receiver, the larger the outage probability, since $\bar{\gamma}$ decreases with distance. If a packet gets lost, the block of readings contained in it can be compressed and transmitted again along with the subsequently generated blocks, as described in Section \[sec:retx\]. After $r$ failed transmission attempts, the block is considered outdated and discarded.
Model of energy dynamics {#sec:en_model}
========================
The energy dynamics of the devices strongly influence the system performance, and both data processing and transmission policies should dynamically adapt to the resource availability. Designing such an energy-aware framework requires to accurately model the energy inflow and consumption.
Energy consumption {#ssec:energy_cons}
------------------
In the following, we describe a parameterized model that tries to capture all the major sources of energy expenditure: communication, data acquisition, processing, and circuitry. We are not interested in the energy spent for packet reception because the data collector does not have energy constraints.
***Data processing.*** In [@Zordan-2014], the energy consumed by processing algorithms is evaluated by mapping the number and type of arithmetic instructions into the corresponding energy drain: $$\begin{aligned}
\label{eq:e_processing}
E_p(k) =
\begin{cases}
E_0 \, L_0 \, N_p(k) & \text{if } 1\le k \le m-1\\
0 & \text{if } k=0,m
\end{cases}
\end{aligned}$$ where $E_0$ is the energy consumption per CPU cycle (which depends on the unit), and $N_p(k)$ is the number of clock cycles required by the compression algorithm per uncompressed bit of the input signal (and depends on the compression ratio). If the packet is not compressed ($k=m$) or is discarded ($k=0$), no energy is consumed.
We assume that the devices employ the Lightweight Temporal Compression (LTC) algorithm, which is a widely used lightweight compression techniques for wireless sensors. In this case, the function $N_p(k)$ is increasing and concave in $k$ [@Zordan-2014]: $$N_p(k) = \alpha_p \frac{k}{m} + \beta_p, \quad 1 \le k \le m-1 ,$$ with $\alpha_p, \beta_p > 0$. Notice that the more compressed the packet, the less the energy spent. This seemingly counterintuitive fact is due to implementation details and we refer the reader to [@Zordan-2014] for explanation. Finally, we do not account for the contribution of channel encoding, because it is typically negligible [@Howard-2006].
***Transmission.*** The energy cost of a wireless transmission with power $P_\ttx$ for a period of length $T$ can be modeled as: $$E_{{\rm tx}} = \frac{T\, P_{{\rm tx}}}{\eta_{\rm A}}, \label{eq:e_tx}$$ where $\eta_{\rm A}\in (0,1]$ is a constant that models the efficiency of the antenna’s power amplifier.
***Sensing and circuitry.*** We assume that the periodical sensing drains a constant amount of energy $\beta_s$ in the window between two transmission slots. Also the energy required for switching between idle and active mode and maintaining synchronization with the receiver can be represented with a constant term $\beta_c$. Finally, we need to account for the additional energy spent by the circuitry during a transmission, $\mathcal{E}_c \,T$, where $\mathcal{E}_c$ is a circuitry power. Hence: $$\begin{aligned}
E_c(k) = \beta_s + \beta_c + \mathcal{E}_c \,T\cdot \chi_{\{k>0\}}, \label{eq:e_circuitry}\end{aligned}$$ where $\chi_{\{k>0\}}$ is the indicator function equal to 1 if $k\!>\!0$ and to zero otherwise (recall that the packet is dropped if $k=0$).
Energy harvesting and battery dynamics {#ssec:eh}
--------------------------------------
The sensor nodes are not connected to the energy grid, but are provided with some energy-scavenging circuitry and can collect energy from the environment. We assume that the energy supply is time-correlated (e.g., solar power); for this case, there exist models in the literature that have been built and validated against real data. In particular, both [@Ku] and [@solarstat] proved that a time-correlated energy supply can be accurately described through a stochastic Markov process, where each possible state entails a different distribution of the energy income. When dealing with energy harvesting, the model is generally discrete, i.e., the energy inflow is quantized.
Accordingly, we track the dynamics of the source through an $X$-state Markov Chain (MC): the source is in state $x\in\mathcal{X}=\{0,\dots, X-1\}$ and scavenges $e\in\{0,\dots,E\}$ quanta of energy from the environment, according to some probability mass function. Our framework is general and can accomodate any assumptions about the number of states of the harvesting process and the harvesting statistics in each state. However, to obtain some representative results, in this paper we consider a 2-state MC ($X=2$). In particular, $x=0$ represents a low energy state (e.g., night) during which no energy can be harvested ($e=0$); when $x=1$, the source is in a high energy state (e.g., day) and the energy income follows a truncated discrete normal distribution, $e \sim \mathcal{N}(\mu, \sigma^2)$ in the discrete interval $\{1,\dots\,E\}$ (analogous to [@zordan] and [@Ku]).
We assume that the battery of a node has a finite size $B$. The temporal evolution of the battery can be modeled as: $$\label{eq:battery}
b' = \min\left\{b + e - u, B\right\} \triangleq \left(b + e - u\right)^\dagger,$$ where $b$ and $b'$ respectively represent the current and next battery level, and $u$ is the energy used in the current slot, which depends on processing, transmission and circuitry, as given in Eqs. , and . Due to the energy causality principle, it is: $$\label{eq:en_causality}
u \le b.$$
An overly aggressive or conservative energy management could either deplete the battery ($b=0$) or fail to use the excess energy and waste it ($b=B$). These situations need to be prevented by designing a scheme that dynamically adapts to the randomness of the energy inflow, so as to ensure acceptable performance on a long-term horizon.
Energy-aware (re)transmission scheme {#sec:retx}
====================================
Here, we extend the joint source-channel coding scheme described in Section \[sec:comm\_model\] to account for packet retransmissions.
![The retransmission mechanism: encoding procedure.[]{data-label="fig:retx"}](retx.pdf){width="\columnwidth"}
The acknowledgment feedback allows the transmitter (i.e., the reporting device) to know whether a packet has been correctly received at its intended destination. When an outage occurs, the packet is lost and needs to be retransmitted. However, if a simple ARQ protocol is employed, the delay for sending newly generated data blocks will increase, depending on the number of the devices participating in the TDMA scheduling as well as on the number of allowed retransmissions, and can potentially become unbounded. Consequently, it may happen that the newly generated data blocks will become outdated, without having a chance to be transmitted. To prevent this, we consider a retransmission scheme in which nodes send their new data blocks together with the previously lost data blocks (if any) in the same time slot, as explained further.
Suppose that the transmission of a data block fails; in the successive time slot, the device will try to send the new block together with the previously lost one. Thus, the device has to compress and transmit two blocks of original size $L_0$ each within the $S$ channel uses available. If this transmission also fails, in the next slot the sensor node will process and transmit the new block and the two that were previously lost. In general, if the last $q$ transmissions failed, the sensor node has to process $q+1$ blocks of $L_0$ bits each and send them together. We assume that a maximum number $r$ of transmission attempts can be made for each piece of data, where the value of $r$ is dictated by the application, e.g., because of latency or QoS considerations.
Recall that data blocks are of a fixed size $L_0$ and are assumed to be independent of each other. Because of the independence assumption, when multiple data blocks are sent in the same time slot, the information they contain is not fused or processed jointly, but the compression is done separately for each of them. For the sake of fairness, we treat the data blocks in the same way, i.e., we apply to them the same compression ratios and obtain $q$ compressed blocks of size $L$, where $q = r + 1 $. These are then joined in a single block of size $q\,L$, which is then encoded, producing a packet of size $S$, and transmitted. Thus, the same distortion is introduced at the source for all data blocks placed in the same packet, but then they are treated as a single entity which is sent over the channel and subject to a certain outage probability that depends on the coding rate $R=q L/S$. Fig. \[fig:retx\] shows the encoding mechanism when a single block (on the left) or two blocks (on the right) are transmitted in the same slot.
![Structure of the MC that models the dynamics of the backlog state $q$ for $r=3$.[]{data-label="fig:queue"}](queue.pdf){width=".9\columnwidth"}
Fig. \[fig:queue\] shows the evolution of the data queue, which behaves like a success-runs MC. If the data queue in a certain slot has size $q$ and the transmitted packet is lost, the queue state becomes $q' = \min\{q+1, r\}$, otherwise the new queue state is $q'\!=\!1$. We denote as $\Delta_q(k)$ the distortion at the receiver when the queue state is $q$ and the chosen compression ratio is $k$. Such distortion depends on the reception outcome: in case of failure (NACK), the distortion at the receiver is set to $\Delta_q(k) = D_\fl$, even though the data blocks could be retransmitted; instead in case of successful transmission (ACK), we sum all the distortions of the $q$ blocks that have been compressed with the same ratio $k/m$, but we also subtract all the penalties that were obtained for the previous packet losses we accounted for, since the blocks were eventually successful. Thus, in this latter case, it is $\Delta_q(k) = \Dack(q,k)$ with: $$\Dack(q,k) = q\,D(k) - (q-1) D_\fl. \label{eq:dack}$$
We denote as $k_q^\star$ the optimal value of $k$ (i.e., the one that determines the optimal source-channel coding scheme to use) when $q$ packets share the same time slot; in Section \[ssec:RDP\] we explain how to determine it.
***Energy consumption.*** Suppose that the data queue state is $q$. Between two consecutive slots, the node consumes energy to process and transmit all $q$ data blocks. However, the energy required by transmission and circuitry does not depend on $q$, because the duration of the transmission is always $T$, i.e., $S$ bits, see Eqs. and . Hence, the energy consumed by the node when there are $q$ data blocks in the queue and it uses a compression ratio equal to $k/m$ is: $$\label{eq:energy_used}
q\,E_p(k) + E_\ttx\cdot\chi_{\{k>0\}} + E_c(k).$$ The optimization described in Section \[ssec:RDP\] determines the optimal source-coding scheme, i.e., the value $k_q^\star$, and the corresponding energy consumption is obtained by substituting $k=k^\star_q$ in Eq. .
Problem formulation {#sec:problem}
===================
Here, we mathematically define the objective of our optimization problem and describe the structure of the MDP, while the solution technique is explained in Section \[sec:solution\].
The optimization objective {#ssec:objective}
--------------------------
Our objective is to maintain for each node an energy-neutral operation mode while minimizing the long-term average distortion at the receiver, which depends on the outcomes of the transmissions, as explained in the previous section. If we consider $q$ data blocks of size $L_0$ that are compressed with the same compression ratio $k/m$ and encoded jointly at rate $R = qL(k)/S$, the expected distortion at the receiver is: $$\begin{aligned}
\mathbb{E}[\Delta_q(k)] = & \; qD(k)\,\left(1- \pr_\out\left(q\frac{L(k)}{S}\right)\right) \nonumber\\
+& \;q D_\fl\,\pr_\out\left(q\frac{L(k)}{S}\right), \label{eq:delta}\end{aligned}$$ where $\pr_\out(\cdot)$ is the outage probability as given in Eq. , $qD(k)$ and $q D_\fl$ are the distortions obtained when the packet is acknowledged or lost, respectively.
In other words, if the packet is successfully received, its distortion corresponds to that introduced at the source for all initial $q$ blocks of measurements, otherwise we account for the maximum distortion level for all packets, as if they had not even been sent. We are interested in the expected distortion at the receiver, thus we weight the two cases with their probabilities. Notice that the distortion at the source $D(k)$ decreases as $k$ increases, whereas the outage probability decreases for smaller coding ratios, i.e., as $k$ decreases. This implies a tradeoff between the distortion introduced by the lossy compression and the probability that the transmitted packet will be successfully received, through the choice of the value of $k$. To minimize $\mathbb{E}[\Delta_q(k)]$ and guarantee self-sufficiency of the network, it is necessary to (i) decide on $k$, and (ii) in each slot, allocate the energy consumption based on the current battery level, the dynamics of the energy source, and the energy consumption profile, in such a way to prevent energy outages (that disrupt the communication) and battery overflows (that waste energy). We formulate the problem by means of an MDP, where the actions correspond to the energy to use operations while the costs are represented by the expected distortion at the receiver. By doing so, the energy self-sufficiency of the node is ensured and the QoS is optimized.
The Markov Decision Process {#ssec:MDP}
---------------------------
The MDP is defined by the tuple $\big(\mathcal{S},\,\mathcal{U},\,P,\,c(\cdot)\big)$, where $\mathcal{S}$ denotes the system state space, $\mathcal{U}$ is the action set space, $P$ is the set of transition probabilities of the system state space and $c(\cdot)$ is the associated cost function for taking an action.
***System state space*** $\mathcal{S} \triangleq \mathcal{X} \times \mathcal{B} \times \mathcal{Q}$, where $\mathcal{X}=\{0,1\}$ represents the set of energy source states, $\mathcal{B}=\{0,\dots,B\}$ the set of energy buffer states, and $\mathcal{Q}=\{1,\dots,r\}$ the set of backlog states, i.e., the number of waiting packets. Notice that we need to keep track of the data queue size, unlike in our previous work [@Pielli-2017] where it was always one.
***Action set space*** $\mathcal{U} \triangleq \{0,\dots, B\}$. In each slot, the device observes the current system state $s\in\mathcal{S}$ and decides how much energy $u \in \mathcal{U}_{s} \subseteq \mathcal{U}$ to use to process and transmit the data it collected. In accordance with , this quantity cannot exceed the battery level, i.e., $\mathcal{U}_{s} = \{0,\dots,b\}$. ***Transition probabilities*** $P$ govern the system dynamics. The probability of going from state $s = (x,\, b,\, q)$ to $s' = (x',\, b',\, q')$ with action $u$ is: $$\begin{split}
\prr(s'|s, u)= &\; \; \: p_x (x'|x) \cdot p_e (e|x)\cdot p_q (q'|q,u)\\
& \cdot \, \delta \left(b' - \left(b + e - u \right)^\dagger \right)
\end{split}$$ where $p_x (x'|x)$ is obtained from the transition probability matrix of the MC that models the source state, $p_e (e|x)$ is the mass distribution function of the energy inflow in state $x$ (see Section \[ssec:eh\]), and $p_q (q'|q,u)$ represents the probability that the backlog size goes from $q$ to $q'$ when action $u$ is taken. The last term $\delta(\cdot)$ is equal to $1$ if its argument is zero, and zero otherwise, and ensures that the transitions between states are consistent with the dynamics of the battery level, see Eq. .
***Cost function*** $c(\cdot)$. When the sensor node is in state $s=(x,b,q)$ and selects action $u$, it implicitly decides upon the source-channel coding scheme that minimizes $\mathbb{E}[\Delta_q(k)]$, i.e., it decides the optimal number $k^\star_{q}$ of information bits to send over the channel per data block, given the available energy $u$. The transition from state $s$ to state $s'$ when action $u$ is taken entails a cost: $$\label{eq:cost}
c(s, u, s') =
\begin{cases}
\Dack(q, k^\star_{q}) \; &\text{if } q' = 1\\
D_\fl \; &\text{otherwise}
\end{cases}$$ and therefore the cost of choosing action $u$ in state $s$ is: $$\label{eq:tot_cost}
\tilde{c}(s, u) = \sum\limits_{s'\in\mathcal{S}} \Pr(s'|s, u) \,c(s, u, s'). $$ Notice that the cost only depends on the data queue component $q$ of the state, while the battery level $b$ affects the set of admissible actions.
When retransmissions are not allowed ($r=1$), the backlog state is always $q\!=\!q'\!=\!1$, whatever the outcome of the transmission is. In this case, the MC of Fig. \[fig:queue\] reduces to a single state, and Eq. is no longer meaningful, thus when $r=1$ we set $c(s, u, s') \equiv \tilde{c}(s, u) \equiv \mathbb{E}[\Delta_1(k^\star_1)]$.
Optimal policy {#sec:solution}
==============
The optimal policy is the one that, based on the statistics of the energy harvesting process and the current battery level, decides how much energy to use in order to guarantee the lowest average distortion at the receiver. We first discuss how to determine the optimal source-channel coding scheme, hence $k^\star_q$ for each possible backlog state $q\le r$, and then describe how to solve the MDP optimally.
Rate-distortion tradeoff {#ssec:RDP}
------------------------
In [@Pielli-2017], we proved that for the case $q=1$ the expected distortion at the receiver given in Eq. exhibits a unique point of minimum $k_{1,R}^\star$ when $D(k)$ and $P_\out(k)$ are characterized as in Eqs. and , respectively. An example of this is shown in Fig. \[fig:k\_r\]. This is still valid when retransmissions are introduced, as can be inferred by looking at Eq. . The $q$ packets are compressed separately but with the same compression ratio, which cannot exceed $m/q$, hence $D(k)$ is truncated for $q>1$. Instead, the outage probability maintains the same shape since the packets are encoded together. Hence, there exists an optimal point $k_{q,R}^\star$, whose value depends on the number of packets $q$ that are sent in the same time slot and that minimizes the expected distortion at the receiver. If the device uses $k < k_{q,R}^\star$, the packet will go through the channel with a higher probability but its distortion will be larger; on the other hand, if $k > k_{q,R}^\star$, the distortion will be smaller, but it is more likely that the packet will be lost.
If the amount of energy allocated $u$ allows it, the device will choose the optimal coding scheme corresponding to $k^\star_{q,R}$, otherwise it simply selects the maximum possible $k$ dictated by the energy constraint, because $\mathbb{E}[\Delta_q(k)]$ is decreasing if $k\le k_{q,R}^\star$, which is due to how $k_{q,R}^\star$ is defined. The energy consumption, see Eq. , is non-increasing in $k$, and we denote as $k_{q,E}^\star(u)$ the largest value of $k$ that solves $q\,E_p(k) + E_\ttx\cdot\chi_{\{k>0\}} + E_c(k) \le u$ for a given $q$, see Eq. . Then, when the backlog state is $q$, the device will choose the source-channel coding scheme corresponding to: $$\label{eq:k_star}
k^\star_{q} = \min\{k_{q,R}^\star, k_{q,E}^\star(u)\}.$$ Such value clearly depends on the energy $u$ that the node decides to employ, but we omit denoting this explicitly in favor of a lighter notation. In [@Pielli-2017] we showed that the expected distortion at the receiver is convex for $k\le k_{q,R}^\star$ and, consequently, is a convex non-increasing function of $u$.
The choice of how much energy to use when in state $s=(x,b,q)$ uniquely determines the joint source-channel coding scheme (i.e., the number of transmitted information bits per data block) that leads to the smallest expected distortion at the receiver. For further details, we refer the reader to [@Pielli-2017].
Solving the MDP {#ssec:RVI}
---------------
When the sensor device is in a certain state, it selects the action to take according to a policy $\pi:\mathcal{S}\rightarrow \mathcal{U}$. The corresponding long-term average cost is: $$\label{eq:J}
J^\pi(s) = \lim\limits_{M\rightarrow +\infty} \frac{1}{M} \mathbb{E}_s\left[ \sum\limits_{m=0}^{M-1} \tilde{c}(s_m, u_m) \middle\rvert s_0 = s \right],$$ where the initial state $s_0$ is given. Notice that each decision affects all subsequent decisions.
We want to determine the optimal policy $\pi^\star$, i.e., the set of rules that maps each system state into the optimal action with respect to the average cost criterion. The MDP defined in Section \[ssec:MDP\] has unichain structure and bounded costs, implying that Eq. does not depend on the initial state: $$J^\pi(s)\equiv J^\pi, \quad \forall s\in\mathcal{S}.$$ Hence, we can restrict our search to Markov policies only [@altman].
To determine $\pi^\star$, we used the Relative Value Iteration Algorithm (RVIA), which is a variant of the well-known Value Iteration algorithm for average long-term problems and provably converges [@bertsekas]. To understand why it works, we first define the $n$-step value-function by induction as: $$\label{eq:v}
v_{n}(s) = \min_{u\in\mathcal{U}_{s}} \left\{ \tilde{c}(s,u) + \sum\limits_{s'\in\mathcal{S}}\,\Pr(s'|s,u)v_{n-1}(s') \right\},$$ where $v_0(s)$ is arbitrarily defined, e.g., $v_0(\cdot)=0$. Function $v_n(s)$ represents the minimum expected $n$-step cost that can be achieved from an initial state $s$, because it sums the immediate cost $\tilde{c}(s,u)$ obtained in the initial state with the expected optimal cost obtained in the $n-1$ subsequent slots (through $v_{n-1}(\cdot)$ and then recursively). Based on this, the optimal policy $\pi^\star$ is such that $J^{\pi^\star}(s) = \lim_{n\rightarrow\infty} v_n(s)/n$, and as $n$ grows the dependence on $s$ fades ($J^{\pi^\star}(s) \equiv J^{\pi^\star}$).
RVIA leverages on this and defines two functions $J$ and $Q$ that are alternatively updated starting from an initial estimate $J_0(\cdot)$ until convergence: $$\begin{aligned}
& Q_{j}(s,u) = \tilde{c}(s,u) + \sum\limits_{s^\prime\in\mathcal{S}} \prr(s^\prime|s, u)\,J_{j-1}(s^\prime) \label{eq:Q}\\
& J_{j}(s) = \min_{u\in \mathcal{U}_s} Q_{j}(s,u), \label{eq:J_Q}
\end{aligned}$$ with $j$ being the iteration index. The convergence criterion is chosen as the span seminorm operator $sp(w)\triangleq \max(w)-\min(w)$, because it guarantees that is a contraction mapping [@bertsekas]. RVIA is stopped at iteration $l$, when $sp(J_{l+1}(s) - J_l(s)) \le t$ for a chosen threshold $t$. The optimal policy dictates the action $u^\star(s)$ to take in each state $s$: $$u^\star(s) = \argmin_{u\in \mathcal{U}_s} Q_{l}(s,u),$$ and entails the following average long-term cost: $$C^\star =\sum\limits_{s\in\mathcal{S}} \rho_s\, \tilde{c}(s,u^\star(s)) \label{eq:final_cost}$$ where $\rho_s$ is the steady state probability of state $s$ induced by the optimal policy (the MDP reduces to a MC when the actions to take in each state are deterministic).
The effects of retransmitting {#ssec:retx_eff}
-----------------------------
When a transmission slot is used for multiple packets combined together, these packets will be more compressed and have a larger distortion than when a slot is reserved to a single packet. However, the convexity of $D(\cdot)$ suggests that combining packets improves the average long-term distortion. This is formalized in the following result:
\[th:retx\] For any value $r$ of the maximum number of transmission attempts that can be dedicated to each packet such that $k_{r,R}^\star > 0$, when the energy constraints are neglected, i.e., $k_{r}^\star = k_{r,R}^\star$ and $k_{1}^\star = k_{1,R}^\star$, the retransmission mechanism described in Section \[sec:retx\] achieves a lower average long-term distortion compared to the base case where packets can be sent only once.
See Appendix \[apx:proof\_retx\].
Notice that, if $k^\star_r = 0$, there would clearly be no gain in using the retransmission scheme as packets are simply discarded; this is unlikely to happen, especially if $r<m$.
Theorem \[th:retx\] proves that the combined retransmission scheme leads to enhanced performance in terms of QoS in the absence of energy constraints. We now discuss what happens when these constraints come into play. The maximum compression ratio allowed by the allocated energy is $k_E^\star(u)/m$, which depends on the energy allocation $u$ and on the backlog state $q$ through Eq. . Unless $k_q^\star=0$, the only contribution to the energy consumption that depends on $q$ is the energy due to processing. Since $q E_P(k)$ increases with $q$, see Eq. , it is $k_{q,E}^\star(u) \le k_{1,E}^\star(u)$, i.e., given the amount of energy available, fitting more packets into a single slot is more expensive than processing a single packet. The actual relation between $k_q^\star$ and $k_1^\star$ is not clear, because it highly depends on the energy the node allocates. If $k_{q,E}^\star(u) \ge k_{q,R}^\star$, then $k^\star_q = k_{q, R}^\star$ and the improvement stated by Theorem \[th:retx\] holds (whatever $k_1^\star$ is). If instead $k_{q,E}^\star(u) < k_{q,R}^\star$, i.e., $k^\star_q = k_{q, E}^\star$, there is a simple relation between the average distortion obtained with and without the retransmission scheme and, according to the actual values of $k^\star_q$ and $k^\star_1$, it may be better to use the retransmission scheme or the single-transmission one[^2]. Anyway, $k_{q,E}^\star(u)$ depends on the energy that the node decides to use in the considered time slot, and it is up to the MDP to manage it in such a way to obtain the lowest average distortion which, as shown in Theorem \[th:retx\], is smaller when the retransmission scheme is adopted. In practice, the retransmission mechanism coupled with an intelligent energy management scheme achieves a better QoS, i.e., lower distortion.
As stated in Theorem \[th:retx\], the retransmission scheme brings enhancements with respect to the single transmission one. However, this improvement is not proportional to the value of $r$, but rather depends on the particular shapes of the distortion and outage probability functions, and on how large $r$ is. In particular, this is formalized as follows:
\[th:retx\_2\] Consider an integer $r>2$. The long-term average distortion achieved when using the retransmission scheme with $r$ may be lower than that obtained with $r-1$.
See Appendix \[apx:proof\_retx\_2\].
In the proof, we showed that the probability of having worse performance with $r$ rather than $r-1$ increases as $r$ becomes larger. Hence, $r$ should be kept reasonably small. This is an intuitive result, since increasing $r$ implies reserving fewer transmitted bits per data block in case of retransmission, thereby introducing a larger compression and, consequently, distortion at the source. A deeper understanding of how the average distortion changes as $r$ increases is given in Appendix \[apx:proof\_retx\_2\].
There are two additional factors that need to be taken into account in the design of the retransmission scheme.
- *Latency*: the information contained in a packet may lose significance as latency increases. In this case, sending a data block long after it has been originated may even be disadvantageous: receiving it brings no benefit to the final application, and, at the same time, reduces the quality of the other data blocks that are transmitted along with it, as it occupies part of the bits available for the transmission.
- *QoS*: the final application may dictate a minimum QoS threshold, i.e., a maximum distortion that can be tolerated on the received information. When data blocks are transmitted together, they are compressed more and have a larger distortion, which may violate the QoS constraints.
Thus, the choice of $r$ should be guided by the specific application constraints and by the values of $k_r^\star$ and $D(k^\star_r)$.
Numerical evaluation {#sec:results}
====================
In the following, we show how the average long-term cost $C^\star$ is affected by the system parameters and compare the performance obtained with the single-transmission scheme (i.e., $r=1$), and the retransmission scheme with $r=2$. We also evaluate the average distortion obtained when an energy-unaware greedy scheme is adopted. This *greedy* policy is myopic and does not optimize the energy consumption according to the expected future availability; i.e., when in state $(x,b,q)$, the node uses all the energy it has in the battery, unless it is more than $u_q^\star$, i.e., the energy needed to achieve $k_{q,R}^\star$. Hence, $u= \min(b, u_q^\star)$.
We investigated the role of the system parameters by running RVIA for the chosen system configurations. In all cases, the original packet size is $L_0=500$ bits and the nodes can decide among $m=30$ different compression ratios. The parameters of the distortion curve of Eq. have been derived from [@Zordan-2017]; in particular, we set $a=0.35$ and $b=19.9$. Notice that with these choices of $a,b$ and $m$, the granularity of the compression ratio (i.e., $1/m$) is such that $D(1) < D_\fl$, which guarantees that the distortion function is decreasing.
The considered transmission power is $P_\ttx=25$ mW, the transmission frequency is $f_0=868.3$ MHz, the used bandwidth is $W = 125$ kHz, and the overall noise power spectral density $N_0=-167$ dBm/Hz. The path loss exponent is $3.5$ and Rayleigh fading is modeled as an exponential random variable with unit mean. The battery dynamics of Eq. assumes that the energy is quantized. Consistently, all the terms of energy consumption, i.e., Eqs. , , , have been mapped into quanta (we ensured an appropriate granularity for this purpose). In our numerical evaluation, the energy due to processing (see also [@Zordan-2014]) is of the same order of magnitude of that needed for a transmission, whereas the circuitry contribution is smaller. We remark that, according to Eq. , a packet cannot be sent if the available energy is below a certain threshold. We also denote as $e_{\max}$ the maximum energy consumption demanded by the processing and transmission of a (single) packet, and introduce the normalized quantities $\overline{\mu} = \mu/e_{\max}$, and $\overline{B}=B/e_{\max}$.
In the energy harvesting process, the probability that the source goes from the bad to the good state is 3 times greater than that of the opposite transition. In the bad state ($x=0$), $e=0$ quanta of energy arrive with probability 1, otherwise the energy inflow follows a Gaussian distribution with mean $\mu$ and variance 10.
Fig. \[fig:dist\] shows the average long-term cost as a function of the distance from the receiver when $\bar{\mu}=1$ and for two different values of the battery size, namely $\bar{B}=0.6$ and $\bar{B}=0.8$. Clearly, when the node is farther from its receiver, the path loss component increases, thereby leading to a larger outage probability. To guarantee that its measurements do not get lost in the transmission, the node will use a stronger coding rate, thus the compression ratio needs to be smaller and a larger distortion is introduced at the source. In both cases, the battery size is quite small and such that the node’s readings generally have to be compressed because the energy stored is not enough to send them as they are. This clearly impinges on the distortion that can be achieved on average. What is interesting to note is the behavior of the retransmission policy when $\bar{B}=0.6$ (Fig. \[fig:dist\], top). As discussed in Section \[ssec:retx\_eff\], the energy required to send two or more packets together is larger than that needed for a single one, since it requires to spend more energy for processing. In this case, the battery size is very small, and the node cannot store enough energy to transmit two packets together. In the end, the node will decide to almost always discard the packets, or compress them a lot to be more robust against failures. This result corroborates the impact of the energy availability on the performance. Also, because of the small battery size, the optimal single-transmission policy performs similarly to the greedy one. When instead $B$ is large enough, the retransmission mechanism leads to improvements with respect to the single-transmission one, as depicted in Fig. \[fig:dist\], bottom.
Fig. \[fig:mu\] shows the performance as a function of the average energy income during the good source state, $\mu$. The distance of the node is fixed to $100$ m for the figures on the left, and $200$ m for those on the right. The normalized battery size is $\bar{B} = 0.6, 0.8,$ and $2$ for the figures on the top, middle, and bottom, respectively. Intuitively, the more the energy that can be harvested (large $\mu$), the lower the average distortion, because the node can choose the optimal point in the rate distortion tradeoff (see Sec. \[ssec:RDP\]) more often. However, after a certain value of $\mu$, the distortion curve tends to remain constant because (i) the battery size is too small and part of the incoming energy needs to be discarded, and (ii) the optimal $k^\star$ is already achievable and the exceess energy is not useful. As already seen in Fig. \[fig:dist\], if the battery size is too small compared to the minimum energy required to send $q=1,\dots,r$ packets, the node will never transmit them, and this effect sharpens with the distance. Fig. \[fig:dist\] also shows that, when the node is closer to the receiver and there is enough energy available (second and third plots on the left), the single-transmission optimal policy behaves similarly to the retransmission one, as the outage probability is low; otherwise, its performance becomes closer to that of the greedy policy and the improvement obtained with the retransmission scheme becomes more significant (see plots on the right).
The effect of the maximum number of retransmissions on the distortion as a function of distance is shown in Fig. \[fig:r\] for $\bar{B}=1$ and $\bar{\mu}=1$. As $r$ increases, the gain obtained becomes smaller: if many packets are sent together in a single slot, the distortion introduced by the lossy compression is large, and this reduces the benefit of retransmitting. If $r$ becomes too large, the number of channel uses available for each packet is so small that the compression ratio that is required is too large and this affects the dynamics of the MDP and thus of the system performance.
Finally, Fig. \[fig:discarded\] shows the packet loss probability as a function of $\mu$ for different values of $r$ and $\bar{B}=1$, where this probability accounts for both the packets lost in the communication and those discarded at the source. Again, it can be observed how the benefits of the retransmission scheme fade as $r$ increases. The gain obtained over the greedy policy by the Value Iteration algorithm is more significant when energy is scarce (small $\mu$), for both the single-transmission scheme ($r=1$) and the retransmission mechanism ($r>1$).
A learning approach {#sec:learning}
===================
The policies found with the RVIA are optimal and have a threshold structure, cf. [@Pielli-2017], which implies limited storage requirements and low implementation complexity. However, this approach requires the knowledge of the statistics of the MDP, which means that the dynamics of the EH process and the average channel gain need to be known, but this information may not be available with enough accuracy. A natural way to determine the optimal policy when the knowledge about the environment is uncertain or limited is to interact with it and learn based on the feedback received. This approach is commonly referred to as *reinforcement learning* [@Sutton-98] and is usually applied to scenarios where the environment is modeled through an MDP, which is the case in this paper. In particular, in this paper we exploit the R-learning algorithm, which is the counterpart of the well-known Q-learning algorithm for the average case rather than for discounted optimization problems.
R-learning {#ssec:R-learning}
----------
R-learning is an RL algorithm introduced by Schwartz in 1993 [@schwartz-1993] for average-reward maximization problems. Notice that it can be promptly adapted to our model by using the negatives of the costs. The basic gist of R-learning is to learn the value of each admissible action for all possible states through exploration. As discussed in Section \[ssec:RVI\], in the long run the average cost, see Eq. , is the same regardless of the initial state. However, there is a transient, and it is this transient that defines the value of a state-action pair: $$q^\pi(s,u) = \sum\limits_{k=1}^\infty \mathbb{E}\left[R_{n+k} - J^\pi \middle | s_n=s,u_n=u \right],$$ where $R_n = -c(s_n, u_n, s_{n+1})$ is the reward obtained in slot $n$; notice that the next state $s_{n+1}$ depends on action $u_n$ and on the environment. The function $q^\pi$ represents the value relative to the average reward under the current policy. The essential approach of RL algorithms is *exploration vs exploitation*: in a state, an agent can choose the action that leads to the highest reward based on current information ([exploitation]{}), or keep trying new actions, hoping that they bring even higher rewards ([exploration]{}).
Similarly to Q-learning, R-learning maintains (i) a behavior policy that simulates experience and dictates the action to choose, (ii) an estimation policy $\pi$ which is the one involved in the policy iteration process and is the policy that is being learned, (iii) an action-value function $Q^L$ that approximates $q^\pi$, and (iv) an estimated average reward $\rho$ that approximates $J^\pi$. Notice that, in order to learn the estimation policy $\pi$, the device has to use an additional policy (the behavior one), that dictates what to do during the learning process. We chose to use an $\varepsilon$-greedy policy as behavior policy: when in state $s$, the node chooses a random action $u\in\mathcal{U}_s$ with probability $\varepsilon$, otherwise it selects the action with the highest $Q^L$ value, i.e., the current best action.
Initialize $\rho$ and $Q^L(s,u) \:\forall s,u$ $s \gets$ current state Choose action $u\in\mathcal{U}_s$ using $\varepsilon$-greedy policy Observe next state $s'$ and reward $R$ $\delta \gets R - \rho + \max_w Q^L(s',w) - Q^L(s,u)$ $Q^L(s,u) \gets Q^L(s,u) + \alpha \delta$ $\rho \gets \rho + \beta \delta$
The complete algorithm is given in Algorithm \[alg:R-learning\] [@Sutton-98]. The scalars $\alpha$ and $\beta$ are step-size hyperparameters. During the learning process, the node continuously updates the $Q^L$ values of all state-action pairs and, at each iteration, chooses either the action corresponding to the highest $Q^L$ value or a random one, according to the $\varepsilon$-greedy policy. After some iterations, the result of $\max_{u\in\mathcal{U}_s} Q^L(s,u)$ will converge and the node learns that the best action to take in state $s$ is: $$u^L(s) = \max_{u\in\mathcal{U}_s} Q^L(s,u). \label{eq:u_L}$$ Hence the policy learned by the device coincides with the one, and in each state $s\in \mathcal{S}$ the action to take is deterministic and given by .
The convergence of the learning algorithm is demonstrated in Fig. \[fig:learning\], which shows the average long-term cost of Eq. as the node is learning the policy. The simulation parameters in this case are $r=2$, $\bar{B}=0.8, \bar{\mu} = 0.7$, and $d=200$ m. Fig. \[fig:learning\] shows the average long-term cost obtained with the policy learned by the learning algorithm vs the number of iterations of the learning phase. The blue line represents the distortion obtained with the optimal policy of Section \[sec:solution\]. The hyperparameters of the learning algorithm were optimized by performing an exhaustive search and selecting the combinations that performed best in the training phase. Both the exploration rate $\varepsilon$ and the learning rate $\alpha$ were chosen to be decreasing during the learning process. It is a reasonable and common practice that, as the learning proceeds and the $Q^L$-values converge, $\varepsilon$ decreases. This happens because initially the device has no knowledge about the environment and makes random moves to maximally explore the state space, so that all possible states can eventually be visited, and the long-term $Q^L$-value of every state-action pair can be determined. Then, it converges to a small exploration rate and exploits the accumulated knowledge [@Sutton-98].
Conclusions {#sec:conclu}
===========
In this paper, we proposed and analyzed a joint retransmission, compression and channel-coding scheme for energy-constrained sensors that access the channel in a TDMA fashion. The goal is to enable the sensors to be energetically self-sufficient and, at the same time, to minimize the long-term average distortion of the data they report to the receiver. The proposed scheme foresees that the data blocks that were not successfully reported are retransmitted along with the newly generated blocks, where the compression is performed separately for all blocks, while the channel-coding procedure is performed for all blocks jointly. The problem has been formulated by means of an MDP, and the optimal policy was derived with a variant of the Value Iteration algorithm. We also proposed the use of a learning approach to determine the policy during the operation of the system.
Both the analytical investigation and the numeral evaluation showed that the retransmission scheme ensures an improved average quality of the received information with respect to the simpler single-transmission scheme, unless the energy available to the node is very scarce (because either the battery size is too small or the energy inflow is insufficient). In this case, the retransmission scheme demands too much energy and the node cannot afford it. When the energy is scarce, a more flexible scheme that decides whether or not to retransmit a packet may improve the performance. For instance, if a packet is lost and the battery charge is low, then the node should opt for transmitting only the new readings and give up on those that were lost, so as to preserve some energy. The gain introduced by the retransmission scheme becomes more significant as the probability of outage increases. As the maximum number of transmission attempts that can be dedicated to a packet increases, the performance improvement tends to fade, while also affecting latency. We also validated the effectiveness of machine learning algorithms, which can be extremely powerful when the information available to the node is imprecise or incomplete, and achieve a performance very close to that of the optimal policies.
In this work, it was assumed that a node can retransmit its lost packets only in the time slots dedicated to the sensor device. An interesting alternative consists in considering some shared time slots, where the devices can send their lost data in a random fashion, i.e., contending for the channel. A study of this scheme is part of our future work.
In our further work, we also intend to adopt a realistic model for the battery consumption, which in practice is not linear, unlike commonly assumed, but depends on the current battery level. Another extension is the investigation of how the actual performance changes when the knowledge about the average channel state is imperfect. Finally, it would be interesting to generalize the model to other compression techniques, possibly with different energy consumption characteristics.
Acknowledgments {#acknowledgments .unnumbered}
===============
The work of C. Pielli and M. Zorzi was partly supported by Intel’s Corporate Research Council, under the project “EC-CENTRIC: an energy- and context-centric optimization framework for IoT nodes”. The work of P. Popovski and C. Stefanovic was in part supported by the European Research Council (ERC Consolidator Grant Nr. 648382 WILLOW) within the Horizon 2020 Program.
Proof of Theorem \[th:retx\] {#apx:proof_retx}
============================
We compare the case in which a packet that is not received cannot be retransmitted ($r=1)$ with that in which a data block that was lost can be combined together with the next data blocks and retransmitted for at most $r>1$ times.
The proof leverages on the fact that the distortion function is convex and decreasing. We will denote $D_\fl$ as $D(0)$. By assumption, we neglect the energy constraints, i.e., $k_r^\star = k_{r,R}^\star$ and $k_1^\star = k_{1,R}^\star$, see Eq. . **Step 1.** First of all, we prove that $k_{r}^\star \ge k_{1}^\star/r$.
When $r>1$ and $r-1$ consecutive transmissions failed, the lost $r-1$ data blocks are retransmitted in the successive time slot along with the most recently generated data block. In this case, the device selects $k_r^\star \in [0,m/r]$ as the optimal point in the rate-distortion tradeoff. According to our assumption, $k_r^\star =k_{r,R}^\star$, which is the point of minimum of function $\mathbb{E}[\Delta_r(k)]$, see Eq. . The outage probability depends on the coding rate $R=rL(k)/S$ and scales with $r$, i.e., it maintains the same shape regardless of $r$ because $k$ cannot exceed $m/r$. The outage probability is represented with a dashed line in Fig. \[fig:proof\_dist\] for $k\in[0,m/r]$ with $r=2$. If the distortion introduced at the source also scaled with $r$, it would be $k_{r}^\star = k_1^\star/r$; instead, it only depends on $L(k)$ and is truncated to $D(m/r)$. Since $D(k)$ is a decreasing function of the compression ratio, the effective distortion in the interval $[0, m/r]$ (light dotted line in Fig. \[fig:proof\_dist\]) is larger than the one that would be obtained by scaling the original distortion in $[0, m]$ to the new range (dark dotted line in Fig. \[fig:proof\_dist\]). This means that the point of minimum of $\mathbb{E}[\Delta_r(k)]$ is shifted right with respect to that obtained when simply scaling both the outage probability and the distortion functions from $[0,m]$ to $[0,m/r]$. Then: $$\label{eq:k_r}
k_{r}^\star \ge k_{1}^\star/r. $$ An example of this can be seen in Fig. \[fig:proof\_dist\]: the light and dark flat curves show $\mathbb{E}[\Delta_2(k)]$ and the scaled version of $\mathbb{E}[\Delta_1(k)]$, respectively, and the markers identify their minima.
**Step 2.** We show that $(r-1)\,D(0) + D(k_1^\star) \ge r\,D(k_r^\star)$.
As a first step, focus on $r=2$. The convexity of $D(k)$ implies that: $$\begin{aligned}
D(x_1) &+ D(x_2) \ge \\
& D\left(\lambda x_1 + (1-\lambda) x_2\right) + D\left((1-\lambda) x_1 + \lambda x_2\right), \label{eq:convexity}
\end{aligned}$$ for any $\lambda\in[0,1]$. In particular, given that $D(k)$ is a decreasing function of $k$, for $x_1 < x_2$ and $\lambda \ge 1/2$, we also have: $$\begin{aligned}
D(x_1) + D(x_2) & \ge \\
& 2 D\left(\lambda x_1 + (1-\lambda) x_2\right) \triangleq 2 D\left(x_3\right),
\end{aligned}$$ This implies that, if we choose $x_1 = 0$ and $x_2 = k_1^\star$, the previous inequality is satisfied for all points $x_3 \ge k_1^\star/2$, and in particular for $k_2^\star$.
This can be extended to a generic value of $r$ by repeatedly applying the convexity property of Eq. for $r-1$ times and choosing $\lambda = 1/r, 1/(r-1), \dots$ until $\lambda\!=\!1/2$ in the last step. This process leads to: $$(r-1) D(x_1) + D(x_2) \ge r D\left(\frac{r-1}{r}x_1 + \frac{1}{r} x_2\right).$$ If we choose $x_1 = 0$ and $x_2 = k_1^\star$, and exploit Eq. , we have that $(r-1)\,D(0) + D(k_1^\star) \ge r\,D(k_r^\star)$.
**Step 3.** Finally, we show that the retransmission scheme leads to lower average distortion.
After $r$ consecutive transmissions, the data generated in the first slot has been either successfully sent or discarded, and the window of the packets involved in the processing-transmission mechanism moves forward. This means that focusing on $r$ transmissions suffices to encompass all possible situations. There can be $2^r$ possible outcomes, each with a corresponding QoS cost. The costs obtained in the two scenarios (retransmissions or single-transmission) differ only when a successful transmission follows $q < r$ consecutive failed transmissions: in this case, if no retransmission mechanism is considered, the total cost is given by $q D(0) + D(k_1^\star)$, otherwise it is $(q+1) D(k_{q+1}^\star)$ (see Section \[sec:retx\]). We focused on the case $q=r-1$ and proved that $r\,D(k_r^\star) \ge (r\!-\!1)\,D(0) + D(k_1^\star)$. All other cases can be straightforwardly traced back to some $r' < r$ and the same result holds. In practice, in all possible combinations of ACK/NACK in $r$ transmission attempts, the distortion obtained when the retransmission mechanism is adopted is not larger than that achieved in the single-transmission scenario. This completes the proof of the theorem.
Proof of Theorem \[th:retx\_2\] {#apx:proof_retx_2}
===============================
The distortion function $D(k)$ is convex and decreasing in $k\in\{0,\dots,m\}$. We want to compare the performance obtained by the retransmission scheme with $r>2$ (case $A$) and with $r-1$ (case $B$) maximum number of retransmissions allowed.
As in the proof of Theorem \[th:retx\], we neglect any energy constraint, so $k_r^\star = k_{r,R}^\star$ and $k_{r-1}^\star = k_{r-1,R}^\star$. If we consider $r$ consecutive transmissions, the only difference between the two cases is obtained when a successful transmission follows $r-1$ failed attempts. The corresponding distortions are, respectively: $$\begin{aligned}
D^A & = r D(k^\star_r) \\
D^B & = (r-1)D(k^\star_{r-1}) + D(0).
\end{aligned}$$ Since $0\le k^\star_{r} \le k^\star_{r-1}$, it is $D(0) \ge D(k^\star_{r}) \ge D(k^\star_{r-1})$. We want to show that $D^A$ is not always lower than $D^B$, but under some circumstances it can be higher.
If $k^\star_{r-1} > k^\star_r = 0$, then $D^A = r D(0) > (r-1) D(k^\star_{r-1}) + D(0) = D^B$. Vice versa, if $k^\star_{r-1} = k^\star_r > 0$, then $D^A < D^B$. This is more likely to happen when $r$ gets closer to $m$, which represents the maximum value that $k$ can take. And, if $r>m$, it surely is $k^\star_q = k^\star_{q-1}$ for some $q\le r$.
This suffices to prove the theorem, but in addition we want to understand what happens if $0 < k^\star_{r} < k^\star_{r-1}$. We introduce the coefficient $\mu \triangleq (r-1)/r$, so that we compare $D_A/r = D(k^\star_r)$ and $D^B/r = \mu D(k^\star_{r-1}) + (1-\mu)D(0)$. Let $k_V$ be the point such that $D(k_V) = D^B/r$, as represented in Fig. \[fig:proof\_2\]. By applying the definition of distortion given in Eq. , we have: $$\label{eq:k_V}
k_V = \left(\mu (k^\star_{r-1})^{-a} + (1-\mu) \frac{b+D_\fl}{b\,m^a} \right)^{-1/a}.$$
The distortion is a decreasing function, hence $D(k)>D(k_V) = D^B/r$ for all $k < k_V$. This implies that: $$\begin{cases}
D^A > D^B &\quad \textrm{if } k^\star_r < k_V \\
D^A = D^B &\quad \textrm{if } k^\star_r = k_V \\
D^A < D^B &\quad \textrm{if } k^\star_r > k_V
\end{cases}$$
Determining when $k^\star_r < k_V$ is a challenging task, hampered by the fact that there is no closed-form expression for the relationship between $r$ and $k^\star_r$, see Section \[ssec:RDP\] and [@Pielli-2017]. However, as $r$ increases, there is a higher probability that $D^A$ is larger than $D^B$. In particular, let $\mathcal{K}^- \triangleq \{k\in\mathbb{N}: 0<k<k_V\}$ and $\mathcal{K} \triangleq \{k\in\mathbb{N}: 0<k<k^\star_{r-1}\}$. As $r$ increases, $\mu = (r-1)/r \rightarrow 1$, and $k_V$ gets closer to $k^\star_{r-1}$ (from the left). Hence, the ratio $|\mathcal{K^-}|/|\mathcal{K}| \rightarrow 1$ and the probability that $k^\star_{r}$ falls in $\mathcal{K^-}$ rather than in $\mathcal{K}\setminus\mathcal{K^-}$ increases.
In practice, for small values of $r$, in general it holds that $k_V<k^\star_r < k^\star_{r-1}$, and increasing $r$ leads to lower distortion at the receiver. But, as $r$ increases, $k_V$ gets very close to $k^\star_{r-1}$ and it is more unlikely that $k^\star_r$ falls in between these two values, thus increasing $r$ brings no benefit.
A curious behavior appears as $r$ gets closer to $m$, because the probability that $k_r=k_{r-1}$ becomes higher. In this case, it may happen that (i) $k^\star_r < k_V$ and therefore it is better to select $r-1$ rather than $r$, but (ii) $k^\star_{r+1} = k^\star_r$ and therefore also $r+1$ is a better choice than $r$. Therefore, it may happen that both decreasing and increasing $r$ lead to a better performance (nevertheless, there is a single optimal choice). Although in general the retransmission mechanism is effective, selecting the appropriate value of $r$ is not trivial, because the attainable gain is not proportional to $r$, but also depends on some system parameters such as the discretization of the compression ratio (hence, $m$) and the energy availability. This goes beyond the scope of the work.
[^1]: We will refer to blocks of readings as data blocks, to distinguish them from the packets sent over the communication channel after processing
[^2]: Following the rasoning of Appendix \[apx:proof\_retx\], if $k^\star_q \ge k^\star_1/q$, then the retransmission scheme leads to some improvement, otherwise nothing can be inferred, in general, because the performance depends on the particular shape of $\mathbb{E}_[\Delta]$ in the two cases.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'This paper is devoted to numerical simulations of the magnetization dynamics driven by a spin-polarized current in extended ferromagnetic multilayers when a point-contact setup is used. We present (i) detailed analysis of methodological problems arising by such simulations and (ii) physical results obtained on a system similar to that studied in Rippard et al., Phys. Rev. Lett., [**92**]{}, 027201 (2004). We demonstrate that the usage of a standard Slonczewski formalism for the phenomenological treatment of a spin-induced torque leads to a qualitative disagreement between simulation results and experimental observations and discuss possible reasons for this discrepancy.'
author:
- 'D.V. Berkov, N.L. Gorn'
title: Micromagnetic simulations of the magnetization precession induced by a spin polarized current in a point contact geometry
---
Introduction
============
Nearly a decade after prediction [@SpInjPred] and subsequent experimental discovery [@SpInjDiscov] of magnetic excitation and magnetization switching induced by a spin-polarized current (SPC) in a thin magnetic film high-quality experiments providing quantitative information concerning the corresponding magnetization dynamics have been performed (see, e.g., Ref. ). For this reason the important problem of the [*quantitative*]{} verification of existing theoretical models for the spin-transfer phenomena and SPC-induced magnetization dynamics [@SpInjTheory] can be addressed. Corresponding research clearly requires full-scale micromagnetic simulations, because single-domain (macrospin) approximation [@Sun2000] by its definition can not incorporate important effects resulting from the inhomogeneity of the magnetization states. Such strongly non-uniform magnetization configurations during the SPC-induced precession are predicted already for thin film elements of very small sizes [@BerkovChaos2005] ($\sim 30$ nm), which are less than typical dimensions of any actually used experimental samples.
Micromagnetic simulations performed up to now deal only with the experiments performed in the so called columnar geometry [@SimulColumnGeom], where an electric current flows through a multilayer magnetic element with very small lateral sizes $\sim 10^2$ nm. Such simulations were able to reproduce many important features of the experimental data obtained on columnar structures, in particular, the existence of regular and quasichaotic oscillation regimes, dependence of the oscillation frequency on the current strength etc. (however, we mention that even the most sophisticated model failed to reproduce experimental data quantitatively [@BerkovEllipse2005]). In contrast to this situation, to the best of our knowledge, simulations of the point-contact experiments [@Rippard2004a; @Rippard2004b], which show some important magnetization dynamics features different from those observed on columnar structures, have not been carried out.
In this paper we present such full-scale micromagnetic simulations of the SPC-induced magnetization dynamics in the point-contact setup. The paper is organized as follows: in Sec. \[sec:NumSimMeth\] we discuss methodological problems of such simulations which are specific for the geometry under study. Next (Sec. \[sec:NumSimRes\]) we present our results, starting from the analysis of the magnetization dynamics without the current-induced magnetic field (which enables a much more transparent presentation of several important effects). Brief comparison with the experimental data is performed in the last section.
Methodological problems of numerical simulations in the point contact setup {#sec:NumSimMeth}
===========================================================================
Numerical simulations were carried out with our software package MicroMagus [@MicroMagus].This package uses the modified Bulirsch-Stoer algorithm with the adaptive step-size control for integrating the Landau-Lifshitz-Gilbert (LLG) equation to calculate the time evolution of the system magnetization configuration. For the additional torque created by the spin polarized current (SPC) we have assumed the symmetric Slonczewski form $\Gamma = (a_J/M_S) \cdot [{\bf M} \times [{\bf M} \times {\bf S}]]$ (${\bf S}$ is the spin polarization direction) in order to study the SPC-induced dynamics in the simplest possible approximation. In the trilayer system under study (see below) the spin torque was assumed to act on the on the magnetization of the ’free’ layer only. The site dependence of the spin torque magnitude $a_J({\rm r})$ (in the first approximation confined within the point contact area) is discussed below.
To enable a comparison with the experimental results reported in the most advanced quantitative studies of magnetization oscillations in the point contact geometry [@Rippard2004a; @Rippard2004b], we have chosen the system parameters as close as possible to those reported in Ref. . We have simulated a trilayer system consisting of two magnetic layers and an interlayer non-magnetic spacer: the lower ’fixed’ layer (underlayer) with the thickness $h_1 = 10$ nm and magnetic parameters typical for Co$_{90}$Fe$_{10}$ (saturation magnetization $M_S = 1500$ G, exchange constant $A = 2 \times 10^{-6}$ erg/cm); the upper (thin) Permalloy-like magnetic layer with the thickness $h_2 = 5$ nm and $M_S = 640$ G (as measured in Ref. ) and $A = 1 \times 10^{-6}$ erg/cm (standard value for Py [@Doering1966]); the spacer thickness was set to $h_{sp}=5$ nm as for the Cu spacer used in Ref. . All results presented below were obtained for an external field $H_0 = 1000$ Oe. Coordinate axes $0x$ and $0z$ lie in the film plane, with the $x$-axis directed along the external field.
The ’fixed’ layer thickness $h_1 = 10$ nm was taken less than the experimental value [@Rippard2004a] $h_1^{\rm exp} = 20$ nm, because proper simulations of such relatively thick layers require not only their in-plane discretization, but also the subdivision into sublayers, which would lead in this case to prohibitively large computational times; the influence of underlayer thickness on the magnetization dynamics will be discussed elsewhere.
We also did not study the effect of the polycrystalline structure of Co$_{90}$Fe$_{10}$ underlayer, although the magnetocrystalline anisotropy of this material is not negligible (cubic anisotropy with $K_1 \approx 5.6 \times 10^5$ erg/cm$^3$ was reported in Ref. ). The magnetization dynamics of the system under study turned out to be very complicated already for ideal layers (without taking into account their polycrystalline structure and corresponding random anisotropy), so we have postponed the study of the random magnetocrystalline anisotropy effects.
In contrast to micromagnetic study of spin injection effects observed in the columnar structures [@Kiselev2003; @Kiselev2004], where standard micromagnetic methods can be applied [@SimulColumnGeom], simulations of the point-contact experiments encounter serious methodological problems. Here we discuss two of them: (i) artificial interference effects occurring both for open and periodic boundary conditions and (ii) artificial short-wave magnetization oscillations arising by the usage of a sharp cut-off of charge and spin currents.
[*Artificial interference effects*]{}. This problem occurs due to the combination of two circumstances: first, for practical purposes magnetic materials with low dissipation rate $\lambda \sim 0.01 - 0.02$ are used (to ensure a small linewidth of the oscillation spectrum and a reasonably low excitation threshold), and second, a lateral size of a system available for simulations is much smaller than that used experimentally. In real experiments the lateral size of a multilayer is about 10 mkm [@Rippard2004a], which is definitely far above the value accessible for numerical simulations, especially taking into account that dynamic simulations require a much finer mesh than quasistatic ones. In particular, we have found out that to obtain a mesh-independent results for a system with geometric and magnetic parameters given above, a lateral discretization as fine as $\Delta x \times \Delta z = 2.5 \times 2.5$ nm$^2$ is necessary, which limits the simulated lateral system size to $\sim 1$ mkm$^2$ (which means more $\sim 10^5$ cells per layer). In a layer made of a material with low dissipation (see above) the decay length of the spin wave with the wavevector corresponding to the inverse size of the point contact has the same order of magnitude as the simulated area size.
When open boundary conditions (OBC) are used (i.e., a finite size element is simulated), this means that the wave emitted by the point contact propagates across the whole element, is reflected at its free borders and returns to the contact location. The strong interference of this reflected wave with primary (SPC-induced) magnetization oscillations leads to unphysical artifacts, especially taking into account that both waves have the same frequency. Another problem arising by using OBC is a complicate pattern of the wave reflection due to the inhomogeneous magnetization configuration on the element edges.
For periodic boundary condition (PBC) the magnetization configuration at the simulated area borders is homogeneous (which is the main reason to use PBC), but the primary wave also propagates across the whole simulated area and due to PBC enters this area from the opposite side, causing the same undesirable interference effects. To eliminate this effects, a method to absorb the wave near the simulation area borders, not affecting the low dissipation at and near the point contact, is required.
To ensure such an absorption, we have embedded in our code an artificial site dependence of the dissipation coefficient $\lambda({\bf r})$. The function describing this dependence should fulfil several conditions: (i) the dissipation within and nearby the point contact area should remain equal to its physical value $\lambda_0$ to preserve the dynamic properties of the system under study, (ii) the dissipation coefficient far from the point contact (near the simulation area borders) should be large enough to ensure the wave energy absorption, (iii) spatial variation $\lambda({\bf r})$ should be smooth enough to prevent the wave reflection from the border between the areas of small and relatively large $\lambda$ (due to the abrupt changes of the media properties). Following these requirements, we have adopted a site-dependent dissipation $\lambda({\bf r})$ in the form $$\label{LamSiteDep}
\lambda({\bf r}) = \lambda_0 + \Delta\lambda
\left(
1+ \tanh {r - R_0 \over \sigma_{\lambda}}
\right)$$
Here it is assumed that the point contact is located at the coordinate origin. The function (\[LamSiteDep\]) provides a gradual increase of the dissipation parameter above the (small) basic value $\lambda_0$ which starts at the distance $\approx (R_0 - \sigma_{\lambda})$ from the point contact and occurs smoothly within a ring of the width $\approx 2\sigma_{\lambda}$. The maximal dissipation value reached outside of this ring is $\lambda_{\rm max} = \lambda_0 + \Delta\lambda$. We have found out that for the system size $L_x \times L_z = 1 \times 1$ mkm$^2$ and basis dissipation values in the range $\lambda_0 = 0.01 - 0.04$ the introduction of the additional dissipation (\[LamSiteDep\]) with $R_0 = 300$ nm, $\sigma_{\lambda} = 40 nm$ and $\Delta\lambda = 0.1$ ensured the wave absorption at the simulation area borders, not changing the magnetization dynamics within and around the point contact area. For simulations considered here we have used the basis dissipation value $\lambda_0 = 0.02$.
[*Spurious magnetization oscillations caused by a sharp spatial cut-off of a point-contact current*]{}. By simulations of the columnar geometry the current is usually assumed to be distributed homogeneously within the layer plane of a nanoelement, which does not lead to any methodical problems because the magnetization is also present only inside the area where the current flows. In contrast to this simple situation, by simulations of a point-contact setup the naive usage of the step-like current density in the form $j(r \le D/2) = j_0$ and $j(r > D/2) = 0$ ($r$ being the distance from the contact center, $D$ - the contact diameter) lead to the development of artificial magnetization oscillations with the smallest wavelength supported by the given lattice. The reason for these oscillations is the non-physical abrupt change of the current density $j({\bf r})$ at $r = D/2$, so that its spatial Fourier image and the Fourier image of the current-induced magnetic field (the Oersted field) $H_{\rm Oe}({\bf k})$ exhibits large tails up to the highest values of the wavevectors available for the simulated discrete system. This leads to artificial instabilities for these wavevectors, resulting in the appearance of corresponding magnetization oscillations.
In order to avoid this problem, we have smoothed the spatial distribution of $H_{\rm Oe}({\bf r})$ obtained in the approximation of a sharp electric current cut-off convolving it with the Gaussian kernel $\exp(-r^2/2\sigma_H^2)$. A physically meaningful choice of the smoothing parameter $\sigma_H$ would require a reliable information about the lateral diffusion of the electric current carriers to calculate the actual spatial distribution of the current density. Lacking such knowledge, we have simply adopted the minimal value $\sigma_H = 2 \Delta x$ (two times larger than the mesh size) which was sufficient to eliminate the artificial oscillations mentioned above. Further increment of this parameter within a reasonable range (up to $\sigma_H = (4 - 5) \cdot \Delta x$) had only a minor influence on physical results.
A similar problem is caused by the sharp spatial cut-off of the [*spin*]{} current density represented by the amplitude $a_J({\bf r})$ of the SPC-induced torque $\Gamma = (a_J/M_S) \cdot [{\bf M} \times [{\bf M} \times {\bf S}]]$, although spatial oscillations caused by this cut-off are weaker than discussed in the previous paragraph. Nevertheless, the same kind of smoothing is also required to solve this problem. The smoothing parameter of the corresponding kernel $\exp(-r^2/2\sigma_S^2)$ is directly related to the spin diffusion length and could in principle be computed from the corresponding theory. In this study, however, we have also simply used the value $\sigma_S = 2 \Delta x$ for the same reasons as explained above for $\sigma_H$. At this point we would like to emphasize, that - in contrast to the Oersted field smoothing parameter - the $\sigma_S$-value significantly influences the system behaviour; in particular, the threshold value of $a_J$ for the onset of steady-state microwave oscillations substantially depends on $\sigma_S$. For this reason the problem of calculating the actual distribution of a [*spin*]{} current in the point contact geometry deserves spatial attention.
All results presented below were obtained employing the site-dependent damping (\[LamSiteDep\]) and smoothing of the Oersted field and spin current distribution with parameters given above. The in-plane discretization of both magnetic layers with the mesh size $\Delta x \times \Delta z = 2.5 \times 2.5$ nm$^2$ and the full size of the simulation area $L_x \times L_z = 1 \times 1$ mkm$^2$ (with PBC) were used. The point contact diameter was set to $D = 40$ nm.
Numerical simulations: Results and discussion {#sec:NumSimRes}
=============================================
In this paper we discuss only simulation results obtained without taking into account the effect of thermal fluctuations ($T = 0$). Even without these effects the system demonstrates very complicated dynamics, which should be understood before thermal fluctuations are included into consideration.
[*Dynamics without taking into account the Oersted field*]{}. We start with the analysis of the magnetization dynamics when the influence of the Oersted field is neglected. We point out already here that in the experimental situation [@Rippard2004a] the current-induced magnetic field is comparable with the externally applied field: For the total current $I \sim 4$ mA flowing through the point contact with the diameter $D \approx 40$ nm the maximal value of the Oersted field is $H_{\rm Oe}^{max} \sim 400$ Oe, whereas the external field used in Ref. to present most detailed results is $H_{\rm ext} = 1000$ Oe. For this reason we do not expect the approximation when $H_{\rm Oe}$ is neglected to be quantitatively correct, but neglecting the Oersted field simplifies the magnetization dynamics of the system under study, preserving most of its qualitative features, which can be thus demonstrated more clearly.
The major feature of the simulated magnetization dynamics in the point contact geometry is the existence of [*two*]{} current regions where the steady-state precession of the magnetization within and nearby the point-contact area exists (Fig. \[fig:TwoPrecRegimes\_noOeField\]).
In the first current region - [*before*]{} the magnetization in the point contact area is switched under the SPC influence, i.e., when the magnetization is (on average) still directed along the external field, the magnetization dynamics is relatively simple. Magnetization configuration of the thin (upper) layer within the point-contact area remains roughly collinear. Spin waves emitted from the area under the contact are smooth and have a simple elliptical wavefront (Fig. \[fig:TwoPrecRegimes\_noOeField\]). The limit oscillation cycle of the magnetization ${\rm m}^{\rm av}$ averaged over this area represents a slightly bended ellipse (Fig. \[fig:PrecBeforeSw\_noOeField\]), time dependencies of the magnetization components are nearly ideal harmonic functions, so that oscillation power spectra consist of a single and very narrow peak. As usual for such in-plane oscillations, the frequency of $m_x^{\rm av}$-oscillations (the component along the external field $H_{\rm ext}$) is twice the $m_z^{\rm av}$-frequency (the in-plane component perpendicular to $H_{\rm ext}$), because $m_x^{\rm av}$ goes back and force twice during a single oscillation cycle. The oscillation frequency decreases monotonically with increasing current ($a_J$ in our formalism), which is mainly due to the increase of the oscillation amplitude (longer limit cycle) with the current strength. The oscillation power sharply increases when the current exceeds the threshold for the oscillation onset and then growth smoothly until the magnetization under the point contact area is switched by the SPC.
It turns out, however, that in the model simulated here a steady-state precession exists also [*after*]{} the point contact area switching caused by spin injection. By the transition to this second regime we observe a large frequency jump - for the system parameters used in this study the frequency drops down from $f_{\rm bef} \approx 9.8$ GHz to $f_{\rm aft} \approx 4.0$ GHz and then remains almost current-independent. Although the limit cycles have in this regime a more complicated form and the time dependencies of the magnetization components considerably differ from ideal sinusoids, spectral lines are still so narrow, that within the physical time corresponding to the longest simulation run performed ($\approx 20$ ns) their width could not be resolved ($\Delta f \le 50$ MHz). Oscillation amplitude slowly decreases with current leading to the corresponding decrease of the oscillation power.
Although spectral lines in this second precession mode remain quite narrow, the magnetization configurations appearing during the precession are extremely complicated even when the Oersted field is neglected. First of all we note that the precession frequency is below the frequency of the homogeneous FMR mode for the layer under consideration ($f_0 = (\gamma/2\pi) \cdot (H_0(H_0 + 4\pi M_S))^{1/2} \approx 8.4$ GHz. For this reason the ’normal’ circular (elliptical) wave can not exist in this regime, so that the energy is emitted in form of the soliton-like wave packages, as shown in Fig. \[fig:PrecAfterSw\_noOeField\]. The magnetization within and nearby the contact area itself forms vortex/antivortex pairs (the latter state is also sometimes called a cross-like configuration), which creation and annihilation is the basic feature of the magnetization dynamics in this regime; a typical example of such a structure is shown in Fig. \[fig:ArrowMagnConf\_noOeField\]. Detailed analysis of these challenging structures will be performed elsewhere.
[*Dynamics with the Oersted field included*]{}. Inclusion of the Oersted field requires the establishing of the relation between the current strength used in the actual experiment and the parameter $a_J$ used in simulations to set the amplitude of the SPC-induced torque. Lacking the exact microscopic theory which could provide such a relation, we have used the same procedure as in Ref. , i.e., we assumed that the oscillation onset threshold $a_J = 2.2$ corresponds to the minimal current $I_{\min} \approx 4$ mA where the magnetization precession is observed experimentally.
The current-induced magnetic field $H_{\rm Oe}$, being strongly inhomogeneous, results in a much more complicated magnetization dynamics than in the absence of $H_{\rm Oe}$. The most obvious change is the appearance of the wave asymmetry in the steady-state precession regime before switching: in that half of the layer where the Oersted field $H_{\rm Oe}$ is directed opposite to the external field $H_0$ (thus partly compensating it), the wave amplitude is significantly larger than in the other half of the film (where the external field is enhanced by $H_{\rm Oe}$). The inhomogeneity of $H_{\rm Oe}$ leads also to further complication of the magnetization states in the ’after-switching’ regime: the number of vortex-antivortex pairs which might exist simultaneously increases and the precession trajectory (limit cycle) of the average magnetization of the point-contact area becomes quasiperiodic. The complete analysis of the corresponding dynamics will be also presented elsewhere.
The major effect of the current-induced field is, however, not the quantitative changes in the ’before-’ and ’after-switching’ precession modes discussed above, but the appearance of a new intermediate regime in-between these two current regions. Corresponding interval is marked in Fig. \[fig:PrecRegimes\_withOeField\] by the legend ’Complicated magnetization dynamics’. For currents within this interval ($2.45 < a_J < 2.9$ for parameters used in our simulations) the $x-$projection of the magnetization under the contact area exhibits relatively rare transitions between the values close to the maximal possible value $m_x^{\rm max} = 1$ and values close to $m_x^{\rm av} = 0$ (Fig. \[fig:ComplicatedMagnDyn\_withFld\]a). Correspondingly, its power spectrum (Fig. \[fig:ComplicatedMagnDyn\_withFld\]b) has a large component at (relatively) low frequencies. The $z-$projection of the magnetization (in-plane projection perpendicular to the external field) oscillates with very different frequencies depending on the $m_x$-value, i.e., on the magnetization configuration in the point contact region: for nearly homogeneous magnetization state ($m_x^{\rm av}$ close to 1) the oscillation frequency of $m_z^{\rm av}$ is much higher than for a strongly inhomogeneous configuration (small values of $m_x^{\rm av}$). Oscillation power spectrum of $m_z^{\rm av}$ consists of several relatively broad lines, which quantitative analysis requires better simulation statistics (longer runs) than those which could be carried out up to now.
Comparison with experimental data {#sec:CompExpData}
=================================
In this last section we briefly compare our simulation data with experimental results from Ref. . First, we note that several important features of experimentally observed magnetization precession in the point contact setup could be reproduced by our simulations. In particular, in the oscillation regime before switching we have obtained, in accordance with Ref. , very narrow spectral lines (in our simulations the linewidth was $\Delta f < 50$ MHz) and nearly linear decay of the oscillation frequency with increasing current (see Fig. \[fig:TwoPrecRegimes\_noOeField\]a and \[fig:PrecRegimes\_withOeField\]a). Such extremely small linewidths can be naturally explained by a smooth variation (in space) of the magnetization configuration, which is due to the absence of strong demagnetizing field effects characteristic for nanoelements in the columnar geometry (see, e.g., Ref. ). The decrease of the oscillation frequency with increasing current is a consequence of the growing precession amplitude when the current (and hence - the spin torque magnitude) is increased, which for non-linear oscillations results in the larger precession period.
More important, however, are the disagreements between simulated and measured data, among which the [*qualitative*]{} one - the existence of at least two oscillation regimes for the simulated precession (whereby only one precession regime was observed experimentally) - is the major problem. There are several possible reasons for this discrepancy, among them (i) the simplicity of the torque term used in simulations (for more complicated forms which should be tested next, see Ref. ), (ii) influence of the SPC-induced effective field recently measured in Ref. , (iii) exchange weakening within magnetic layers resulting from the local Joule heating (in the point contact geometry this effect may be especially pronounced due to high current densities required to induce magnetization oscillations) and (iv) substantial contribution of the layer regions outside the point contact area to the measured microwave oscillation spectra. In particular, for the last mentioned reason, the signal from these regions would be present in the regime before switching, but nearly absent for the ’after-switching’ mode (as it can be seen from the comparison of right panels of Fig. \[fig:PrecBeforeSw\_noOeField\] and \[fig:PrecAfterSw\_noOeField\]), thus strongly enhancing the oscillation power in the regime before switching compared to the second regime. However, to clarify whether it is really the case, calculations of the current distribution for the concrete experimental setup are required.
Another interesting problem is the existence of a strong second harmonic in the experimentally measured spectrum. Its presence could be caused by the local (under the point contact) deviation of the underlayer magnetization from the external field direction. This would lead to the contributions from the oscillations of both longitudinal ($m_x$) and transverse ($m_z$) in-plane magnetization components, thus providing the required second harmonic ($m_x$) and basic ($m_z$) oscillation frequency [@BerkovEllipse2005]. Such a magnetization deviation from the $H_0$ direction could exist due, e.g., to the random magnetic anisotropy of $Co_{90}Fe_{10}$ crystallites. Another explanation of the second harmonic presence could be the above mentioned contribution of the area around the point contact, because due to the local conservation of the magnetic moment magnitude the waves of both $m_x$- and $m_z$-components contain both frequencies (see Fig. \[fig:PrecBeforeSw\_noOeField\], right panels).
Conclusion
==========
In conclusion we point out that full-scale micromagnetic simulations (performed in frames of the Slonczewski formalism) of the magnetization dynamics induced by a spin-polarized current in the point contact geometry recover several important features of experimental observations, like very narrow spectral lines and current dependence of the oscillation frequency. However, if we assume that the measured signal comes solely from the magnetization oscillation under the point contact area, simulations results exhibit serious qualitative disagreements with experimental data, the main of which is the existence of at least two precession modes in the two current intervals corresponding to (i) the precession around the external field direction (’before-switching’ mode) and (ii) around the direction opposite to $H_0$ (’after-switching’ mode). This disagreement clearly shows that further refinement of theoretical models is required for understanding of the spin torque induced magnetization excitations in point-contact experiments.
[99]{}
J.C. Slonczewski, J. Magn. Magn. Mat., [**159**]{}, L1, (1996); L. Berger, Phys. Rev. B, [**54**]{}, 9353 (1996)
J.Z. Sun, J. Magn. Magn. Mat., [**202**]{}, 157 (1999); M. Tsoi, A.G.M. Jansen, J. Bass, W.-C. Chiang, M. Seck, V. Tsoi, P. Wyder, Phys. Rev. Lett., [**80**]{}, 4281 (1998)
S.I. Kiselev, J.C. Sankey, I.N. Krivorotov, N.C. Emley, R.J. Schoelkopf, R.A. Buhrman, D.C. Ralph, Nature, [**425**]{}, 380 (2003);
S.I. Kiselev, J.C. Sankey, I.N. Krivorotov, N.C. Emley, M. Rinkoski, C. Perez, R.A. Buhrman, D.C. Ralph, Phys. Rev. Lett., [**93**]{}, 036601 (2004).
W.H. Rippard, M.R. Pufall, T.J. Silva, Applied Physics Letters [**82**]{}, 1260 (2003).
W.H. Rippard, M.R. Pufall, S. Kaka, S.E. Russek, T.J. Silva, Phys. Rev. Lett., [**92**]{}, 027201 (2004)
W.H. Rippard, M.R. Pufall, S. Kaka, T.J. Silva, S.E. Russek, Phys. Rev. B, [**70**]{}, 100406(R) (2004)
X. Waintal, E.B. Myers, P.W. Brouwer, D.C. Ralph, Phys. Rev. B, [**62**]{} 12317 (2000), M.D.Stiles, A.Zangwill, Phys. Rev. B, [**66**]{} 014407 (2002) M.D.Stiles and A.Zangwill, J. Appl. Phys. [**91**]{} 6812 (2002)
J.Z. Sun, Phys. Rev. B, [**62**]{} 570 (2000)
D.V. Berkov, N.L. Gorn, Phys. Rev. B, [**71**]{} 052403 (2005)
J. Miltat, G. Albuquerque, A. Thiaville, C. Vouille, J. Appl. Phys., [**89**]{}, 6982 (2001), Z. Li, S. Zhang, Phys. Rev. B, [**68**]{} 024404-1 (2003), J.-G. Zhu, X. Zhu, IEEE Tarns. Magn., [**40**]{}, 182 (2004), X. Zhu, J.-G. Zhu, R.M. White, J. Appl. Phys., [**95**]{}, 6630 (2004), K.J. Lee, A. Deac, O. Redon, J.P. Nozieres, B. Dieny, Nature Materials, [**3**]{} (2004) 877
D.V. Berkov, N.L. Gorn, to appear in: Phys. Rev. B, Sept. 2005; see also: cond-mat/0503754
D.V. Berkov, N.L. Gorn, MicroMagus - package for micromagnetic simulations, http:$\backslash\backslash$www.micromagus.de
M.D. Stiles, J. Xiao, A. Zangwill, Phys. Rev., [**B69**]{} (2004) 054408; M.L. Polianski, P.W. Brouwer, Phys. Rev. Lett., [**92**]{} (2004) 026602
W. Doering, [*Micromagnetismus*]{}, in: [*Handbook of Physics*]{}, ed. by S. Flügge, Springer-Verlag Berlin, 1966
J. Pelzl, R. Meckenstock, D. Spoddig, F. Schreiber, J. Pflaum, Z. Frait, J. of Phys.: Cond. Matt. [**15**]{} (2002) S451
J.C. Slonczewski, J. Magn. Magn. Mat., [**247**]{} 324 (2002)
J. Xiao, A. Zangwill, M.D. Stiles, Phys. Rev. B, [**70**]{}, 172405 (2004)
M.A. Zimmler, B. Özyilmaz, W. Chen, A.D. Kent, J.Z. Sun, M.J. Rooks, R.H. Koch, Phys. Rev. B [**70**]{} (2004) 184438
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We study timelike and spacelike virtual Compton scattering in the generalized Bjorken scaling regime at next to leading order in the strong coupling constant, in the medium energy range which will be studied intensely at JLab12 and in the COMPASS-II experiment at CERN. We show that the Born amplitudes get sizeable $O(\alpha_s)$ corrections and, even at moderate energies, the gluonic contributions are by no means negligible. We stress that the timelike and spacelike cases are complementary and that their difference deserves much special attention.'
author:
- 'H. Moutarde'
- 'B. Pire'
- 'F. Sabatié'
- 'L. Szymanowski'
- 'J. Wagner'
title: On timelike and spacelike deeply virtual Compton scattering at next to leading order
---
\#1[ ]{} CPHT-RR002.0113
IRFU-13-01
Introduction
============
Spacelike Deeply Virtual Compton Scattering (DVCS) [@historyofDVCS] $$\gamma^*(q_{in}) N(P) \to \gamma(q_{out}) N'(P'=P+\Delta) \,,~~~~~q_{in}^2 =-Q^2,~~~~~q_{out}^2 =0\,,$$ has been the model reaction for studying QCD collinear factorization in exclusive processes in terms of Generalized Parton Distributions (GPDs) [@gpdrev1; @gpdrev2], which access correlated information about the light cone momentum fraction and transverse location [@Impact] of partons in hadrons. The specific kinematical regime where this factorization property is proven at leading twist is the generalized Bjorken regime of large energy and large $Q^2$ but finite and fixed momentum transfer $\Delta^2$. A number of experimental results at various energies [@expdvcs] have now established the relevance of this approach at accessible kinematical conditions. Detailed phenomenological studies [@fitting] are under way to quantify to which degree one may in a foreseeable future extract from experimental data the physical information encoded in GPDs. The DVCS process contributes to the leptoproduction of a real photon $$l^\pm (k) N(P) \to l^\pm (k') \gamma(q_{out}) N'(P'=P+\Delta)\,\,.$$ It interferes with the Bethe-Heitler process $$l^\pm (k) \gamma^*(-\Delta) \to l^\pm (k') \gamma(q_{out}) \,\,,$$ where the hadronic interaction is entirely determined by the nucleon (spacelike) electromagnetic form factors $F_1(\Delta^2)$ and $F_2(\Delta^2)$.
Timelike Compton Scattering (TCS) [@BDP] $$\gamma(q_{in}) N(P)\to \gamma^*(q_{out}) N'(P'=P+\Delta)\,,~~~~~q_{in}^2 =0,~~~~~q_{out}^2 =Q^2\,,$$ which contributes to the photoproduction of a lepton pair $$\gamma(q_{in}) N(P) \to l^- (k) l^+ (k') N'(P'=P+\Delta)\,,~~~~~k +k' =q_{out}\,,$$ and interferes with the Bethe-Heitler process $$\gamma(q_{in}) \gamma^*(-\Delta) \to l^- (k) l^+ (k') \,\,,$$ shares many features with DVCS and allows in principle to access the same GPDs. The experimental situation [@exptcs] is not as encouraging as in the DVCS case but progress is expected in the next few years. The amplitudes of these two reactions are related at Born order by a simple complex conjugation but they significantly differ at next to leading order (NLO) in the strong coupling constant $\alpha_s$ [@MPSW].
Complete NLO calculations [@JiO; @Mankiewicz; @Belitsky:1999sg; @NLOphenoDVCS; @Pire:2011st] are available for both the DVCS and TCS reactions and there is no indication that NLO corrections are negligible in the kinematics relevant for current or near future experiments.
In this paper, we explore the consequence of including NLO gluon coefficient functions and NLO corrections to the quark coefficient functions entering the DVCS and TCS amplitudes, firstly in the calculation of spacelike and timelike Compton form factors with two models of GPDs. Then we proceed to the calculation of specific observables in the kinematical conditions which will soon be accessible in lepton nucleon collisions.
We focus on kinematics relevant to the next JLab and COMPASS measurements. The case of very large energy, and subsequently very small skewness, deserves its own study that will be addressed elsewhere. Ultraperipheral collisions at hadron colliders may already open the access to timelike Compton scattering in this domain [@PSW]. Also the proposed next generation of electron-ion colliders [@eic; @eicwp] will access in a detailed way this interesting domain with small skewness.
In this paper we concentrate on the influence of the NLO corrections to the DVCS and TCS observables. Because of that, although required to ensure QED gauge invariance [@twist3], we do not discuss here twist 3 effects. Neither do we implement the new results on target mass and finite $t$ corrections [@Braun:2012bg] nor the recently proposed resummation formula [@APSW] which focuses on the regions near $x=\pm \xi$.
Kinematics and amplitudes
=========================
Kinematics
----------
We introduce two light-like vectors $p$ and $n$ satisfying $p^2=0$, $n^2=0$ and $np=1$. We decompose the momenta in this light-cone basis as $k^\mu=\alpha n^\mu +\beta p^\mu +k_T^\mu$, with $k_T^2<0$, and we further note $k^+ = kn$. We also introduce the standard kinematical variables: $\Delta=P'-P$, $t=\Delta^2$ and $W^2=(q_{in}+P)^2$.
We define kinematics separately for spacelike (sl) and timelike (tl) Compton scattering. We denote the (positive) skewness variable as $\xi$ in the DVCS case, and as $\eta$ in the TCS case. In the DVCS case, where $-q_{in}^2=Q^2>0$ and $q_{out}^2=0$, we parametrize the momenta as follows $$\begin{aligned}
\label{kinDVCS}
&& q_{in}^\mu = \frac{Q^2}{4\xi}n^\mu -2\xi p^\mu \;, \;\;\;\;\;q_{out}^\mu = \alpha_{sl}n^\mu -\frac{\Delta_T^2}{2\alpha_{sl}} p^\mu -\Delta_T^\mu\;,
\nonumber \\
&& P^\mu=(1+\xi)p^\mu+\frac{M^2}{2(1+\xi)}n^\mu\;,\;\;\;\; P'^\mu=\beta_{sl}p^\mu +\frac{M^2-\Delta^2_T}{2\beta_{sl}}n^\mu +\Delta_T^\mu \;,\end{aligned}$$ where $M$ is the nucleon mass. The coefficients $\alpha_{sl}$, $\beta_{sl}$ in (\[kinDVCS\]) satisfy the following system of equations $$\label{eqsDVCS}
\alpha_{sl} = \frac{Q^2}{4\xi}+\frac{M^2}{2}\left(\frac{1}{1+\xi}-\frac{1}{\beta_{sl}} \right) +\frac{\Delta_T^2}{2\beta_{sl}}\;,\;\;\;\;\beta_{sl}= 1-\xi +\frac{\Delta_T^2}{2\alpha_{sl}}\;,$$ which in the limit $M=0$ and $\Delta_T=0$ relevant for calculation of the coefficient function leads to the standard values $\alpha_{sl} = Q^2/4\xi$ and $\beta_{sl}= 1-\xi$.
In the TCS case, where $q_{in}^2=0$ and $q_{out}^2=Q^2>0$, we parametrize the momenta as $$\begin{aligned}
\label{kinTCS}
&& q_{in}^\mu = \frac{Q^2}{4\eta}n^\mu \;, \;\;\;\;\;q_{out}^\mu = \alpha_{tl}n^\mu +\frac{Q^2-\Delta_T^2}{2\alpha_{tl}}p^\mu -\Delta_T^\mu\;,
\nonumber \\
&& P^\mu=(1+\eta)p^\mu+\frac{M^2}{2(1+\eta)}n^\mu\;,\;\;\;\; P'^\mu=\beta_{tl}p^\mu +\frac{M^2-\Delta^2_T}{2\beta_{tl}}n^\mu +\Delta_T^\mu \;.\end{aligned}$$ The coefficients $\alpha_{tl}$, $\beta_{tl}$ in (\[kinTCS\]) are solutions of the following system of equations $$\label{eqsTCS}
\alpha_{tl} = \frac{Q^2}{4\eta}+\frac{M^2}{2}\left(\frac{1}{1+\eta}-\frac{1}{\beta_{tl}} \right) +\frac{\Delta_T^2}{2\beta_{tl}}\;,\;\;\;\;\beta_{tl}= 1+\eta-\frac{Q^2-\Delta_T^2}{2\alpha_{tl}}\;,$$ which, again in the limit $M=0$ and $\Delta_T=0$ relevant for calculation of the coefficient function, take the standard values $\alpha_{tl} = Q^2/4\eta$ and $\beta_{tl}= 1-\eta$.
The DVCS and TCS amplitudes
---------------------------
After proper renormalization, the full Compton scattering amplitude[^1] reads in its factorized form (at factorization scale $\mu_F$) $$\begin{aligned}
\mathcal{A}^{\mu\nu} &=& -g_T^{\mu\nu}\int_{-1}^1 dx
\left[
\sum_q^{n_F} T^q(x) F^q(x)+T^g(x) F^g(x)
\right] \nonumber \\
&+& i\epsilon_T ^{\mu\nu}\int_{-1}^1 dx
\left[
\sum_q^{n_F} \widetilde{T}^q(x) \widetilde{F}^q(x)+\widetilde{T}^g(x) \widetilde{F}^g(x)
\right] \,,
\label{eq:factorizedamplitude}\end{aligned}$$ where we omitted the explicit skewness dependence. Renormalized coefficient functions are given by $$\begin{aligned}
T^q(x)&=& \left[ C_{0}^q(x) +C_1^q(x) +\ln\left(\frac{Q^2}{\mu^2_F}\right) \cdot C_{coll}^q(x)\right] - ( x \to -x ) \,,\nonumber\\
T^g(x) &=& \left[ C_1^g(x) +\ln\left(\frac{Q^2}{\mu^2_F}\right) \cdot C_{coll}^g(x)\right] +( x \to -x )\,,\nonumber\\
\widetilde{T}^q(x)&=& \left[
\widetilde{C}_{0}^q(x) +\widetilde{C}_1^q(x) +\ln\left(\frac{Q^2}{\mu^2_F}\right) \cdot \widetilde{C}_{coll}^q (x)\right]
+( x \to -x )
\,,\nonumber\\
\widetilde{T}^g(x) &=& \left[
\widetilde{C}_1^g(x) +\ln\left(\frac{Q^2}{\mu^2_F}\right) \cdot \widetilde{C}_{coll}^g(x)\right] - ( x \to -x )\,.
\label{eq:ceofficients}\end{aligned}$$ Results of the NLO calculations [@JiO; @Mankiewicz; @Belitsky:1999sg; @NLOphenoDVCS; @Pire:2011st] of the quark coefficient functions, based on the standard definitions of GPDs given in the Diehl’s review [@gpdrev1], read in the DVCS case $$\begin{aligned}
C_0^q(x,\xi) &=& -e_q^2 \frac{1}{x+\xi-i\varepsilon} \,, \nonumber \\
C_1^q(x,\xi) &=& \frac{e_q^2\alpha_SC_F}{4\pi} \frac{1}{x+\xi-i\varepsilon}
\bigg[
9 -3\frac{x+\xi}{x-\xi}\log(\frac{x+\xi}{2\xi}-i\varepsilon)
-\log^2(\frac{x+\xi}{2\xi}-i\varepsilon )
\bigg]
\,, \nonumber\\
C_{coll}^q (x,\xi)&=&\frac{e_q^2\alpha_SC_F}{4\pi}
\frac{1}{x+\xi-i\varepsilon}
\bigg[
-3-2\log(\frac{x+\xi}{2\xi}-i\varepsilon) \bigg]
\,, \nonumber\\
\widetilde{C}_0^q (x,\xi)&=& -e_q^2 \frac{1}{x+\xi-i\varepsilon} \,,\nonumber \\
\widetilde{C}_1^q (x,\xi)&=& \frac{e_q^2\alpha_SC_F}{4\pi} \frac{1}{x+\xi-i\varepsilon}
\bigg[
9 -\frac{x+\xi}{x-\xi}\log(\frac{x+\xi}{2\xi}-i\varepsilon)
-\log^2(\frac{x+\xi}{2\xi}-i\varepsilon )
\bigg]
\,, \nonumber\\
\widetilde{C}_{coll}^q (x,\xi)&=&\frac{e_q^2\alpha_SC_F}{4\pi}
\frac{1}{x+\xi-i\varepsilon}
\bigg[
-3-2\log(\frac{x+\xi}{2\xi}-i\varepsilon) \bigg]
\,,
\label{eq:QDVCS}\end{aligned}$$ where $C_F= (N_c^2-1) / (2N_c)$, $N_c$ being the number of colors, and $e_q$ is the quark electric charge in units of the proton charge. Using the same conventions, gluon coefficient functions read in the DVCS case $$\begin{aligned}
C_1^g (x,\xi)&=& \frac{\Sigma e_q^2\alpha_ST_F}{4\pi}\frac{1}{(x+\xi-i\varepsilon)(x-\xi+i\varepsilon)}\times \nonumber \\
&&\bigg[
2\frac{x+3\xi}{x-\xi}\log\left(\frac{x+\xi}{2\xi}- i \varepsilon\right)
-\frac{x+\xi}{x-\xi}\log^2\left(\frac{x+\xi}{2\xi} - i \varepsilon\right)
\bigg] \,,\nonumber\\
C_{coll}^g (x,\xi)&=& \frac{\Sigma e_q^2\alpha_ST_F}{4\pi}\frac{2}{(x+\xi-i\varepsilon)(x-\xi+i\varepsilon)}\left[
-\frac{x+\xi}{x-\xi}\log\left(\frac{x+\xi}{2\xi} - i \varepsilon\right)
\right]\,, \nonumber \\
\widetilde{C}_1^g (x,\xi)&=& \frac{\Sigma e_q^2\alpha_ST_F}{4\pi}\frac{1}{(x+\xi-i\varepsilon)(x-\xi+i\varepsilon)}\times \nonumber \\
&&\bigg[
-2\frac{3x+\xi}{x-\xi}\log\left(\frac{x+\xi}{2\xi}- i \varepsilon\right)
+\frac{x+\xi}{x-\xi}\log^2\left(\frac{x+\xi}{2\xi} - i \varepsilon\right)
\bigg] \,,\nonumber\\
\widetilde{C}_{coll}^g (x,\xi)&=& \frac{\Sigma e_q^2\alpha_ST_F}{4\pi}\frac{2}{(x+\xi-i\varepsilon)(x-\xi+i\varepsilon)}\left[
\frac{x+\xi}{x-\xi}\log\left(\frac{x+\xi}{2\xi} - i \varepsilon\right)
\right]\,,
\label{eq:GDVCS}\end{aligned}$$ where $T_F =\frac{1}{2}$. The results for the TCS case are simply [@MPSW] related to these expressions $$\begin{aligned}
^{TCS}T(x,\eta) = \pm \left(^{DVCS}T(x,\xi=\eta) + i \pi C_{coll}(x,\xi = \eta)\right)^* \,,
\label{eq:TCSvsDVCS}\end{aligned}$$ where $+$ $(-)$ sign corresponds to vector (axial) case.
Models for GPDs {#GPD}
================
In our analysis we will use two GPD models based on Double Distributions (DDs) [@historyofDVCS; @MR]. DDs allow to trivially achieve one of the strongest constraints on GPDs : the polynomiality of the Mellin moments of GPDs. They also automatically restore usual PDFs in the forward limit at $\xi,~t \to 0$. The GPDs are expressed as a two-dimensional integral over $\alpha$ and $\beta$ of the double distribution $f_i$ $$F_i (x,\xi,t) = \int_{-1}^1 d\beta \, \int_{-1+|\beta |}^{1-|\beta |} d\alpha \, \delta (\beta+\xi\alpha - x) \, f_i(\beta,\alpha,t) + D^F_i\left(\frac{x}{\xi},t\right)\, \Theta(\xi^2-x^2)\, ,
\label{Eq:GPD_DD}$$ where $F = H, E, \widetilde{H}, \widetilde{E}$ and $i$ denotes the flavor ($val$ for valence quarks, $sea$ for sea quarks and $g$ for gluons). In our analysis we only take into account the contribution of $H$ and $\widetilde{H}$. Indeed, $E$ and $\widetilde{E}$ are mostly unknown, and recent phenomenological studies of Ref. [@KMS] showed that most of existing DVCS observables are sensitive mostly to $H$ and $\widetilde{H}$.
The DD $f_i$ reads $$f_i(\beta,\alpha,t) = g_i(\beta,t) \, h_i(\beta ) \, \frac{\Gamma (2n_i+2)}{2^{2n_i+1}\Gamma^2 (n_i+1)} \, \frac{[(1-|\beta |)^2-\alpha^2]^{n_i}}{(1-|\beta |)^{2n_i+1}} \, ,
\label{Eq:DD}$$ where $\Gamma$ is the gamma function, $n_i$ is set to 1 for valence quarks, 2 for sea quarks and gluons. $g_i(\beta,t)$ parametrizes the $t$-dependence of GPDs, the $h_i(\beta)$’s in case of GPDs $H$ and $\widetilde{H}$ denote their forward limit and are related to the usual polarized and unpolarized PDFs in the following way $$\begin{aligned}
h_g( \beta ) & = & | \beta | \, g(| \beta |) \nonumber \, , \\
h^q_{\rm sea}( \beta ) & = & q_{\rm sea} ( | \beta | ) \, {\rm sign}(\beta ) \, , \nonumber \\
h^q_{\rm val}( \beta ) & = & q_{\rm val}( \beta ) \, \Theta(\beta) \, ,\nonumber\\
\tilde{h}_g( \beta ) & = & \beta \, \Delta g(| \beta |) \nonumber \, , \nonumber\\
\tilde{h}^q_{\rm sea}( \beta ) & = & \Delta q_{\rm sea} ( | \beta | ) \, , \nonumber \\
\tilde{h}^q_{\rm val}( \beta ) & = & \Delta q_{\rm val}( \beta ) \, \Theta(\beta) \, .\end{aligned}$$ $D^F_i$ in Eq. (\[Eq:GPD\_DD\]) denotes the Polyakov-Weiss D-term [@Polyakov:1999gs]. In our estimates we will use parametrizations obtained by a fit to the chiral soliton model [@KivPolVan]: $$\begin{aligned}
D^H_q(\frac{x}{\xi},t) = - D^E_q(\frac{x}{\xi},t) &=& \frac{1}{3}D^q(\frac{x}{\xi})F_D(t) \, , \nonumber\\
D^H_g(\frac{x}{\xi},t) = - D^E_g(\frac{x}{\xi},t) &=& \xi D^g(\frac{x}{\xi})F_D(t) \, , \end{aligned}$$ where $$\begin{aligned}
D^q(x,t) &=& - (1-x^2)\sum_{n=1}^{\infty(odd)} d_n^q~ C_n^{\frac{3}{2}}(x) \, , \nonumber \\
D^g(x,t) &=& - \frac{3}{2}(1-x^2)^2\sum_{n=1}^{\infty(odd)} d_n^g~ C_{n-1}^{\frac{5}{2}}(x) \, , \end{aligned}$$ where at $\mu =0.6 {\,{\rm GeV}}$ matching to chiral soliton model gives: $d_1^q =4.0$, $d_3^q =1.2$, $d_5^q = 0.4$, and we make an assumption that $d^g_n$ at input scale vanishes. In the QCD evolution of $d^q_n, d^g_n$ we switch from 3 to 4 flavours at $\mu = 1.5 {\,{\rm GeV}}$ and $\Lambda_3 =0.232 {\,{\rm GeV}}$, $\Lambda_4 =0.200 {\,{\rm GeV}}$.
The Goloskokov-Kroll model for the GPDs {#secGK}
----------------------------------------
As described in details in Refs. [@GK1; @GK2; @GK3; @KMS], the GPDs of the so-called Goloskokov-Kroll (or GK) model is constructed using CTEQ6m PDFs [@cteq6m]. The low-$x$ behavior of PDFs is well reproduced by power-law, with the power assumed to be generated by Regge poles. In this GPD model, the Regge behavior with linear trajectories of the DD encoded in the function $g_i(\beta,t)$ (see Eq. (\[Eq:DD\])) is assumed $$g_i(\beta,t) = {\rm e}^{b_i t} \, | \beta |^{-\alpha_i' t} \, .
\label{Eq:KG_Regge}$$ For the unpolarized GPD $H(x,\xi,t)$ the values of the Regge trajectory slopes and residues $\alpha_i', b_i$ as well as the expansions of the CTEQ6m PDFs [@cteq6m] used for $h_i$ may be found in Refs. [@GK1; @GK2; @GK3]. Finally, this model uses simple relations to parameterize the quark sea $$\begin{aligned}
H^u_{\rm sea} &=& H^d_{\rm sea} = \kappa_s H^s_{\rm sea} \, , \nonumber \\
{\rm with} \;\;\;\;\;\; \kappa_s&=&1+0.68/(1+0.52 \ln Q^2/Q_0^2 ) \, ,\end{aligned}$$
with the initial scale of the CTEQ6m PDFs $Q_0^2=4$ GeV$^2$. Similarly to the GPD $H$, polarized GPD $\widetilde H$ is constructed using the Blümlein - Böttcher (BB) polarized PDF parametrization [@BB] to fix the forward limit. Meson electroproduction data from HERA and HERMES have been considered to fix parameters for this GPD in the GK model, $\widetilde H$ for valence quarks and gluons have been parametrized, however we have neglected $\widetilde H$ for sea quarks. The values of the Regge trajectory slopes and residues as well as the expansions of the BB PDFs used in the GK model may be found in Refs. [@GK1; @GK5] .
The MSTW08 based model with factorized $t$ - dependence
-------------------------------------------------------
For the second model we use double distribution with MSTW08 PDFs [@MSTW08]. In that case we take simple factorizing ansatz for $t$ - dependence $$\begin{aligned}
g_u(\beta , t) &=& \frac{1}{2}F_1^u(t)\, , \\
g_d(\beta , t) &=& F_1^d(t)\, , \\
g_s(\beta , t) &=& g_g(\beta , t) = F_D(t)\, ,\end{aligned}$$ where $$\begin{aligned}
F_1^u(t) &=& 2 F_1^p(t) + F_1^n(t)\, , \\
F_1^d(t) &=& F_1^p(t) +2 F_1^n(t)\, , \\
F_D(t) &=& (1 -t/M_V^2)^{-2}\,,\end{aligned}$$ with $M_V =0.84 {\,{\rm GeV}}$, $F_1^p$ and $F_1^n$ are electromagnetic Dirac spacelike form factors of the proton and neutron. We use that model to construct only $\mathcal{H}$.
Compton Form Factors
====================
Let us now present the results for spacelike and timelike Compton Form Factors (CFF) at NLO, $\mathcal{H}$ and $\widetilde \mathcal{H}$, defined in the DVCS case as $$\begin{aligned}
\mathcal{H}(\xi,t) &=& + \int_{-1}^1 dx \,
\left(\sum_q T^q(x,\xi)H^q(x,\xi,t)
+ T^g(x,\xi)H^g(x,\xi,t)\right) \nonumber \\
\widetilde \mathcal{H}(\xi,t) &=& - \int_{-1}^1 dx \,
\left(\sum_q \widetilde T^q(x,\xi)\widetilde H^q(x,\xi,t)
+\widetilde T^g(x,\xi)\widetilde H^g(x,\xi,t)\right).
\label{eq:CFF}\end{aligned}$$
These CFFs are the GPD dependent quantities which enter the amplitudes. For DVCS they are defined through relations such as [@BDP] $$\mathcal{A}^{\mu\nu}(\xi,t) = - e^2 \frac{1}{(P+P')^+}\, \bar{u}(P^{\prime})
\left[\,
g_T^{\mu\nu} \, \Big(
{\mathcal{H}(\xi,t)} \, \gamma^+ +
{\mathcal{E}(\xi,t)} \, \frac{i \sigma^{+\rho}\Delta_{\rho}}{2 M}
\Big)
+i\epsilon_T^{\mu\nu}\, \Big(
{\widetilde{\mathcal{H}}(\xi,t)} \, \gamma^+\gamma_5 +
{\widetilde{\mathcal{E}}(\xi,t)} \, \frac{\Delta^{+}\gamma_5}{2 M}
\Big)
\,\right] u(P) \, ,
\label{eq:amplCFF}$$ Similar relation holds for TCS with $\xi$ replaced by $\eta$. We now present our results for the spacelike and timelike virtual Compton scattering, for the $\xi$ and $\eta$ values which include kinematical regimes of the JLab and Compass experiments. As we present our results for $Q^2 = \mu_F^2 = \mu_R^2 = 4 $ GeV$^2$, throughout the whole paper we use the value of $\alpha_S = 0.3$ .
Spacelike Compton Form Factors
------------------------------
Let us first discuss the importance of including NLO effects in CFFs related to DVCS observables. We show on Fig. \[fig:DVCSRe2x2\] (resp. Fig. \[fig:DVCSIm2x2\]) the results of our calculations for the real (resp. imaginary) parts of the CFF $\mathcal{H}(\xi)$ (multiplied by $\xi$ for a better legibility of the figure, since the CFFs are roughly proportional to $1/\xi$), with the two GPD models described in Sect. \[GPD\]. The dotted curves are the LO results and the solid lines show the results including all NLO effects. Although the results are naturally dependent of the choice of model GPDs, the main conclusions are quite universal. One can first observe that NLO corrections are by no means small, as exemplified by the ratio of these NLO corrections to the LO result shown in the lower part of Figs. \[fig:DVCSRe2x2\] and \[fig:DVCSIm2x2\]. One can also observe that the NLO corrections tend to diminish the real part of the CFF, and even change its sign for $\xi \gtrsim 0.01$). NLO corrections also decrease the imaginary part of the CFF. These are not new results [@Mankiewicz; @Belitsky:1999sg; @NLOphenoDVCS]. To quantify the main source of the NLO contribution, we show on the same plots with dashed lines the real and the imaginary parts of the CFF including quark NLO effects but not the gluon effects. Gluon effects are the most important part of the NLO correction, even at quite large values of $\xi$ (up to around $\xi \approx 0.3$) and they contribute as a very significant part to the full CFF $\mathcal{H}(\xi)$ including NLO effects. Since the CFF $\mathcal{H}(\xi)$ dominates the DVCS amplitude, this means that extracting quark GPDs from a leading order analysis of DVCS data is, to say the least, questionable. More positively, this result indicates that DVCS experiments even in the low energy regime of JLab12 provide us with a nice way to measure gluon GPDs. This fact has, to our knowledge, never been clearly spelled out before.
For completeness, we show on Fig. \[fig:DVCStH\] the real and imaginary parts of the CFF $\xi \widetilde \mathcal{H} (\xi)$ only for KG model. As for the case of $ \mathcal{H} (\xi)$, although the effects are less dramatic here, the NLO corrections are by no means small.
Let us now comment on the D-term contribution to the CFFs. Since the GPD originating from a D-term is a function of the ratio $x/\xi$ with a support in the ERBL region $-\xi < x < \xi$, it results in a constant real contribution to CFFs in the spacelike case. For $Q^2= \mu_F^2 = 4$ GeV$^2$ and $t = -0.1$ GeV$^2$ , the values of different contributions to $\mathcal{H_D}$ are shown in the second column of the Tab. \[tableDTCS\]. One can see that this D-term contribution is very important for $\xi \sim 0.1$, since it modifies the the full NLO result by about $50\%$. It is much less important at lower values of $\xi$ (around $3\%$ for $\xi \sim 0.01$ and only a few per mil at $\xi\sim 0.001$).
Timelike Compton Form Factors
-----------------------------
Let us now discuss the corresponding results for the timelike CFFs. On Fig. \[fig:TCSRe2x2\] and Fig. \[fig:TCSIm2x2\], we show the real and imaginary parts of the CFF $\mathcal H$ for the KG and the MSTW08 models of GPDs described in Sect. \[GPD\], for the invariant mass of the lepton pair $Q'^2=4 \textrm{~GeV}^2$, $t=-.1{\,{\rm GeV}}^2$ and factorization scale $\mu_F=Q$. For the imaginary part the correction does not exceed $40\%$. In the real part, the correction is of the order of few hundred percent. We observe that the main part of that large correction comes from the contribution of gluonic GPDs. To quantify the sensitivity of this statement on the uncertainties of the input PDF parametrizations we replot on the Fig. \[fig:TCSerror\] the upper right panel of Fig. \[fig:TCSRe2x2\] and the upper right panel of Fig. \[fig:TCSIm2x2\], now with the shaded bands showing the effect of a one sigma uncertainty of the input MSTW08 fit on the full NLO result, compared to the LO.
The D-term contribution to the CFF is a $\eta$-independent quantity and it has both a real and an imaginary parts at NLO in the TCS case. We show in Tab. \[tableDTCS\] the values of this D-term contribution in the LO and NLO cases. Its relative effect on the imaginary part of the CFF decreases significantly when $\eta$ decreases, from 10 to 1 and 0.1% when $\eta$ decreases from 0.1 to 0.01 and to 0.001.
Re$\mathcal{H_D}$ Im$\mathcal{H_D}$
------------------------ ------------------- -------------------
LO -2.59 0
NLO quark contribution -0.16 -0.85
NLO gluon contribution 0.18 0.16
Full NLO -2.57 -0.69
We then compare TCS and DVCS by plotting the ratio of NLO corrections on Fig. \[fig:TCStoDVCS\]. There is a striking difference in the magnitude of the corrections to the real part of CFFs, mostly insensitive to the choice of GPD parametrizations. As discussed in Ref. [@MPSW], this is a consequence of Eq. (\[eq:TCSvsDVCS\]) which by adding a phase to the dominant imaginary part of the spacelike CFF at small skewness, gives rise to a sizeable real part of the corresponding CFF in the timelike case. Such large corrections to the real part of CFFs will have a significant influence on observables which depend on the interference of the TCS process with the Bethe-Heitler amplitude, [*i.e.*]{} connected to the azimuthal angular distribution of the leptons. We shall discuss this in the next section.
![The real (left) and imaginary(right) parts of the TCS Compton Form Factor $\mathcal{H}$ multiplied by $\eta$, as a function of $\eta$ in the double distribution model based on MSTW08 parametrization, for $\mu_F^2=Q^2=4$ GeV$^2$ and $t= -0.1$ GeV$^2$. The dotted line shows the LO result and shaded bands around solid lines show the effect of a one sigma uncertainty of the input MSTW08 fit to the full NLO result.[]{data-label="fig:TCSerror"}](ReHerr_nice.eps "fig:"){width="7cm"}
For completeness, we show on Fig. \[fig:TCStH\] the real and imaginary parts of $\eta \widetilde \mathcal{H} (\eta)$. The NLO corrections are here smaller than 20% in the whole domain.
Cross sections and asymmetries
==============================
Deeply virtual Compton scattering
---------------------------------
Let us first briefly review the effects of including NLO corrections on the DVCS observables. On Fig. \[fig:c1\] we show the total DVCS cross section, the difference of cross sections for opposite lepton helicities and the beam spin asymmetry $A_{LU}^-$, defined by Eq. (48) of Ref. [@KMS]. We choose a set of values of kinematic variables representative for JLab12, namely $Q^2 = 4$ GeV$^2$, $E_{e}=11$ GeV and $t= -0.2$ GeV$^2$. The observables are shown as a function of the azimuthal angle $\phi$ (in the Trento convention). The Born order result is shown as the dotted line, the full NLO result by the solid line and the NLO result without the gluonic contribution as the dashed line. We see that the effects of the NLO corrections are quite large in both GPD models. Although the value of $\xi$ is quite large, we see that the gluon contribution ([*i.e.*]{} the difference between the dashed and the solid curve) is by no means negligible.
On Fig. \[fig:compass\] we show the DVCS observables relevant to the COMPASS experiment at CERN, namely (from left to right) the mixed charge-spin asymmetry, the mixed charge-spin difference and the mixed charge-spin sum defined in Eq. (59) of Ref. [@KMS], at the kinematics $\xi=0.05$, $Q^2=4$ GeV$^2$ and $-t=0.2$ GeV$^2$. The upper part of Fig. \[fig:compass\] uses the GK parametrization and the lower part the MSTW08 parametrizations with $1\sigma$ errors. We display here only the contribution from the GPD $H$ . The lower value of the skewness $\xi=0.05$ allows to test a complementary regime with respect to JLab measurements. Note the dramatic difference in the real part once gluon GPDs are in, inducing for instance a change of sign of the mixed charge-spin asymmetry and the mixed charge-spin difference for the MSTW case for instance. Note that the change is also huge in the case of GK parametrization but there is no sign change, indicating a significant model dependence. At any rate, NLO effects should be highly visible at COMPASS, which probes processes occuring at higher energy than JLab.
![From left to right, the total DVCS cross section in pb/GeV$^4$, the difference of cross sections for opposite lepton helicities in pb/GeV$^4$, the corresponding asymmetry, all as a function of the usual $\phi $ angle (in the Trento convention [@Bacchetta:2004jz]) for $E_e=11 {\,{\rm GeV}}, \mu_F^2=Q^2=4$ GeV$^2$ and $t= -0.2$ GeV$^2$. On the first line, the GPD $H(x,\xi,t)$ is parametrized by the GK model, on the second line $H(x,\xi,t)$ is parametrized in the double distribution model based on the MSTW08 parametrization. The contributions from other GPDs are not included. In all plots, the LO result is shown as the dotted line, the full NLO result by the solid line and the NLO result without the gluonic contribution as the dashed line. The Bethe-Heitler contribution appears as the dash-dotted line in the cross section plots (left part).[]{data-label="fig:c1"}](jlab.eps){width="12.5cm"}
![The DVCS observables for the COMPASS experiment, from left to right, mixed charge-spin asymmetry, mixed charge-spin difference and mixed charge-spin sum (in nb/GeV$^4$). The kinematical point is chosen as $ \xi=0.05, Q^2=4$ GeV$^2$, $t=-0.2$ GeV$^2$. On the first line, the GPD $H(x,\xi,t)$ is parametrized by the GK model, on the second line $H(x,\xi,t)$ is parametrized in the double distribution model based on the MSTW08 parametrization. The contributions from other GPDs are not included. In all plots, the LO result is shown as the dotted line, the full NLO result by the solid line and the NLO result without the gluonic contribution as the dashed line.[]{data-label="fig:compass"}](compass.eps){width="12.5cm"}
Timelike Compton scattering
---------------------------
Now we pass to predictions for the observables in the timelike counterpart of DVCS, namely TCS. On the left part of Fig. \[fig:TCS\_LO\_NLO\] we show the TCS contribution to the differential cross section as a function of the skewness $\eta$ for $Q^2 = \mu^2 = 4$ GeV$^2$, and $t= -0.2$ GeV$^2$ integrated over $\theta \in (\pi/4,3\pi/4)$ and over $\phi \in (0,2\pi)$. We see that the inclusion of the NLO corrections is more important at small skewness. On the right panel of Fig. \[fig:TCS\_LO\_NLO\] we show that the Bethe-Heitler dominates the integrated cross-section for this kinematics. In consequence, more differential observables, as the azimuthal $\phi$ dependence (with angles $\theta$ and $\phi$ defined in Ref. [@BDP] ) reveal in a better way the different contributions. Moreover simple $\phi$ dependence of the interference term allows for an easy access to the real part of the CFFs, which as we observed on the Fig. \[fig:TCSRe2x2\], are subject to the big NLO corrections. We indeed observe that effect on the Fig. \[fig:xsec\_phidep\], which shows the $\phi$ dependence of the unpolarized differential cross sections for pure BH process, and with a LO and NLO corrections to the interference term.
![left : LO (dotted) and NLO (solid) the TCS contribution to the cross-section as a function of $\eta$ for $Q^2 = \mu^2 = 4$ GeV$^2$, and $t= -0.2$ GeV$^2$ integrated over $\theta \in (\pi/4,3\pi/4)$ and over $\phi \in (0,2\pi)$. Right : LO (dotted) and NLO (solid) TCS and Bethe-Heitler (dashed) contributions to the cross-section as a function of $t$ for $Q^2 = \mu^2$ GeV$^2$ integrated over $\theta \in (\pi/4,3\pi/4)$ and over $\phi \in (0,2\pi)$ for $E_\gamma=10$ GeV ($\eta \approx 0.11$).[]{data-label="fig:TCS_LO_NLO"}](TCS_LO_NLO_xi_t02.eps "fig:"){width="7"} ![left : LO (dotted) and NLO (solid) the TCS contribution to the cross-section as a function of $\eta$ for $Q^2 = \mu^2 = 4$ GeV$^2$, and $t= -0.2$ GeV$^2$ integrated over $\theta \in (\pi/4,3\pi/4)$ and over $\phi \in (0,2\pi)$. Right : LO (dotted) and NLO (solid) TCS and Bethe-Heitler (dashed) contributions to the cross-section as a function of $t$ for $Q^2 = \mu^2$ GeV$^2$ integrated over $\theta \in (\pi/4,3\pi/4)$ and over $\phi \in (0,2\pi)$ for $E_\gamma=10$ GeV ($\eta \approx 0.11$).[]{data-label="fig:TCS_LO_NLO"}](xsec.eps "fig:"){width="7"}
![The $\phi$ dependence of the cross-section at $E_\gamma = 10$ GeV, $Q^2 = \mu ^2 = 4$ GeV$^2$, and $t= -0.1$ GeV$^2$ integrated over $\theta \in (\pi/4,3\pi/4)$: pure Bethe-Heitler contribution (dashed), Bethe-Heitler plus interference contribution at LO (dotted) and NLO (solid). []{data-label="fig:xsec_phidep"}](xsec_phidep.eps){width="10"}
![Ratio R defined by Eq. (\[eq:R\_ratio\]) as a function of $\eta$, for $Q^2 = \mu_F^2 = 4$ GeV$^2$ and $t= -0.1$ GeV$^2$. The dotted line represents LO contribution and the solid line represents NLO result. []{data-label="fig:Ratio_R"}](R.eps){width="10"}
To quantify how big is the deviation from pure Bethe-Heitler process in the unpolarized cross section we calculate (see Fig. \[fig:Ratio\_R\]) the ratio R, defined in Ref. [@BDP] by $$R(\eta) = \frac{2\int_0^{2 \pi} d \varphi \,\cos \varphi\, \frac{dS}{dQ'^2dtd\varphi}}{\int_0^{2 \pi} d \varphi\frac{dS}{dQ'^2dtd\varphi}}
\,,
\label{eq:R_ratio}$$ where $S$ is a weighted cross section given by Eq. (43) of Ref. [@BDP]. It is plotted on Fig. \[fig:Ratio\_R\] as a function of the skewness $\eta$ for $Q^2 = \mu^2 = 4$ GeV$^2$, and $t= -0.2$ GeV$^2$. In the leading twist the numerator is linear in the real part of the CFFs, and the denominator, for the kinematics we consider, is dominated by the Bethe - Heitler contribution. The inclusion of NLO corrections to the TCS amplitude is indeed dramatic for such an observable and includes also change of sign.
Imaginary parts of the CFFs are accesible through observables making use of photon circular polarizations [@BDP]. The photon beam circular polarization asymmetry $$A= \frac{\sigma^+ - \sigma^-}{\sigma^+ + \sigma^-}\,,$$ is shown on the left part of Fig. \[fig:Asymmetry\_xi\], as a function of $\phi$ for the kinematic variables relevant for JLab: $Q^2 =4$ GeV$^2$= $\mu_F^2$, $t=-0.1$ GeV$^2$ and $E_\gamma = 10$ GeV (which corresponds to $\eta \approx 0.11$). The same quantity is shown on the right panel of Fig. \[fig:Asymmetry\_xi\] as a function of $\eta$ for $\phi = \pi/2$ and $Q^2 =4$ GeV$^2$= $\mu_F^2$. The effect of the NLO corrections on that observable is rather large, ranging from $10\%$ at the $\eta=0.1$ (relevant for JLab) through $30\%$ at $\eta=0.05$ (relevant for COMPASS) up to $100\%$ at very small $\eta$’s.
![ (Left) Photon beam circular polarization asymmetry as a function of $\phi$, for $t=-0.1$ GeV$^2$, $Q^2 = \mu^2 = 4$ GeV$^2$, integrated over $\theta \in (\pi/4,3\pi/4)$ and for $E_\gamma = 10$ GeV ($\eta \approx 0.11$). (Right) The $\eta$ dependence of the photon beam circular polarization asymmetry for $Q^2 = \mu^2 = 4$ GeV$^2$, and $t= -0.2$ GeV$^2$ integrated over $\theta \in (\pi/4,3\pi/4)$. The LO result is shown as the dotted line, the full NLO result by the solid line.[]{data-label="fig:Asymmetry_xi"}](Asymmetry_E_10_t_01_Q_2_nice.eps "fig:"){width="40.00000%"} ![ (Left) Photon beam circular polarization asymmetry as a function of $\phi$, for $t=-0.1$ GeV$^2$, $Q^2 = \mu^2 = 4$ GeV$^2$, integrated over $\theta \in (\pi/4,3\pi/4)$ and for $E_\gamma = 10$ GeV ($\eta \approx 0.11$). (Right) The $\eta$ dependence of the photon beam circular polarization asymmetry for $Q^2 = \mu^2 = 4$ GeV$^2$, and $t= -0.2$ GeV$^2$ integrated over $\theta \in (\pi/4,3\pi/4)$. The LO result is shown as the dotted line, the full NLO result by the solid line.[]{data-label="fig:Asymmetry_xi"}](Asymmetry_xi_nice.eps "fig:"){width="40.00000%"}
Conclusion.
===========
Deeply Virtual Compton Scattering, both in its spacelike and timelike realizations, is the golden channel to extract GPDs from experimental observables. This extraction may be seen as a two step process: firstly, Compton form factors may be separated from a careful analysis of various differential cross sections and asymmetries. Secondly, convolutions of coefficient functions with model GPDs may be confronted to these CFFs. We have demonstrated here, in the case of medium energy kinematics which will be explored in the near future at JLab and COMPASS, that the inclusion of NLO corrections to the coefficient function was an important issue, and that the difference of these corrections between the spacelike and timelike regimes was so sizeable that they can be promoted to the status of direct tests of the QCD understanding of the reactions.
Let us stress again a feature that was largely overlooked in previous studies, namely the importance of gluon contributions to the DVCS amplitude, even when the skewness variable $\xi$ is in the so-called valence region. This is not a real surprise when one recalls that gluons (in terms of distribution functions) are by no means restricted to the very low $x$ region and that gluon CFFs at a given $\xi$ value also depend on gluon PDFs at lower values of $x$. This effect is particularly big when one considers the real part of CFFs in the timelike case. This promotes the observables related to this quantity as sensitive probes of the 3-dimensional gluon content of the nucleon.
We did not extend our study to the very high energy regime which will be explored (both for spacelike and timelike Compton scattering) at future electron-ion colliders, nor to the case of ultraperipheral collisions at present hadron colliders. This will be addressed in separate studies. We did not discuss the rich scope of factorization scale dependence issues, which deserves special attention. Contrarily to the case of inclusive reactions where various strategies have been built to optimize the factorization scale, it has been shown in Ref. [@Anikin] that it was quite impossible to find a recipe to minimize higher order corrections both for the real and the imaginary part of a CFF. Moreover we find it difficult to advocate a different choice of scale for the two cases of timelike and spacelike reactions.
Acknowledgements {#acknowledgements .unnumbered}
================
We are grateful to Markus Diehl, Dieter Müller, Stepan Stepanyan, Pawe[ł]{} Nadel-Turoński and Samuel Wallon for useful discussions and correspondence. This work is partly supported by the Polish Grants NCN No DEC-2011/01/D/ST2/02069, by the Joint Research Activity “Study of Strongly Interacting Matter” (acronym HadronPhysics3, Grant Agreement n.283286) under the Seventh Framework Programme of the European Community, by the GDR 3034 PH-QCD, and the ANR-12-MONU-0008-01, and by the COPIN-IN2P3 Agreement.
[99]{} D. M[ü]{}ller [*et al.*]{}, Fortsch. Phys. [**42**]{}, 101 (1994); X. Ji, Phys. Rev. Lett. [**78**]{}, 610 (1997); A. V. Radyushkin, Phys. Rev. [**D56**]{}, 5524 (1997);
M. Diehl, T. Gousset, B. Pire and J. P. Ralston, Phys. Lett. B [**411**]{}, 193 (1997); J. C. Collins and A. Freund, Phys. Rev. [**D59**]{}, 074009 (1999). M. Diehl, Phys. Rept. [**388**]{} (2003) 41.
A. V. Belitsky and A. V. Radyushkin, Phys. Rept. [**418**]{}, 1 (2005); S. Boffi and B. Pasquini, Riv. Nuovo Cim. [**30**]{}, 387 (2007); K. Goeke, M. V. Polyakov and M. Vanderhaeghen, Prog. Part. Nucl. Phys. [**47**]{} (2001) 401 \[hep-ph/0106012\].
M. Burkardt, Phys. Rev. D [**62**]{} (2000) 071503; J. P. Ralston and B. Pire, Phys. Rev. D [**66**]{} (2002) 111501; M. Diehl, Eur. Phys. J. C [**25**]{} (2002) 223 \[Erratum-ibid. C [**31**]{} (2003) 277\]. C. Mu[ñ]{}oz Camacho *et al.*, Phys. Rev. Lett [**97**]{}, 262002 (2006);
F.X. Girod *et al.*, Phys. Rev. Lett. [**100**]{}, 162002 (2008); A. Aktas [*et al.*]{} \[H1 Collaboration\], Eur. Phys. J. C [**44**]{} (2005) 1;
F. D. Aaron [*et al.*]{} \[H1 Collaboration\], Phys. Lett. B [**681**]{}, 391 (2009);
S. Chekanov [*et al.*]{} \[ZEUS Collaboration\], Phys. Lett. B [**573**]{}, 46 (2003) and JHEP [**0905**]{}, 108 (2009);
A. Airapetian [*et al.*]{} \[HERMES Collaboration\], JHEP [**0806**]{}, 066 (2008), JHEP **1006**, 019 (2010), JHEP [**1207**]{}, 032 (2012) and arXiv:1206.5683 \[hep-ex\]; S. Chen [*et al.*]{} \[CLAS Collaboration\], Phys. Rev. Lett. [**97**]{}, 072002 (2006); S. Stepanyan [*et al.*]{} \[CLAS Collaboration\], Phys. Rev. Lett. [**87**]{}, 182002 (2001) \[arXiv:hep-ex/0107043\]. K. Kumericki, D. Mueller and K. Passek-Kumericki, Nucl. Phys. B [**794**]{}, 244 (2008) \[arXiv:hep-ph/0703179\]; M. Guidal, Eur. Phys. J. A [**37**]{}, 319 (2008) \[Erratum-ibid. A [**40**]{}, 119 (2009)\] \[arXiv:0807.2355 \[hep-ph\]\]; M. Guidal and H. Moutarde, Eur. Phys. J. A [**42**]{}, 71 (2009) \[arXiv:0905.1220 \[hep-ph\]\]; K. Kumericki and D. Mueller, Nucl. Phys. B [**841**]{} (2010) 1 \[arXiv:0904.0458 \[hep-ph\]\]; H. Moutarde, Phys. Rev. D [**79**]{} (2009) 094021 \[arXiv:0904.1648 \[hep-ph\]\]; M. Guidal, Phys. Lett. B [**693**]{} (2010) 17 \[arXiv:1005.4922 \[hep-ph\]\]; M. Guidal, Phys. Lett. B [**689**]{} (2010) 156 \[arXiv:1003.0307 \[hep-ph\]\]; H. Moutarde, arXiv:1010.4521 \[hep-ph\]; G. R. Goldstein, J. O. Hernandez and S. Liuti, Phys. Rev. D [**84**]{} (2011) 034007 \[arXiv:1012.3776 \[hep-ph\]\]; K. Kumericki, D. Muller and A. Schafer, JHEP [**1107**]{} (2011) 073 \[arXiv:1106.2808 \[hep-ph\]\]; K. Kumericki, D. Mueller and M. Murray, arXiv:1301.1230 \[hep-ph\]. E. R. Berger, M. Diehl and B. Pire, Eur. Phys. J. C [**23**]{}, 675 (2002). P. Nadel-Turonski [*et al.*]{}, AIP Conf. Proc. [**1182**]{}, 843 (2009); T. Horn [*et al.*]{}, AIP Conf. Proc. [**1374**]{}, 542 (2011).
D. Mueller, B. Pire, L. Szymanowski and J. Wagner, Phys. Rev. D [**86**]{}, 031502 (2012).
X. D. Ji and J. Osborne, Phys. Rev. D [**57**]{}, 1337 (1998) and Phys. Rev. D [**58**]{}, 094018 (1998). L. Mankiewicz, G. Piller, E. Stein, M. Vanttinen and T. Weigl, Phys. Lett. B [**425**]{}, 186 (1998). A. V. Belitsky, D. Mueller, L. Niedermeier and A. Schafer, Phys. Lett. B [**474**]{}, 163 (2000).
A. Freund and M. F. McDermott, Phys. Rev. D [**65**]{}, 074008 (2002) \[arXiv:hep-ph/0106319\]; A. Freund and M. F. McDermott, Phys. Rev. D [**65**]{}, 091901 (2002) \[arXiv:hep-ph/0106124\]; A. Freund and M. McDermott, Eur. Phys. J. C [**23**]{}, 651 (2002) \[arXiv:hep-ph/0111472\]. B. Pire, L. Szymanowski and J. Wagner, Phys. Rev. D [**83**]{}, 034009 (2011). B. Pire, L. Szymanowski and J. Wagner, Phys. Rev. D [**79**]{}, 014010 (2009), Nucl. Phys. Proc. Suppl. [**179-180**]{}, 232 (2008) and Acta Phys. Polon. Supp. [**2**]{}, 373 (2009). D. Boer [*et al.*]{}, arXiv:1108.1713 \[nucl-th\]. A. Deshpande [*et al.*]{}, arXiv:1212.1701 \[nucl-ex\]; J. L. Abelleira Fernandez [*et al.*]{} \[LHeC Study Group Collaboration\], J. Phys. G [**39**]{}, 075001 (2012) \[arXiv:1206.2913 \[physics.acc-ph\]\].
I. V. Anikin, B. Pire and O. V. Teryaev, Phys. Rev. D [**62**]{}, 071501 (2000);
A. V. Radyushkin and C. Weiss, Phys. Rev. D [**63**]{}, 114012 (2001); N. Kivel, M. V. Polyakov and M. Vanderhaeghen, Phys. Rev. D [**63**]{}, 114014 (2001); A. V. Belitsky and D. Mueller, Nucl. Phys. B [**589**]{}, 611 (2000); I. V. Anikin, R. S. Pasechnik, B. Pire and O. V. Teryaev, Eur. Phys. J. C [**72**]{}, 2055 (2012) \[arXiv:1112.1849 \[hep-ph\]\]. V. M. Braun and A. N. Manashov, Phys. Rev. Lett. [**107**]{}, 202001 (2011);
V. M. Braun, A. N. Manashov and B. Pirnay, Phys. Rev. D [**86**]{}, 014003 (2012) \[arXiv:1205.3332 \[hep-ph\]\]. T. Altinoluk, B. Pire, L. Szymanowski and S. Wallon, JHEP [**1210**]{} (2012) 049;
T. Altinoluk, B. Pire, L. Szymanowski and S. Wallon, arXiv:1206.3115 \[hep-ph\]. A. V. Belitsky and D. Mueller, Phys. Lett. B [**486**]{}, 369 (2000) \[hep-ph/0005028\].
I. V. Musatov and A. V. Radyushkin, Phys. Rev. D [**61**]{}, 074027 (2000) \[hep-ph/9905376\]. P. Kroll, H. Moutarde and F. Sabatié, arXiv:1210.6975 \[hep-ph\]. M. V. Polyakov and C. Weiss, Phys. Rev. D [**60**]{}, 114017 (1999) \[hep-ph/9902451\]. N. Kivel, M. V. Polyakov and M. Vanderhaeghen, Phys. Rev. D [**63**]{}, 114014 (2001). S. V. Goloskokov and P. Kroll, Eur. Phys. J. C [**42**]{}, 281 (2005) \[hep-ph/0501242\]. S. V. Goloskokov and P. Kroll, Eur. Phys. J. C [**50**]{}, 829 (2007) \[hep-ph/0611290\]. S. V. Goloskokov and P. Kroll, Eur. Phys. J. C [**53**]{}, 367 (2008) arXiv: 0708.3569 \[hep-ph\].
J. Pumplin, D. R. Stump, J. Huston, H. L. Lai, P. M. Nadolsky and W. K. Tung, JHEP [**0207**]{}, 012 (2002) \[hep-ph/0201195\]. J. Blümlein and H. Böttcher, Nucl. Phys. B [**636**]{}, 225 (2002) \[hep-ph/0203155\]. S. V. Goloskokov and P. Kroll, Eur. Phys. J. C [**65**]{}, 137 (2010) \[arXiv:0906.0460 \[hep-ph\]\]. A. D. Martin, W. J. Stirling, R. S. Thorne and G. Watt, Eur. Phys. J. C [**63**]{}, 189 (2009) \[arXiv:0901.0002 \[hep-ph\]\]. A. Bacchetta, U. D’Alesio, M. Diehl and C. A. Miller, Phys. Rev. D [**70**]{} (2004) 117504 \[hep-ph/0410050\]. I. V. Anikin, B. Pire, L. Szymanowski, O. V. Teryaev and S. Wallon, Eur. Phys. J. C [**42**]{}, 163 (2005).
[^1]: We do not consider the photon helicity changing amplitude coming from the transversity gluon GPD [@transversity].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We give a trichotomy on the complexity of integer reachability in vector addition systems with states extended with affine operations (affine VASS). Namely, we show that integer reachability in affine VASS is -complete for VASS with resets, -complete for VASS with (pseudo-)transfers and VASS with (pseudo-)copies, and undecidable otherwise. We further present a dichotomy for standard reachability in affine VASS: it is decidable for VASS with permutations, and undecidable otherwise. This yields a complete and unified complexity landscape of reachability in affine VASS.'
author:
- Michael Blondin
- Mikhail Raskin
bibliography:
- 'references.bib'
title: The Complexity of Reachability in Affine Vector Addition Systems with States
---
Introduction {#sec:introduction}
============
Preliminaries {#sec:preliminaries}
=============
A complexity trichotomy for integer reachability {#sec:trichotomy}
================================================
A complexity dichotomy for reachability {#sec:dichotomy}
=======================================
Conclusion and further work {#sec:conclusion}
===========================
Appendix {#sec:appendix}
========
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We study calculus of variations problems, where the Lagrange function depends on the Caputo-Katugampola fractional derivative. This type of fractional operator is a generalization of the Caputo and the Caputo–Hadamard fractional derivatives, with dependence on a real parameter $\r$. We present sufficient and necessary conditions of first and second order to determine the extremizers of a functional. The cases of integral and holomonic constraints are also considered.'
author:
- |
Ricardo Almeida\
[[email protected]]{}
date: |
Center for Research and Development in Mathematics and Applications (CIDMA)\
Department of Mathematics, University of Aveiro, 3810–193 Aveiro, Portugal
title: 'Variational Problems Involving a Caputo-Type Fractional Derivative'
---
[**Mathematics Subject Classification 2010:** ]{}26A33,34A08,34K28.
[**Keywords:** ]{}fractional calculus, Caputo–Katugampola fractional derivative, variational problems.
Introduction
============
Fractional calculus of variations was first studied by Riewe in [@Riewe], where he showed that nonconservative forces such as friction are modeled by non-integer order derivatives. In fact, although most known methods deal with conservative systems, natural processes are nonconservative and so the usual Lagrange formulation is inadequate to characterize these phenomena. It turns out that fractional calculus, due to its non-local character, may better describe the behavior of the certain real processes. For this reason nowadays it is an important research area, which has attracted the attention not only of mathematicians, but also of physicists and engineers. Generally speaking, fractional calculus deals with integrals and derivatives of arbitrary real order, and it was considered since the very beginning of calculus, but only on the past decades it has proven to be applicable in real problems. We can find several definitions for fractional operators, and to decide which one is more efficient to model the problem is a question whose answer depends on the system. Thus, we find several works that deal with similar subjects, for different kinds of fractional operators (see e.g. [@Almeida; @Atanackovic; @Baleanu10; @MR2411429; @MR2563910; @Cresson; @El-Nabulsi; @Frederico:Torres10; @Jumarie; @Jumarie3b; @comBasia:Frac1; @MT; @book:MT; @withTatiana:Basia] and references therein). In this paper we intend to present a more general theory, that includes the Caputo and the Caputo–Hadamard fractional derivatives, following the work started in [@Almeida1].
The text is organized as follows. In Section \[sec:FC\] we review the necessary definitions and results for the Caputo–Katugampola fractional derivative. The main results of the paper are presented in Section \[FCV\]. First, in §\[sec:fund\] we consider the fundamental problem, where we present necessary and sufficient conditions that every minimizer of the functional must fulfill. In §\[sec:legendre\] we prove a second order condition to determine if the extremals are in fact minimizers for the functional. Then, we consider variational problems subject to an integral constraint in §\[sec:iso\] and to a holomonic constraint in §\[sec:holo\]. We end with a generalization of the variational problem, known in the literature as Herglotz problem (§\[sec:herglotz\]).
Caputo–Katugampola fractional derivative {#sec:FC}
========================================
We find several definitions for fractional derivatives, each of them presenting its advantages and disadvantages. One of those, considered mainly by engineers, is the Caputo derivative exhibiting two important features: the derivative of a constant is zero and the Laplace transform depends only on integer-order derivatives. Given a function $x:[a,b]\to\mathbb R$, the Caputo fractional derivative of $x$ of order $\a\in(0,1)$ is defined by [@Kilbas] $${^CD_{a+}^\a}x(t)=\frac{1}{\Gamma(1-\a)}\frac{d}{dt}\int_a^t\frac{1}{(t-\t)^\a}[x(\t)-x(a)]\,d\t,$$ where $\Gamma$ denotes the well-known Gamma function, $$\Gamma(z)=\int_0^\infty t^{z-1}\exp(-t)\, dt, \quad z>0,$$ and if $x$ is of class $C^1$, then we have the equivalent form $${^CD_{a+}^\a}x(t)=\frac{1}{\Gamma(1-\alpha)}\int_a^t \frac{1}{(t-\t)^{\a}}x'(\t) \,d\tau.$$ From the definition, it is clear that if $x$ is a constant function, then ${^CD_{a+}^\a}x(t)=0$ and that, if $x(a)=0$, then the Caputo fractional derivative coincides with the Riemann–Liouville fractional derivative. The Caputo–Hadamard derivative is a very recent concept [@Baleanu1; @Baleanu2], and it combines the Caputo derivative with the Hadamard fractional operator. Given a function $x$, the Caputo–Hadamard fractional derivative of order $\a$ is defined as $${^HD_{a+}^\a}x(t)=\frac{1}{\Gamma(1-\a)}t\frac{d}{dt}\int_a^t\left(\ln\frac{t}{\t}\right)^{-\a}\frac{x(\t)-x(a)}{\t}d\t,$$ and again if $x$ is of class $C^1$, then $${^HD_{a+}^\a}x(t)=\frac{1}{\Gamma(1-\alpha)}\int_a^t\left(\ln\frac{t}{\t}\right)^{-\a}x'(\t)\, d\tau.$$
In [@Almeida1], a new type of operator is presented, that generalizes the two previous operators, by introducing a new parameter $\r>0$ in the definition. The same idea has already been done in [@Katugampola1; @Katugampola2], where a new operator is defined which generalizes the Riemann–Liouville and the Hadamard fractional derivatives.
Let $0<a<b<\infty$, $x:[a,b]\rightarrow\mathbb{R}$ be a function, and $\alpha\in(0,1)$ and $\rho>0$ two fixed reals. The Caputo–Katugampola fractional derivative of order $\a$ is defined as $$\LD x(t) =\frac{\r^\a}{\Gamma(1-\alpha)}t^{1-\r}\frac{d}{dt}\int_a^t \frac{\t^{\r-1}}{(t^\r-\t^\r)^\a}[x(\t)-x(a)] \, d\tau.$$
This was motivated by the recent notion due to Katugampola in [@Katugampola1], where a new fractional integral operator ${I_{a+}^{\a,\r}}x(t)$ is presented, $${I_{a+}^{\a,\r}}x(t)=\frac{\r^{1-\a}}{\Gamma(\alpha)}\int_a^t \frac{\t^{\r-1}}{(t^\r-\t^\r)^{1-\a}}x(\t) d\tau.$$ When $\a\in\mathbb N$, the fractional integral reduces to a n-fold integral of the form $$\int_a^t \t_1^{\r-1}\, d\t_1\,\int_a^{\t_1} \t_2^{\r-1}\, d\t_2\,\ldots \int_a^{\t_{n-1}} \t_n^{\r-1}x(\t_n)\, d\t_n.$$
If $x$ is continuously differentiable, then the fractional operator can be written in an equivalent way [@Almeida1]: $$\LD x(t) =\frac{\r^\a}{\Gamma(1-\alpha)}\int_a^t \frac{1}{(t^\r-\t^\r)^\a}x'(\t)\, d\tau,$$ that is, we have $$\LD x(t) ={I_{a+}^{1-\a,\r}}\left(\t^{1-\r}\frac{d}{d\t}x\right)(t).$$ The main results of [@Almeida1] are:
1. the operator is linear and bounded from $C([a,b])$ to $C([a,b])$,
2. if $x\in C^1[a,b]$, then the map $t\mapsto \LD x(t)$ is continuous in $[a,b]$ and $\LD x(a)=0$,
3. if $x$ is continuous, then $\LD {I_{a+}^{\a,\r}} x(t)=x(t)$,
4. if $x\in C^1[a,b]$, then ${I_{a+}^{\a,\r}}\, \LD x(t)=x(t)-x(a)$.
We remark that, taking $\r=1$, we obtain the Caputo fractional derivative, $${^CD_{a+}^\a}x(t)={^CD_{a+}^{\a,1}}x(t),$$ and doing $\r\to0^+$, then we get the Caputo–Hadamard fractional derivative: $${^HD_{a+}^\a}x(t)={^CD_{a+}^{\a,0^+}}x(t).$$
One crucial result for our present work is an integration by parts formula. For that, we need the two following auxiliary operators, a fractional integral type $${I_{b-}^{\a,\r}} x(t)=\frac{\r^{1-\a}}{\Gamma(\alpha)}\int_t^b \frac{1}{(\t^\r-t^\r)^{1-\a}}x(\t)\,d\tau,$$ and a fractional differential type $$\RD x(t)=\frac{\r^\a}{\Gamma(1-\alpha)}\frac{d}{dt}\int_t^b \frac{1}{(\t^\r-t^\r)^\a}x(\t)\,d\tau.$$
\[IntegrationParts\] Let $x$ be a continuous function and $y$ be a function of class $C^1$. Then, the following equality holds: $$\int_a^b x(t) \LD y(t) \, dt=\left[y(t) \RI x(t)\right]_{t=a}^{t=b}-\int_a^b y(t)\RD x(t) \, dt.$$
Starting with the definition, we have $$\int_a^b x(t) \LD y(t) \, dt=\frac{\r^\a}{\Gamma(1-\alpha)}\int_a^b\int_a^t x(t) \frac{1}{(t^\r-\t^\r)^\a}y'(\t)\, d\tau\,dt.$$ Using the Dirichlet’s formula, we get $$\int_a^b\int_a^t x(t) \frac{1}{(t^\r-\t^\r)^\a}y'(\t)\, d\tau\,dt=\int_a^b\int_t^b x(\t) \frac{1}{(\t^\r-t^\r)^\a}y'(t)\, d\tau\,dt.$$ Integrating by parts, considering $$u'(t)=y'(t) \quad \mbox{and} \quad v(t)=\int_t^b x(\t) \frac{1}{(\t^\r-t^\r)^\a}\, d\tau,$$ we obtain the desired formula.
The variational problem {#FCV}
=======================
Fractional calculus of variations appeared in 1996, with the work ok Riewe [@Riewe], since as he explained, “Traditional Lagrangian and Hamiltonian mechanics cannot be used with nonconservative forces such as friction”. Since then, several studies have appeared, for different types of fractional derivatives and/or fractional integrals, namely to determine necessary and sufficient conditions that any extremal for the variational functional must satisfy.
To start, we present the concept of minimizer for a given functional. On the space $C^1[a,b]$, consider the norm $\|\cdot\|$ given by $$\|x\| =\max_{t\in[a,b]}|x(t)|+ \max_{t\in[a,b]}\left|\LD x(t)\right|.$$ Let $D\subseteq C^1[a,b]$ be a nonempty set and $J$ a functional defined on $D$. We say that $x$ is a local minimizer of $J$ in the set $D$ if there exists a neighborhood $V_\delta(x)$ such that for all $x^* \in V_\delta(x)\cap D$, we have $J(x)\leq J(x^*)$. Note that any function $x^*\in V_\delta(x)\cap D$ can be represent in the form $x^*=x+\epsilon h$, where $|\epsilon| \ll 1$ and $h$ is such that $x+\epsilon h\in D$.
The purpose of this work is to study fractional calculus of variations problems, where the integral functional depends on the Caputo–Katugampola fractional derivative. Given $x\in C^1[a,b]$, consider the functional $$\label{functional}J(x)=\int_a^b L(t,x(t),\LD x(t))\,dt,$$ with the following assumptions:
1. $L:[a,b]\times \mathbb R^2\to\mathbb R$ is continuously differentiable with respect to the second and third arguments;
2. given any function $x$, the map $t\mapsto \RD(\partial_3L (t,x(t),\LD x(t)))$ is continuous.
Here, and along the work, given a function with several independent variables $f:A\subseteq \mathbb R^n\to\mathbb R$, we denote $$\partial_i f(x_1,\ldots,x_n):=\frac{\partial f}{\partial x_i}(x_1,\ldots,x_n).$$
We remark that $x$ may be fixed or free at $t=a$ and $t=b$. Both cases will be considered later.
The fundamental problem {#sec:fund}
-----------------------
We seek necessary and sufficient conditions that each extremizers of the functional must fulfill. In order to obtain such equations we consider variations of the solutions and use the fact that the first variation of the functional must vanish at the minimizer.
Let $x$ be a minimizer of the functional $J$ as in , defined on the subspace $$U=\left\{ x\in C^1[a,b]\, : \, x(a)=x_a \quad \mbox{and} \quad x(b)=x_b \right\},$$ where $x_a,x_b\in\mathbb R$ are fixed. Then, $x$ is a solution for the fractional differential equation $$\label{ELequation}\partial_2L(t,x(t),\LD x(t))-\RD \left(\partial_3L(t,x(t),\LD x(t))\right)=0$$ on $[a,b]$.
Let $x+\epsilon h$ be a variation of $x$, with $|\epsilon| \ll 1$ and $h\in C^1[a,b]$. Since $x+\epsilon h$ must belong to the set $U$, the boundary conditions $h(a)=0=h(b)$ must hold. Define the function $j$ in a neighborhood of zero as $$j(\epsilon)=J(x+\epsilon h).$$ Since $x$ is a minimizer of $J$, then $\epsilon=0$ is a minimizer of $j$ and so $j'(0)=0$. Computing $j'(0)$ and using Theorem \[IntegrationParts\], we get $$\int_a^b\left[\partial_2L(t,x(t),\LD x(t))-\RD \left(\partial_3L(t,x(t),\LD x(t))\right)\right] h(t)\,dt$$ $$\label{aux1}+\left[h(t) \RI \left(\partial_3L(t,x(t),\LD x(t))\right)\right]_{t=a}^{t=b}=0.$$ Since $h(a)=0=h(b)$ and $h$ is arbitrary elsewhere, we conclude that $$\partial_2L(t,x(t),\LD x(t))-\RD \left(\partial_3L(t,x(t),\LD x(t))\right)=0, \quad \forall t\in[a,b].$$
Eq. is called the Euler–Lagrange equation associated to functional $J$. The solutions of this fractional differential equation are called extremals of $J$.
The case of several dependent variables is similar, and the Euler–Lagrange equations are easily deduced. Let $$J(\overline x)=\int_a^b L(t,\overline x(t),\LD \overline x(t))\,dt,$$ where
1. $\overline x(t)=(x_1(t),\ldots,x_n(t))$ and $\LD\overline x(t)=(\LD x_1(t),\ldots,\LD x_n(t))$;
2. $L:[a,b]\times \mathbb R^{2n}\to\mathbb R$ is continuously differentiable with respect to its $i$th argument, for all $i\in\{2,\ldots, 2n+1\}$;
3. given any function $x$, the maps $t\mapsto \RD(\partial_{n+i}L (t,x(t),\LD x(t)))$ is continuous, for all $i\in\{2,\ldots, n+1\}$.
In this case, if $x$ is a minimizer of the functional $J$, subject to the restrictions $\overline x(a)= \overline x_a$ and $\overline x(b)= \overline x_b$, where $ \overline x_a,\overline x_b\in\mathbb R^n$ are fixed, then $x$ is a solution of $$\partial_{i}L(t,x(t),\LD x(t))-\RD \left(\partial_{n+i}L(t,x(t),\LD x(t))\right)=0,$$ for all $i\in\{2,\ldots, n+1\}$.
We remark that Eq. gives only a necessary condition. To deduce a sufficient condition, we recall the notion of convex function. Given a function $L(\underline t,x,y)$ continuously differentiable with respect to the second and third arguments, we say that $L$ is convex in $S\subseteq\mathbb R^3$ if condition $$L(t,x+x_1,y+y_1)-L(t,x,y)\geq\partial_2 L(t,x,y)x_1+\partial_3 L(t,x,y)y_1$$ holds for all $(t,x,y),(t,x+x_1,y+y_1)\in S$.
\[Teo:suffi\] If the function $L$ as in is convex in $[a,b]\times\mathbb R^2$, then each solution of the fractional Euler–Lagrange equation minimizes $J$ in $U$.
Let $x$ be a solution of Eq. and $x+\epsilon h$ be a variation of $x$, with $|\epsilon| \ll 1$ and $h\in C^1[a,b]$ with $h(a)=0=h(b)$. Then $$\begin{array}{ll}
J(x+\epsilon h)-J(x)&= \displaystyle\int_a^b \left[L(t,x(t)+\epsilon h(t),\LD x(t)+\epsilon \LD h(t))-L(t,x(t),\LD x(t))\right] \, dt \\
& \geq \displaystyle \int_a^b \left[ \partial_2 L(t,x(t),\LD x(t))\epsilon h(t)+ \partial_3 L(t,x(t),\LD x(t))\epsilon \LD h(h) \right] \, dt\\
&= \displaystyle\int_a^b \left[ \partial_2 L(t,x(t),\LD x(t)) -\RD ( \partial_3 L(t,x(t),\LD x(t)))\right] \epsilon h(t)\,dt\\
&=0
\end{array}$$ since $x$ is a solution of (\[ELequation\]). Therefore, $x$ is a local minimizer of $J$.
If we do not impose any restrictions on the boundaries, we obtain two transversality conditions.
Let $x$ be a minimizer of the functional $J$ as in . Then, $x$ is a solution for the fractional differential equation $$\partial_2L(t,x(t),\LD x(t))-\RD \left(\partial_3L(t,x(t),\LD x(t))\right)=0, \quad t\in[a,b].$$ If $x(a)$ is free, then $$\RI \left(\partial_3L(t,x(t),\LD x(t))\right) \quad \mbox{at} \quad t=a.$$ If $x(b)$ is free, then $$\RI \left(\partial_3L(t,x(t),\LD x(t))\right) \quad \mbox{at} \quad t=b.$$
Since $x$ is a minimizer, then $$\partial_2L(t,x(t),\LD x(t))-\RD \left(\partial_3L(t,x(t),\LD x(t))\right)=0,$$ for all $t\in[a,b]$. Therefore, using Eq. , we obtain $$\left[h(t) \RI \left(\partial_3L(t,x(t),\LD x(t))\right)\right]_{t=a}^{t=b}=0.$$ If $x(a)$ is free, then $h(a)$ is also free and taking $h(a)\not=0$ and $h(b)=0$, we get $$\RI \left(\partial_3L(t,x(t),\LD x(t))\right) \quad \mbox{at} \quad t=a.$$ The second case is similar.
Observe that in the previous results, the initial point of the cost functional coincides with the initial point of the fractional derivative. Next we consider a more general type of problems, by considering $A\in(a,b)$ and the functional $$\label{funcA}J(x)=\int_A^b L(t,x(t),\LD x(t))\,dt,$$ with the same assumptions on $L$ as previously, defined on the set $$U_A=\left\{ x\in C^1[a,b]\, : \, x(A)=x_A \quad \mbox{and} \quad x(b)=x_b \right\}.$$
If $x$ is a minimizer of the functional $J$ as in , then $x$ satisfies the fractional differential equations $${D_{A-}^{\a,\r}}\left(\partial_3L(t,x(t),\LD x(t))\right)-\RD \left(\partial_3L(t,x(t),\LD x(t))\right)=0 \quad \mbox{on} \, [a,A],$$ $$\partial_2L(t,x(t),\LD x(t))-\RD \left(\partial_3L(t,x(t),\LD x(t))\right)=0 \quad \mbox{on} \, [A,b],$$ and the transversality condition $${I_{A-}^{1-\a,\r}}\left(\partial_3L(t,x(t),\LD x(t))\right)-\RI \left(\partial_3L(t,x(t),\LD x(t))\right)=0 \quad \mbox{at} \quad t=a.$$
Let $x+\epsilon h$ be a variation of $x$, with $|\epsilon| \ll 1$, and $h\in C^1[a,b]$ with $h(A)=0=h(b)$. Computing the first variation of $J$, we deduce the following $$\begin{array}{ll}
0&=\displaystyle \int_A^b\left[\partial_2L(t,x(t),\LD x(t))h(t)+\partial_3L(t,x(t),\LD x(t))\LD h(t)\right]\,dt\\
&=\displaystyle \int_a^b\left[\partial_2L(t,x(t),\LD x(t))h(t)+\partial_3L(t,x(t),\LD x(t))\LD h(t)\right]\,dt\\
&\displaystyle\quad -\int_a^A\left[\partial_2L(t,x(t),\LD x(t))h(t)+\partial_3L(t,x(t),\LD x(t))\LD h(t)\right]\,dt.
\end{array}$$ Integrating by parts, and using the fact that $h(A)=0=h(b)$, we arrive at $$\int_a^A\left[{D_{A-}^{\a,\r}}\left(\partial_3L(t,x(t),\LD x(t))\right)-\RD \left(\partial_3L(t,x(t),\LD x(t))\right)\right] h(t)\,dt$$ $$+ \int_A^b\left[\partial_2L(t,x(t),\LD x(t))-\RD \left(\partial_3L(t,x(t),\LD x(t))\right)\right] h(t)\,dt$$ $$+h(a)\left[{I_{A-}^{1-\a,\r}}\left(\partial_3L(t,x(t),\LD x(t))\right)-\RI \left(\partial_3L(t,x(t),\LD x(t))\right)\right]^{t=a}=0.$$ Since $h$ is an arbitrary function, we obtain the three necessary conditions.
One interesting question is to determine the best type of fractional derivative for which the functional attains the minimum possible value. The Caputo–Katugampola fractional derivative depends on an extra parameter $\r$, and we can obtain e.g. the Caputo and the Caputo–Hadamard fractional derivatives when $\r=1$ and $\r\to0^+$, respectively. Thus, we are interested now to determine not only the minimizer $x$ but also the value of $\r$ for which $J$ attains its minimum value.
Let $(x,\r)$ be a minimizer of the functional $J_\r$ given by $$J_\r(x,\r)=\int_a^b L(t,x(t),\LD x(t))\,dt,$$ defined on the subspace $U\times\mathbb R^+$, where $$U=\left\{ x\in C^1[a,b]\, : \, x(a)=x_a \quad \mbox{and} \quad x(b)=x_b \right\}.$$ Then, $(x,\r)$ is a solution for the fractional differential equation $$\partial_2L(t,x(t),\LD x(t))-\RD \left(\partial_3L(t,x(t),\LD x(t))\right)=0$$ on $[a,b]$, and satisfies the integral equation $$\int_a^b \partial_3L(t,x(t),\LD x(t)) \psi'(\r)\,dt=0,$$ where $\psi(\r)=\LD x(t)$.
Let $(x+\epsilon h,\r+\epsilon\r_0)$ be a variation of $(x,\r)$, with $|\epsilon| \ll 1$, $h\in C^1[a,b]$ with $h(a)=0=h(b)$ and $\r_0\in\mathbb R$. If we define $j$ as $$j(\epsilon)=J(x+\epsilon h,\r+\epsilon\r_0),$$ then $$0=j'(0)=\int_a^b\left[\partial_2L(t,x(t),\LD x(t))-\RD \left(\partial_3L(t,x(t),\LD x(t))\right)\right] h(t)\,dt$$ $$+\int_a^b \partial_3L(t,x(t),\LD x(t)) \psi'(\r)\r_0\,dt.$$ If we consider $\r_0=0$, by the arbitrariness of $h$ on $(a,b)$, we conclude that $$\partial_2L(t,x(t),\LD x(t))-\RD \left(\partial_3L(t,x(t),\LD x(t))\right)=0, \quad \forall t\in[a,b].$$ So, $$\int_a^b \partial_3L(t,x(t),\LD x(t)) \psi'(\r)\r_0\,dt=0.$$ Taking $\r_0=1$, we get the second necessary condition.
The Legendre condition {#sec:legendre}
----------------------
The Legendre condition is a second-order condition which an extremal must fulfill in order to be a minimizer of the functional. We suggest [@Lazo] where a similar problem is solved for functionals depending on the Riemann–Liouville fractional derivative. To simplify notation, we introduce the following $$\partial^2_{ij} f(x_1,\ldots,x_n):=\frac{\partial^2f}{\partial x_i\partial x_j}(x_1,\ldots,x_n).$$
Let $x$ be a minimizer of the functional $J$ as in , defined on the subspace $U$. If $\partial^2_{ij}L $ exists and is continuous for $i,j\in\{2,3\}$, then $x$ satisfies $$\label{2nd_equation}\partial^2_{33}L(t,x(t),\LD x(t))\geq0$$ on $[a,b]$.
Let $x+\epsilon h$ be a variation of $x$, with $h\in C^1[a,b]$ such that $h(a)=0=h(b)$. If we define $$j(\epsilon)=J(x+\epsilon h),$$ then we must have $j''(\epsilon)\geq0$, that is $$\int_a^b\left[\partial^2_{22}L(t,x(t),\LD x(t))h^2(t)+2\partial^2_{23}L(t,x(t),\LD x(t))h(t)\LD h(t)\right.$$ $$\label{2nd}\left.+\partial^2_{33}L(t,x(t),\LD x(t))(\LD h(t))^2\right] \,dt\geq0.$$ Assume that there exists some $t_0 \in[a,b]$ for which $$\partial^2_{33}L(t_0,x(t_0),\LD x(t_0))<0.$$ Thus, there exists a subinterval $[c,d]\subseteq[a,b]$ and three real constants $C_1,C_2,C_3$ with $C_3<0$ such that $$\label{2nd_C}\partial^2_{22}L(t,x(t),\LD x(t))<C_1, \quad \partial^2_{23}L(t,x(t),\LD x(t))<C_2, \quad \partial^2_{33}L(t,x(t),\LD x(t))<C_3,$$ for all $t\in[c,d]$. Define the function $\underline h:[c,d]\to\mathbb R$ by $$\underline h(t)=(\a+2)(t^\r-c^\r)^{1+\a}-2\frac{\a+4}{d^\r-c^\r}(t^\r-c^\r)^{2+\a}+\frac{\a+10}{(d^\r-c^\r)^2}(t^\r-c^\r)^{3+\a}
-\frac{4}{(d^\r-c^\r)^3}(t^\r-c^\r)^{4+\a}.$$ Then, $\underline h$ is of class $C^1$, $\underline h(c)=0=\underline h(d)$, $\underline{h}'(c)=0=\underline h'(d)$ and ${^CD_{c+}^{\a,\r}}\underline h(c)=0={^CD_{c+}^{\a,\r}}\underline h(d)$. Also, we have the following upper bounds: $$\underline h(t)\leq 28(d^\r-c^\r)^{1+\a} \quad \mbox{and} \quad {^CD_{c+}^{\a,\r}}\underline h(t)\leq C(d^\r-c^\r), \, C=50 \r^\a\Gamma(2+\a),$$ for all $t\in[c,d]$. Thus, the function $h:[a,b]\to\mathbb R$ defined by $$h(t)=\left\{\begin{array}{ll}
\underline h(t) & \quad \mbox{ if } \, t\in[c,d]\\
0 & \quad \mbox{otherwise}
\end{array}\right.$$ is of class $C^1$, $h(a)=0=h(b)$ and its fractional derivative $$\LD h(t)=\left\{\begin{array}{ll}
{^CD_{c+}^{\a,\r}} \underline h(t) & \quad \mbox{ if } \, t\in[c,d]\\
0 & \quad \mbox{otherwise}
\end{array}\right.$$ is continuous. Inserting this variation $h$ into the second order condition , and using relations , we obtain $$\int_a^b\left[\partial^2_{22}L(t,x(t),\LD x(t))h^2(t)+2\partial^2_{23}L(t,x(t),\LD x(t))h(t)\LD h(t)\right.$$ $$\left.+\partial^2_{33}L(t,x(t),\LD x(t))(\LD h(t))^2\right] \,dt$$ $$=\int_c^d\left[\partial^2_{22}L(t,x(t),\LD x(t))h^2(t)+2\partial^2_{23}L(t,x(t),\LD x(t))h(t)\LD h(t)\right.$$ $$\left.+\partial^2_{33}L(t,x(t),\LD x(t))(\LD h(t))^2\right] \,dt$$ $$\leq \int_c^d \left[C_1 28^2(d^\r-c^\r)^{2+2\a}+C_2 56 C(d^\r-c^\r)^{2+\a}+C_3 C^2(d^\r-c^\r)^{2}\right]\,dt$$ $$=(d^\r-c^\r)^{2}(d-c)\left[C_1 28^2(d^\r-c^\r)^{2\a}+C_2 56 C(d^\r-c^\r)^{\a}+C_3 C^2\right],$$ which is negative if $d^\r-c^\r$ is chosen arbitrarily small, and thus we obtain a contradiction.
The isoperimetric problem {#sec:iso}
-------------------------
The isoperimetric problem is an old question, and was considered first by the ancient Greeks. It is related to finding a closed plane curve with a fixed perimeter $l$ which encloses the greatest area, that is, if we consider the place curve with parametric equations $(x(t),y(t))$, $t \in [a,b]$, then we wish to maximize the functional $$J(x,y)=\frac12\int_a^b (x(t)y'(t)-x'(t)y(t))\,dt,$$ under the boundary restrictions $$x(a)=x(b) \quad \mbox{and} \quad y(a)=y(b),$$ and the isoperimetric constraint $$\int_a^b\sqrt{(x')^2(t)+(y')^2(t)}\,dt=l.$$ Only in 1841, a rigorous proof that the solution is a circle was obtained by Steiner.
Nowadays, any variational problem that involves an integral constraint is called an isoperimetric problem. For the following, let $l\in\mathbb R$ be fixed, $g:[a,b]\times \mathbb R^2\to\mathbb R$ be continuously differentiable with respect to the second and third arguments such that, for any function $x\in C^1[a,b]$, the map $t\mapsto \RD(\partial_3g (t,x(t),\LD x(t)))$ is continuous. The integral constraint that we will consider is the following: $$\label{isoperi}I(x)=\int_a^b g(t,x(t),\LD x(t))\,dt=l.$$
Let $x$ be a minimizer of the functional $J$ as in , defined on the subspace $$U=\left\{ x\in C^1[a,b]\, : \, x(a)=x_a \quad \mbox{and} \quad x(b)=x_b \right\},$$ subject to the additional restriction . If $x$ is not an extremal of $I$, then there exists a real $\lambda$ such that, defining the function $K:[a,b]\times \mathbb R^2\to\mathbb R$ by $K=L+\lambda g$, $x$ is a solution of the equation $$\label{eq:EL_ISO}\partial_2K(t,x(t),\LD x(t))-\RD \left(\partial_3K(t,x(t),\LD x(t))\right)=0$$ on $[a,b]$.
Consider a variation of $x$ with two parameters $x+\epsilon_1 h_1+\epsilon_2 h_2$, with $|\epsilon_i| \ll 1$ and $h_i\in C^1[a,b]$ satisfying $h_i(a)=0=h_i(b)$, for $i=1,2$. Define the functions $i$ and $j$ with two parameters $(\epsilon_1,\epsilon_2)$ in a neighborhood of zero as $$i(\epsilon_1,\epsilon_2)=I(x+\epsilon_1 h_1+\epsilon_2 h_2)-l,$$ and $$j(\epsilon_1,\epsilon_2)=L(x+\epsilon_1 h_1+\epsilon_2 h_2).$$ Using Theorem \[IntegrationParts\], we get $$\frac{\partial i}{\partial \epsilon_2}(0,0)=\int_a^b\left[\partial_2g(t,x(t),\LD x(t))-\RD \left(\partial_3g(t,x(t),\LD x(t))\right)\right] h_2(t)\,dt$$ $$+\left[h_2(t) \RI \left(\partial_3g(t,x(t),\LD x(t))\right)\right]_{t=a}^{t=b}=0.$$ Since $h_2(a)=0=h_2(b)$ and $x$ is not an extremal of $I$, there exists some function $h_2$ such that $$\frac{\partial i}{\partial \epsilon_2}(0,0)\not=0.$$ Thus, by the Implicit Function Theorem, there exists an unique function $\epsilon_2(\cdot)$ defined in a neighborhood of zero such that $i(\epsilon_1,\epsilon_2(\epsilon_1))=0$, that is, there exists a subfamily of variations that satisfy the isoperimetric constraint .
On the other hand, $(0,0)$ is a minimizer of $j$, under the restriction $i(\cdot, \cdot)=0$, and we just proved that $\nabla i(0,0)\not=0$. Appealing to the Lagrange Multiplier Rule, there exists a real $\lambda$ such that $\nabla(j+\lambda i)(0,0)=0$. Differentiating the map $$\epsilon_1\mapsto j(\epsilon_1,\epsilon_2)+\lambda i(\epsilon_1,\epsilon_2),$$ and putting $(\epsilon_1,\epsilon_2)=(0,0)$, we get $$\int_a^b\left[\partial_2K(t,x(t),\LD x(t))-\RD \left(\partial_3K(t,x(t),\LD x(t))\right)\right] h_1(t)\,dt$$ $$+\left[h_1(t) \RI \left(\partial_3K(t,x(t),\LD x(t))\right)\right]_{t=a}^{t=b}=0.$$ Using the boundary conditions $h_1(a)=0=h_1(b)$, we prove the desired.
We can include the case where $x$ is an extremal of $I$. In this case, we apply the general form of the Lagrange Multiplier Rule, that is, there exist two reals $\lambda_0$ and $\lambda$, not both zeros, such that if we define the function $K_0:[a,b]\times \mathbb R^2\to\mathbb R$ by $K_0=\lambda_0L+\lambda g$, $x$ is a solution of the equation $$\partial_2K_0(t,x(t),\LD x(t))-\RD \left(\partial_3K_0(t,x(t),\LD x(t))\right)=0$$ on $[a,b]$.
\[teo:suffi\_ISO4\] Suppose that the functions $L$ and $g$ as in and are convex in $[a,b]\times\mathbb R^2$, and let $\lambda\geq0$ be a real. Define the function $K=L+\lambda g$. Then, each solution $x$ of the fractional Euler–Lagrange equation minimizes $J$ in $U$, subject to the integral constraint .
First, observe that the function $K$ is convex. So, by Theorem \[Teo:suffi\], we conclude that $x$ minimizes $K$, that is, for all variations $x+\epsilon h$, we have $$\int_a^b L(t,x(t)+\epsilon h(t),\LD x(t)+\epsilon \LD h(t))\, dt+\int_a^b \lambda g(t,x(t)+\epsilon h(t),\LD x(t)+\epsilon \LD h(h)) \, dt$$ $$\geq \int_a^b L(t,x(t),\LD x(t))\, dt+\int_a^b \lambda g(t,x(t),\LD x(t)) \, dt.$$ Using the integral constraint, we obtain $$\int_a^b L(t,x(t)+\epsilon h(t),\LD x(t)+\epsilon \LD h(t))\, dt+l\geq \int_a^b L(t,x(t),\LD x(t))\, dt+l,$$ and so $$\int_a^b L(t,x(t)+\epsilon h(t),\LD x(t)+\epsilon \LD h(t))\, dt\geq \int_a^b L(t,x(t),\LD x(t))\, dt,$$ proving the desired.
Theorem \[teo:suffi\_ISO4\] remains valid under the assumptions that $L$ is convex in $[a,b]\times\mathbb R^2$, $\lambda\leq0$ and $g$ is concave in $[a,b]\times\mathbb R^2$, that is, $$g(t,x+x_1,y+y_1)-g(t,x,y)\leq\partial_2 g(t,x,y)x_1+\partial_3 g(t,x,y)y_1$$ holds for all $(t,x,y),(t,x+x_1,y+y_1)\in [a,b]\times\mathbb R^2$.
Holonomic constraints {#sec:holo}
---------------------
Consider the functional $J$ defined by $$\label{funct2}J(x_1,x_2)=\int_a^b L(t,x_1(t),x_2(t),\LD x_1(t),\LD x_2(t))\,dt,$$ on the space $$U=\left\{ (x_1,x_2)\in C^1[a,b]\times C^1[a,b]\, : \, (x_1(a),x_2(a))=x_a \quad \mbox{and} \quad (x_1(b),x_2(b))=x_b \right\},$$ with $x_a,x_b\in\mathbb R^2$ fixed. We are assuming that the Lagrangian verifies the two following conditions
1. $L:[a,b]\times \mathbb R^4\to\mathbb R$ is continuously differentiable with respect to its $i$th argument, for $i=2,3,4,5$;
2. given any function $x$, the maps $t\mapsto \RD(\partial_iL (t,x(t),\LD x(t)))$ are continuous, $i=4,5$.
We consider in this new variational problem an extra constraint (called in the literature as holomonic constraint). Assume that the admissible functions lie on the surface $$\label{rest2}g(t,x_1(t),x_2(t))=0,$$ where $g:[a,b]\times \mathbb R^2\to\mathbb R$ is continuously differentiable with respect to the second and third arguments. For simplicity, we denote $$x=(x_1,x_2) \quad \mbox{and} \quad \LD x=(\LD x_1,\LD x_2).$$
Let $x\in U$ be a minimizer of $J$ as in , under the restriction . If $$\partial g_3(t,x(t))\not=0 \quad \forall t\in[a,b],$$ then there exists a continuous function $\lambda:[a,b]\to\mathbb R$ such that $x$ is a solution of the two next equations $$\label{eq:EL:holomo}\partial_2L(t,x(t),\LD x(t))-\RD \left(\partial_4L(t,x(t),\LD x(t))\right)+\lambda(t) \partial _2 g(t,x(t))=0$$ and $$\partial_3L(t,x(t),\LD x(t))-\RD \left(\partial_5L(t,x(t),\LD x(t))\right)+\lambda(t) \partial _3 g(t,x(t))=0$$ on $[a,b]$.
Consider a variation of $x$ of type $x+\epsilon h$, with $|\epsilon| \ll 1$, and $h\in C^1[a,b]\times C^1[a,b]$ satisfying the boundary conditions $h(a)=(0,0)=h(b)$. Since $$\partial g_3(t,x(t))\not=0, \quad \forall t \in[a,b],$$ by the Implicit Function Theorem, there exists a subfamily of variations that satisfy the restriction , that is, there exists an unique function $h_2(\epsilon,h_1)$ such that $(x_1(t)+\epsilon h_1(t),x_2(t)+\epsilon h_2(t))$ satisfies . So, since for all $t\in[a,b]$, we have $$g(t,x_1(t)+\epsilon h_1(t),x_2(t)+\epsilon h_2(t))=0,$$ differentiating with respect to $\epsilon$ and putting $\epsilon=0$, we get $$\label{aux2} \partial g_2(t,x(t))h_1(t)+\partial g_3(t,x(t))h_2(t)=0 .$$ Define the function $$\label{aux3}\lambda(t)=-\frac{\partial_3L(t,x(t),\LD x(t))-\RD \left(\partial_5L(t,x(t),\LD x(t))\right)}{\partial g_3(t,x(t))}.$$ Using equations and , we obtain $$\label{aux4}\lambda(t)\partial g_2(t,x(t))h_1(t)=\left[\partial_3L(t,x(t),\LD x(t))-\RD \left(\partial_5L(t,x(t),\LD x(t))\right)\right]h_2(t).$$ On the other hand, since $x$ is a minimizer of $J$, the first variation of $J$ must vanish: $$\int_a^b \partial_2L(t,x(t),\LD x(t))h_1(t)+\partial_3L(t,x(t),\LD x(t))h_2(t)$$ $$+\partial_4L(t,x(t),\LD x(t))\LD h_1(t)+\partial_5L(t,x(t),\LD x(t))\LD h_2(t)\,dt=0.$$ Integrating by parts, we obtain $$\int_a^b\left[\partial_2L(t,x(t),\LD x(t))-\RD \left(\partial_4L(t,x(t),\LD x(t))\right)\right] h_1(t)$$ $$+\left[\partial_3L(t,x(t),\LD x(t))-\RD \left(\partial_5L(t,x(t),\LD x(t))\right)\right] h_2(t)\,dt=0.$$ Inserting Eq. into the integral, we get $$\int_a^b\left[\partial_2L(t,x(t),\LD x(t))-\RD \left(\partial_4L(t,x(t),\LD x(t))\right)+\lambda(t)\partial g_2(t,x(t))\right] h_1(t)\,dt=0,$$ and since $h_1$ is arbitrary, we have that for all $t\in[a,b]$, $x$ is a solution of the equation $$\partial_2L(t,x(t),\LD x(t))-\RD \left(\partial_4L(t,x(t),\LD x(t))\right)+\lambda(t)\partial g_2(t,x(t))=0.$$ Also, using Eq. , we obtain the second condition $$\partial_3L(t,x(t),\LD x(t))-\RD \left(\partial_5L(t,x(t),\LD x(t))\right)+\lambda(t)\partial g_3(t,x(t))=0.$$
Suppose that the function $L(\underline t,x_1,x_2,y_1,y_2)$ as in is convex in $[a,b]\times\mathbb R^4$, $g:[a,b]\times \mathbb R^2\to\mathbb R$ is continuously differentiable with respect to the second and third arguments with $$\partial g_3(t,x(t))\not=0, \quad \forall t \in[a,b],$$ and let $\lambda$ be given by Eq. . If $x$ is a solution of the fractional Euler–Lagrange equation , then $x$ minimizes $J$ in $U$, subject to the constraint .
If $x+\epsilon h$ is a variation of $x$, then $$\begin{array}{ll}
\displaystyle J(x+\epsilon h)-J(x)&\geq \displaystyle\int_a^b \left[ \partial_2 L(t,x(t),\LD x(t)) -\RD ( \partial_4 L(t,x(t),\LD x(t)))\right] \epsilon h_1(t)\\
\displaystyle&\quad+\left[ \partial_3 L(t,x(t),\LD x(t)) -\RD ( \partial_5 L(t,x(t),\LD x(t)))\right] \epsilon h_2(t)\,dt.\end{array}$$ Since the variation functions must satisfy the constraint , we have the following relation $$h_2(t)=-\frac{\partial g_2(t,x(t))h_1(t)}{\partial g_3(t,x(t))},$$ and using Eq. , we deduce $$\begin{array}{ll}
\displaystyle J(x+\epsilon h)-J(x)&\geq \displaystyle\int_a^b \left[\partial_2L(t,x(t),\LD x(t))-\RD \left(\partial_4L(t,x(t),\LD x(t))\right)\right.\\
\displaystyle&\quad\left.+\lambda(t) \partial _2 g(t,x(t))\right] \epsilon h_1(t)\, dt,\end{array}$$ which is zero by hypothesis.
The Herglotz problem {#sec:herglotz}
--------------------
The fractional Herglotz problem is described in the following way. Determine a curve $x\in C^1[a,b]$ subject to the boundary conditions $x(a)=x_a$ and $x(b)=x_b$, such that the solution $z$ of the system $$\label{Herglotz}
\left\{\begin{array}{l}
z'(t)=L(t,x(t),\LD x(t),z(t)), \quad t\in[a,b]\\
z(a)=z_a\\
\end{array}\right.$$ attains a minimum at $t=b$. The assumptions are
1. given any function $x$, the map $t\mapsto {^C_aD^\alpha_t}x(t)$ is continuously differentiable;
2. $L:[a,b]\times \mathbb R^3\to\mathbb R$ is continuously differentiable with respect to its $i$th argument, for $i=2,3,4$;
3. given any function $x$, the map $t\mapsto \RD(\lambda(t)\partial_3L (t,x(t),\LD x(t)),z(t))$ is continuous, where $$\label{lambda}\lambda(t)=\exp\left(-\int_a^t \partial _4L(\t,x(\t),\LD x(\t),z(\t))d\tau\right).$$
We note that, given $x$, system becomes an initial value problem $$\left\{\begin{array}{l}
z'(t)=f(t,z(t)), \quad t\in[a,b]\\
z(a)=z_a\\
\end{array}\right.$$ and so the solution depends on $t$ and on $x$, that is, $z=z[t,x]$. If we consider variations of $x$ of type $x+\epsilon h$ into Eq. , then the solution also depends on $\epsilon$, and it is differentiable with respect to $\epsilon$ (see Section 2.6 in [@Anosov]).
Let $x$ be such that $z(b)$ as in Eq. attains a minimum. Then, $x$ is a solution of the fractional differential equation $$\lambda(t)\partial_2L(t,x(t),\LD x(t),z(t))-\RD\left(\lambda(t)\partial_3L(t,x(t),\LD x(t),z(t))\right)=0$$ on $[a,b]$.
Let $x+\epsilon h$ be a variation of $x$, with $|\epsilon| \ll 1$ and $h\in C^1[a,b]$ such that $h(a)=0=h(b)$. The variation of $z$ is given by $$\theta (t)=\frac{d}{d\epsilon} \left.z[t,x+\epsilon h]\right|_{\epsilon=0}.$$ Inserting the variations into the differential equation , we obtain $$\frac{d}{dt}z[t,x+\epsilon h]=L(t,x(t)+\epsilon h(t),\LD x(t)+\epsilon\LD h(t),z[t,x+\epsilon h]).$$ Since $$\begin{array}{ll}
\theta'(t)&=\displaystyle \left.\frac{d}{d\epsilon} \frac{d}{dt}z[t,x+\epsilon h]\right|_{\epsilon=0}\\
&=\displaystyle\frac{d}{d\epsilon }
\left. L(t,x(t)+\epsilon h(t),\LD x(t)+\epsilon\LD h(t),z[t,x+\epsilon h])\right|_{\epsilon=0}\\
&=\displaystyle \partial _2L (t,x(t),\LD x(t),z(t))h(t)+\partial _3L (t,x(t),\LD x(t),z(t))\LD h(t)\\
&\quad\displaystyle+\partial _4L (t,x(t),\LD x(t),z(t))\theta(t),\\
\end{array}$$ we obtain a linear differential equation whose solution is given by $$\theta(t)\lambda(t)-\theta(a)$$ $$=\int_a^t \lambda(\t)\left[ \partial _2L (\t,x(\t),\LD x(\t),z(\t))h(\t)+\partial _3L (\t,x(\t),\LD x(\t),z(\t))\LD h(t) \right]d\tau,$$ and integrating by parts, we get $$\label{Herglotz2}\theta(t)\lambda(t)-\theta(a)$$ $$=\int_a^t \left[\lambda(\t) \partial _2L (\t,x(\t),\LD x(\t),z(\t))-\RD\left(\lambda(\t)\partial _3L (\t,x(\t),\LD x(\t),z(\t))\right)\right]h(\t)\,d\tau.$$ Since $z(a)$ is fixed, we get $\theta(a)=0$. Also, since $z(b)$ is minimum, then $\theta (b)=0$. Replacing $t=b$ in Eq. , and by the arbitrariness of $h$ in $(a,b)$, we obtain that $$\lambda(t)\partial_2L(t,x(t),\LD x(t),z(t))-\RD\left(\lambda(t)\partial_3L(t,x(t),\LD x(t)z(t))\right)=0$$ must hold for all $t\in[a,b]$.
Acknowledgments {#acknowledgments .unnumbered}
===============
I would like to thank Tatiana Odzijewicz, for a careful and thoughtful reading of the manuscript. This work was supported by Portuguese funds through the CIDMA - Center for Research and Development in Mathematics and Applications, and the Portuguese Foundation for Science and Technology (FCT-Fundação para a Ciência e a Tecnologia), within project UID/MAT/04106/2013.
[99]{}
R. Almeida, A.B. Malinowska and T. Odzijewicz, Fractional differential equations with dependence on the Caputo–Katugampola derivative, submitted.
R. Almeida and D.F.M. Torres, Calculus of variations with fractional derivatives and fractional integrals, Appl. Math. Lett. [**22**]{} (2009), 1816–1820.
D.V. Anosov, S.K. Aranson, V.I. Arnold, I.U. Bronshtein, V.Z. Grines and Y.S. Ilyashenko, Ordinary Differential Equations and Smooth Dynamical Systems, Encyclopaedia of Mathematical Sciences, Vol. 1, 1988.
T.M. Atanacković, S. Konjik and S. Pilipović, Variational problems with fractional derivatives: Euler-Lagrange equations, J. Phys. A [**41**]{} (2008), 095201, 12 pp.
D. Baleanu, T. Maaraba, F. Jarad, Fractional variational principles with delay, J. Phys. A [**41**]{} (2008), 315403, 8 pp.
D. Baleanu, S.I. Muslih and E. M. Rabei, On fractional Euler-Lagrange and Hamilton equations and the fractional generalization of total time derivative, Nonlinear Dynam. [**53**]{} (2008), 67–74.
D. Baleanu and J.I. Trujillo, A new method of finding the fractional Euler-Lagrange and Hamilton equations within Caputo fractional derivatives, Commun. Nonlinear Sci. Numer. Simul. [**15**]{} (2010), 1111–1115.
J. Cresson, Fractional embedding of differential operators and Lagrangian systems, J. Math. Phys. [**48**]{} (2007) 033504, 34 pp.
R.A. El-Nabulsi, D.F.M. Torres, Fractional actionlike variational problems, J. Math. Phys. [**49**]{} (2008) 053521, 7 pp.
G.S.F. Frederico and D.F.M. Torres, Fractional Noether’s theorem in the Riesz-Caputo sense, Appl. Math. Comput. [**217**]{} (2010), 1023–1033.
Y.Y Gambo, F. Jarad, D. Baleanu and T. Abdeljawad, On Caputo modification of the Hadamard fractional derivatives, Adv. Difference Equ. (2014), 2014:10 doi:10.1186/1687–1847–2014–10.
F. Jarad, T. Abdeljawad and D. Baleanu, Caputo-type modification of the Hadamard fractional derivatives, Adv. Difference Equ. (2012), 2012:142 doi:10.1186/1687–1847–2012–142.
G. Jumarie, From Lagrangian mechanics fractal in space to space fractal Schrödinger’s equation via fractional Taylor’s series, Chaos Solitons Fractals [**41**]{} (2009), 1590–1604.
G. Jumarie, An approach via fractional analysis to non-linearity induced by coarse-graining in space, Nonlinear Anal. Real World Appl. [**11**]{} (2010), 535–546.
U.N. Katugampola, New approach to a generalized fractional integral, Appl. Math. Comput. [ **218**]{} (2011), 860–865.
U.N. Katugampola, New approach to generalized fractional derivatives, Bull. Math. Anal. App. [**6**]{} (2014) 1–15.
A.A. Kilbas, H.M. Srivastava and J.J. Trujillo, Theory and Applications of Fractional Differential Equations. North-Holland Mathematics Studies, 204. Elsevier Science B.V., Amsterdam, 2006.
M.J. Lazo and D.F.M. Torres, The Legendre condition of the fractional calculus of variations, Optimization [**63**]{} (2014) 1157–1165.
A.B. Malinowska and D.F.M. Torres, Generalized natural boundary conditions for fractional variational problems in terms of the Caputo derivative, Comput. Math. Appl. [**59**]{} (2010), 3110–3116.
A.B. Malinowska and D.F.M. Torres, Multiobjective fractional variational calculus in terms of a combined Caputo derivative, Appl. Math. Comput. [**218**]{} (2012), 5099–5111.
A.B. Malinowska and D.F.M. Torres, Introduction to the Fractional Calculus of Variations, Imp. Coll. Press, London, 2012.
T. Odzijewicz, A.B. Malinowska and D.F.M. Torres, Fractional variational calculus with classical and combined Caputo derivatives, Nonlinear Anal. [**75**]{} (2012), 1507–1515.
F. Riewe, Nonconservative Lagrangian and Hamiltonian mechanics. Phys. Rev. E [**53**]{} (1996), 1890–1899.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Facial expression analysis based on machine learning requires large number of well-annotated data to reflect different changes in facial motion. Publicly available datasets truly help to accelerate research in this area by providing a benchmark resource, but all of these datasets, to the best of our knowledge, are limited to rough annotations for action units, including only their absence, presence, or a five-level intensity according to the Facial Action Coding System. To meet the need for videos labeled in great detail, we present a well-annotated dataset named FEAFA for Facial Expression Analysis and 3D Facial Animation. One hundred and twenty-two participants, including children, young adults and elderly people, were recorded in real-world conditions. In addition, 99,356 frames were manually labeled using Expression Quantitative Tool developed by us to quantify 9 symmetrical FACS action units, 10 asymmetrical (unilateral) FACS action units, 2 symmetrical FACS action descriptors and 2 asymmetrical FACS action descriptors, and each action unit or action descriptor is well-annotated with a floating point number between 0 and 1. To provide a baseline for use in future research, a benchmark for the regression of action unit values based on Convolutional Neural Networks are presented. We also demonstrate the potential of our FEAFA dataset for 3D facial animation. Almost all state-of-the-art algorithms for facial animation are achieved based on 3D face reconstruction. We hence propose a novel method that drives virtual characters only based on action unit value regression of the 2D video frames of source actors.'
address:
- 'School of Engineering Science, University of Chinese Academy of Sciences, Beijing 100049, China'
- 'Information Engineering College, Capital Normal University, Beijing 100048, China'
author:
- Yanfu Yan
- Ke Lu
- Jian Xue
- Pengcheng Gao
- Jiayi Lyu
bibliography:
- 'refs.bib'
title: 'FEAFA: A Well-Annotated Dataset for Facial Expression Analysis and 3D Facial Animation'
---
Facial action units ,facial expression dataset ,blendshape model ,facial animation
Introduction {#sec:introduction}
============
Humans communicate through the exchange of verbal and non-verbal messages, and facial expressions in fact play a key role in conveying non-verbal information between interactions. Therefore, facial expression analysis has become an increasingly important research area that can be used for many practical applications such as human-machine interaction [@Vinciarelli2009Social] and facial animation [@Cao2016Real]. Publicly available datasets are fundamental for accelerating research in facial expression analysis, because they not only require much effort to label massive numbers of video frames but also provide a benchmark through which researchers can compare their algorithm more objectively.
There are quite a few publicly available datasets for facial expression, such as the extended Cohn-Kanade (CK+) [@Lucey2010The], DIFSA [@Mavadati2013DISFA], MMI [@Bartlett2006Fully], AM-FED [@Mcduff2013Affectiva], SEMINA [@Mckeown2012The], BP4D [@Zhang2013A], and FERA [@Valstar2015FERA] datasets. All of these labeled datasets are based on the Facial Action Coding System (FACS) , which was initially developed in 1978 and was informed by earlier research. The FACS presents the most comprehensive catalogue of unique facial action units (AUs) and facial action descriptors (ADs) to describe all possible facial expressions. Each AU is anatomically based on facial actions that depend on a few facial muscles and may occur individually or in combination. In particular, ADs are not defined using specific facial muscles but can still represent some common facial motions. A recent version of FACS [@Ekman2002Facial] assigns the letters A through E to represent AUs intensity variation, from barely detectable to maximum intensity, thus enabling the measurement and scoring of facial expression in an objective and quantitative way. However, the majority of the available FACS-coded datasets only label the presence or absence for AUs. Only few datasets provide the AU intensity at five levels. Even though the intensity information is accessible, these AU annotations with five-level can only suitable for detection tasks or rough intensity estimation tasks. There have been lots of state-of-the-art algorithms for AU detection [@Wu2017Deep; @Li2017Action] and AU intensity estimation [@Walecki2016Copula; @Walecki2017Deep] based on existing datasets. However, these existing datasets are not accurate enough to describe subtle intensity variation and enable an AU value regression task.
![Sample images from collected videos in the FEAFA dataset. These video sequences have various lightning, scale, position and pose. Participants express emotions including happiness, sadness, surprise, fear, contempt, anger and disgust.[]{data-label="fig:sample"}](sample.png){width="0.95\columnwidth"}
The FACS has become a standard not only for the anatomically-based description of facial expression but also for muscle-based facial animation such as MPEG-4 facial animation parameters. Three-dimensional (3D) facial animation aims to capture the facial expression of source actors and then map the expression information to virtual characters. Impressive results in this area have been achieved, both based on RGB [@Cao2014Displaced; @Cao2016Real; @Thies2016Face2Face] and RGB-D [@Li2013Realtime] video data. However, almost all of these state-of-the-art techniques require a 3D shape reconstruction process, which is truly a time-consuming and tedious work. Because the most important information we want to obtain is facial expression parameters (to be more specific, AU values), we can in fact calculate them more conveniently from the 2D frames of a monocular video stream. Then we can use the AU values of the source actor and the corresponding avatar’s blendshape for each AU to generate animations using linear combinations based on a blendshape model. The regression of AU values is thus necessary. Given that the facial appearance of some AU changes depending on the other AUs and some AUs co-occur more often than others, we provide a baseline system to achieve joint AU value regression.
To meet the need of AU values regression, in this paper, we present a facial expression dataset named FEAFA with well-annotated AUs for the intensity with float numbers. The FEAFA dataset contains:
- **Facial videos:** 123 webcam videos of 122 participants recorded in real-world conditions.
- **Labeled Frames:** 99,356 frames manually labeled using the Expression Quantification Tool to obtain the intensity as a floating point number between 0 and 1, for 9 symmetrical FACS AUs, 10 asymmetrical (unilateral) FACS AUs, 2 symmetrical FACS ADs and 2 asymmetrical FACS ADs.
- **Baseline System:** The Baseline performance of a joint AU value regression algorithm using deep Convolutional Neural Networks (CNNs) on this dataset.
- **New technique for 3D facial animation:** 3D facial animation of a virtual character that mimics the user’s performance only using the AU value regression of a 2D image rather than 3D shape reconstruction.
To facilitate future research on facial expression related algorithms, we have made the database and some demos available at <http://www.iiplab.net/feafa/>. Some sample images of the dataset are shown in [Figure \[fig:sample\]]{}. The rest of this paper is organized as follows: Section \[sec:existingdatasets\] reviews some existing datasets related to our work; Section \[sec:thefeafa\] describes the redefined AUs, data collection and annotation; Section \[sec:baseline\] presents a baseline system to regress AU values; and Section \[sec:experiments\] shows training, testing, and baseline performance of 24 regressed AU values on FEAFA. In Section \[sec:applications\], we demonstrate two major applications based on this dataset. And finally Section \[sec:conclusions\] concludes the paper.
Existing Datasets {#sec:existingdatasets}
=================
The Cohn-Kanade dataset [@Lucey2010The] plays a key role in the facial expression analysis. An extension of the Cohn-Kanade dataset (called CK+) contains 123 subjects and 593 posed and spontaneous sequences (10,708 frames), whose peak frames are AU coded with 30 AUs and given either intensity degree on a 7-point ordinal scale or an unspecified intensity. These videos are recorded under controlled light conditions and head motions.
The DISFA dataset [@Mavadati2013DISFA] contains 54 sequences of 27 young adults and participants. The participants generate facial expressions while watching emotive video clips. Each frame of the sequences (130,000 frames) was manually labeled for 12 AUs as presence, absence and intensity according the FACS. The naturalistic and spontaneous expressions of the facial behavior is imaged using a high resolution stereo-vision system under uniform illumination.
The MMI dataset [@Bartlett2006Fully] contains 25 subjects and 1,395 manually AU-coded sequences. The MMI labels consist of AU intensity for onset (start frame), apex (maximum intensity), and offset (end frame) but the dataset does not include intensity labels for each frame. MMI has many posed expressions and some spontaneous expressions, and both are recorded in a lab setting.
**AU** **Our Definition** **FACS No. and Definition** **AU** **Our Definition** **FACS No. and Definition**
-------- ---------------------- -------------------------------------- -------- -------------------------- -----------------------------------
1 Left Eye Close AU43[ ]{}[ ]{}Eye Closure 13 Right Lip Corner Pull AU12[ ]{}[ ]{}Lip Corner Puller
2 Right Eye Close AU43[ ]{}[ ]{}Eye Closure 14 Left Lip Corner Stretch AU20[ ]{}[ ]{}Lip stretcher
3 Left Lid Raise AU 5[ ]{}[ ]{}[ ]{}Upper lid raiser 15 Right Lip Corner Stretch AU20[ ]{}[ ]{}Lip strecher
4 Right Lid Raise AU 5[ ]{}[ ]{}[ ]{}Upper lid raiser 16 Upper Lip Suck AU28[ ]{}[ ]{}Lip Suck
5 Left Brow Lower AU 4[ ]{}[ ]{}[ ]{}Brow lowerer 17 Lower Lip Suck AU28[ ]{}[ ]{}Lip Suck
6 Right Brow Lower AU 4[ ]{}[ ]{}[ ]{}Brow lowerer 18 Jaw Thrust AD29[ ]{}[ ]{}Jaw Thrust
7 Left Brow Raise AU 2[ ]{}[ ]{}[ ]{}Outer brow raiser 19 Upper Lip Raise AU10[ ]{}[ ]{}Upper Lip Raiser
8 Right Brow Raise AU 2[ ]{}[ ]{}[ ]{}Outer brow raiser 20 Lower Lip Depress AU16[ ]{}[ ]{}Lower Lip Depressor
9 Jaw Drop AU26[ ]{}[ ]{}Jaw Drop 21 Chin Raise AU17[ ]{}[ ]{}Chin Raiser
10 Jaw Slide Left AD30[ ]{}[ ]{}Jaw Sideways 22 Lip Pucker AU18[ ]{}[ ]{}Lip Pucker
11 Jaw Slide Right AD30[ ]{}[ ]{}Jaw Sideways 23 Cheeks Puff AD34[ ]{}[ ]{}Puff
12 Left Lip Corner Pull AU12[ ]{}[ ]{}Lip Corner Puller 24 Nose Wrinkle AU 9[ ]{}[ ]{}[ ]{}Nose wrinkler
The AM-FED dataset [@Mcduff2013Affectiva] contains 242 AU related sequences (168,359 frames) of 242 subjects labeled frame by frame for the presence of 10 symmetrical AUs and four asymmetrical AUs. The AM-FED dataset records ecologically spontaneous facial responses online in real-world conditions. The SEMINA dataset [@Mckeown2012The] consists of 150 participants recorded in laboratory settings for a total of 959 conversations with individual SAL (Sensitive Artificial Listener) characters. Moreover, 180 images from eight sessions are FACS coded. The BP4D dataset [@Zhang2013A] contains 41 young adults generating spontaneous expression from stimulus tasks conducted in a lab environment. These expression sequences are labeled with the onset and offset of 27 AUs independently for each condition. The RU-FACS dataset [@Bartlett2006Automatic] contains video of the spontaneous behavior of 100 participants engaging in a 2.5 minute interview. The Bosphorus dataset [@Aly20083D] consists of 105 participants and the frames are labeled using FACS. The UNBC-McMaster Pain Archive [@Lucey2011Painful] provides naturalistic and spontaneous facial expressions with pain. This dataset is labeled for 10 AUs and the pain intensity is coded. The camera position, lighting, and resolution are controlled.
Most of the datasets reviewed above are coded for AU presence or absence and were collected in laboratory settings; only a few are coded with the five intensity levels introduced by the FACS and recorded in natural settings. To provide more accurate AU annotations, we hence create the FEAFA dataset based on videos recorded in real-world conditions.
FEAFA Dataset {#sec:thefeafa}
=============
In this section, we redefine AUs based on the FACS, describe the collected video sequences, and show how the video frames are annotated in great detail.
Action Unit Based on FACS
-------------------------
We selected nine symmetrical FACS AUs, 10 asymmetrical (unilateral) FACS AUs, two symmetrical FACS ADs, and 2 asymmetrical FACS ADs to describe most expressions of the human face. To facilitate facial expression analysis based on our database, especially the blendshape process for 3D facial animation, we reorganized and renumbered the FACS AUs and ADs. Moreover, we refer to all of these facial actions as AUs for convenience. We also renamed the asymmetrical AUs, and, in particular, we also subdivided some AUs into upper and lower ones (e.g., Lip Suck). We hence obtained neutral expression and the 24 AUs to cover common facial expressions that vary among different individuals. The definition of each AU and its corresponding FACS AU or FACS AD are showed in Table \[tab:aulist\].
Data Collection
---------------
The facial behavior of 122 individuals (62 men and 60 women) was recorded by ordinarily monocular web cameras. We made Facial Expression Data Recording and Labeling Protocol (available at our website) for the participants to follow when their facial behaviors were recorded by the web cameras. This protocol, which will be available with the dataset at our website later, not only includes how to label the AUs in every image but also illustrates how to elicit the required expression. Besides, we trained every participant in order to avoid the problem of the mixture of children and adults. We required that facial area of the participants account for at least 100,000 pixels of the frame and the orientation of the faces should be constrained, with at most a 30 degree deviation from a frontal view. In addition, these participants were all from Asia, and the facial expressions they presented were specified only through the already defined AUs regardless of whether the expression was a simple single expression or complex combination of expressions. Furthermore, these videos were recorded under real-world conditions and thus exhibit non-uniform lighting and non-uniform frame rates. We obtained a total of 123 final sequences which last 7 - 112 seconds; each of the sequences consists of 4 - 29 different facial expressions. From an AU value regression perspective, there are some challenges in our dataset such as the lighting of varied environment and differentiation of individuals’ expressions, even if they present completely similar expressions. Additional details of our dataset are provided in Table \[tab:detail\].
**[ ]{}Glasses[ ]{}** **Facial hair**
---------- ----------------------- ----------------- --------- ---------
(0, 15\] (15, 60\] (60, 80\] Present Present
4 110 8 40 38
: Age distribution, glasses wearing, and facial hair occurrences in the video sequences in our FEAFA dataset.[]{data-label="tab:detail"}
{width="98.00000%"}
AU Annotation
-------------
Each of the posed sequences were independently labeled frame by frame. Every frame was coded by at least three highly trained coders chosen from 20 coders. These coders manually labeled each AU we already defined and they worked independently. The reliability was also tested through coding to ensure consistency. AU annotations were then subsequently labeled by another independent FACS-trained individual and discrepancies within the coding were reviewed.
We require labels that are more precise than those of FACS which only has five levels for each action units. Hence, we use floating point numbers from 0 to 1 and accurate to two decimal places to quantify each AU. A facial action state that is close to a neutral state is given a corresponding AU value close to 0; larger deviations of the facial action state from the neutral state are given a corresponding AU value that is closer to 1. This special annotation method is instrumental in identifying AU values with expression coefficients. [Figure \[fig:annotation\]]{} gives examples of AU2 (Right Eye Close) and AU9 (Jaw Drop) annotation for cropped facial images.
To label AUs faster and more conveniently, we developed a software called Expression Quantification Tool (ExpreQuantTool), which is specifically designed for labeling data more efficiently. [Figure \[fig:screenshot\]]{} shows a screenshot of our software. On the right side of the software’s GUI, the generated expression blendshape of a virtual character mimicking the expression of the subject of the frame in the Expression Visualization window is shown.
![Screenshot of the ExpreQuantTool used to label the video sequences on the FEAFA dataset.[]{data-label="fig:screenshot"}](screenshot.png){width="0.98\columnwidth"}
Similar to many facial animation techniques, we represent facial expressions in terms of expression blendshapes. To generate any possible expression of the source actors, we need a neutral face blendshape and 24 expression blendshapes according to 24 predefined AUs as our base poses. [Figure \[fig:blendshapes\]]{} shows an example of the expression blendshapes selected from FaceWarehouse [@Cao2014FaceWarehouse]. Using these blendshapes $B = \{B_{0} , B_{1} , ..., B_{24}\}$, any facial expression of the subject of each recorded sequence can be easily replicated through linear combinations using the following equation. $$\label{eq:blendshape}
M=B_{0}+ \sum_{i=1}^{24}\beta_{i}(B_{i}-B_{0}),
$$ where $\beta =\{\beta_{1}, \beta_{2}, ..., \beta_{24}\}$ is a vector of expression coefficient and $B_{0}$ denotes the neutral face.
![Example of the corresponding expression blendshape for each AU.[]{data-label="fig:blendshapes"}](blendshapes.png){width="0.98\columnwidth"}
When the coder labels the frame of certain expression of the video subjects displayed in the middle window, we can then obtain the AU values of the source actors, which actually are the expression coefficients of the virtual character, as well. Using the blendshapes and expression coefficients, we then generate the virtual character’s specific expression model. This tool is of immense help to the coders to annotate images more precisely. During labeling, our ExpreQuantTool presents a corresponding 3D facial expression model once coders set AU values for a video frame. We can hence compare the expression on the 3D facial model with the expression on the source image directly so that the influence of the subjectivity on the AU labels accuracy can be minimized. Meanwhile, this tool can help to control differences of annotated AU values for one image between coders to a certain range. Besides, the AU values will be further validated to eliminate obviously wrong annotations and then we take the average of valid AU values among coders. After the coders finish labeling all the frames of one video sequence, they can play this video and the generated expression model based on the blendshapes will play automatically and is thus instrumental in double checking the labels of the frames.
This process yielded a total of 99,356 well-annotated frames. The distribution of each AU in four intervals is listed in Table \[tab:distribution\].
**AU** **(0, 0.25\]** **(0.25, 0.5\]** **(0.5, 0.75\]** **(0.75, 1\]**
-------- ---------------- ------------------ ------------------ ----------------
1 13052 7105 2967 16199
2 12551 7163 2958 16263
3 1391 1190 960 2340
4 1384 1126 956 2396
5 3193 3044 2635 3359
6 2761 3248 2690 3423
7 2138 1816 888 3083
8 2141 1693 920 3210
9 7033 2581 2010 5720
10 921 889 552 1043
11 815 695 490 985
12 4726 4995 3023 2300
13 4800 4770 3082 2305
14 3492 3102 865 1186
15 3488 3084 895 1219
16 3237 2334 1649 3418
17 2860 1924 1285 3466
18 813 517 414 810
19 5321 3862 2319 3057
20 6942 5097 2512 3635
21 1706 1585 838 1730
22 2060 1917 1443 3275
23 657 669 435 2805
24 911 522 823 948
: Distribution of each AU on the FEAFA dataset. The range (0,1\] is evenly divided into four intervals to present our dataset in great detail.\[tab:distribution\]
Baseline System {#sec:baseline}
===============
Our baseline system consists of the complete pipeline of image pre-processing, feature representation, and AU value regression. We first use the dlib detector based on histogram of oriented gradient (HoG) features to track the face. We then use the dlib C++ library to detect crucially fiducial points and crop the video frames. Each cropped face is then passed through deep CNN architectures to compute the features required for the regression of AU values. Finally, we use a 3-layer neural network based on the regression model to estimate the AUs. An overview of our baseline system is given in [Figure \[fig:baseline\]]{}. These modules are described in following subsections.
{width="0.98\linewidth"}
Image Pre-processing
--------------------
HoG-based object detectors have become a classic computer vision method for detecting semi-rigid objects and have attracted attention in the face analysis field. Dlib’s HoG detector [@King2009Dlib] can represent both appearance and shape information. It is relatively accurate at multiple resolutions and multiple pose angles and detecting faces in unconstrained scenarios. Hence, we use the HoG-based dlib detector to track the face.
After obtaining facial bounding box, we then use the dlib C++ library of [@Kazemi2014One], which uses an ensemble of regression trees to locate the face’s landmark positions from a sparse subset of intensity values in milliseconds with high quality predictions. It optimizes an appropriate loss function and performs feature selection in a data-driven manner, specifically using the gradient tree boosting method while minimizing the same loss function during training and test. The cascade regression can be summarized using following equation. $$\label{eq:lib}
S^{(t+1)}=S^{(t)}+r_{t}(I, S^{(t)}),
$$ where $S$ denotes the facial shape. The shape estimate $S^{(t+1)}$ at the $(t + 1)^{th}$ stage can be written in terms of the shape estimate at the previous stage $S^{(t)}$ and shape regressor $r_{t}$, making its predictions based on features such as pixel intensity values.
Once the facial landmarks have been detected, we calculate the distance between the center of two pupils, which is used to locate the face ROI. We then crop the face image and rescale it according to the input of the network for the feature representation.
Feature Representation
----------------------
Various publicly available CNN architectures that have been trained on enormous datasets can be used as feature extractors. Once we obtain the cropped face image, we compute the facial features using AlexNet, VGGNet, and GoogleNet. We describe these CNNs in detail in the following subsections. For all three networks, we extract deep features from the last fully connected layer, which results in three feature vectors of length 1,000.
### AlexNet
AlexNet [@Krizhevsky2012ImageNet] requires 227 $\times$ 227 RGB images as input, which are obtained in the image pre-processing step. AlexNet is a relatively shallow network, consisting five convolutional (Conv) layers and three fully-connected (FC) layers. It employs kernels with large receptive fields in layers close to the input and smaller kernels close to the output, incorporates rectified linear units(ReLU) instead of the hyperbolic tangent as the activation function and uses normalization layers and the overlapping pooling technique to aid generalization.
### VGGNet
VGGNet [@Simonyan15] not only performs well in image classification tasks but also generates better visual representations. VGGNet is a large network that requires input size 224 $\times$ 224 $\times$ 3 pixels. Here, We use a 16-layer model often referred to as VGG-16 to represent facial features. VGG-16 includes 13 Conv layers, five max-pooling layers and three FC layers. It employs a much deeper network than AlexNet by stacking smaller kernels (3 $\times$ 3) instead of using a single layer of kernels with a large receptive field. The convolution stride is fixed to 1 pixel and the max-pooling layers are performed over a 2 $\times$ 2 pixel window with a stride of 2.
### GoogleNet
GoogleNet [@7298594] employs a complex 22-layer deep network and uses inception modules to replace the mapping of convolutions of different sizes. The size of the receptive field is 224 $\times$ 224 taking RGB color channels with mean subtraction. Modules of the above type are stacked on each other, with occasional max-pooling layers with of stride 2 to halve the resolution of the grid. In addition, 1 $\times$ 1 convolutions compute reductions before expensive 3 $\times$ 3 and 5 $\times$ 5 convolutions. In addition, moving from FC layers to average pooling while retaining the use of dropout increases the top-1 accuracy.
Regression of AU Values
-----------------------
The values of facial AUs can be regressed independently and jointly using deep learning techniques, as done in many AU recognition methods . In contrast to the independent AU values regression, which ignores the relationship between AUs, the joint method can achieve better performance by adding AU relationships or dynamic dependencies.
We thus use a 3-layer neural network to jointly learn the AU values regression function. To balance real-time calculation with test accuracy, the number of layers was determined experimentally to be three. We use the back propagation algorithm with batch gradient descent to train the network weights and biases. Besides, unlike many AU detection and AU intensity estimation algorithms that evaluate the loss independently for each action unit, we regard them as a whole and evaluate all the loss from 24 AUs together. According to the Mean Square Error (MSE), which is commonly used to measure regression performance [@Rudovic2015Context], we optimize the following loss function to learn the regression. $$\label{eq:loss}
L_{train}=\frac{1}{N}\sum_{i=1}^{N}\Vert\textbf{x}_{i}-\hat{\textbf{x}}_{i}\Vert^{2},
$$ where $L_{train}$ is the average loss for all the training samples, $\hat{x}_{i}$ denotes the ground truth values of the AUs and $x_{i}$ denotes the predicted values of the AUs and is a vector of length 24. The first two FC layers are equipped with ReLU non-linearity. Because not all the dimensions are necessary for obtaining the information of the AUs, dimensionality reduction is needed, which is also beneficial to generalization. We thus add dropout [@Srivastava2014Dropout] after each FC layer so that the network itself can determine the best number of dimensions for the AU value regression rather than reducing the dimension manually. The dropout ratios are 0.4 and 0.3 for the input layer and the hidden layer, respectively. Over-fitting is thus reduced. The input layer accepts the 1000-dimensional arrays obtained from the facial feature representation task. After the ReLU and dropout operations, the output from the first layer is a hidden layer consisting of 512 hidden units. The following layer is an output layer containing 24 units that represent the regressed values of the 24 AUs and the loss between the ground truth labels and predicted labels is subsequently calculated. The 3-layer neural network and its parameters are showed in [Figure \[fig:regression\]]{}.
![Structure and parameters of the 3-layer neural network used for predicting the AU values.[]{data-label="fig:regression"}](regression.png){width="0.9\linewidth"}
{width="0.98\linewidth"}
Experiments {#sec:experiments}
===========
In this paper, we provide a baseline performance for the regression of the AU values conducted on our FEAFA database. Our method was implemented using the public Caffe framework [@Jia2014Caffe]. Training was performed on a NVIDIA Titan X Pascal GPU machine with 12GB of video memory. As described above, we used three feature representations from AlexNet, VGG-16, and GoogleNet and then used a 3-layer neural network to regress the values of the 24 AUs.
We adopted a standard five-fold subject-exclusive cross-validation protocol [@Eidinger2014Age] to augment the numbers of training and testing data. Each AU in our dataset was distributed uniformly within these five folds. In the validation stage, the network parameters were selected by minimizing the loss function mentioned in . During validation, the hyperparameters, such as learning rate, weight decay, and number of iterations, were optimized.
Network parameter (weight and bias) initialization strategies do have a significant effect on not only the outcome of the optimization procedure but also the ability of the network to generalize. Because the commonly used strategy, which initialize randomly from Gaussian or uniform distribution, only employ less information about the image, we believe that the pre-trained CNN architectures provide us with more information. Thus, the deep features extracted from the highest fully connected layers of AlexNet, VGG-16 and GoogleNet pre-trained for image classification were used for our AU value regression task. We directly used the bvlc\_alexnet model and bvlc\_googlenet model publicly provided in Caffe framework and downloaded the publicly available pre-trained network weights for the VGG\_ILSVRC\_16\_layers model provided by Visual Geometry Group. We then fine-tuned the network and obtained the final results of the values of AU regression. The regression results of using three CNN features respectively with or without pre-trained models are showed in Table \[tab:result\]. The results show that pre-trained models give us a better result with lower loss than the original models.
Most importantly, the baseline performance indicates that the regression of AU values is feasible on this well-annotated and naturalistic dataset. However, there also remains room for improvement, such as how to regress AU values more accurately in low illumination conditions.
-- ------------------------------------------------------ ---------------------------------------------------------------------------------------------------------------
**Feature Extractors**
ALexNet[ ]{}[ ]{}[ ]{}VGG-16[ ]{}[ ]{}[ ]{}GoogleNet
not pre-trained 0.454[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}**0.398**[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}0.412
pre-trained 0.364[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}**0.289**[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}[ ]{}0.314
-- ------------------------------------------------------ ---------------------------------------------------------------------------------------------------------------
: Regression results pretained using three CNN features respectively on the FEAFA dataset for all 24 AUs. VGG-16 performs better than GoogleNet and AlexNet, whether the model is pre-trained or not.[]{data-label="tab:result"}
Applications {#sec:applications}
============
The FEAFA dataset and its baseline system can be employed in various facial expression applications. In this section, we introduce two major applications: one is 3D facial animation from video sequences and the other is AU detection and AU intensity estimation.
3D Facial Animation from Video Sequences
----------------------------------------
In this application, a virtual character is driven by the continuously changing facial performance of the source video sequence. We first use the algorithm described above to extract the AU values of the 2D video frames. The AU values of the source actor are later used as the blendshape coefficients of a digital avatar. In particular, we do not need to estimate the face identity of the participants in the video clip. Then, we can use an avatar’s neutral face blendshape and the 24 corresponding expression blendshapes as basis poses. These blendshapes can be easily found in the FaceWarehouse [@Cao2014FaceWarehouse], a database of 3D facial expression models containing the data of 150 individuals. Moreover, any virtual characters with 25 corresponding blendshapes can be used in this application. After obtaining the values of AUs from the 2D video frames and the corresponding blendshapes of the digital avatar, the animation of the digital avatar is generated by using a blendshape model through the linear combination given in . In addition, we can also easily estimate the rigid transformation of head movements [@Lepetit2009EP]. Then the 3D facial mesh is represented as a linear combination of the basis blendshapes plus a rotation $R$ and translation $t$. It can be summarized by the following equation. $$\label{eq:genavatar}
F=R(B_{0}+\sum_{i=1}^{24}\beta_{i}(B_{i}-B_{0}))+t.
$$
This whole process is more convenient using AU value regression instead of 3D facial shape reconstruction. We also developed a software using features from AlexNet for facial animation. Our software runs at over 18 fps on an macOS PC with an Intel (R) Core (TM) i5 [email protected] GHz processor. In addition, because the AU value regression process takes AU relationships and dynamic dependencies into consideration, animation of the avatars is more verisimilar and naturalistic. In [Figure \[fig:expression\]]{}, we present the results of the facial animation of simple virtual characters with combined AUs.
AU Detection and AU Intensity Estimation
----------------------------------------
There are quite a few state-of-the-art techniques for AU detection [@Wu2017Deep; @Li2017Action] and AU intensity estimation [@Walecki2016Copula; @Walecki2017Deep] based on deep learning. Binary AU detection clarifies whether facial motions are present or absent. AU intensity estimation recognizes not only whether the facial motions are present or not but also five increasing levels of facial expression intensity. Our FEAFA dataset is also suitable for these two tasks. For AU detection, we first use our baseline system to regress the AU values. And we regard each AU as presence only if its value is larger than 0.1. With regard to AU intensity estimation, after obtaining the regressed AU values, we divide the range (0,1\] of AU values into (0, 0.2\], (0.2, 0.4\], (0.4, 0.6\], (0.6, 0.8\], and (0.8, 1\] to represent the letters A through E, which describe an AUs intensity variation, from barely detectable to maximum intensity according to the FACS. Here, a certain AU is not present if its value is exact 0.
Conclusions and Future Work {#sec:conclusions}
===========================
The majority of previous datasets for facial expression analysis are limited to rough AU annotations or were collected in lab settings. We present a facial expression dataset of FACS-labeled data that is exact in decimal. And we provide a parameter system with redefined AUs for facial expression quantization in order to describe facial expression in great detail. Our FEAFA dataset contains 122 participants from Asia, 123 videos recorded in real-world conditions, and 99,356 frames. Each frame is anatomically labeled by the presence of 9 symmetrical FACS AUs, 10 asymmetrical (unilateral) FACS AUs, 2 symmetrical FACS ADs and 2 asymmetrical FACS ADs. The intensity of each AU is annotated with a floating point number between 0 and 1. The FEAFA dataset thus provides continuous annotations of changes in facial expression. Through FEAFA, continuous AU values of any video frames could be regressed based on trained CNNs. These AU values can then be used to generate 3D facial animation easily and smoothly with great reality and perform expression recognition and classification faster and more precisely. Continuous AU values can quantify facial expression more accurately than discrete AU levels and enable more precise AU detection and AU intensity estimation. In the future, we hope the release of this dataset will encourage researchers to test new AU value regression algorithms that performs better than the benchmark we have already provided. For AU detection and AU intensity estimation, we expect that researchers can compare their state-of-the-art algorithms with each other using FEAFA. We also expect that many other applications other than 3D facial animation can benefit from our dataset, such as facial expression recognition and facial image manipulation.
Acknowledgements {#acknowledgements .unnumbered}
================
This work is supported by National Key R&D Program of China under contract No. 2017YFB1002203, National Natural Science Foundation of China (NSFC, Grant No. 61671426, 61731022, 61871258, 61572077, 61471150), the Beijing Natural Science Foundation (Grant No. 4182071), the Instrument Developing Project of the Chinese Academy of Sciences (Grant No. YZ201670), and the Innovation Practice Training Program for College Students of Chinese Academy of Sciences.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The performance analysis of free space optical communication (FSO) system using relays and spatial diversity at the source is studied in this paper. The effects of atmospheric turbulence and attenuation, caused by different weather conditions and geometric losses, have also been considered for analysis. The exact closed-form expressions are presented for bit error rate (BER) of M-ary quadrature amplitude modulation (M-QAM) technique for multi-hop multiple-input and single-output (MISO) FSO system under log-normal fading channel. Furthermore, the link performance of multi-hop MISO and multi-hop single-input and single-output (SISO) FSO systems are compared to the different systems using on-off keying (OOK), repetition codes (RCs) and M-ary pulse amplitude modulation (M-PAM) techniques. A significant performance enhancement in terms of BER analysis and SNR gains is shown for multi-hop MISO and multi-hop SISO FSO systems with M-QAM over other existing systems with different modulation schemes. Moreover, Monte-Carlo simulations are used to validate the accuracy and consistency of the derived analytical results. Numerical results show that M-QAM modulated multi-hop MISO and multi-hop SISO FSO system with relays and spatial diversity outperforms other systems while having the same spectral efficiency for each system.'
author:
-
bibliography:
- 'references.bib'
title: 'Performance analysis of FSO using relays and spatial diversity under log-normal fading channel'
---
Free Space Optical Communications, Log-Normal Fading Channel, Quadrature Amplitude Modulation, Spatial Diversity, Bit Error Rate, Decode and Forward Relay
Introduction
============
Free space optical communication (FSO) is a rapidly evolving technology to handle high data rate and has very huge information handling capacity. FSO systems are proven an alternative to the fiber optics technology. In many cases, only the free space transmission is possible to establish a connection between the source and the destination over a line-of-sight (LOS) path. Optical signal transmission in free space has explored the unexplored areas in wireless communications for decades and appearing as the key of many systems for varieties of applications, such as radio frequency (RF) wireless transmission, satellite communications, long-haul connections and optical fiber back up, etc. FSO systems have to face many challenges, mainly atmospheric turbulence, beam wandering, beam attenuation, weather attenuation, geometric losses and scintillation [@navidpour2007ber].
Related Works
-------------
In literature, different types of channel models have been proposed to exactly model the atmospheric turbulence which matches well with the experimental results. Log-normal, gamma-gamma and negative exponential channel models are used for weak, strong-moderate and strong turbulence conditions, respectively [@luong2013effect; @rodrigues2013evaluation; @peppas2012simple]. Fading and attenuation has affected the performance of FSO systems in large extent. Turbulence and weather conditions, such as snow, fog and haze also affecting the FSO transmission. Further, geometric loss has also lessened the performance of FSO systems. In previous works, varieties of spatial diversity techniques have been proposed to mitigate the effect of atmospheric turbulence in FSO communications [@abaza2014diversity; @tsiftsis2009optical] where three primary spatial diversity schemes are more popular among the available diversity schemes, such as orthogonal space time block codes (OSTBCs), transmit laser selection (TLS) and repetition codes (RCs). Here, TLS is considered as the best spatial diversity scheme with channel state information (CSI) at the source end, but at the cost of increased system complexity [@garcia2009selection; @safari2008relay]. With moderate system complexity in different channel conditions, RCs outperforms its counterpart OSTBCs. Apart from spatial diversity schemes, some other techniques have also been proposed to mitigate turbulence effects such as relay assisted techniques [@safari2008relay], error-correcting codes [@abaza2011ber; @zhu2002free] and maximum likelihood estimation [@zhu2000maximum], out of which relay assisted technique appears to be prominent as several short links are used instead of a long communication link for efficient data transmission over FSO channels.
Motivation and Contribution
---------------------------
In previous works, most of the systems introduced are using pulse amplitude modulation (PAM) as a high level modulation technique, but very few have considered to use quadrature amplitude modulation (QAM) scheme, which is the motivation behind this work. In this paper, with the help of decode and forward (DF) relay for intensity modulation and direct detection (IM/DD), M-ary QAM (M-QAM) modulation is used against a log-normal fading channel to study the link performance of different architectures of FSO systems for short distance communication links under different weather conditions. Here, bit error rate (BER) is considered as a performance metric which is used to measure the performance of the FSO system. For analysis purposes, the effect of turbulence, path loss, scattering and scintillation have also been taken into account. Assuming the correlation effects among the transmitters, the performance of multi-hop multi-input and single-output (MISO) and multi-hop single-input and single-output (SISO) system with M-QAM is compared with multi-hop MISO and SISO system using M-PAM, MISO with RCs and OOK and SISO with on-off keying (OOK) [@abaza2015performance]. In all cases, it is assumed that the spectral efficiency is same. At the receiver, maximum likelihood (ML) decoding technique is employed for receiving the M-QAM signals.
Organization
------------
The rest of the paper is organized as follows: in Section \[sec2\], the DF relay FSO system is described for multi-hop MISO configurations under various modulation techniques and the mathematical expressions of BER have been presented for M-QAM and $\textrm{M}^2$-QAM modulation schemes. Section \[sec3\] presents the numerical results where the parameters, considered for simulation purposes, are tabulated in Table \[tab:1\]. Section \[sec4\] concludes the work.
System Description {#sec2}
==================
Multi-hop DF relay System
-------------------------
### M-QAM Transmission
For transmission of signals, a MISO system is considered with M transmitters in the form of lasers and a photo detector as receiver, which is shown in Fig. \[fig:3\]. At the source, transmission of signals is in the form of binary data, modulated by a $M_I \times M_Q$ electrical-QAM modulator. For the transmission through M paths, mapping of data symbols into QAM symbols is done with the help of Gray coding. In addition, to mitigate the spatial correlation effect, sufficient channel distance is considered.
The probability of error of QAM signals can be easily calculated from the probability of error of PAM signals as the phase-quadrature signal components can be perfectly extracted at the QAM demodulator. For the M-ary QAM signals, the probability of a correct decision is written as [@proakis2001digital] $$\label{eq22}
P_C= (1-P_{\sqrt M} )^2,$$ where $P_{\sqrt M}$ is the probability of error for $\sqrt {M}$-ary PAM with half the average power in each quadrature component of the signal of the corresponding QAM system. The probability of error of M-ary PAM can be written as $$\label{eq23}
P_{\sqrt M}=2\bigg(1-\frac{1}{\sqrt M} \bigg)Q\bigg(\frac{3log_2 M\gamma}{M-1}\bigg),$$ where $\gamma$ represents the average SNR per symbol. Consequently, the probability of a symbol error for M-ary QAM signals can be written as $$\label{eq24}
P_M=1-(1-P_{\sqrt M})^2,$$ which is valid for $M=2^k$ and even $k$. By using the optimum detector, the tightly upper bounded symbol error probability can be written as $$\label{eq25}
\begin{split}
P_M\leq 1-\bigg[1-2Q\bigg(\sqrt{\frac{3 \gamma}{M-1}}\bigg)\bigg]\\
\leq 4Q\bigg(\sqrt{\frac{3 log_2{M}\gamma}{M-1}}\bigg).
\end{split}$$
![A MISO multi-hop DF relay FSO System, which is not benefited by the use of transmit diversity.[]{data-label="fig:3"}](Figure_3){width="\linewidth"}
Here, introduction of relays creates some effect in FSO system for SISO and MISO configurations with the help of DF relays, which helps to achieve significant improvement in the link performance. For the system under consideration, $(K-1)$ DF relays are set up in between the source and the destination for $K$ hops. The received signal at the $k$th hop can be written as $$\label{eq26}
r_k(t) = x(t)\eta I_k + n_k(t), \ k=1,2,\ldots K.$$ In this scheme, $K$ time slots are required for transmission of signals from source to destination, unlike MISO and SISO systems. For the benefit of achieving similar spectral efficiency, $2^K$-ary M-QAM modulation technique is employed. DF relays used in between the transceiver pairs create shorter communication links which help to reduce the effects of turbulence and path losses, significantly. The upper bound BER for multi-hop DF relay system can be written as [@abaza2015performance] $$\label{eq27}
\textrm{BER} \leq 1-\prod_{k=1}^{K}(1-\textrm{BER}_k).$$ As the identical statistical properties are considered for all hops, the above expression (\[eq27\]) can be approximated by $$\label{eq28}
\textrm{BER} \approx \frac{1}{2}[1-(1-2\textrm{BER}_k )^K].$$ The conditional BEP of M-QAM is given as [@singh4performance] $$\label{eq29}
P(e|\gamma)\approx \frac{2\Big(1-\frac{1}{\sqrt M}\Big)}{\textrm{log}_2 M}\Bigg[Q\Bigg(\sqrt{\frac{3\textrm{log}_2M \gamma}{2(M-1)}} \Bigg)\Bigg],$$ and the instantaneous SNR $\gamma$ is given by $$\label{eq30}
\gamma = \frac{2P^2}{\sigma_n^2 R},$$ where $P$ implies the signal power and $R$ signifies the bit rate. Now, by applying the approximate $Q$ function [@abaza2015performance] on (\[eq29\]), the conditional BEP of M-QAM can be written as $$\label{eq31}
\begin{split}
P(e|\gamma) & \approx \frac{2\Big(1-\frac{1}{\sqrt M}\Big)}{\textrm{log}_2 M}\\
& \times\bigg[\frac{1}{12}\textrm{exp}\bigg(-\frac{3 \textrm{log}_2 M \gamma}{4(M-1)}\bigg) + \frac{1}{4}\textrm{exp}\bigg(-\frac{ \textrm{log}_2 M \gamma}{(M-1)}\bigg)\bigg],
\end{split}$$ which is the general expression of the conditional BEP for FSO system using M-QAM. Now, the BER expression of the $k$th hop can be derived using Hermite polynomial and $Q$-function approximation as in [@abaza2015performance]. The closed-form expression of BER for M-QAM modulation can be written as [@abaza2015performance] $$\label{eq32}
\begin{split}
\textrm{BER}_k & \approx \\ & \frac{G}{12}\sum_{i=1}^{N}w_i \textrm{exp}\bigg(-\frac{3\textrm{log}_2M \beta_{kn}^2 \overline\gamma e^{-4\sigma_k^2+x_i\sqrt{32\sigma_k^2}}}{4(M-1)}\bigg)
\\
& +\frac{G}{4}\sum_{i=1}^{N}w_i \textrm{exp}\bigg(-\frac{\textrm{log}_2M \beta_{kn}^2 \overline\gamma e^{-4\sigma_k^2+x_i\sqrt{32\sigma_k^2}}}{(M-1)}\bigg),
\end{split}$$ where $G =2\big(1-\frac{1}{\sqrt M}\big)/( \textrm{log}_2 (M) \sqrt{\pi})$, $\beta_{kn}$ represents the normalized path loss coefficient, $x_i$ and $w_i$ are the zeros and the weights of the Hermite polynomial, $\sigma_k^2$ is the channel variance and $\overline\gamma$ is the average SNR. Substituting (\[eq32\]) in (\[eq27\]) and (\[eq28\]), the upper bound and the average BER expressions of multi-hop system can be obtained, respectively. The BER expressions of M-PAM, SISO with OOK, and MISO with RCs and OOK are taken from [@abaza2015performance] for analysis and comparison purpose.
### $M^2$-QAM Transmission
The $M^2$ symbols of $M^2$-QAM made up of an in-phase and quadrature phase component basis function, orthogonal to each other. In each symbol duration, the two basis functions are modulated with the independent data resulting a multiplication by a series of M amplitude values to each basis function to comprise the $M^2$ symbols [@hranilovic2006wireless]. The constellation of QAM shows a two dimensional regular array of points and the minimum spacing between the points is prescribed by the amount of DC bias added by virtue of the non-negativity constraint, which is [@hranilovic2006wireless] $$d_{min}=\frac{P}{M-1}\sqrt{\frac{2\textrm{log}_2 M}{R}},$$ where R signifies the bit rate. The probability of symbol error is estimated by using the union bound approximation [@abaza2015performance]. By evaluating the average number of neighbors $d_{min}$ away from every constellation point, $P_{esym}$ is approximated as $$P_{esym}=\frac{4M-1}{m}\cdot Q\Bigg(\frac{P}{M-1}\sqrt{\frac{1}{4 R_s \sigma^2}}\Bigg),$$ where $R_s=R/\textrm{log}_2 M^2$. By applying the Gray coding approximation, the conditional BEP of $M^2$-QAM is given by [@hranilovic2006wireless] $$\label{eq33}
P(e|\gamma)\approx \frac{2\Big(M-1\Big)}{M\textrm{log}_2M}\Bigg[Q\Bigg(\sqrt{\frac{\textrm{log}_2M \gamma}{4{(M-1)}^2}}\Bigg)\Bigg],$$ where instantaneous SNR $
\gamma = \frac{2P^2}{\sigma_n^2 R}.
$ Now, applying the approximate $Q$ function on (\[eq33\]), the conditional BEP of M-QAM is given as $$\label{eq34}
\begin{split}
P(e|\gamma) & \approx
\frac{2\Big(M-1\Big)}{M\textrm{log}_2M}\\
& \times \bigg[\frac{1}{12}\textrm{exp}\bigg(-\frac{ \textrm{log}_2 M \gamma}{8(M-1)^2}\bigg) + \frac{1}{4}\textrm{exp}\bigg(-\frac{\textrm{log}_2 M \gamma}{6(M-1)^2}\bigg)\bigg],
\end{split}$$ which represents the typical expression of the conditional BEP for the FSO system with $M^2$-QAM. The closed form expression of BER of $M^2$-QAM modulation derived as discussed earlier, can be written as $$\label{eq35}
\begin{split}
\textrm{BER}_k & \approx \frac{G}{12}\sum_{i=1}^{N}w_i \textrm{exp}\bigg(-\frac{\textrm{log}_2M \beta_{kn}^2 \overline\gamma e^{-4\sigma_k^2+x_i\sqrt{32\sigma_k^2}}}{8(M-1)^2}\bigg)
\\
& + \frac{G}{4}\sum_{i=1}^{N}w_i \textrm{exp}\bigg(-\frac{\textrm{log}_2M \beta_{kn}^2 \overline\gamma e^{-4\sigma_k^2+x_i\sqrt{32\sigma_k^2}}}{6(M-1)^2}\bigg),
\end{split}$$ where $G = 2\big(M-1\big)/M\textrm{log}_2M \sqrt{\pi}$ and $\beta_{kn}$ is the normalized path loss coefficient with reference to the direct link of the multi-hop system. Substituting (\[eq35\]) in (\[eq27\]) and (\[eq28\]), the upper bound BER expression and the average BER expression for the multi-hop system can be retrieved, respectively.
![A hybrid MISO multi-hop DF relay FSO System which is benefited by relays and transmit diversity, both.[]{data-label="fig:4"}](Figure_4){width="\linewidth"}
MISO multi-hop DF relay system
------------------------------
This is a hybrid technique which uses the advantages of FSO systems with relays and transmit diversity to deteriorate the channel attenuation and turbulence effects. The block diagram of this system model is shown in Fig. \[fig:4\]. Here, $K-1$ relays are employed to receive the signals from $N_t$ transmitters which, simultaneously, re-transmit those signals with the help of RCs. For this system, the number of hops are considered identical with the number of transmitters per relay. The signal received at each hop can be given as $$\label{eq36}
r_k (t) = x(t)\eta\sum_{i=1}^{N_t}I_{ki} +n_k (t), \ k=1,2 \ldots K.$$ Similarly, using the conditional BEP expression for M-QAM, the BER expression for $k$th hop can be written as [@abaza2015performance] $$\label{eq37}
\begin{split}
\textrm{BER}_k
& \approx \sum_{n_1=1}^{N}\ldots \sum_{n_{N_t}}^{N}\Bigg[\prod_{i=1}^{N_t}\frac{w_{n_i}}{\sqrt{\pi}}\Bigg]\frac{F}{12}\textrm{exp}\\
& \times \Bigg(-\frac{3\textrm{log}_2(M) \beta_{kn}^2\overline\gamma}{4(M-1)N_t}
\sum_{i=1}^{N_t}\Bigg[\textrm{exp}\bigg(\sqrt{32}\sum_{j=1}^{N_t}{c_{ij}^{'}x_{nj}-4\sigma_k^2}\bigg)\Bigg]\Bigg)\\
& + \sum_{n_1=1}^{N}\ldots \sum_{n_{N_t}}^{N}\Bigg[\prod_{i=1}^{N_t}\frac{w_{n_i}}{\sqrt{\pi}}\Bigg]\frac{F}{4}\textrm{exp}\\
& \times \Bigg(-\frac{\textrm{log}_2(M) \beta_{kn}^2\overline\gamma}{(M-1)N_t}
\sum_{i=1}^{N_t}\Bigg[\textrm{exp}\bigg(\sqrt{32}\sum_{j=1}^{N_t}{c_{ij}^{'}x_{nj}-4\sigma_k^2}\bigg)\Bigg]\Bigg),
\end{split}$$ and using the conditional BEP equation of $M^2$-QAM, the BER expression for the $k$th hop has been given as $$\label{eq38}
\begin{split}
\textrm{BER}_k & \approx
\sum_{n_1=1}^{N}\ldots \sum_{n_{N_t}}^{N}\Bigg[\prod_{i=1}^{N_t}\frac{w_{n_i}}{\sqrt{\pi}}\Bigg]
\frac{F}{12}\textrm{exp}\\
& \times\Bigg(-\frac{\textrm{log}_2(M) \beta_{kn}^2\overline\gamma}{8(M-1)N_t}
\sum_{i=1}^{N_t}\Bigg[\textrm{exp}\bigg(\sqrt{32}\sum_{j=1}^{N_t}{c_{ij}^{'}x_{nj}-4\sigma_k^2}\bigg)\Bigg]\Bigg)\\
& + \sum_{n_1=1}^{N}\ldots \sum_{n_{N_t}}^{N}\Bigg[\prod_{i=1}^{N_t}\frac{w_{n_i}}{\sqrt{\pi}}\Bigg]
\frac{F}{4}\textrm{exp}\\
& \times \Bigg(-\frac{\textrm{log}_2(M) \beta_{kn}^2\overline\gamma}{6(M-1)N_t}
\sum_{i=1}^{N_t}\Bigg[\textrm{exp}\bigg(\sqrt{32}\sum_{j=1}^{N_t}{c_{ij}^{'}x_{nj}-4\sigma_k^2}\bigg)\Bigg]\Bigg),
\end{split}$$ where $F = 2\big(1-\frac{1}{\sqrt M}\big)/\textrm{log}_2 (M)$ for (\[eq37\]) and $F=2\big(M-1\big)/M\textrm{log}_2M$ for (\[eq38\]) and $c_{ij}^{'}$ represents the $(i,j)$th coefficients of the spatial covariance matrix Γ$\Gamma_{sq}^{'}=\Gamma^{'1/2}$. By substituting these results in (\[eq27\]) and (\[eq28\]), expressions of the upper bound BER and the closed-form average BER of the multi-hop MISO system can be obtained, respectively. The BER expressions for other configurations of FSO system are shown in [@abaza2015performance].
Numerical Analysis and Discussions {#sec3}
==================================
Numerical results are presented in this section which verify the derived analytical results using Monte-Carlo simulations. For simulation purposes, a minimum of $10^6$ bits are relayed for respective SNR values and increases with increasing SNR. The parameters needed to achieve a target BER of $10^{-9}$ are defined and tabulated in Table \[tab:1\]. Further, all hops are considered equidistant with each other. The scintillation index (SI) is $\leq$ 0.75 for log-normal channel and $\sigma_x=\sqrt{\textrm{ln}(\textrm{SI})+1/2}$ [@moradi2011switched]. Hence, for SI = 0.75, $\sigma_x\leq0.374$ is required, which represents the maximal value assumed for the analysis purpose.
![BER of multi-hop FSO system with 8-QAM, 8-PAM, SISO-OOK and MISO-RCs and OOK in clear weather condition.[]{data-label="fig:7"}](Figure_7){width="\linewidth"}
**FSO Parameters** **Numerical Values**
------------------------------------------------------ ------------------------------------------------------- --
Relay Spacing 400 m
Wavelength ($\lambda$) 1550 nm
Link Distance $(l)$ 1200 m
Beam Divergence Angle $(\theta_T)$ 2 mrad
Correlation Coefficient $(\rho)$ 0.3
Tx and Rx Aperture Diameter $(D_R\textrm{ and }D_T)$ 20 cm
Attenuation Constant $(\alpha)$ 0.43 dB/km (Clear Weather)
20 dB/km (Light Fog)
Refractive Index Constant $(C_n^2)$ $5\times10^{-14} \textrm{m}^{-(2/3)}$ (Clear Weather)
$1.7\times 10^{-14} \textrm{m}^{-2/3}$ (Light Fog)
: Simulation Parameters
\[tab:1\]
Fig. \[fig:7\] represents the BER of multi-hop FSO system, where the number of relays is two and the number of transmit lasers is one for SISO scheme or three for MISO scheme, wherever required with 8-QAM, 8-PAM, MISO with RCs and OOK and SISO-OOK modulation techniques. It demonstrates the effect on the performance gain of MISO and SISO multi-hop FSO systems for different modulation techniques under clear weather conditions. Further, at the target BER, SNR gains of 19.52 dB and 27.65 dB is achieved for multi-hop MISO and multi-hop SISO systems as compared to MISO with OOK and RCs and SISO with OOK, respectively, for 8-PAM modulation scheme, whereas with 8-QAM modulation, SNR gains of 13.21 dB and 13.44 dB is achieved for multi-hop MISO and multi-hop SISO systems over PAM modulated multi-hop MISO and multi-hop SISO systems, respectively, maintaining the same number of transmitters. Furthermore, results show that, multi-hop MISO system outperforms multi-hop SISO system by 3.1 dB and 3.33 dB, respectively, for 8-QAM and 8-PAM.
![BER of multi-hop FSO system with 8-QAM, 8-PAM, SISO-OOK and MISO-RCs and OOK in light fog condition.[]{data-label="fig:8"}](Figure_8){width="\linewidth"}
In Fig. \[fig:8\], BER for multi-hop FSO system with two relays and one or three transmit lasers has been shown under light fog conditions for 8-QAM, 8-PAM, MISO with RCs and OOK and SISO with OOK modulation schemes. For two relay multi-hop system, 8-QAM modulation scheme provides an equal spectral efficiency for SISO and MISO systems. Derived analytical results show that multi-hop system still outperforms other systems and the increment in the performance of the system is significant. For 8-PAM modulation scheme, SNR gains of 43.13 dB and 47.64 dB are achieved at the target BER of multi-hop MISO and multi-hop SISO systems in comparison with MISO with RCs and OOK and SISO-OOK modulated systems, whereas using 8-QAM technique, system outperforms the 8-PAM system with 13.34 dB and 13.66 dB of SNR gains for multi-hop MISO and multi-hop SISO systems, respectively. In addition, 8-QAM and 8-PAM multi-hop MISO systems provide the performance gain of 1.16 dB and 1.48 dB over the 8-QAM and 8-PAM multi-hop SISO system, respectively.
Conclusions {#sec4}
===========
The effects of moderate turbulence under clear weather condition and weak turbulence under light fog conditions on the performance of different types of FSO systems are studied in this paper. It is evident from the derived analytical results that diverse conditions in the atmosphere have a huge effect on the performance of MISO and SISO multi-hop FSO systems. Furthermore, multi-hop MISO system with M-QAM modulation outperforms other systems, such as SISO with OOK, MISO with RCs and OOK and M-PAM multi-hop SISO and MISO systems in terms of SNR gains and BER performance considering the same spectral efficiency for all the systems. Moreover, the results also demonstrate that multi-hop systems are capable of mitigating the effects of turbulence and path losses caused by geometric losses and attenuations, whereas MISO systems can counteract turbulence effects only. Consequently, it can also be stated that the overall performance of the FSO system can be improved significantly by increasing the number of relays where the introduction of spatial diversity could also be of great importance as well.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We study the radial distribution of supernova Ia (SNe Ia) in morphologically selected early–type host galaxies from the *Sloan Digital Sky Survey* (SDSS) and discuss its implications for the progenitor systems of SNe Ia. While new observations of early-type galaxies suggest that they contain small fractions of young stellar populations, they are also the most likely hosts for long time delay SNe Ia. We find that there is no statistically significant difference between the radial distribution of SNe Ia and the light profile of their early–type host galaxies, which are dominated by old, metal-rich stellar populations. This confirms the commonly accepted idea that some SN Ia progenitors have time delays of the order of several Gyr.'
author:
- |
Francisco Förster$^{1}$[^1] and Kevin Schawinski$^{1}$[^2]\
$^{1}$Dept. of Physics, University of Oxford, Denys Wilkinson Building, Keble Road, Oxford, OX1 3RH, United Kingdom
date: 'Accepted 2008 May 18. Received 2008 May 12; in original form 2008 April 29.'
title: 'The radial distribution of type Ia supernovae in early-type galaxies: implications for progenitor scenarios'
---
\[firstpage\]
supernovae: general, chemical: general.
Introduction
============
Type Ia supernovae (SNe Ia) have been extensively used as distance indicators for cosmology [see e.g. @rie98; @per99] using the *empirical* Phillips relation [@phi93] and related methods. They are also important for understanding galaxy evolution, affecting the chemistry and energy budget of their hosts, and the star formation and metallicity evolution of the Universe. Uncertainties in both the progenitor scenarios and the explosion mechanism [for a review see @HN00] prevent us from having a convincing physical picture that justifies their use as *standardisable* candles at high redshift and is able to reproduce their observed diversity in the local Universe.
The SN Ia time delay distribution
---------------------------------
Linking SN Ia progenitor scenarios with cosmology, galaxy evolution and the cosmic star formation history requires information on the time between their formation and explosion, or time delay distribution. Thermonuclear explosions seem to originate in white dwarf stars (WDs), with main sequence (MS) lifetimes ranging from $\sim 30$ Myr to several billion years, thus the minimum time delay must be of the order of $\sim 30$ Myr. To narrow the constraints details of possible **progenitor systems need to be used.**
Most proposed progenitor scenarios involve mass transfer on to a CO WD that reaches the Chandrasekhar limit in a binary system. In these scenarios the WD grows through either the expansion and Roche lobe overflow of an evolved companion (*single degenerate \[SD\] scenarios*), or the slow release of gravitational waves, orbital shrinking, Roche lobe overflow and merging of a compact double WD system (*double-degenerate \[DD\] scenarios*).
For the SD scenario it has been suggested that only systems that can transfer enough mass under a critical rate can bring the WD star to the Chandrasekhar limit [@nom91]. In binary population synthesis (BPS) simulations most of the WDs that reach the Chandrasekhar mass accrete matter from a slightly evolved MS star, the so called CO WD + MS – SD scenario [@HKN99; @lan00; @HP04]. In this channel the mass of the companion determines both when accretion starts and the rate of accretion. As a consequence, the time-delay distribution of this channel seem to be relatively narrow, peaking at $\sim 670$ Myr and rapidly becoming negligible after $\sim 1.5$ Gyr.
Several Gyr time–delay progenitors can also be produced in the SD scenario when a relatively low mass red-giant (RG) star transfers matter on to a CO WD star, in the CO WD + RG – SD scenario [@HKN96]. This scenario appears to be supported by observational evidence [@2007Sci...317..924P], although its relative contribution is unclear [@HP04].
The characteristic time delay in the DD scenario [@IT84; @web84] is the coalescence time–scale of the binary system, which depends roughly on the fourth power of the separation of the double-degenerate system [@sha83]. The time-delay distribution can be described by a low time-delay cutoff ($\sim 30-100$ Myr), an approximately power-law decline up to the age of the Universe and some secondary peaks depending on the details of the BPS simulation.
However, the expected accretion rates in the DD scenario are thought to be too big to lead to successful thermonuclear explosions, with most simulations leading to accretion-induced collapse (AIC) and the formation of compact objects instead [see @SN98 and references therein]. Rotation or more sophisticated neutrino physics may give this channel a more significant role in the future [@2007MNRAS.380..933Y].
Observationally, a comparison between the cosmic supernova rate and star formation history has led to indirect constraints on the time delay distribution [see e.g. @GYM04; @str04; @2006ApJ...651..142H], but uncertainties in both functional quantities seem unlikely to give constraints with the required accuracy to distinguish between different channels [@2006MNRAS.368.1893F; @2008arXiv0803.3793B].
A better probe for clues regarding the SN Ia time delay distribution might be to look at their individual host galaxies [see e.g. @man05; @2006ApJ...648..868S; @2008ApJ...673..999P] and their galaxy environments [see e.g. @2007ApJ...660.1165S]. With this approach uncertainties in the star formation histories of individual galaxies can be better controlled if accurate stellar population reconstructions are used.
Early–type galaxies
-------------------
The dominant stellar populations of early–type galaxies are old and must have formed in the high redshift universe [see @2005ApJ...621..673T and references therein].
However, the GALEX UV space telescope has led to the discovery that a significant fraction of early–type galaxies in the low redshift universe formed a few percent of their stellar mass in the last Gyr [@2005ApJ...619L.111Y; @2006Natur.442..888S; @2007ApJS..173..619K; @2007ApJS..173..512S]. These episodes of residual star formation are not apparent in the optical, but become prominent in the UV. This opens up the possibility that the SNe Ia observed in apparently passive early–type galaxies may be due to these small young populations and so might have short time delays.
The bulk stellar population of massive early-type galaxies follow a $r^{1/4}$ or de Vaucouleurs law [@1948AnAp...11..247D] and the dynamical relaxation time–scale of these systems is much higher than the age of the Universe [@1987gady.book.....B]. High resolution H$\beta$ maps have been used to trace young stellar populations in early-type galaxies and they tend not to follow the typical de Vaucouleurs profile of their hosts [@2006MNRAS.366.1151S; @2006MNRAS.369..497K]. Evidence for central molecular disks in early-type galaxies supports this idea [@2007AJ....134.2148L; @2008MNRAS.tmp..490C].
If the radial distribution of SNe Ia follows the light of the bulk stellar population in early–type galaxies, then their progenitors must have an age of at least the minimum between the relaxation time–scale and the age of the bulk of the stellar population. We attempt to quantitatively test whether the radial distribution of SNe Ia follows a de Vaucouleurs law.
Sample Selection {#sec:sample}
================
We create a sample of SN Ia early-type host galaxies by selecting all SNe Ia in the CfA list of supernovae[^3] whose hosts have been covered by the Sloan Digital Sky Survey (SDSS, @2000AJ....120.1579Y [@2002AJ....123..485S]). The SDSS is a survey of half the Northern Sky, providing us with photometry in the five optical filters *u,g,r,i* and *z* [@1996AJ....111.1748F; @1998AJ....116.3040G]. We use the current data release DR6 [@2008ApJS..175..297A].
We visually inspect every SN Ia host galaxy observed by SDSS and select only those with early-type morphology. Visual classification avoids the introduction of biases associated with proxies for morphology such as colour, concentration or structural parameters [@2008arXiv0804.4483L]. From the remaining galaxies we select those with half–light radii bigger than 2.4$\arcsec$, or six pixels in the SDSS images. The final sample contained galaxies with a mean and median half–light radius of 17.6$\arcsec$ and 12.9$\arcsec$, or about 44 and 32 pixels, respectively. Their redshifts were in the range 0.001 to 0.3, with a mean of 0.06 and median of 0.05.
About three quarters of the SNe in the final sample were dicovered by the KAIT/LOSS SN survey, the SDSS SN survey and amateur astronomers in almost equal proportions. The rest were discovered by various surveys including the Nearby SN Factory, among others.
Analysis {#sec:analysis}
========
We use elliptical projected radial coordinates and model the surface brightness profiles of the early–type galaxies in our sample as [@1968adga.book.....S]: $$\begin{aligned}
\label{eq:sersic}
I(r) \equiv I_{\rm e}~ \exp \biggl\lbrace - b_{\rm n} \biggl[
\biggl(\frac{r}{r_{\rm e}}\biggr)^{1/n} - 1 \biggr] \biggr\rbrace,\end{aligned}$$ where $n$ is known as the Sérsic index (normally assumed to be four in early–type galaxies) and by definition half of the light comes from $r < r_{\rm e}$, with $I_{\rm e} \equiv I(r_{\rm e})$. These requirements define $b_{\rm n}$, approximately $2 n - 1/3$.
Now, using the change of variable $u \equiv b_{\rm n}
\biggl(\frac{r}{r_{\rm e}}\biggr)^{1/n}$ one can show that the probability density per unit variable $u$ of a photon coming at a radius $r = r_{\rm e} (u / b_{\rm n})^n$ is: $$\begin{aligned}
\label{eq:fu}
f_u(u) = \frac{u^{2n - 1}
e^{-u}}{\int_0^\infty t^{2 n - 1} e^{-t} dt} = \frac{u^{2 n -
1}}{\Gamma(2 n)} e^{-u},\end{aligned}$$ where $\Gamma(u) \equiv \int_0^\infty t^{u-1} e^{-t} dt$. This allows one to define an equivalent SN Ia rate surface magnitude brightness: $$\begin{aligned}
\label{eq:muia}
\mu_{\rm Ia} \equiv -2.5~ \ln \biggl(\frac{f_u(u)~ \Gamma(2n)}{u^{\rm
2n - 1}}\biggr) = 2.5 ~ b_{\rm n}
\biggl(\frac{r}{r_{\rm e}}\biggr)^{1/n},\end{aligned}$$ which, for a given $n$, should have the same slope of the log of the true surface brightness in $(r/r_{\rm e})^{1/n}$ space, but differ by a constant.
Thus, the cumulative probability distribution per unit variable $u$ that a photon comes from inside a region of radius $r = r_{\rm e} (u /
b_{\rm n})^n$ is: $$\begin{aligned}
\label{eq:Fx}
F_u(u) = P(2n,~u),\end{aligned}$$ where $P(a,~u)$ is the regularised incomplete Gamma function defined as $P(a,~u) \equiv \int_0^u t^{a - 1} e^{-t} dt~ /~ \Gamma(a)$.
With the distribution of normalised distances of observed SNe Ia it is possible to test whether the SN Ia rate is proportional to the surface brightness intensity in early–type galaxies.
Systematic effects
------------------
Since core collapse SNe are very rare in early-type hosts, SN type misclassification is unlikely. However, early–type galaxies have extended radial profiles that easily overlap with other galaxies in the line of sight, making host galaxy identification problems important at large radii. Moreover, several effects will occur at small radii: different point–spread functions, deviations from the Sérsic profile close to galaxy cores, astrometric errors between SDSS and SN discovery coordinates, or the Shaw effect , i.e. undetected SNe near the core of galaxies due to the saturation of old photographic plates. To minimise these systematic errors we limit our sample to SNe with intermediate radii.
SDSS positions are accurate to 0.1$\arcsec$ [@2003AJ....125.1559P] and only 15% of the SNe in our sample were discovered before 2000, which means that astrometric problems or the Shaw effect are unlikely to be significant. We define a minimum SN distance of 0.2 half–light radii, since this central zone is well resolved in more than 80% of the images. For the smallest allowed half–light radii this corresponds to about five times the SDSS astrometric accuracy, but about two times the SDSS resolution. We also define an arbitrary maximum distance of four half-light radii, which is the radius at which approximately 85% of the light of the galaxy is contained in a de Vaucouleurs profile. The resulting mean and median SN distances from the galaxy cores were 18.2$\arcsec$ and 9.6$\arcsec$, about 45 and 24 pixels, or 13 and 7 times de resolution, respectively.
Thus, we limit our sample to SNe with $0.2 < r / r_{\rm e} < 4$. When more than one host candidate met these requirements we arbitrarily chose the one closer to the SNe. This could bias our results to a centrally concentrated distribution, but dim host galaxies that are not seen in SDSS images could have the opposite effect.
In Fig. \[fig:SDSS\_power\] we plot the equivalent of the surface brightness profile using type Ia SNe and assuming a de Vaucouleurs law with the restricted normalised distances defined above. As a first approximation it shows that SNe Ia follow their early–type host light profiles. Other authors have studied the radial distribution of type Ia SNe, but not using half–light radii normalised distances [@2000ApJ...542..588I; @2007HiA....14..316B]. Recently, [@2008arXiv0804.0909T] have also used half–light radii normalised distances, but with photometrically typed SNe Ia at high redshift and with a much lower relative resolution.
Goodness of Fit test
--------------------
To quantitatively test the hypothesis that SNe Ia follow their early–type hosts light profiles in this limited sample we normalise the cumulative probability distribution in equation (\[eq:Fx\]), so that $b_{\rm n}~ 0.2^{1/n} < u < b_{\rm n}~ 4^{1/n}$, and contrast it with the observed cumulative distribution using a Kolmogorov-Smirnov (KS) goodness of fit test. We conclude that for this restricted region the SN Ia rate is statistically consistent with a de Vaucouleurs profile (see Fig. \[fig:test\]).
![Observed (error bars) values of $\mu_{\rm Ia}$, defined by eq. (\[eq:muia\]), vs the expected values (line) if SNe Ia follow a de Vaucouleurs law in early–type galaxies. The binning is robust at intermediate radii, with 5 or more elements in more than 80% of the bins.[]{data-label="fig:SDSS_power"}](figure1){width="0.8\hsize"}
However, it is known that early–type galaxies light profiles have a continuous distribution of Sérsic indices connecting dwarf to giant ellipticals, increasing from approximately $n = 2$ to $n = 6$ [@GG03]. We have included a variable Sérsic index in our fitting procedure and found that the best–fitting Sérsic indices are distributed in approximately the same range, but have a slight preference for lower values, with an average of 3.4 and a median of 3.3. This distribution might be a result of the way SNe Ia have historically been discovered, which in many cases depends on pre–selected galaxy samples.
To allow for this diversity to be included in our analysis we use a new change of variable suggested by eq. (\[eq:Fx\]), namely $v \equiv
P(2n, u)$, or the fraction of the integrated light of the galaxy at the position of the SNe if the profile was well described by equation (\[eq:sersic\]). The new probability density and cumulative probability density will be simply: $$\begin{aligned}
f_v(v) = 1 ~~~~~~~~{\rm and} ~~~~~~~~ F_v(v) &= v,\end{aligned}$$ which are independent of $n$. This allows us to compare all galaxies irrespective of their Sérsic indices.
Hence, after finding the best–fitting values of $n$ for each individual galaxy we set a cut for the SN sample based on the new variable $v$ rather than $u$. Given that the half–light radii resulting from the Sérsic profile fitting were on average smaller, we limit our sample to SNe within $0.2 < u < 0.85$, equivalent to $0.3
< r / r_{\rm e} < 4$ for the average Sérsic index. We also select galaxies were the best-fitting Sérsic indices were in the expected range $2 < n < 6$. The distribution of the variable $v$ contrasted to $F_v(v)$ is in Fig. \[fig:test\].
Conclusions and discussion {#sec:conclusions}
==========================
We have found that there is no statistically significant difference between the radial dependence of the SN Ia rate in morphologically selected early–type galaxies and their host surface brightness profiles at intermediate radii. The probabilities of not rejecting the hypotheses that SNe Ia follow a de Vaucouleurs or Sérsic profile is 24.0% and 27.0%.
In this work we use host half–light radii ($r_{\rm e}$) normalised distances, obtained after fitting de Vaucouleurs or Sérsic profiles on $r$ band SDSS images, and quantified any deviations from a de Vaucouleurs or Sérsic profile at intermediate radii. Intermediate radii are defined in Section \[sec:analysis\].
The main implications of this work is that the time delay of type Ia SNe in early–type galaxies should be of the order of the age of their host galaxies, or several Gyr. This assumes that young stellar populations of early–type galaxies do not follow the light distribution of their hosts, which is supported by recent observations of central molecular disks and high–resolution SAURON H$\beta$ maps and stellar populations reconstructions, showing that when young ($\lesssim$1 Gyr) stellar populations are present, they are mainly in the central regions of their hosts. Even after including small radii in our sample we did not see an excess of the SN Ia rate in the central parts of early–type galaxies, but given all the possible systematic errors of our method we cannot conclude if there is any real deficit.
Our result does not exclude either the SD or DD scenarios in a picture of multiple progenitor scenarios, but confirms the idea that very young SN Ia progenitor scenarios are probably not dominant in early–type galaxies. They rather indicate that long time delays of several Gyrs for SNe Ia in early–type galaxies are dominant.
Interestingly, early-type galaxies are known to exhibit radial gradients in metallicity, but not in age or $\alpha$-enhancement . The metallicity gradients range all the way to the outer parts of early-type galaxies, beyond two effective radii [@2005ApJ...627..767M]. Given this, our result also suggests that within the metallicity range exhibited by massive early-type galaxies, the SN Ia rate appears not to be strongly affected by metallicity, which hints that long time–delay SN Ia progenitor scenarios may not have a very strong metallicity dependence. Note that a significant dependence has been predicted for metallicities below solar [see @kob98; @2008arXiv0802.2471M].
If the metallicity gradients and surface brightness profiles are relics of the process of galaxy formation in early–type galaxies, our result would suggest that the progenitors of long time–delay SN Ia are as old as their host galaxies.
Future studies of the radial distribution of sub–classes of SNe Ia, e.g. SN1991bg and SN1991T–like events, should provide further insights into whether there are multiple progenitor scenarios [see e.g. @2005ApJ...629L..85S].
Acknowledgements {#acknowledgements .unnumbered}
================
We thank an anonymous referee for relevant comments that significantly improved this manuscript. We are indebted to professional and amateur astronomers that have made their SN discoveries public. We also thank James Binney, Martin Bureau, Alison Crocker, Stephen Justham, Sadegh Khochfar, Phillip Podsiadlowski, Katrien Steenbrugge, Mark Sullivan and Christian Wolf for useful discussions, and Andrés Jordan for providing the package PDL::Minuit. F.F. was supported by a Fundación Andes – PPARC Gemini studentship. K.S. was supported by the Henry Skynner Junior Research Fellowship at Balliol College, Oxford. Parts of the analysis presented here made use of the Perl Data Language[^4]. The SN types and coordinates were obtained from the Central Bureau for Astronomical Telegrams (CBAT) at the Harvard-Smithsonian Center for Astrophysics[^5]. This publication makes use of data from the Sloan Digital Sky Survey. The full acknowledgements can be found at http://www.sdss.org/collaboration/credits.html. We acknowledge the use of NASA’s SkyView facility http://skyview.gsfc.nasa.gov. This work was supported in part through a European Research & Training Network on Type Ia Supernovae (HPRN-CT-20002-00303).
Adelman-McCarthy J. K., et al., 2008, ApJS, 175, 297
Bartunov O. S., Tsvetkov D. Y., Pavlyuk N. N., 2007, HiA, 14, 316
Binney J., Tremaine S., 1987, Princeton Series in Astrophysics
Blanc G., Greggio L., 2008, arXiv, 803, arXiv:0803.3793
Crocker A. F., Bureau M., Young L. M., Combes F., 2008, MNRAS, 490
de Vaucouleurs G., 1948, AnAp, 11, 247
F[ö]{}rster F., Wolf C., Podsiadlowski P., Han Z., 2006, MNRAS, 368, 1893
Fukugita M., Ichikawa T., Gunn J. E., Doi M., Shimasaku K., Schneider D. P., 1996, AJ, 111, 1748
Gal-Yam A., Maoz D., 2004, MNRAS, 347, 942
Graham, A. W., & Guzm[á]{}n, R. 2003, AJ, 125, 2936
Gunn J. E., et al., 1998, AJ, 116, 3040
Hachisu I., Kato M., Nomoto K., 1996, ApJ, 470, L97 Hachisu I., Kato M., Nomoto K., 1999, ApJ, 522, 487
Han Z., Podsiadlowski Ph., 2004, MNRAS, 350, 1301
Hillebrandt W., Niemeyer J. C., 2000, ARA&A, 38, 191
Hopkins A. M., Beacom J. F., 2006, ApJ, 651, 142
Iben I., Tutukov A. V., 1984, ApJS, 54, 335
Ivanov V. D., Hamuy M., Pinto P. A., 2000, ApJ, 542, 588
Kaviraj S., et al., 2007, ApJS, 173, 619
Kobayashi C., Tsujimoto T., Nomoto K., Hachisu I., Kato, M., 1998, ApJ, 503L, 155K
Kuntschner H., et al., 2006, MNRAS, 369, 497
Langer N., Deutschmann A., Wellstein S., H[ö]{}flich P., 2000, A&A, 362, 1046
Lintott C. J., et al., 2008, arXiv, 804, arXiv:0804.4483
Lucero D. M., Young L. M., 2007, AJ, 134, 2148
Mannucci F., della Valle M., Panagia N., Cappellaro E., Cresci G., Maiolino R., Petrosian A., Turatto M., 2005, A&A, 433, 807
Mehlert, D., Thomas, D., Saglia, R. P., Bender, R., & Wegner, G. 2003, A&A, 407, 423
M[é]{}ndez, R. H., Thomas, D., Saglia, R. P., Maraston, C., Kudritzki, R. P., & Bender, R. 2005, ApJ, 627, 767
Meng X., Chen X., Han Z., 2008, arXiv, 802, arXiv:0802.2471
Nomoto K., Kondo Y., 1991, ApJ, 367, L19
Patat F., et al., 2007, Science, 317, 924
Perlmutter S. et al., 1999, ApJ, 517, 565
Phillips M. M., 1993, ApJ, 413, L105
Pier J. R., Munn J. A., Hindsley R. B., Hennessy G. S., Kent S. M., Lupton R. H., Ivezi[ć]{} [Ž]{}., 2003, AJ, 125, 1559
Prieto J. L., Stanek K. Z., Beacom J. F., 2008, ApJ, 673, 999
Riess A. G. et al., 1998, AJ, 116, 1009
Saio H., Nomoto K., 1998, ApJ, 500, 388
Saglia, R. P., Maraston, C., Greggio, L., Bender, R., & Ziegler, B. 2000, A&A, 360, 911
Sarzi M., et al., 2006, MNRAS, 366, 1151
Scannapieco E., Bildsten L., 2005, ApJ, 629, L85
Schawinski K., et al., 2006, Nature, 442, 888
Schawinski K., et al., 2007, ApJS, 173, 512
Sérsic J. L., 1968, Atlas de Galaxias Australes. Observatorio Astronómico, Córdoba, Argentina.
Shapiro S. L., Teukolsky S. A., Black holes, white dwarfs and neutron stars: The physics of compact objects
Sharon K., Gal-Yam A., Maoz D., Filippenko A. V., Guhathakurta P., 2007, ApJ, 660, 1165
Shaw R. L., 1979, A&A, 76, 188
Stoughton C., et al., 2002, AJ, 123, 485
Strolger L.-G. et al., 2004, ApJ, 613, 200
Sullivan M., et al., 2006, ApJ, 648, 868
Thomas, D., Maraston, C., Bender, R., & Mendes de Oliveira, C. 2005, ApJ, 621, 673
Totani T., Morokuma T., Oda T., Doi M., Yasuda N., 2008, arXiv, 804, arXiv:0804.0909
Webbink R. F., 1984, ApJ, 277, 355
Wu, H., Shao, Z., Mo, H. J., Xia, X., & Deng, Z. 2005, ApJ, 622, 244
Yi S. K., et al., 2005, ApJ, 619, L111
Yoon S.-C., Podsiadlowski P., Rosswog S., 2007, MNRAS, 380, 933
York, D. G., et al. 2000, AJ, 120, 1579
[^1]: [[email protected]]{}
[^2]: [[email protected]]{}
[^3]: See: `http://www.cfa.harvard.edu/iau/lists/Supernovae.html`
[^4]: The Perl Data Language (PDL) has been developed by K. Glazebrook, J. Brinchmann, J. Cerney, C. DeForest, D. Hunt, T. Jenness, T. Luka, R. Schwebel, and C. Soeller and can be obtained from http://pdl.perl.org
[^5]: `http://www.cfa.harvard.edu/iau/lists/Supernovae.html`
| {
"pile_set_name": "ArXiv"
} | ArXiv |
CERN-PH-TH/2014-009\
OUTP-14-10p\
[**[aMCfast]{}: automation of fast NLO computations for PDF fits**]{}
Valerio Bertone$^{1}$, Rikkert Frederix$^{1}$, Stefano Frixione$^{1}$, Juan Rojo$^{1,2}$, and Mark Sutton$^{3}$
[**Abstract**]{}
We present the interface between [[MadGraph5\_aMC@NLO]{}]{}, a self-contained program that calculates cross sections up to next-to-leading order accuracy in an automated manner, and [[APPLgrid]{}]{}, a code that parametrises such cross sections in the form of look-up tables which can be used for the fast computations needed in the context of PDF fits. The main characteristic of this interface, which we dub [[aMCfast]{}]{}, is its being fully automated as well, which removes the need to extract manually the process-specific information for additional physics processes, as is the case with other matrix-element calculators, and renders it straightforward to include any new process in the PDF fits. We demonstrate this by studying several cases which are easily measured at the LHC, have a good constraining power on PDFs, and some of which were previously unavailable in the form of a fast interface.
Introduction {#sec:intro}
============
The accurate determination of the Parton Distribution Functions (PDFs) of the proton [@Forte:2013wc; @Ball:2012wy; @Perez:2012um; @Forte:2010dt; @DeRoeck:2011na] is one of the most important tasks for precision phenomenology at the Large Hadron Collider (LHC). PDFs are a dominant source of theoretical uncertainty in the predictions for Higgs boson production, where the errors that affect them degrade the accuracy of the Higgs characterization in terms of couplings and branching fractions [@Dittmaier:2011ti]; they induce large uncertainties in the cross sections for processes with very massive new-physics particles [@AbelleiraFernandez:2012ty]; and they substantially affect Standard Model (SM) precision measurements such as those of the mass of the $W$ boson [@Bozzi:2011ww] and of the effective lepton mixing angle $\sin^2\theta^l_{\rm eff}$ [@Rojo:2013wba]. For these reasons, an active program towards better PDFs is being carried out by different groups [@Martin:2009iq; @Ball:2010de; @Gao:2013xoa; @Alekhin:2013nda; @CooperSarkar:2011aa], which emphasise the use of new experimental inputs, more accurate theoretical calculations, and improved fitting methodology.
Modern global PDF analyses are based on a variety of hard-scattering data such as deep-inelastic scattering (DIS) structure functions at fixed-target experiments, lepton-proton cross sections from the HERA collider, and inclusive $W$, $Z$, and jet production at hadron colliders. Since the beginning of the LHC data taking, these and many other processes for which measurements have become available and can be used to constrain PDFs. For example, LHC data that might provide information on PDFs encompass inclusive electroweak vector boson production [@Aad:2011dm; @Chatrchyan:2013mza], inclusive jet and dijet production [@Aad:2011fc; @Chatrchyan:2012bja; @Aad:2013tea; @Aad:2013lpa; @Watt:2013oha], direct photon production [@d'Enterria:2012yj], top quark pair production cross sections [@Czakon:2013tha], $W$ production in association with charm quarks [@Chatrchyan:2013uja; @Chatrchyan:2013mza], low and high-mass Drell-Yan production [@Chatrchyan:2013tia], the $W$ and $Z$ bosons large-${p_{{\scriptscriptstyle}T}}$ distributions and their ratios [@Malik:2013kba], high-mass $W$ production, and single-top production, as well as ratios of cross sections measured at different center-of-mass energies [@Mangano:2012mh; @Aad:2013lpa].
It is thus clear that a wide variety of high-quality measurements sensitive to PDFs are already available; more data will follow in the coming years. Therefore, in order to improve the accuracy of PDF determinations, it is essential to include in PDF fits as many of these measurements as possible, in order to constrain different combinations of PDFs in a wide range of Bjorken $x$’s. The most serious difficulty in doing so is due to the fact that next-to-leading order (NLO) QCD calculations of hadron collider processes with realistic acceptance cuts are much slower than what is needed in the iterative PDF fitting procedure, which requires computing the theoretical predictions a very large number of times. In order to bypass this problem, a popular solution in the past has been that of performing leading order (LO) computations, supplemented by bin-by-bin $K$ factors. Unfortunately, such a solution is not sufficiently accurate to match the precision of present and future LHC data; in particular, it has the very undesirable feature of neglecting the combinations of initial-state partons that do not appear at the LO.
In order to overcome this problem, several solutions have been proposed. The underlying idea common to all these approaches is: interpolating the PDFs in the ${\left(}x,Q^2{\right)}$-plane with some suitable polynomial basis; precomputing the hadronic cross section by using the basis members as input; and finally reconstructing the original calculation with the numerical convolution of the precomputed cross sections and the actual PDFs. Note that the information about the latter is only required at a [*finite*]{} number of points $(x_i,Q^2_j)$, which are called the interpolating-grid nodes. Therefore, the time-consuming task of the precomputation of the cross section with basis members is performed only once, and the reconstruction of the full result associated with arbitrary PDFs is extremely fast.
The strategy sketched above, which is closely related to the one adopted by $x$-space PDF evolution codes [@Salam:2008qg; @Botje:2010ay; @Bertone:2013vaa], is what underpins the two best-known fast interpolators of NLO QCD cross sections, [[FastNLO]{}]{} [@Kluge:2006xs; @Wobisch:2011ij] and [[APPLgrid]{}]{} [@Carli:2010rw]. [[FastNLO]{}]{} is interfaced to the jet-production code [NLOjet++]{} [@Nagy:2001fj], and is thus able to provide fast computations of multijet production at lepton-proton and hadron-hadron colliders[^1]. [[APPLgrid]{}]{} is interfaced to various programs, including [NLOjet++]{} and [MCFM]{} [@Campbell:2002tg]. The main drawback of these tools, namely that of being capable of handling a relatively small number of processes, is a consequence of the fact that adding new ones requires an ad-hoc procedure, which is rather time-consuming and error-prone. This is also the principal reason why they are only interfaced to calculations which are of NLO (and, in the one case mentioned above, NNLO) in QCD, but neither to (N)LO results matched to parton showers, nor to NLO results in the electroweak theory.
The goal of this work is that of solving all of these problems in a general manner, which is possible thanks to the fact that NLO(+PS) calculations can now be routinely done by means of automated codes. In fact, among the many features of such codes there are two which are directly relevant to the problems at hand: firstly, given a sufficient amount of CPU power the cross section for any process, however complicated, can be computed; and secondly, the way in which these cross sections are handled and the form in which they are written are completely standardised. This is what renders it possible to construct a generic and automated interface between an automated cross section calculator and a fast interpolator. The main result of this paper is the construction of such an automated interface, that we call [[aMCfast]{}]{}, and which bridges the automated cross section calculator [[MadGraph5\_aMC@NLO]{}]{} [@Alwall:2014hca] with the fast interpolator [[APPLgrid]{}]{}. Thus, the chain [[MadGraph5\_aMC@NLO]{}]{} – [[aMCfast]{}]{} – [[APPLgrid]{}]{} will allow one to include, in a straighforward manner, any present or future LHC measurement in an NLO global PDF analysis. We point out that a strategy analogous to that pursued in the present paper has motivated the recent construction of [MCgrid]{} [@DelDebbio:2013kxa], whereby [[APPLgrid]{}]{} has been interfaced to [Rivet]{} [@Buckley:2010ar].
We remind the reader that [[MadGraph5\_aMC@NLO]{}]{} contains all ingredients relevant to the computations of LO and NLO cross sections, with or without matching to parton showers. NLO results not matched to parton showers (called fNLO [@Alwall:2014hca]) are obtained by adopting the FKS method [@Frixione:1995ms; @Frixione:1997np] for the subtraction of the singularities of the real-emission matrix elements (automated in the module [[MadFKS]{}]{} [@Frederix:2009yq]), and the OPP integral-reduction procedure [@Ossola:2006us] for the computation of the one-loop matrix elements (automated in the module [[MadLoop]{}]{} [@Hirschi:2011pa], which makes use of [[CutTools]{}]{} [@Ossola:2007ax] and of an in-house implementation of the optimisations proposed in ref. [@Cascioli:2011va] ([[OpenLoops]{}]{})). Matching with parton showers is achieved by means of the MC@NLO formalism [@Frixione:2002ik]. In the present public version, [[MadGraph5\_aMC@NLO]{}]{} is restricted to computing NLO QCD corrections to SM processes; however, as discussed in ref. [@Alwall:2014hca], all obstacles that enforce such a limitation have been cleared, paving the way to higher-order calculations in the context of arbitrary renormalisable theories in the near future. As far as [[APPLgrid]{}]{} is concerned, only a few high-level routines have been extended in view of its interface with [[aMCfast]{}]{}. All of these modifications are of a technical character, except one which is related to the computation of the factorisation and renormalisation scale dependences, as we shall explain in more detail later.
The scope of this paper is that of fNLO QCD computations. Alternatively, in [[aMCfast]{}]{} the specific nature of the perturbative expansion is used only in a rather trivial way, since it determines a number of interpolating grids and their linear combination that defines the physical cross sections. Therefore, when e.g. electroweak corrections will become publicly available in [[MadGraph5\_aMC@NLO]{}]{}, that feature will be immediately inherited by [[aMCfast]{}]{} through some straighforward generalisation. Furthermore, what is done here will allow one, with only a few minimal extensions, to construct fast interfaces to NLO+PS predictions. This is expected to have several beneficial effects in the context of PDF fits: a closer connection to the experimentally accessible observables, a wider range of data that can be used to constrain PDFs, and the possibility of eventually extract PDFs that are specifically tailored to their use with NLO event generators.
This paper is organised as follows: in sect. \[sec:amcfast\] we give a short introduction to the interpolating-grid techniques employed here, review the computation of cross sections in [[MadGraph5\_aMC@NLO]{}]{}, and discuss how the separation between PDFs, partonic cross sections, and ${\alpha_{{\scriptscriptstyle}S}}$ dependence can be exploited to construct a fast interface using the [[APPLgrid]{}]{} routines. The flexibility of [[aMCfast]{}]{}is illustrated in sect. \[sec:pheno\], where we present results for many relevant LHC processes, some of which are obtained for the first time with a fast NLO interface, and whose numerical performance and accuracy are analysed. Finally, in sect. \[sec:outlook\] we summarise our findings and briefly discuss the plans for the future developments of [[aMCfast]{}]{}. An appendix reports some additional information.
Automation of fast NLO computations\[sec:amcfast\]
==================================================
In this section we first outline the basics of an interpolation technique based on the expansion of a given function onto a basis of polynomials; we then discuss how short-distance cross sections can be represented, in the most general manner, in terms of interpolating grids, that allow them to be quickly computed with arbitrary PDFs, and factorisation and renormalisation scales. Finally, we show how these formulae can be employed in the construction of the [[aMCfast]{}]{} bridge that interfaces [[APPLgrid]{}]{} to [[MadGraph5\_aMC@NLO]{}]{}.
The construction of interpolating grids\[sec:gen\]
--------------------------------------------------
The basic idea used by [[APPLgrid]{}]{} is that of a Lagrange-polynomial expansion. Given a function $F(z)$ of a real variable $z$, one has the representation: F(z)=\_[i=0]{}\^s F(( +i)) I\_i\^[(s)]{}(- ), \[frepr\] where $s$ is a given integer (the interpolation order), $\delta$ is a small number (the grid spacing or binning; $p\delta$ for some integer $p$ is called a grid node), the interpolating functions are: I\_i\^[(s)]{}(u)= \_[k=0,ki]{}\^s (u-k), \[Ibasis\] and we have denoted by $\floor{u}$ the largest integer smaller than or equal to $u$: , u-1<u, u. The equality in eq. (\[frepr\]) holds up to functional terms of order $I_i^{(s+1)}$ and higher. Such terms vanish identically when the argument of $F$ coincides with a grid node, so that eq. (\[frepr\]) is an identity in that case; this is straightforward to prove, and follows directly from the values that the interpolating functions take when computed with an integer argument: I\_i\^[(s)]{}(k)=\_[ik]{}, 0ks, k. When $z$ is not a grid node (i.e., $z\ne p\delta$ for any integer $p$), eq. (\[frepr\]) tells one that $F(z)$ is reconstructed by using the values that $F$ takes in the $(s+1)$ grid nodes which are nearest to $z$; the number of relevant nodes to the left of $z$ is equal to number of nodes to the right of $z$, possibly up to one.
For any given function $S(z)$, let us now compute the simple example integral J=\_a\^b dz S(z) F(z) \[Idef\] by means of its corresponding Riemann sums (or, equivalently, by Monte Carlo methods). This implies J=\_[k=1]{}\^M \_k S(z\_k) F(z\_k), \[IMC\] with $M$ points $z_k\in (a,b)$, and $\Phi_k$ suitable normalisation factors. By using eq. (\[frepr\]), we obtain J=\_[k=1]{}\^M \_k S(z\_k) \_[i=0]{}\^s F((p\_(z\_k)+i)) I\_i\^[(s)]{}(-p\_(z\_k)), \[I0\] where we have defined: p\_(z)=, \[pdeldef\] which is the integer associated with the leftmost grid node in the set of the $(s+1)$ nearest neighbours of $z$. By means of a change of the summation variable $i$, eq. (\[I0\]) becomes: J&=&\_[k=1]{}\^M \_k S(z\_k) \_[j=p\_(z\_k)]{}\^[s+p\_(z\_k)]{} F(j) I\_[j-p\_(z\_k)]{}\^[(s)]{}(-p\_(z\_k)) \[I1\]\
&=& \_[j=-]{}\^ F(j) G\_j, \[I2\] with G\_j=\_[k=1]{}\^M \_k S(z\_k) I\_[j-p\_(z\_k)]{}\^[(s)]{}(-p\_(z\_k)) (p\_(z\_k)js+p\_(z\_k)). \[griddef\] Equation (\[griddef\]) defines the grid values $G_j$. Owing to the ${\Theta}$ function it contains, the sum in eq. (\[I2\]) features a finite number of non-null contributions (if the range $(a,b)$ is finite). Thus, the meaning of eq. (\[I2\]) is that the integral $J$ can be computed a posteriori by knowing a finite number of grid values, and the values of the function $F$ at the grid nodes; more importantly, this a-posteriori computation can be done for any function $F$, because the grid values are independent of $F$, and can therefore be pre-evaluated and stored for a given function $S$. This also explains the reason for writing the integrand of eq. (\[Idef\]) as a product of two functions: this is convenient whenever the time spent in evaluating $F$ (the “fast” function) and $S$ (the “slow” function) is small and large respectively. When this is the case, the computation of the grid $\{G_j\}$ may be time-consuming, but it is an operation that has to be carried out only once; on the other hand, the subsequent evaluations of eq. (\[I2\]) will be quick. We also point out that the derivation above is unchanged in the case where $z$ is not the integration variable, but a function itself of one (or more) integration variable(s) for the problem at hand. This is because the starting point is actually eq. (\[IMC\]), and not eq. (\[Idef\]), and in the former the role of $z_k$ as integration variable can be fully ignored.
Generalities on short-distance cross sections\[sec:obs\]
--------------------------------------------------------
In what follows we use the expressions derived within the FKS subtraction formalism. The same notations as in ref. [@Frederix:2011ss] are adopted; this is particularly convenient in view of the fact that, in that paper, cross sections were represented in terms of PDF- and scale-independent coefficients, and these will be the main ingredients in the definition of the interpolating grids. We point out, however, that the procedure outlined below remains valid regardless of the subtraction method chosen to perform the NLO computations. Lest we clutter the notation with details which are irrelevant here, we work by fixing the partonic process. This implies, in particular, that we do not need to write explicitly the identities of the incoming partons; we shall reinstate them later.
The NLO short-distance cross section relevant to a $2\to n+1$ process consists of four terms: d[\^[(NLO)]{}]{}&&{d[\^[[(NLO]{},]{}]{}}\_[=E,S,C,SC]{}, \[fact3aAPP\]\
d[\^[[(NLO]{},]{}]{}&=&[f\_1]{}([x\_1\^[()]{}]{},[\_[F]{}\^[()]{}]{})[f\_2]{}([x\_2\^[()]{}]{},[\_[F]{}\^[()]{}]{}) [W\^[()]{}]{}d\_[Bj]{}d\_[n+1]{}, \[fact3\] called event ($\alpha=E$), and soft, collinear, and soft-collinear ($\alpha=S, C, SC$) counterevents, respectively. The quantities $d{\chi}_{Bj}$ and $d{\chi}_{n+1}$ are the integration measures over the Bjorken $x$’s and the $(3n-1)$ phase-space variables respectively, while ${f_1}$ and ${f_2}$ are the PDFs relevant to the colliding partons coming from the left and from the right. The ${W^{(\alpha)}}$’s can be parametrised as follows: [W\^[()]{}]{}&=&[g\_[S]{}]{}\^[2b+2]{}([\_[R]{}\^[()]{}]{})\
&+& [g\_[S]{}]{}\^[2b]{}([\_[R]{}\^[()]{}]{})[\_[B]{}]{}\_, \[Wadef\] where the coefficients $\widehat{W}$ are (renormalisation and factorisation) scale- and PDF-independent; the last term on the r.h.s. of eq. (\[Wadef\]) is the Born contribution which, as the notation suggests, factorises ${\alpha_{{\scriptscriptstyle}S}}^b$. In eqs. (\[fact3\]) and (\[Wadef\]) we have denoted by $x_i^{(\alpha)}$, ${\mu_{{\scriptscriptstyle}F}^{(\alpha)}}$, and ${\mu_{{\scriptscriptstyle}R}^{(\alpha)}}$ the Bjorken $x$’s, factorisation scale, and renormalisation scale respectively; in general, they may assume different values in the event and various counterevent configurations. Finally, $Q$ is the Ellis-Sexton scale[^2]. The integration of the above cross section leads to a set of $4N$ weighted events: {{[[[K]{}]{}\_[n+1;k]{}\^[()]{}]{},[x\_[1;k]{}\^[()]{}]{},[x\_[2;k]{}\^[()]{}]{},[\_k\^[()]{}]{}}\_[=E,S,C,SC]{} }\_[k=1]{}\^N, \[NLOsetApp\] with ${{{\cal K}}_{n+1;k}^{(\alpha)}}$ an $(n+1)$-body kinematic configuration (possibly degenerate, in which case it is effectively an $n$-body configuration that can be used to compute Born matrix elements), and the event weights defined as follows: [\_k\^[()]{}]{}= ([[[K]{}]{}\_[n+1;k]{}\^[()]{}]{},[x\_[1;k]{}\^[()]{}]{},[x\_[2;k]{}\^[()]{}]{}), \[wgtNLO\] where we understand possible normalization pre-factors (such as $1/N$, or adaptive-integration jacobians). Given an observable $O$, its $h^{th}$ histogram bin defined by will assume, at the end of the run, the value: \_O\^[(h)]{}&=& \_[k=1]{}\^N\_[\_k\^[(h,)]{}]{}[\_k\^[()]{}]{}, \[histoNLO\]\
[\_k\^[(h,)]{}]{}&=& (O([[[K]{}]{}\_[n+1;k]{}\^[()]{}]{})-O\_[LOW]{}\^[(h)]{}) (O\_[UPP]{}\^[(h)]{}-O([[[K]{}]{}\_[n+1;k]{}\^[()]{}]{})). \[binNLO\] In view of eq. (\[Wadef\]), it is convenient to recast eq. (\[histoNLO\]) as follows: \_O\^[(h)]{}=\_[=0,F,R,B]{}\_[O,]{}\^[(h)]{}, \[histoNLO2\] where, using eqs. (\[fact3\]) and (\[wgtNLO\]), one has: \_[O,0]{}\^[(h)]{}&=&\_[k=1]{}\^N\_[\_k\^[(h,)]{}]{} [f\_1]{}([x\_[1;k]{}\^[()]{}]{},[\_[[F]{};k]{}\^[()]{}]{})[f\_2]{}([x\_[2;k]{}\^[()]{}]{},[\_[[F]{};k]{}\^[()]{}]{})[g\_[S]{}]{}\^[2b+2]{}([\_[[R]{};k]{}\^[()]{}]{})[\^[()]{}\_[0;k]{}]{}, \[histoNLOz\]\
\_[O,F]{}\^[(h)]{}&=&\_[k=1]{}\^N\_[\_k\^[(h,)]{}]{} [f\_1]{}([x\_[1;k]{}\^[()]{}]{},[\_[[F]{};k]{}\^[()]{}]{})[f\_2]{}([x\_[2;k]{}\^[()]{}]{},[\_[[F]{};k]{}\^[()]{}]{})[g\_[S]{}]{}\^[2b+2]{}([\_[[R]{};k]{}\^[()]{}]{}) [\^[()]{}\_[[F]{};k]{}]{}[()\^2]{}, \[histoNLOF\]\
\_[O,R]{}\^[(h)]{}&=&\_[k=1]{}\^N\_[\_k\^[(h,)]{}]{} [f\_1]{}([x\_[1;k]{}\^[()]{}]{},[\_[[F]{};k]{}\^[()]{}]{})[f\_2]{}([x\_[2;k]{}\^[()]{}]{},[\_[[F]{};k]{}\^[()]{}]{})[g\_[S]{}]{}\^[2b+2]{}([\_[[R]{};k]{}\^[()]{}]{}) [\^[()]{}\_[[R]{};k]{}]{}[()\^2]{}, \[histoNLOR\]\
\_[O,B]{}\^[(h)]{}&=&\_[k=1]{}\^N\_[\_k\^[(h,)]{}]{} [f\_1]{}([x\_[1;k]{}\^[()]{}]{},[\_[[F]{};k]{}\^[()]{}]{})[f\_2]{}([x\_[2;k]{}\^[()]{}]{},[\_[[F]{};k]{}\^[()]{}]{})[g\_[S]{}]{}\^[2b]{}([\_[[R]{};k]{}\^[()]{}]{}) [\^[()]{}\_[[B]{};k]{}]{}. \[histoNLOB\] Here we have defined: [\^[()]{}\_[;k]{}]{}&=&[\^[()]{}\_]{}([[[K]{}]{}\_[n+1;k]{}\^[()]{}]{})=0,F,R,\
[\^[()]{}\_[[B]{};k]{}]{}&=&[\_[B]{}]{}([[[K]{}]{}\_[n+1;k]{}\^[(S)]{}]{})\_,\
[\_[[F]{};k]{}\^[()]{}]{}&=&[\_[F]{}]{}([[[K]{}]{}\_[n+1;k]{}\^[()]{}]{}),\
[\_[[R]{};k]{}\^[()]{}]{}&=&[\_[R]{}]{}([[[K]{}]{}\_[n+1;k]{}\^[()]{}]{}), namely, the values of the short-distance coefficients and of the scales computed at the kinematic configurations associated with each of the events and counterevents of eq. (\[NLOsetApp\]).
We now apply the method outlined in sect. \[sec:gen\] to eqs. (\[histoNLOz\])–(\[histoNLOB\]). In order to do so, we recall that the main result of that section is that of computing the number $J$, whose original expression is that of eq. (\[IMC\]) (regardless of the fact that such an expression was in turn obtained from the integral of eq. (\[Idef\]), for the reason explained at the end of sect. \[sec:gen\]), by means of the interpolating grid given in eq. (\[griddef\]) and of eq. (\[I2\]).
One starts by observing that the $h^{th}$ bin value $\sigma_{O,\beta}^{(h)}$ of eqs. (\[histoNLOz\])–(\[histoNLOB\]) is written in the same form as $J$ of eq. (\[IMC\]): the double sums over $k$ and $\alpha$ in eqs. (\[histoNLOz\])–(\[histoNLOB\]) can easily be recast in the form of a single sum as in eq. (\[IMC\]). The goal is therefore that of representing $\sigma_{O,\beta}^{(h)}$ with an interpolating grid, as is done for $J$ in eq. (\[I2\]). In order to do so, one may easily proceed by analogy. First of all, since we are interested in a quick recomputation of the cross section after changing the scales and the PDFs, and given that the latter depend on the Bjorken $x$’s and the factorisation scale, the implication is that the role of the variable $z$ of eqs. (\[IMC\])–(\[griddef\]) needs to be played by the four variables: [x\_1]{},[x\_2]{},[\_[F]{}]{},[\_[R]{}]{}. \[varset\] In other words, we have the correspondence: {z\_k}\_[k=1,M]{} {[x\_[1;k]{}\^[()]{}]{},[x\_[2;k]{}\^[()]{}]{},[\_[[F]{};k]{}\^[()]{}]{},[\_[[R]{};k]{}\^[()]{}]{}}\_[k=1,N]{}\^[=E,S,C,SC]{}. The polynomial expansion of eq. (\[frepr\]), whence the interpolating grid is derived, is valid for each of the variables in eq. (\[varset\]); therefore, the only change w.r.t. the case of the computation of $J$ is the fact that the one-dimensional grid $\{G_j\}$ defined in eq. (\[griddef\]) will be replaced by a four-dimensional grid, corresponding to the variables in eq. (\[varset\]). For the construction of the latter, it is sufficient to compare directly eq. (\[IMC\]) with eqs. (\[histoNLOz\])–(\[histoNLOB\]). The role of the “slow” and “fast” functions $S$ and $F$ will be played by the short-distance coefficients $\widehat{W}$ and by the scale- and PDF-dependent terms respectively; that of the factor $\Phi_k$ by the bin-defining ${{\Theta}_k^{(h,\alpha)}}$ of eq. (\[binNLO\]). In other words, we have the following identifications: =0 (S,F)&& ([\_0]{},[f\_1]{}([x\_1]{},[\_[F]{}]{})[f\_2]{}([x\_2]{},[\_[F]{}]{})[g\_[S]{}]{}\^[2b+2]{}([\_[R]{}]{})), \[NLOgr1\]\
=F (S,F)&& ([\_[F]{}]{},[f\_1]{}([x\_1]{},[\_[F]{}]{})[f\_2]{}([x\_2]{},[\_[F]{}]{})[g\_[S]{}]{}\^[2b+2]{}([\_[R]{}]{})[()\^2]{}), \[NLOgr2\]\
=R (S,F)&& ([\_[R]{}]{},[f\_1]{}([x\_1]{},[\_[F]{}]{})[f\_2]{}([x\_2]{},[\_[F]{}]{})[g\_[S]{}]{}\^[2b+2]{}([\_[R]{}]{})[()\^2]{}), \[NLOgr3\]\
=B (S,F)&& ([\_[B]{}]{},[f\_1]{}([x\_1]{},[\_[F]{}]{})[f\_2]{}([x\_2]{},[\_[F]{}]{})[g\_[S]{}]{}\^[2b]{}([\_[R]{}]{})), \[NLOgr4\] and \_k[\_k\^[(h,)]{}]{}. \[NLOgr5\] It is now sufficient to use eqs. (\[NLOgr1\])–(\[NLOgr5\]) to rewrite eqs. (\[I2\]) and (\[griddef\]). We denote by $j_1$, $j_2$, $j_3$, and $j_4$ the indices that run over the grid nodes relevant to the variables ${x_1}$, ${x_2}$, ${\mu_{{\scriptscriptstyle}F}}$, and ${\mu_{{\scriptscriptstyle}R}}$, respectively; $\delta_i$ and $s_i$, $i=1,\ldots 4$, are the corresponding grid spacings and interpolating orders. We obtain: \_[O,0]{}\^[(h)]{}&=&\_[j\_1,j\_2,j\_3,j\_4]{} [f\_1]{}(j\_1\_1,j\_3\_3)[f\_2]{}(j\_2\_2,j\_3\_3) [g\_[S]{}]{}\^[2b+2]{}(j\_4\_4)G\_[j\_1j\_2j\_3j\_4]{}\^[(h,0)]{}, \[gridNLOz\]\
\_[O,F]{}\^[(h)]{}&=&\_[j\_1,j\_2,j\_3,j\_4]{} [f\_1]{}(j\_1\_1,j\_3\_3)[f\_2]{}(j\_2\_2,j\_3\_3) [g\_[S]{}]{}\^[2b+2]{}(j\_4\_4) ()\^2 G\_[j\_1j\_2j\_3j\_4]{}\^[(h,F)]{}, \[gridNLOF\]\
\_[O,R]{}\^[(h)]{}&=&\_[j\_1,j\_2,j\_3,j\_4]{} [f\_1]{}(j\_1\_1,j\_3\_3)[f\_2]{}(j\_2\_2,j\_3\_3) [g\_[S]{}]{}\^[2b+2]{}(j\_4\_4) ()\^2 G\_[j\_1j\_2j\_3j\_4]{}\^[(h,R)]{}, \[gridNLOR\]\
\_[O,B]{}\^[(h)]{}&=&\_[j\_1,j\_2,j\_3,j\_4]{} [f\_1]{}(j\_1\_1,j\_3\_3)[f\_2]{}(j\_2\_2,j\_3\_3) [g\_[S]{}]{}\^[2b]{}(j\_4\_4)G\_[j\_1j\_2j\_3j\_4]{}\^[(h,B)]{}, \[gridNLOB\] with the interpolating grids: G\_[j\_1j\_2j\_3j\_4]{}\^[(h,)]{}&=& \_[k=1]{}\^N\_[\_k\^[(h,)]{}]{}[\^[()]{}\_[;k]{}]{} \[4dimgrid\]\
&&I\_[j\_1-p\_[\_1]{}([x\_[1;k]{}\^[()]{}]{})]{}\^[(s\_1)]{} (-p\_[\_1]{}([x\_[1;k]{}\^[()]{}]{})) (p\_[\_1]{}([x\_[1;k]{}\^[()]{}]{})j\_1s\_1+p\_[\_1]{}([x\_[1;k]{}\^[()]{}]{}))\
&&I\_[j\_2-p\_[\_2]{}([x\_[2;k]{}\^[()]{}]{})]{}\^[(s\_2)]{} (-p\_[\_2]{}([x\_[2;k]{}\^[()]{}]{})) (p\_[\_2]{}([x\_[2;k]{}\^[()]{}]{})j\_2s\_2+p\_[\_2]{}([x\_[2;k]{}\^[()]{}]{}))\
&&I\_[j\_3-p\_[\_3]{}([\_[[F]{};k]{}\^[()]{}]{})]{}\^[(s\_3)]{} (-p\_[\_3]{}([\_[[F]{};k]{}\^[()]{}]{})) (p\_[\_3]{}([\_[[F]{};k]{}\^[()]{}]{})j\_3s\_3+p\_[\_3]{}([\_[[F]{};k]{}\^[()]{}]{}))\
&&I\_[j\_4-p\_[\_4]{}([\_[[R]{};k]{}\^[()]{}]{})]{}\^[(s\_4)]{} (-p\_[\_4]{}([\_[[R]{};k]{}\^[()]{}]{})) (p\_[\_4]{}([\_[[R]{};k]{}\^[()]{}]{})j\_4s\_4+p\_[\_4]{}([\_[[R]{};k]{}\^[()]{}]{})). Equations (\[gridNLOz\])–(\[4dimgrid\]) give the most general representation of NLO cross sections in terms of interpolating grids. We shall employ them (in a simplified form) in the next subsection, in the construction of the [[aMCfast]{}]{} bridge to [[APPLgrid]{}]{}.
The interface of [APPLgrid]{} with [MadGraph5\_aMC@NLO]{}\[sec:int\]
--------------------------------------------------------------------
[[APPLgrid]{}]{} is a general-purpose [C++]{} library that provides a suitable number of interpolation and convolution routines that can be used to construct fast interfaces to NLO calculations of lepton-proton and hadron-hadron collider processes. [[APPLgrid]{}]{} is widely used by various PDF fitting collaborations as well as by ATLAS and CMS in their own PDF studies. Our purpose is that of exploiting [[APPLgrid]{}]{} for the construction of the interpolating grids of eq. (\[4dimgrid\]), and their subsequent use in the calculation of the histogram bins, eqs. (\[gridNLOz\])–(\[gridNLOB\]). As those equations show, [[APPLgrid]{}]{} needs to be given, in an initialisation phase, the observables to be computed and their respective binnings (we stress again that each bin of each observable corresponds to a set of four grids per type of parton luminosity); and, on an event-by-event basis, the values of those observables and the short-distance coefficients $\widehat{W}$. These tasks are essentially what [[aMCfast]{}]{} is responsible for: it extracts the relevant information from [[MadGraph5\_aMC@NLO]{}]{}, and feeds them to [[APPLgrid]{}]{}, in the format required by the latter.
The first observation is that, in its current version, [[APPLgrid]{}]{} supports three-dimensional grids, while those in eq. (\[4dimgrid\]) are four-dimensional. The reason for the latter is that in our derivation we have left the freedom of choosing different functional forms for the factorisation and renormalisation scales. On the other hand, such a flexibility is seldom exploited, and typically one chooses ${\mu_{{\scriptscriptstyle}F}}$ and ${\mu_{{\scriptscriptstyle}R}}$ equal in the whole phase space, up to an overall constant. This is equivalent to setting: [\_[F]{}]{}=[\_[F]{}]{}, [\_[R]{}]{}=[\_[R]{}]{}, \[FeqR\] with $\mu$ a function of the kinematics. When doing this, two of the four variables in eq. (\[varset\]) are degenerate, and therefore one must simply consider: [x\_1]{},[x\_2]{},, \[varset2\] which reduces the number of grid dimensions from four to three. In this case, it is also convenient to set[^3]: Q=, \[Qeqmu\] so that ()\^2 [\_[F]{}]{}\^2, ()\^2 [\_[R]{}]{}\^2, in eqs. (\[gridNLOF\]) and (\[gridNLOR\]) respectively. The other changes to eqs. (\[gridNLOz\])–(\[4dimgrid\]) due to the reduction of the grid dimensionality are all trivial: one eliminates the sums over $j_4$, formally replaces: j\_3\_3j\_3\_3[\_[F]{}]{}, j\_4\_4j\_3\_3[\_[R]{}]{}, \[repl\] and ${\mu_{{\scriptscriptstyle}F}}$ with $\mu$ in the next-to-last line on the r.h.s. of eq. (\[4dimgrid\]) (this and eq. (\[repl\]) are due to the fact that now it is the variable $\mu$ which corresponds to one of the dimensions of the grid), and finally eliminates the last line of that equation.
The next thing to consider is that [[APPLgrid]{}]{} does not use the variables in eq. (\[varset2\]) directly, but rather constructs the grids using: [y\_1]{},[y\_2]{},, \[varset3\] which are defined through a change of variables, several of which are available but is usually set to: &&y\_i=Y(x\_i) Y(x)=+(1-x), \[ydef\]\
&&=T()=. \[taudef\] This is because the grids feature a linear spacing, and given the PDFs and ${\alpha_{{\scriptscriptstyle}S}}$ dependences on Bjorken $x$’s and scales a linear sampling in terms of the variables of eq. (\[varset3\]) turns out to be more effective than one based the variables of eq. (\[varset2\]). In eq. (\[ydef\]) the parameter $\kappa$ controls the relative density of grid nodes in the large- w.r.t. the small-$x$ region; in eq. (\[taudef\]), the parameter $\Lambda$ is typically chosen to be of the order of $\Lambda_{\rm{\scriptscriptstyle}QCD}$.
Finally, one needs to consider the fact that the formulae presented above are obtained for a given partonic process. In order to compute the actual hadronic cross section, one needs to sum over all possible such processes, so that eq. (\[fact3\]) must be generalised as follows: d[\^[[(NLO]{},]{}]{}=\_[rsq]{}[f\_r]{}([x\_1\^[()]{}]{},[\_[F]{}\^[()]{}]{})[f\_s]{}([x\_2\^[()]{}]{},[\_[F]{}\^[()]{}]{}) [W\_[rsq]{}\^[()]{}]{}d\_[Bj]{}d\_[n+1]{}, \[fact4\] where $r$ and $s$ are the identities of the incoming partons, and $q$ collectively denotes the identities of all outgoing partons. Note that, for simple-enough cases, the sum over $q$ is trivial, since $(r,s)$ are sufficient to determine unambiguously a partonic process; however, the notation used in eq. (\[fact4\]) is general, and encompasses all possible situations. Since the interpolating grids are additive, the most straightforward solution is that of following the procedure outlined so far, and of creating a grid for each possible partonic process. There is however a better strategy, that helps save disk space and decrease memory footprint, in that it reduces the number of interpolating grids. This is based on the observation that one may find pairs of parton indices $(r,s)$ and $(r^\prime,s^\prime)$ such that: [W\_[rsq]{}\^[()]{}]{}=[W\_[r\^s\^q\^]{}\^[()]{}]{} (r,s)(r\^,s\^) q,q\^. \[WeqW\] This suggests to rewrite eq. (\[fact4\]) in the following way: d[\^[[(NLO]{},]{}]{}=\_[l]{}[[F]{}\^[(l)]{}]{}([x\_1\^[()]{}]{},[x\_2\^[()]{}]{},[\_[F]{}\^[()]{}]{}) \_[q]{}[W\_[lq]{}\^[()]{}]{}d\_[Bj]{}d\_[n+1]{}, \[fact5\] where [[F]{}\^[(l)]{}]{}([x\_1\^[()]{}]{},[x\_2\^[()]{}]{},[\_[F]{}\^[()]{}]{})=\_[rs]{}T\_[rs]{}\^[(l)]{}[f\_r]{}([x\_1\^[()]{}]{},[\_[F]{}\^[()]{}]{})[f\_s]{}([x\_2\^[()]{}]{},[\_[F]{}\^[()]{}]{}), \[lumldef\] and the values of $T_{rs}^{(l)}$ are either zero or one; we have implicitly defined ${W_{lq}^{(\alpha)}}={W_{rsq}^{(\alpha)}}$ for $(r,s,q)$ and $l$ such that $T_{rs}^{(l)}=1$. We point out that, while there may be more than one way to write the r.h.s. of eq. (\[fact5\]) (in other words, the luminosity factors ${{\cal F}^{(l)}}$ may not be uniquely defined), it is always possible to arrive at such a form, thanks to the fact that the r.h.s.’s of eqs. (\[fact4\]) and (\[fact5\]) are strictly identical. In fact, eq. (\[fact5\]) is what is used internally by [[MadGraph5\_aMC@NLO]{}]{}; one of the tasks of [[aMCfast]{}]{} is that of gathering this piece of information, and of using it to construct the luminosity factors ${{\cal F}^{(l)}}$ to be employed at a later stage. In eq. (\[fact5\]) one factors out identical matrix elements; since the computation of these is the most time-consuming operation, this procedure helps increase the overall efficiency, which is larger the larger the number of terms in each luminosity factor ${{\cal F}^{(l)}}$. The fact that, in general, the number of terms in the sum over $l$ in eq. (\[fact5\]) is smaller than that in the sums over $(r,s)$ in eq. (\[fact4\]) is ultimately what allows one to reduce the number of interpolating grids. The counting of the terms in such sums is easy when $(r,s)$ are sufficient to determine uniquely the partonic process. Denoting by $n_l$ the largest value assumed by the luminosity index $l$ for a given process: 1ln\_l, and by ${N_{{\scriptscriptstyle}F}}$ the number of active light flavours, one has: 1n\_l(2[N\_[F]{}]{}+1)\^2, with the two limiting cases being either that where all of the allowed PDF combinations are associated with the same partonic matrix element, or that where each PDF combination corresponds to a different partonic matrix element. In appendix \[sec:lum\] we shall give the explicit form of eq. (\[lumldef\]) for all of the processes studied.
By putting everything together, we finally arrive at the representation of the $h^{th}$ bin value and of its corresponding grids as they are constructed by the [[MadGraph5\_aMC@NLO]{}]{} – [[aMCfast]{}]{} – [[APPLgrid]{}]{} chain. We denote by $\delta_y$ and $s_y$ the grid spacing and interpolating order relevant to the variables $y_1$ and $y_2$; the analogous quantities relevant to the variable $\tau$ are denoted by $\delta_\tau$ and $s_\tau$ respectively. We have: \_[O,0]{}\^[(h)]{}&=&\_[j\_1,j\_2,j\_3]{}\_[l=1]{}\^[n\_l]{} [\^[(l)]{}]{}(j\_1\_y,j\_2\_y,j\_3\_) [\_[S]{}]{}\^[2b+2]{}(j\_3\_)G\_[j\_1j\_2j\_3]{}\^[(h,0,l)]{}, \[appNLOz\]\
\_[O,F]{}\^[(h)]{}&=&\_[j\_1,j\_2,j\_3]{}\_[l=1]{}\^[n\_l]{} [\^[(l)]{}]{}(j\_1\_y,j\_2\_y,j\_3\_) [\_[S]{}]{}\^[2b+2]{}(j\_3\_) [\_[F]{}]{}\^2G\_[j\_1j\_2j\_3]{}\^[(h,F,l)]{}, \[appNLOF\]\
\_[O,R]{}\^[(h)]{}&=&\_[j\_1,j\_2,j\_3]{}\_[l=1]{}\^[n\_l]{} [\^[(l)]{}]{}(j\_1\_y,j\_2\_y,j\_3\_) [\_[S]{}]{}\^[2b+2]{}(j\_3\_) [\_[R]{}]{}\^2G\_[j\_1j\_2j\_3]{}\^[(h,R,l)]{}, \[appNLOR\]\
\_[O,B]{}\^[(h)]{}&=&\_[j\_1,j\_2,j\_3]{}\_[l=1]{}\^[n\_l]{} [\^[(l)]{}]{}(j\_1\_y,j\_2\_y,j\_3\_) [\_[S]{}]{}\^[2b]{}(j\_3\_) G\_[j\_1j\_2j\_3]{}\^[(h,B,l)]{}, \[appNLOB\] with the grids: G\_[j\_1j\_2j\_3]{}\^[(h,,l)]{}&=& \_[k=1]{}\^N\_\_q [\_k\^[(h,)]{}]{}[\^[()]{}\_[,lq;k]{}]{} \[3dimgrid\]\
&&I\_[j\_1-p\_[\_y]{}([y\_[1;k]{}\^[()]{}]{})]{}\^[(s\_y)]{} (-p\_[\_y]{}([y\_[1;k]{}\^[()]{}]{})) (p\_[\_y]{}([y\_[1;k]{}\^[()]{}]{})j\_1s\_y+p\_[\_y]{}([y\_[1;k]{}\^[()]{}]{}))\
&&I\_[j\_2-p\_[\_y]{}([y\_[2;k]{}\^[()]{}]{})]{}\^[(s\_y)]{} (-p\_[\_y]{}([y\_[2;k]{}\^[()]{}]{})) (p\_[\_y]{}([y\_[2;k]{}\^[()]{}]{})j\_2s\_y+p\_[\_y]{}([y\_[2;k]{}\^[()]{}]{}))\
&&I\_[j\_3-p\_[\_]{}([\_k\^[()]{}]{})]{}\^[(s\_)]{} (-p\_[\_]{}([\_k\^[()]{}]{})) (p\_[\_]{}([\_k\^[()]{}]{})j\_3s\_+p\_[\_]{} ([\_k\^[()]{}]{})). In eq. (\[3dimgrid\]) we have inserted the partonic indices $l$ and $q$ in the short-distance quantities $\widehat{W}$, and we have used eqs. (\[ydef\]) and (\[taudef\]) to introduce: [y\_[i;k]{}\^[()]{}]{}&=&Y([x\_[i;k]{}\^[()]{}]{}),\
[\_k\^[()]{}]{}&=&T(([[[K]{}]{}\_[n+1;k]{}\^[()]{}]{})), and we have defined (note the factors ${\xi_{{\scriptscriptstyle}F}}$ and ${\xi_{{\scriptscriptstyle}R}}$): [\^[(l)]{}]{}([y\_1]{},[y\_2]{},)&=& [[F]{}\^[(l)]{}]{}(Y\^[-1]{}([y\_1]{}),Y\^[-1]{}([y\_2]{}),[\_[F]{}]{} T\^[-1]{}()),\
[\_[S]{}]{}()&=&[g\_[S]{}]{}([\_[R]{}]{} T\^[-1]{}()).
Equations (\[appNLOz\])–(\[3dimgrid\]) are the main results of this paper. In the initialisation phase, [[aMCfast]{}]{} provides [[APPLgrid]{}]{} with the total number of grids needed (equal to the sum over all observables of the number of bins relevant to each observable, times four), the grid spacings $\delta_y$ and $\delta_\tau$, the interpolation orders $s_y$ and $s_\tau$, and the interpolation ranges in $y_i$ and $\tau$. These information are under the user’s control; in particular, one must make sure that the latter ranges are sufficiently wide for the process under consideration, and one will want to be careful in the case of a dynamical scale choice for $\mu$. Then, during the course of the run and event-by-event, [[aMCfast]{}]{} gets ${{\Theta}_k^{(h,\alpha)}}$, ${\widehat{W}^{(\alpha)}_{\beta,lq;k}}$, ${y_{i;k}^{(\alpha)}}$, and ${\tau_k^{(\alpha)}}$ from [[MadGraph5\_aMC@NLO]{}]{} and feeds[^4] them to [[APPLgrid]{}]{}, whose grid-filling internal routines iteratively construct $G_{j_1j_2j_3}^{(h,\beta,l)}$ as defined in eq. (\[3dimgrid\]).
We conclude this section with two remarks. Firstly, the sums over $l$ and $q$ in the formulae above achieve the sum over subprocesses which is necessary in order to obtain the hadronic cross section. In practice, in [[MadGraph5\_aMC@NLO]{}]{} the number of contributions which are integrated separately and eventually summed to give the physical cross section is often larger than the number of subprocesses, owing to the FKS dynamic partition and to the use of multi-channel techniques (see ref. [@Alwall:2014hca] for more details). When interfacing [[MadGraph5\_aMC@NLO]{}]{} with [[APPLgrid]{}]{}, each of these contributions generates temporary grids, which are then combined by [[aMCfast]{}]{} at the end of the run to give the actual grids of eq. (\[3dimgrid\]) that are to be used in fast computations (i.e., in eqs. (\[appNLOz\])–(\[appNLOB\])). Secondly, thanks to the fact that the information on the cross section is (also) given in terms of the scale- and PDF-independent coefficients $\widehat{W}$, [[MadGraph5\_aMC@NLO]{}]{} is capable of computing a-posteriori PDF and scale uncertainties independently of its interface to [[APPLgrid]{}]{}, by means of a reweighting procedure (see ref. [@Frederix:2011ss]). While this feature renders the recomputation of the cross section with alternative PDFs and/or scales much faster than the original calculation, it is still not fast enough to be employed in PDF fits, for which the only viable solution is that of using the interpolating grids discussed here. The reason is related to the sum over events ($1\le k\le N$ and $\alpha=E,S,C,SC$ in the previous formulae): while such sum is performed only once in the case of interpolating grids (when the grid is constructed: see eq. (\[3dimgrid\])), it must be carried out for each new choice of PDFs and scales in the context of reweighting. This difference is crucial, because $N$ must be a large number (especially so in fNLO computations), in order to obtain a good statistical precision[^5].
Scale choices\[sec:scale\]
--------------------------
As one can see from eq. (\[Wadef\]), the grids $G_{j_1j_2j_3}^{(h,0,l)}$ and $G_{j_1j_2j_3}^{(h,B,l)}$ are all that is needed when one is interested in computing the cross section that corresponds to setting the factorisation and renormalisation scales equal to their reference value $\mu$, i.e. with ${\xi_{{\scriptscriptstyle}F}}={\xi_{{\scriptscriptstyle}R}}=1$ (see eq. (\[FeqR\])). When ${\xi_{{\scriptscriptstyle}F}}\ne 1$ and/or ${\xi_{{\scriptscriptstyle}R}}\ne 1$, then the grids $G_{j_1j_2j_3}^{(h,F,l)}$ and/or $G_{j_1j_2j_3}^{(h,R,l)}$, respectively, are also necessary (see eqs. (\[appNLOF\]) and (\[appNLOR\])). We point out that these two grids are not constructed when [[APPLgrid]{}]{} is interfaced to codes other than [[MadGraph5\_aMC@NLO]{}]{}. In that case, [[APPLgrid]{}]{} is still capable of computing the cross section for non-central scales, since the form of the latter can be determined by using renormalisation group equations (RGEs). By doing so, one arrives at an expression (e.g., eq. (14) of ref. [@Carli:2010rw]) which is in one-to-one correspondence with the results derived here (in order to see this, one must use the explicit expressions of the $\widehat{W}$ coefficients given in ref. [@Frederix:2011ss]); this is not surprising, since both are ultimately a direct consequence of RGE invariance.
Although fully equivalent from a physics viewpoint, the two-grid and four-grid approaches do not give pointwise-identical results (in other words, their outcomes are strictly equal only in the limit of infinite statistics). The main advantage of using only two grids is that of a smaller memory footprint. Conversely, the two-grid approach has two drawbacks that we consider significant, and are the reason why [[aMCfast]{}]{} works with four grids. Firstly, the ${\xi_{{\scriptscriptstyle}F}}$-dependent term features a convolution (i.e., a one-dimensional integral) of a one-loop Altarelli-Parisi kernel [@Altarelli:1977zs] with a PDF, $P^{(0)}\otimes f$. This convolution is effectively performed when summing over events in the four-grid approach, in eq. (\[appNLOF\]). The absence of $G_{j_1j_2j_3}^{(h,F,l)}$ in the two-grid procedure implies that [[APPLgrid]{}]{} needs to perform this convolution explicitly, and it does so by resorting to an [*external*]{} code, the PDF-evolution package [HOPPET]{} [@Salam:2008qg]. [[APPLgrid]{}]{} does work without [HOPPET]{} being installed, but in this case it is not possible to choose a non-central factorisation scale. Secondly, and related to the previous point: the representation of a cross section in terms of interpolating grids is exact up to terms of higher order in the interpolation-order parameter; this implies, for example, that eq. (\[appNLOF\]) is an identity up to terms that contain the interpolating functions $I^{(s_y+1)}$ and $I^{(s_\tau+1)}$ – as we shall show later, in all practical cases such missing terms are totally negligible. The crucial point here is that this is an identity regardless of the statistics used in the computation of the cross section (i.e., it is independent of $N$, up to fluctuations that may affect the grid construction in the case of very small $N$’s; again, we shall document this later). However, and for the specific case of the ${\xi_{{\scriptscriptstyle}F}}$-dependent eq. (\[appNLOF\]), this is true only because the convolution integral $P^{(0)}\otimes f$ implicit in there is carried out by the very procedure (i.e., the same number of events, the same seeds) that fills the interpolating grid. Any other way of computing $P^{(0)}\otimes f$, and in particular one that uses an external program such as [HOPPET]{}, implies that this property does not hold. Therefore, in the two-grid approach the difference between a cross section computed with non-central scales, and its representation in term of grids, is much larger than formally guaranteed by the chosen interpolation orders; this difference tends to zero only in the limit of infinite statistics $N\to\infty$. We did explicitly verify that this is indeed the case.
In summary, we believe that working with four grids in [[APPLgrid]{}]{}gives a couple of clear advantages over the two-grid procedure: one does not need to install any external convolution package, and the cross section and its grid representation are identical for any scale choices, regardless of the statistics used. The latter feature is beneficial also in view of the fact that results for central and non-central scale choices in [[MadGraph5\_aMC@NLO]{}]{} are fully correlated: hence, the ratios of predictions obtained with different scales are more stable than each of them individually. We point out that both of these advantages are relevant to the factorisation scale dependence. As far as the renormalisation scale is concerned, the situation is simpler, since no convolution integral is involved. Therefore, the optimal approach[^6] would be that of using three grids, and construct $G_{j_1j_2j_3}^{(h,R,l)}$ (relevant to the ${\xi_{{\scriptscriptstyle}R}}$ dependence) on the fly, using $G_{j_1j_2j_3}^{(h,B,l)}$. However, the gain of doing so w.r.t. our four-grid implementation is extremely marginal, and thus we did not consider it when constructing [[aMCfast]{}]{}.
Results and validation\[sec:pheno\]
===================================
As an illustration of the flexibility of [[aMCfast]{}]{}, in this section we present predictions for a variety of LHC processes, which either are currently or might soon become relevant for PDF determinations. Predictions are given for a 14 TeV LHC, using as inputs the NNPDF2.3 NLO PDF set [@Ball:2012cx] (associated with ${\alpha_{{\scriptscriptstyle}S}}(M_Z)=0.1180$) in the case of five or four light flavours, and the NNPDF2.1 NLO PDF set [@Ball:2011mu] (associated with ${\alpha_{{\scriptscriptstyle}S}}(M_Z)=0.1190$) in the case of three light flavours. We have made this choice of input parameters in order to be definite: the pattern of our findings would be unchanged had we used e.g. different PDF sets [@Ball:2012wy; @Martin:2009iq; @Gao:2013xoa]. For the same reason, we do not consider PDF uncertainties in our illustrative study, and thus we only employ the central PDF member of the NNPDF2.3 and NNPDF2.1 sets. All such sets are taken from the [LHAPDF5]{} [@Bourilkov:2006cj] library.
General strategy
----------------
The main idea is the following: given a process, an observable $O$, and a binning for the differential distribution in $O$, we compute the value of its $h^{th}$ bin (for all bins) $\sigma_O^{(h)}$, in two different ways: directly, by means of [[MadGraph5\_aMC@NLO]{}]{}, and a posteriori, using the grids constructed with the [[MadGraph5\_aMC@NLO]{}]{} – [[aMCfast]{}]{} – [[APPLgrid]{}]{} chain. For the latter chain to be validated, these two results must be identical up to numerical inaccuracies; we shall call the former the [*reference*]{} result, and the latter the [*reconstructed*]{} result. In other words, we regard the l.h.s. of eq. (\[histoNLO2\]) as computed directly with [[MadGraph5\_aMC@NLO]{}]{}, and its r.h.s. as obtained from the interpolating grids, eqs. (\[appNLOz\])–(\[3dimgrid\]), so as to establish numerically the accuracy of the equality in eq. (\[histoNLO2\]). As we have already stressed, we expect that inaccuracies are solely due the interpolating procedure, and not to a (possible) lack of statistics in the simulations; importantly, as was discussed in sect. \[sec:scale\], in our approach this property must hold for any non-central scale choices as well. In order to document this, all of our results will be presented for three different scale choices[^7]: [\_[F]{}]{}=,&&[\_[R]{}]{}=, \[sc1\]\
[\_[F]{}]{}=2,&&[\_[R]{}]{}=/2, \[sc2\]\
[\_[F]{}]{}=/2,&&[\_[R]{}]{}=2, \[sc3\] and as obtained from two different runs: one with “low” statistics, and one with “high” statistics. For each choice of scales and type of run, we shall plot the reference result, and the ratio of the reconstructed over the reference result. From that ratio, we shall see that the typical numerical inaccuracies due to the use of the interpolating grids are in the lower $10^{-4}$, with occasional larger differences in the case of the low-statistics runs, due to a still-insufficient amount of information stored in the grids[^8]. The crucial point is the following: even with very limited statistics, the reconstruction of the reference results is extremely good, and this for distributions whose quality is largely insufficient for any phenomenological application. This proves that, for all practical purposes, the level of agreement between reconstructed and reference results is independent of the statistics employed.
In keeping with the strategy at the core of all applications of [[MadGraph5\_aMC@NLO]{}]{}, the only operations to be performed by the user when running with [[aMCfast]{}]{} are related to the process-specific analysis. In particular, in view of the fast interface to fNLO results, the user’s analysis (to be stored, as for any other kind of fNLO user utility, into the directory [MYPROC/FixedOrderAnalysis]{} – see sect. 3.2.2 of ref. [@Alwall:2014hca] for more details) will need to contain, for each observable that he/she wants to consider, the instructions for filling the histogram that will contain the reference result (such instructions are identical to those of an fNLO analysis used when no interface with [[aMCfast]{}]{} is considered), and those for setting up the associated interpolating grids. Several templates of analyses that can be used with [[aMCfast]{}]{} will be provided in [MYPROC/FixedOrderAnalysis]{} (starting from the next version of [[MadGraph5\_aMC@NLO]{}]{}), and the user should be able to easily adapt them to suit the needs of any given calculation. For the whole procedure to work, one will need to install, on top of [[MadGraph5\_aMC@NLO]{}]{} and of the bridge code [[aMCfast]{}]{}, also [[APPLgrid]{}]{} version 1.4.56 or higher.
Parameter value Parameter value
----------------------- --------------------------- --------------------------- ---------------
$\kappa$ 5 $\Lambda$ 0.250 GeV
$(x_{\min},x_{\max})$ $(2{\!\cdot\!}10^{-7},1)$ $(\mu_{\min},\mu_{\max})$ (10,3162) GeV
$N_y$ 50 $N_\tau$ 30
$s_y$ 3 $s_\tau$ 3
: \[tab:param\] Grid parameters used in the [[MadGraph5\_aMC@NLO]{}]{} – [[aMCfast]{}]{} – [[APPLgrid]{}]{} validation runs presented here. In the case of $t{\bar{t}}$ production (sect. \[sec:tt\]), the values of $\mu_{\max}$ and $N_\tau$ have been set equal to $10^4$ GeV and $50$, respectively.
All of the calculations reported below are carried out either with the default physics model of fNLO computations in [[MadGraph5\_aMC@NLO]{}]{}, which corresponds to the SM with ${N_{{\scriptscriptstyle}F}}=4$ light flavours, or with its ${N_{{\scriptscriptstyle}F}}=5$ or ${N_{{\scriptscriptstyle}F}}=3$ variants; the quarks which are not light (the top and, depending on the value of ${N_{{\scriptscriptstyle}F}}$, the bottom and the charm) are given a non-zero mass. The parameters used to initialise the interpolating grids are given in table \[tab:param\]: from the first and second lines there, and by using eqs. (\[ydef\]) and (\[taudef\]), one immediately obtains the actual interpolation ranges in the variables $y_i$ and $\tau$, and the corresponding grid spacings by taking the ratio of those over $(N_y-1)$ and $(N_\tau-1)$, respectively. In the low- and high-statistics runs we employ a number of phase-space points per integration channel of the order of $10^3$ and $10^6$ respectively. For each process we shall consider a couple of observables to be definite, but we emphasise again that the number and the nature of the observables that one can compute are the user’s choice. The kinematical cuts will in general be similar to those used by ATLAS or CMS in some analogous experimental analyses; however, a comparison with experimental data is beyond the scope of this work. In all cases, we shall set $\mu=H_{{\scriptscriptstyle}T}/2$, with $H_{{\scriptscriptstyle}T}$ the scalar sum of the transverse energies of all the final-state particles. We point out that the choice of scales may have some implications on the settings of the corresponding interpolation parameters. For example, as one can see from table \[tab:param\], in the case of $t{\bar{t}}$ production the values of $\mu_{\max}$ and $N_\tau$ have been increased w.r.t. those used for all the other processes, and this in order to attain a comparable numerical precision. On the other hand, we have verified that, by choosing scales which do not depend on the event kinematics, and are of the order of the hardness of the given process at threshold, setting $N_\tau=10$ is amply sufficient. To put things in perspective, however, we emphasise that even by using $N_\tau=30$ in $t{\bar{t}}$ production with $\mu=H_{{\scriptscriptstyle}T}/2$, numerical inaccuracies are at most $5{\!\cdot\!}10^{-4}$ and generally smaller than that, i.e. better than any PDF-fit application will ever need.
We start by studying top-quark pair production, which provides information on the gluon PDF in the large-$x$ region complementary to those provided by jet-production data. We shall then consider isolated-photon production in association with jets, a process that allows a variety of tests of perturbative QCD and that is also directly sensitive to the gluon PDF. We shall next discuss the production of a lepton pair in association with one extra hard jet. We shall then turn to considering $Zb\bar{b}$ production, a case-study of how the complete automation of fast NLO QCD calculations allows one to include in global PDF fits arbitrarily complicated processes at no significant extra cost. Finally, we shall discuss in some details another important process at the LHC for PDF extraction, namely the production of a $W$ boson in association with charm quarks, which is directly sensitive to the poorly known strange-quark PDF, and for which data from ATLAS and CMS have recently become available.
Top-quark pair production\[sec:tt\]
-----------------------------------
Top-quark pair production provides one with unique information on the poorly known large-$x$ gluon PDF [@Czakon:2013tha; @Alekhin:2013nda; @Guzzi:2014wia]. Unlike the case of the Tevatron, where $t{\bar{t}}$ production is mostly driven by $q\bar{q}$ scattering, at the LHC the $gg$-initiated process is dominant, contributing to more that 85% of the total inclusive cross section. The recent calculation of the total cross section to NNLO [@Czakon:2013goa], as well as the upcoming NNLO results for differential distributions, make top-quark pair production an essential ingredient for present and future global PDF analyses. A wide range of data from the LHC for this process are available, both for inclusive cross-sections [@Collaboration:2010ey; @Chatrchyan:2013faa; @Aad:2012mza] and for differential distributions [@Chatrchyan:2012saa; @Aad:2012hg], and much more will be measured in the near future.
As usual, the production of stable $t{\bar{t}}$ pairs in [[MadGraph5\_aMC@NLO]{}]{} is simulated by starting with the generation of the process. Since in the default model used by [[MadGraph5\_aMC@NLO]{}]{} there are four light flavours, one needs first to import a proper massless-bottom model. This is done by means of the following commands:
[[MG5\_aMC>]{}]{} [ import model loop\_sm-no\_b\_mass]{}
[[MG5\_aMC>]{}]{} [ define p = g u d s c b u\~ d\~ s\~ c\~ b\~]{}
Here, the second command instructs [[MadGraph5\_aMC@NLO]{}]{} that the proton contains five light flavours, and does include the bottom quark; by doing so, partonic subprocesses will be generated that feature bottom quarks in the initial state. After this, one may proceed with the generation command;
[[MG5\_aMC>]{}]{} [ generate p p > t t\~ \[QCD\]]{}
followed by the standard [output]{} and [launch]{} commands (see sect. 3 of ref. [@Alwall:2014hca] for more details). There are $n_l=7$ contributing parton luminosities. For each of them, we report in table \[ttlist\]
----- ---------- -------------------- -------------------- -------------------- --------------------- --------------------
$l$ $n_{rs}$ $(r,s)$
1 1 $(g\,, g)$
2 5 $({\bar{b}}\,, g)$ $({\bar{c}}\,, g)$ $({\bar{s}}\,, g)$ $({\bar{u}}\,, g)$ $({\bar{d}}\,, g)$
3 5 $(d\,, g)$ $(u\,, g)$ $(s\,, g)$ $(c \,,g)$ $(b \,,g)$
4 5 $(g\,, {\bar{b}})$ $(g\,, {\bar{c}})$ $(g \,,{\bar{s}})$ $( g\,, {\bar{u}})$ $(g \,,{\bar{d}})$
5 5 $(g \,,d)$ $(g \,,u)$ $(g \,,s)$ $(g \,,c)$ $(g \,,b)$
6 5 $(d \,,{\bar{d}})$ $(u \,,{\bar{u}})$ $(s \,,{\bar{s}})$ $(c \,,{\bar{c}})$ $(b \,,{\bar{b}})$
7 5 $({\bar{b}}\,,b)$ $({\bar{c}}\,,c)$ $({\bar{s}}\,,s )$ $({\bar{u}}\,,u)$ $({\bar{d}}\,,d)$
----- ---------- -------------------- -------------------- -------------------- --------------------- --------------------
: \[ttlist\] Values of $l$ and $(r,s)$ for which $T_{rs}^{(l)}=1$, as generated by [[aMCfast]{}]{} in $t{\bar{t}}$ production. For a given $l$, the number $n_{rs}$ of non-null terms in the sum on the r.h.s. of eq. (\[lumldef\]) is also given.
the corresponding $(r,s)$ contributions, which are the terms on the r.h.s. of eq. (\[lumldef\]) with $T_{rs}^{(l)}=1$; the number of such terms, denoted by $n_{rs}$, is also reported in this table. As was already mentioned in sect. \[sec:int\], the assignments of table \[ttlist\] are automatically determined by [[aMCfast]{}]{}using the information provided by [[MadGraph5\_aMC@NLO]{}]{}. We point out that each of the lines in table \[ttlist\] corresponds to the set of four grids of eqs. (\[appNLOz\])–(\[appNLOB\]). It is possible to use them separately, e.g. in order to determine the relative contribution of each parton luminosity to the different bins of the kinematical distributions studied, although this feature has not been used in this paper.
The results of our validation are presented in figs. \[fig:ytt\_ttbar\] and \[fig:mtt\_ttbar\], where we show the rapidity of the top quark ($y(t)$), and the $t{\bar{t}}$ invariant mass ($M(t{\bar{t}})$), respectively. Each figure consists of two panels, with that of the left (right) obtained with a low-statistics (high-statistics) run. Each panel has a main frame, where the solid histogram is the [[MadGraph5\_aMC@NLO]{}]{} reference result obtained with the central scales of eq. (\[sc1\]), and the dashed histograms are the [[MadGraph5\_aMC@NLO]{}]{} reference results relevant to the two non-central scale choices of eqs. (\[sc2\]) and (\[sc3\]) – incidentally, we note that the latter two predictions are obtained by using the reweighting technique presented in ref. [@Frederix:2011ss] (i.e., [[MadGraph5\_aMC@NLO]{}]{} has been run only once). The three lower insets show the accuracy obtained with interpolating grids, since they display as dashed histograms the ratio of the reconstructed over the reference result, for each of the three scale choices. These histograms thus represent the validation of the [[MadGraph5\_aMC@NLO]{}]{} – [[aMCfast]{}]{} – [[APPLgrid]{}]{} chain, with values equal to one equivalent to a fast computation of the cross section with zero interpolation errors. As one can see from the plots, we obtain in all bins and for all scale choices an accuracy of $3\cdot 10^{-4}$ at worst in the case of the high-statistics run. In keeping with the general discussion given at the beginning of sect. \[sec:pheno\], the reconstruction accuracy of the low-statistics results is also quite good, in spite of the fact that the corresponding differential cross sections display extremely large fluctuations, and are unsuitable for phenomenology.
Photon production in association with one jet\[sec:gamma\]
----------------------------------------------------------
The production of direct photons in hadronic collisions is sensitive to the gluon PDF, owing to the dominance of QCD Compton-scattering-like diagrams. Thanks to tight isolation requirements applied to both theory prediction and data, one is able to significantly reduce or even eliminate completely the cross section of non-direct photons, due to poorly-known fragmentation-fuction contributions in the former, and to the contaminating $\pi^0\to\gamma\gamma$ decays in the latter. A recent re-analysis of all available isolated-photon collider data (inclusive in additional QCD radiation) has shown that NLO QCD gives a good description of all the data sets, and in addition it can provide one with some constraints on the gluon PDF in the region of Bjorken-$x$ relevant to the calculation of the gluon-fusion Higgs production cross section [@d'Enterria:2012yj]. Therefore, there is a solid case to include photon data in the next generation of global PDF analyses. Data for isolated-photon production has been released by ATLAS and CMS [@Aad:2013zba; @Chatrchyan:2011ue; @Khachatryan:2010fm]; in particular, the latest ATLAS measurement [@Aad:2013zba], that makes use of the complete 7 TeV run statistics, extends their kinematical coverage in the photon transverse energy up to a value of ${E_{{\scriptscriptstyle}T}}(\gamma)\sim 600$ GeV (see also ref. [@ATL-PHYS-PUB-2013-018] for a study of the sensitivity of this measurement to various sets of PDFs).
On top of inclusive isolated-photon production, that of isolated photons in association with extra jets allows one to extend the sensitivity to the gluon PDF to a wider range of Bjorken-$x$, and also provides one with some information on the quark PDFs. Some measurements of isolated photon plus jets have been reported by ATLAS and CMS [@Aad:2013gaa; @Chatrchyan:2013oda; @Chatrchyan:2013mwa]; while these data are still not precise enough to directly constrain PDFs [@Carminati:2012mm], the increase of the statistics collected in future LHC runs will dramatically change the picture, so that this process will become a useful addition to global PDF fits.
We have thus computed $\gamma+$jet production with [[MadGraph5\_aMC@NLO]{}]{}. The computation is performed with five light flavours, so the generation command must be preceded by the same [import]{} and [define]{} commands which have been used in the case of $t{\bar{t}}$ production (see sect. \[sec:tt\]). On top of those, we must also use here:
[[MG5\_aMC>]{}]{} [ define j = g u d s c b u\~ d\~ s\~ c\~ b\~]{}
since a light jet must have the same parton contents as the proton. The generation command is then:
[[MG5\_aMC>]{}]{} [ generate p p > a j \[QCD\]]{}
We have imposed the photon to be hard and central, ${p_{{\scriptscriptstyle}T}}(\gamma)\ge 80$ GeV and $\abs{\eta(\gamma)}\le 3$. We use the Frixione photon-isolation prescription [@Frixione:1998jh], since in this way one sets identically equal to zero the contribution of the fragmentation-function component in an infrared-safe manner.
The parameters of the isolation (which in [[MadGraph5\_aMC@NLO]{}]{} can be easily modified at the run-card level) are set as follows: $\epsilon_{\gamma}=1.0$ and $n=1$ (see eq. (3.4) of ref. [@Frixione:1998jh]; note that, in an hadronic environment, the energies and angles of that equation have to be formally replaced by transverse energies and distances in the $(\eta,\varphi)$ plane); we use an isolation cone of radius $R_0=0.4$. Jets are reconstructed with the anti-${k_{{\scriptscriptstyle}T}}$ algorithm [@Cacciari:2008gp], as implemented in [FastJet]{} [@Cacciari:2011ma], with a jet radius of $R=0.4$, and subject to the conditions ${p_{{\scriptscriptstyle}T}}(j)\ge 30$ GeV and $\abs{\eta(j)}\le 4.4$.
There are $n_l=33$ parton luminosities that contribute to the current process; their list is given in table \[pjlist\] (see appendix \[sec:lum\]). That table shows that the definition of the parton luminosities in [[MadGraph5\_aMC@NLO]{}]{} (i.e., eq. (\[fact5\])) is not maximally efficient. For example, the matrix elements associated with $l=30$ and $l=31$ in table \[pjlist\] are not equal, but differ by a trivial overall factor (the charge squared of a up-type vs an down-type quark). The combination of these two contributions could therefore be achieved by simply relaxing the condition that all non-zero $T_{rs}^{(l)}$ elements be equal to one. While this can easily be done in the context of a specific computation, its general implementation in an automated code able to deal with any user-defined process is not straightforward.
The validation plots, presented in figs. \[fig:ptgamma\_gammajet\] and \[fig:etadist\_gammajet\], follow the same pattern as those for $t{\bar{t}}$ production. The representative observables chosen here are the photon transverse momentum (${p_{{\scriptscriptstyle}T}}(\gamma)$, fig. \[fig:ptgamma\_gammajet\]) and the photon pseudorapidity ($\eta(\gamma)$, fig. \[fig:etadist\_gammajet\]). As in sect. \[sec:tt\], the agreement between the reference and reconstructed results is at the level of $10^{-4}$ or better for the high-statistics run, and only slightly worse for the low-statistics one.
Dilepton production in association with one jet\[sec:Zj\]
---------------------------------------------------------
We now turn to studying the hadroproduction of a lepton pair in association with one jet; the leptons originate from an intermediate virtual $Z$ or $\gamma$ boson, and we shall denote the pair with the shorthand notation $Z^\star$ in what follows. Inclusive $Z$-boson production has been used for a long time in PDF fits, since it provides one with a clean handle on the quark flavour separation in the proton, and various quality measurements from the Tevatron [@Aaltonen:2010zza; @Abazov:2007jy] and the LHC [@Aad:2011dm; @Chatrchyan:2011wt] are available. On the other hand, by requiring the presence of an additional extra jet one enters in a different domain, since at the Born level one is dominated by $qg$ channels, which thus allow one to probe directly the gluon PDF. In this context, it is convenient to study the large-${p_{{\scriptscriptstyle}T}}(Z^\star)$ region: in fact, the fraction of gluon-initiated events increases[^9] with ${p_{{\scriptscriptstyle}T}}(Z^\star)$, and furthermore by doing so one is safely away from the small-${p_{{\scriptscriptstyle}T}}(Z^\star)$ region, where fixed-order calculations do not give physically-meaningful predictions. Another advantage associated with this kinematic region is that ratios of cross sections for $W$ over $Z$ production in association with jets can result in clean constraints on quarks and antiquarks PDFs at large-$x$ [@Malik:2013kba], where they are poorly known (and especially so for the antiquarks).
There are various ATLAS and CMS measurements available for $Z$+jets production, which typically focus on the comparison between theory and data as a test of perturbative-QCD predictions which are capable of describing a significant amount of hard radiation [@Aad:2013ysa; @Aad:2011qv; @Chatrchyan:2013tna; @CMS-PAS-SMP-12-017]. However, from the PDF point of view, it is more interesting to perform the analysis as inclusively as possible (bar for the first jet, which one needs in order to be sensitive to the gluon PDF, as mentioned above), and to concentrate on the precise measurement of the $Z^\star$ transverse momentum, possibly in bins of different rapidity $y(Z^\star)$. It is particularly convenient to perform such a measurement with a leptonically-decaying $Z$, which offers a cleaner scenario w.r.t. that of a hadronic decay; there are ongoing analyses in ATLAS and CMS – see e.g. [@CMS-PAS-SMP-13-013].
We have computed dilepton$+$jet production at fNLO with [[MadGraph5\_aMC@NLO]{}]{} in the five-flavour scheme, following the generation command[^10]:
[[MG5\_aMC>]{}]{} [ generate p p > mu+ mu- j \[QCD\]]{}
which has been preceded by the same [import]{} and [define]{} commands relevant to sect. \[sec:gamma\]. We have also imposed the following cuts: [p\_[T]{}]{}(\^)10 [GeV]{}, 2.5, M(\^+\^-)20 [GeV]{}.
The list of parton luminosities relevant to this process is the same as that for $\gamma+$jet production, and is given in table \[pjlist\]. As in the $\gamma+$jet case, jets are reconstructed by using the anti-$k_t$ algorithm with $R = 0.4$, ${p_{{\scriptscriptstyle}T}}(j)\ge 30$ GeV, and $\abs{\eta(j)}\le 4.4$.
For the [[aMCfast]{}]{} validation we have chosen the transverse momentum of the lepton pair and its invariant mass, ${p_{{\scriptscriptstyle}T}}(\mu^+\mu^-)$ and $M(\mu^+\mu^-)$, which we display in figs. \[fig:ptz\_zjet\] and \[fig:etaz\_zjet\] respectively. The latter clearly shows the small-$M$ enhancement due to the presence of an intermediate photon. In the former, the feature around ${p_{{\scriptscriptstyle}T}}(\mu^+\mu^-)\sim 30$ GeV is due to the presence of a sharp kinematic threshold only at the Born level (induced by the ${p_{{\scriptscriptstyle}T}}(j)$ cut), which causes this region to be infrared-sensitive [@Catani:1997xc]. Although this is an issue from the physics point of view for perturbative calculations, it does not pose any problems for the reconstruction through interpolating grids. We observe again an excellent agreement between the reference and the reconstructed results.
$Z$ production in association with a $b\bar{b}$ pair\[sec:Zbb\]
---------------------------------------------------------------
The production of a $Z$ boson in association with a bottom-antibottom quark pair is an interesting example of how the automation of a fast interface to NLO QCD computations achieved by [[aMCfast]{}]{} can be a valuable tool to easily include arbitrarily complicated process into a global PDF fit. A related application could be that of studying to which degree the recent discrepancies between theory and experiment (see e.g. ref. [@Chatrchyan:2014dha]) for the production of $Z$ bosons in association with bottom quarks might be absorbed by a change in PDFs. The hadroproduction of $Wb{\bar{b}}$ and $(Z/\gamma^*)b{\bar{b}}$ in the [[MadGraph5\_aMC@NLO]{}]{} framework has been previously discussed in ref. [@Frederix:2011qg], without however considering any aspects relevant to PDFs.
The generation of the current process in [[MadGraph5\_aMC@NLO]{}]{} is achieved through the command:
[[MG5\_aMC>]{}]{} [ generate p p > z b b\~ \[QCD\]]{}
Note that the default model adopted by [[MadGraph5\_aMC@NLO]{}]{} for NLO computations treats the bottom quarks as massive particles, and one works in a four-flavour scheme: hence, no models are explicitly imported, and no multi-particle labels redefined, at variance with was done previously. This is important in order to be able to perform the computation of $Zb{\bar{b}}$ production without resorting to any kind of jet-reconstruction algorithm, to probe the $b$ quarks down to zero transverse momentum, and to define a fully-inclusive cross section. The list of partonic luminosities that contribute to this process is given in table \[zbblist\].
We validate the results of the present simulation by considering the rapidity distribution of the $Zb{\bar{b}}$ system (shown in fig. \[fig:yzbb\_zbb\]), and the transverse momentum of the $b{\bar{b}}$ pair (shown in fig. \[fig:mzbb\_zbb\]). As in the previous cases, we observe no significant deviations of the reconstructed from the reference predictions.
$W$ production in association with charm quarks\[sec:Wc\]
---------------------------------------------------------
The final representative example of an LHC process of interest for PDF fits that we discuss in this work is the hadroproduction of $W$ bosons in association with charm quarks. This process is directly sensitive to the PDF of the strange quark (see e.g. refs. [@Baur:1993zd; @Stirling:2012vh]), which among the light-quark PDFs is the most poorly determined. In the majority of global PDF fits, the strange PDF is constrained by neutrino data [@MasonPhD; @Mason:2007zz; @Goncharov:2001qe] (in particular by the so-called dimuon process, i.e. charm production in charged-current DIS). The use of neutrino DIS data in PDF fits has several drawbacks, such as the need to model charm fragmentation, and to understand charm-quark mass effects at low scales.
It is therefore quite interesting that $Wc$ hadron-collider data provide one with a clean and robust independent check of the strange PDF determined elsewhere. In addition to that, the differences in the production rates of $W^+$ and $W^-$ bosons in association with charm quarks give information on the strange asymmetry in the proton, that plays an important role in the explanation [@Ball:2009mk] of the NuTeV anomaly [@Goncharov:2001qe]. Data on $W$ production in association with either charm jets or $D$ mesons have been made public by ATLAS [@Aad:2014xca] and CMS [@Chatrchyan:2013uja], based on the 7 TeV 2011 data sets, and analogous measurements from the 8 TeV run are ongoing. In particular, the CMS $Wc$ data has been recently used in a QCD analysis [@Chatrchyan:2013mza], together with HERA data, to show that the strange-quark PDF from collider-only data can be determined with a precision comparable to that of global fits which include neutrino data. Another recent analysis of the compatibility of the LHC $Wc$ data with existing fixed-target DIS and Drell-Yan data has been presented in ref. [@Alekhin:2014sya].
Once again, we now compare a couple of reference distributions from [[MadGraph5\_aMC@NLO]{}]{} to those obtained with the a-posteriori [[aMCfast]{}]{} – [[APPLgrid]{}]{} convolution. For simplicity, we restrict ourselves to considering $W^+\bar{c}$ production; on the other hand, the process we shall actually deal with is that where an electron and a neutrino, rather than a $W$, are present in the final state, so as to make sure that all production spin correlations are correctly taken into account. Incidentally, this also shows that one is not restricted to performing computations in the narrow width approximation; in particular, the complex mass scheme [@Denner:2005fg] is fully supported in [[MadGraph5\_aMC@NLO]{}]{}. Similarly to the case of the $b$ quark in $Zb{\bar{b}}$ production discussed in sect. \[sec:Zbb\], in the present case it is best to treat the charm as a massive particle. Since in the default model used by [[MadGraph5\_aMC@NLO]{}]{} the charm is massless, one needs to import a proper massive-charm model before proceeding with the generation of the process. This is done by means of the following commands:
[[MG5\_aMC>]{}]{} [ import model loop\_sm-c\_mass]{}
[[MG5\_aMC>]{}]{} [ define p = g u d s u\~ d\~ s\~]{}
[[MG5\_aMC>]{}]{} [ generate p p > e+ ve c\~ \[QCD\]]{}
In this list, the second command instructs [[MadGraph5\_aMC@NLO]{}]{} that the proton contains three light flavours, and does not feature the charm quark; by doing so, no partonic subprocesses will be generated that feature a charm in the initial state. The following cuts have been imposed: [p\_[T]{}]{}(e\^+)10 [GeV]{}, 2.5. The list of partonic luminosities that contribute to this process in given in table \[wclist\].
The two representative distributions that we use for validation are the pseudorapidity of the positron, ($\eta(e^+)$, presented in fig. \[fig:yw\_wcharm\]), and its transverse momentum (${p_{{\scriptscriptstyle}T}}(e^+)$ , presented in fig. \[fig:ptw\_wcharm\]). The same conclusions as in all previous cases apply here.
Conclusions and outlook\[sec:outlook\]
======================================
In this paper we have presented [[aMCfast]{}]{}, a tool that serves to achieve the complete automation of fast NLO QCD computations for arbitrary processes, by constructing a bridge between [[MadGraph5\_aMC@NLO]{}]{}, which performs the core automated cross section calculations, and [[APPLgrid]{}]{}, which allows one to represent a given observable in term of interpolating grids that can be used for a fast and a-posteriori evaluation, and one in which the possibility is given of changing the scales and PDFs w.r.t. those used in the original computation. Since fast calculations of NLO cross sections are an absolute necessity in the context of global PDF fits, the combination of the automated programs discussed in this paper covers all the present and, more importantly, future needs relevant to PDF analyses.
We have validated our approach by considering a representative sample of production processes at the LHC (some of which not previously available in the form of fast computations), and by comparing for several observables the bin-by-bin agreement between the original results of [[MadGraph5\_aMC@NLO]{}]{}, and those reconstructed by means of [[aMCfast]{}]{} and [[APPLgrid]{}]{}. The typical accuracy that we find is of the order of some parts in $10^{-4}$, which is in fact far more than sufficient for any kind of phenomenological application. Moreover, our approach allows one to consider scale variations without any loss of precision, and without the need for interfacing [[APPLgrid]{}]{}with an external program to perfom PDF evolution.
A natural outcome of this work is the use of the grids produced using [[aMCfast]{}]{} in actual global PDF analyses, in view of the possible inclusion into the latter of LHC data, and especially of those which have not been considered so far, for lack of either sufficiently precise measurements, or of theoretical NLO calculations suited to the task. Given that the format of the results of the [[aMCfast]{}]{} – [[APPLgrid]{}]{} interface is a trivial extension of that produced by [[APPLgrid]{}]{} interfaced with other codes, [[aMCfast]{}]{} should be straighforward to employ by the global-PDF fitters already experienced with the former code.
The outlook for the near future features two important developments.
Firstly, there is no conceptual difference between what has been done here, and its analogue in the case of a perturbative expansion in either the electroweak coupling constant ${\alpha}$, or simultaneously in ${\alpha_{{\scriptscriptstyle}S}}$ and ${\alpha}$ (mixed-coupling expansion). From the technical viewpoint, the only implication of a mixed-coupling expansion is the necessity of introducing additional interpolating grids, which is trivial. This point is important in view of the fact that [[MadGraph5\_aMC@NLO]{}]{}is expected to be able to handle soon both kind of computations; thus, any extension of the [[MadGraph5\_aMC@NLO]{}]{} capabilities will be made immediately available for fast computations, thanks to straightforward modifications to [[aMCfast]{}]{}. This will pave the way to a systematic inclusion of electroweak NLO corrections into global PDF fits; cases where such an inclusion may prove particularly relevant include high-${p_{{\scriptscriptstyle}T}}$ jet production, the high-mass Drell-Yan cross section, as well as the consistent treatment of photon-initiated processes that are necessary for the determination of the photon PDF [@Ball:2013hta]. Secondly, the scale- and PDF-independent coefficients computed by [[MadGraph5\_aMC@NLO]{}]{}, and used by [[aMCfast]{}]{} to set up the interpolating grids at fixed order, are quite similar to those relevant to the MC@NLO short-distance cross sections. Therefore, the current structure of [[aMCfast]{}]{} will need only a minor upgrade to be able to deal with NLO+PS computations. Apart from its obvious phenomenological spinoffs, and the possibility of obtaining PDF fits that include resummation effects well tuned to experimental data, such a project is also interesting from a more theoretical viewpoint, in that it would help spur a thorough investigation of the effects of the PDFs in initial-state showers; there is some evidence that these effects are small, but more systematic studies should be carried out, in view of the fact that they cannot be taken into account by any method based solely on short-distance cross section information.
The public version of the [[aMCfast]{}]{} code will soon be available at:\
$\phantom{aaa}$[http://amcfast.hepforge.org/]{}\
Meanwhile, potential users are kindly requested to contact the authors.
Acknowledgments {#acknowledgments .unnumbered}
===============
We are grateful to Tancredi Carli and Pavel Starovoitov for discussions on the [[APPLgrid]{}]{} framework. The work of J. R. has been partially supported by an STFC Rutherford Fellowship ST/K005227/1. The work of V. B. has been supported by the ERC grant 291377, “LHCtheory: Theoretical predictions and analyses of LHC physics: advancing the precision frontier”.
Luminosity factors\[sec:lum\]
=============================
In this appendix, we report the luminosity indices $l$ and parton identities $(r,s)$ for which the quantity $T_{rs}^{(l)}$ that enters eq. (\[lumldef\]) is equal to one. We do so for all of the processes studied in sect. \[sec:pheno\], except for $t{\bar{t}}$ production, whose analogous information have already been given in sect. \[sec:pheno\].
----- ---------- ---------------------------- --------------------------- ---------------------------- ----------------------------- ---------------------------- ----------------------------
$l$ $n_{rs}$ $(r,s)$
1 2 $(g \,, u)$ $(g \,,c)$
2 6 $({\bar{b}}\,,u)$ $({\bar{b}}\,, c)$ $({\bar{s}}\,,u)$ $( {\bar{s}}\,, c)$ $({\bar{d}}\,, u)$ $({\bar{d}}\,, c)$
3 6 $(d \,,u)$ $(d \,, c)$ $(s \,,u)$ $( s \,, c)$ $( b \,, u)$ $( b \,, c)$
4 2 $({\bar{c}}\,,c)$ $({\bar{u}}\,, u)$
5 2 $(u \,,u)$ $(c \,, c)$
6 2 $({\bar{c}}\,, u)$ $({\bar{u}}\,, c)$
7 2 $(u \,,c)$ $(c \,, u)$
8 1 $(g \,,g)$
9 3 $(g \,,d)$ $(g \,, s)$ $(g \,, b)$
10 3 $( {\bar{b}}\,,b)$ $({\bar{s}}\,, s)$ $( {\bar{d}}\,, d)$
11 3 $(d \,,d)$ $(s \,,s)$ $(b \,, b)$
12 6 $({\bar{c}}\,,d)$ $({\bar{c}}\,, s)$ $({\bar{c}}\,,b)$ $({\bar{u}}\,, d)$ $({\bar{u}}\,, s)$ $({\bar{u}}\,, b)$
13 6 $(u \,,d)$ $(u \,,s)$ $(u \,, b)$ $(c \,, d)$ $(c \,,s)$ $(c \,, b)$
14 6 $({\bar{b}}\,, d)$ $({\bar{b}}\,, s)$ $({\bar{s}}\,, d)$ $({\bar{s}}\,, b)$ $({\bar{d}}\,,s)$ $({\bar{d}}\,, b)$
15 6 $(d \,,s)$ $(d \,, b)$ $(s \,,d)$ $(s \,, b)$ $(b \,, d)$ $(b \,,s)$
16 2 $(g \,, {\bar{c}})$ $(g \,,{\bar{u}})$
17 6 $({\bar{b}}\,, {\bar{c}})$ $({\bar{b}}\,,{\bar{u}})$ $({\bar{s}}\,,{\bar{c}})$ $({\bar{s}}\,,{\bar{u}})$ $({\bar{d}}\,, {\bar{c}})$ $({\bar{d}}\,, {\bar{u}})$
18 6 $(d \,,{\bar{c}})$ $(d \,,{\bar{u}})$ $(s \,, {\bar{c}})$ $(s \,,{\bar{u}})$ $(b \,,{\bar{c}})$ $(b \,,{\bar{u}})$
19 2 $({\bar{c}}\,,{\bar{c}})$ $({\bar{u}}\,,{\bar{u}})$
20 2 $(u \,,{\bar{u}})$ $(c \,,{\bar{c}})$
21 2 $({\bar{c}}\,,{\bar{u}})$ $({\bar{u}}\,,{\bar{c}})$
22 2 $(u \,, {\bar{c}})$ $(c \,,{\bar{u}})$
23 3 $(g \,,{\bar{b}})$ $(g \,, {\bar{s}})$ $(g \,,{\bar{d}})$
24 3 $({\bar{b}}\,,{\bar{b}})$ $({\bar{s}}\,,{\bar{s}})$ $({\bar{d}}\,,{\bar{d}})$
25 3 $(d \,,{\bar{d}})$ $(s \,,{\bar{s}})$ $(b \,,{\bar{b}})$
26 6 $({\bar{c}}\,, {\bar{b}})$ $({\bar{c}}\,,{\bar{s}})$ $({\bar{c}}\,,{\bar{d}})$ $({\bar{u}}\,, {\bar{b}})$ $({\bar{u}}\,,{\bar{s}})$ $({\bar{u}}\,, {\bar{d}})$
27 6 $(u \,,{\bar{b}})$ $(u \,,{\bar{s}})$ $(u \,,{\bar{d}})$ $(c \,,{\bar{b}})$ $(c \,,{\bar{s}})$ $(c \,,{\bar{d}})$
28 6 $({\bar{b}}\,,{\bar{s}})$ $({\bar{b}}\,,{\bar{d}})$ $({\bar{s}}\,, {\bar{b}})$ $( {\bar{s}}\,, {\bar{d}})$ $({\bar{d}}\,,{\bar{b}})$ $({\bar{d}}\,,{\bar{s}})$
29 6 $(d \,,{\bar{b}})$ $(d \,,{\bar{s}})$ $(s \,,{\bar{b}})$ $(s \,,{\bar{d}})$ $(b \,, {\bar{s}})$ $(b \,,{\bar{d}})$
30 2 $(u \,,g)$ $(c \,, g)$
31 3 $(d \,,g)$ $(s \,, g)$ $(b \,,g)$
32 2 $({\bar{c}}\,,g)$ $({\bar{u}}\,,g)$
33 3 $({\bar{b}}\,,g)$ $({\bar{s}}\,, g)$ $({\bar{d}}\,,g)$
----- ---------- ---------------------------- --------------------------- ---------------------------- ----------------------------- ---------------------------- ----------------------------
: \[pjlist\] As in table \[ttlist\], for $\gamma+$jet and dilepton$+$jet production.
----- ---------- -------------------- --------------------
$l$ $n_{rs}$ $(r,s)$
1 1 $(g\,, g)$
2 2 $({\bar{s}}\,,g)$ $({\bar{d}}\,,g)$
3 2 $(d \,,g )$ $(s \,,g)$
4 2 $({\bar{c}}\,,g)$ $({\bar{u}}\,,g)$
5 2 $(u \,,g )$ $(c \,,g)$
6 2 $(g \,,{\bar{s}})$ $(g \,,{\bar{d}})$
7 2 $(g \,,d )$ $(g \,,s)$
8 2 $(g \,,{\bar{c}})$ $(g \,,{\bar{u}})$
9 2 $(g \,,u)$ $(g \,,c)$
10 2 $(u \,,{\bar{u}})$ $(c \,,{\bar{c}})$
11 2 $(d \,,{\bar{d}})$ $(s \,,{\bar{s}})$
12 2 $({\bar{c}}\,,c)$ $({\bar{u}}\,,u)$
13 2 $({\bar{s}}\,,s)$ $({\bar{d}}\,,d)$
----- ---------- -------------------- --------------------
: \[zbblist\] As in table \[ttlist\], for $Zb{\bar{b}}$ production.
----- ---------- --------------------------- ---------------------------
$l$ $n_{rs}$ $(r,s)$
1 1 $(g \,, {\bar{s}})$
2 2 $(u \,,{\bar{s}})$ $(d \,, {\bar{s}})$
3 2 $({\bar{d}}\,,{\bar{s}})$ $({\bar{u}}\,,{\bar{s}})$
4 1 $(s \,,{\bar{s}})$
5 1 $({\bar{s}}\,,{\bar{s}})$
6 1 $(g \,,g)$
7 1 $({\bar{s}}\,,g)$
8 2 $({\bar{s}}\,,u)$ $({\bar{s}}\,,d)$
9 2 $({\bar{s}}\,,{\bar{d}})$ $({\bar{s}}\,,{\bar{u}})$
10 1 $({\bar{s}}\,,s)$
----- ---------- --------------------------- ---------------------------
: \[wclist\] As in table \[ttlist\], for $e^+\nu_e \bar{c}$ production.
[^1]: Recently, the [[FastNLO]{}]{} interface has been generalised to processes other than jet production, for instance to the approximate NNLO calculation of differential distributions in top-pair production of ref. [@Guzzi:2014wia].
[^2]: We remind the reader that the Ellis-Sexton scale $Q$, originally introduced in ref. [@Ellis:1985er], is any scale that may be used in one-loop computations to express the arguments of all the logarithms appearing there as rather than as , where $s_{ij}$ and $s_{kl}$ are two invariants constructed with the four-momenta of the particles that enter the hard process.
[^3]: Although we did not indicate this explicitly in sect. \[sec:obs\], the Ellis-Sexton scale is in general a function of the kinematics, whence the possibility of using eq. (\[Qeqmu\]).
[^4]: [[aMCfast]{}]{}-specific input routines have been added to [[APPLgrid]{}]{}.
[^5]: On the other hand, reweighting is more flexible than grid interpolation, in that it gives one the possibility of recomputing the cross section by adopting a different [*functional*]{} form for the scales w.r.t. that used in the original computation, which is not feasible when using grids. Such a possibility is however not available in the public version of [[MadGraph5\_aMC@NLO]{}]{}. More importantly, reweighting does not need the prior knowledge of the observables which one is going to study.
[^6]: This is only the case if one does not need to deal with processes that feature ${\mu_{{\scriptscriptstyle}R}}$-dependent Yukawa; otherwise, the fourth grid is not trivially related to the Born one.
[^7]: One should be careful not to interpret the envelope resulting from eqs. (\[sc1\])–(\[sc3\]) as representative of the higher-order theoretical uncertainty: these choices are made for the sole purpose of validating [[aMCfast]{}]{} with arbitrary ${\xi_{{\scriptscriptstyle}F}}$ and ${\xi_{{\scriptscriptstyle}R}}$.
[^8]: This inaccuracy includes a contribution from the internal interpolation in [LHAPDF5]{}, which is significantly reduced with [LHAPDF6]{}.
[^9]: This comment applies to an intermediate hardness region; asymptotically, $qg$-initiated contributions are strongly suppressed by the fastly decreasing gluon PDF at large $x$’s.
[^10]: The muons are treated as massless particles.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We present GalMC (Acquaviva et al 2011), our publicly available Markov Chain Monte Carlo algorithm for SED fitting, show the results obtained for a stacked sample of Lyman Alpha Emitting galaxies at z $\sim$ 3, and discuss the dependence of the inferred SED parameters on the assumptions made in modeling the stellar populations. We also introduce SpeedyMC, a version of GalMC based on interpolation of pre-computed template libraries. While the flexibility and number of SED fitting parameters is reduced with respect to GalMC, the average running time decreases by a factor of 20,000, enabling SED fitting of each galaxy in about one second on a 2.2GHz MacBook Pro laptop, and making SpeedyMC the ideal instrument to analyze data from large photometric galaxy surveys.'
author:
- 'Viviana Acquaviva$^1$, Eric Gawiser$^1$ and Lucia Guaita$^2$'
title: 'SED fitting with MCMC: methodology and application to large galaxy surveys'
---
SED fitting with MCMC: a two-step process
=========================================
SED fitting is the process of extracting information on the physical properties of galaxies, such as stellar population age, mass, star formation rate, dust content, metallicity, and redshift, starting from a set of templates that predict how galaxy spectra look like as a function of these properties, which are the SED fitting parameters. This process relies on the simple but powerful idea that since the properties of the models are known, if we can find models that resemble the observations we can infer the properties of the data. A more rigorous way of comparing models with observations – in other words, deciding whether a model resembles the data or not - is the $\chi^2$ statistics. For each set of parameters, we need to compute the prediction of what the observations would be if that model was the true one. This step is conceptually simple but complicated in practice, because of the many astrophysical processes that needs to be modeled. GalMC implements this process through the sequence described in Fig. 1. GalMC is based on Bayesian statistics. Therefore, a second step of the inference process requires to reconstruct the probability distribution of the SED fitting parameters, which are treated as random variables. This is done by exploring the parameter space with a random walk biased so that the frequency of visited locations is proportional to the probability density function. This path through parameter space is the Markov Chain. Once these probabilities are known, one can compute the desired credible intervals for each of the parameters; and because of how visited locations are chosen, integrating the probability distribution functions (PDFs) becomes a simple matter of summing over the points in the chains.
Probability distributions, degeneracies, and the impact of systematics
======================================================================
Detailed results for two stacked samples of Lyman Alpha Emitting galaxies at $z \sim 3$ were presented in [@2011arXiv1101.2215A Acquaviva et al (2011)]. Here we just show that GalMC is able to capture multi-modal probability distributions, such as the double peak in the Age vs Stellar Mass distribution, which is due to the degeneracy between these two parameters. We also want to highlight the impact of the assumptions made in modeling the stellar populations, shown for the case of Stellar Mass in the right panel of Fig. 2. The different curves all refer to models commonly used in the literature: the BC03 ([@BC03]) and CB07 ([@CB07]) stellar population templates, at Solar or variable metallicity, and with or without including nebular emission. The corresponding scatter in the estimate of stellar mass (which does not include the possibility of different initial mass functions, IMFs) is a factor of $\sim$ 2.5, significantly larger than the statistical uncertainty for the same data.
{width="\linewidth"}
SpeedyMC: MCMC for large galaxy catalogs
========================================
MCMC algorithms are much more efficient ways of exploring high-dimensional parameter spaces with respect to algorithms where the probability distribution is sampled at a set of fixed locations on a grid. In fact, the “interesting" region of parameter space (the one where data and models look like each other) often occupies a small fraction of the total volume. While grid-based models need to explore all of it, Markov Chains are able to “recognize" the interesting regions and will spend most of the time visiting (sampling) those locations. Yet, the complicated process described in Fig. 1, which leads to the computation of the $\chi^2$ value corresponding to a set of parameters, usually needs to be repeated tens of thousands of times. The computational bottlenecks in this case are the generation of a stellar population template at the right age, and the convolution with the filter transmission curves. To alleviate the first problem, GalMC uses our modified version of GALAXEV (Bruzual and Charlot 2011), developed in collaboration with the authors, which is $\sim$ 20 times faster than the official release. However, the typical time per iteration is still about 0.4 seconds on a 2.2GHz MacBook Pro laptop (for simplicity, all quoted running times will be referred to this machine), and therefore the typical chain per object takes a few hours to run. This becomes impractical for catalogs comprising thousands of objects. The basic idea of SpeedyMC is to find a different (faster) way to compute the $\chi^2$ corresponding to a certain set of parameters. To achieve this objective, we take the following four steps:
{width="5cm"} {width="4.2cm"} {width="3.56cm"} \[sed\]
1. We compute the spectra on a grid of locations exploring the entire parameter space, saving the final product of the sequence of steps described in Fig 1 ([*after*]{} convolution with the filter transmission curves, so we retain only a handful of numbers corresponding to the flux densities in the observations’ bands);
2. We read the grid into memory;
3. We run MCMC as usual, but to compute the $\chi^2$ at each location we use [*multi-linear interpolation*]{} between the pre-computed spectra;
4. We enjoy the speed up factor of 20,000, which allows us to fit the SED of each galaxy in a few seconds (even assuming to run several chains per object).
A couple of caveats are in order. First, this method doesn’t have the flexibility of GalMC; because it is difficult to perform interpolation in more than three dimensions, not more than four SED fitting parameters can be used (the fourth parameter being stellar mass, which is a normalization and therefore is excluded from the interpolation process). Second, there is an “overhead" cost in computing the grid; for 50 values of age and E(B-V), and 100 values of redshift, running the initial grid takes about 24 hours, and this need to be repeated for a different survey (since the set of utilized filters change), or to use, , a different star formation history or IMF. Still, for large surveys the set of filters is fixed, the number of different modeling options one might want to try is limited, and four parameters are enough to capture the general physical properties of a population of galaxies. Finally, let us observe that the resolution of the initial grid [*does not*]{} correspond to the resolution with which the PDF is sampled, as is the case in grid-based models. MCMC is still free to sample any desired location in parameter space, and the accuracy of the predicted spectrum corresponds to the accuracy of the linear interpolation between the points of the grid. This is illustrated in Fig. 3. The accuracy can be improved by increasing the number of points in the grid. A test conducted on the LAEs at $z = 3.1$ revealed that 50 points in age between 0 and the age of the Universe and 50 values of E(B-V) between 0 and 1 are enough to produce the estimate and credible intervals as the original GalMC, and using 100 values rather than 50 does not produce any appreciable difference.
SpeedyMC is currently being used for the analysis of data from the Cosmic Assembly Near-Infrared Deep Extragalactic Legacy Survey (CANDELS), ([@grogin Grogin et al 2011], [@koekemoer Koekemoer et al 2011], [@Acq2012 Acquaviva et al 2012]). The algorithm is not yet public, but you are welcome to contact the author for discussion on how to implement it, starting from GalMC.
{width="8cm"}
2011, *ApJ*, 737, 47. 2012, [*in preparation*]{}. 2003, *MNRAS*, 344, 1000. Charlot, S. and Bruzual, G., private communication, 2011. Grogin, N. A. et al 2011, *ApJS* 197, 35. Koekemoer, A. M. et al 2011, *ApJS* 197, 36. Lai, K., et al 2008, *ApJ* 674, 70. Lewis, A. M. and Bridle, S. 2002, *PRD*, D66, 103511.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We study equilibrium shapes, stability and possible bifurcation diagrams of fluids in higher dimensions, held together by either surface tension or self-gravity. We consider the equilibrium shape and stability problem of self-gravitating spheroids, establishing the formalism to generalize the MacLaurin sequence to higher dimensions. We show that such simple models, of interest on their own, also provide accurate descriptions of their general relativistic relatives with event horizons. The examples worked out here hint at some model-independent dynamics, and thus at some universality: smooth objects seem always to be well described by both “replicas” (either self-gravity or surface tension). As an example, we exhibit an instability afflicting self-gravitating (Newtonian) fluid cylinders. This instability is the exact analogue, within Newtonian gravity, of the Gregory-Laflamme instability in general relativity. Another example considered is a self-gravitating Newtonian torus made of a homogeneous incompressible fluid. We recover the features of the black ring in general relativity.'
author:
- Vitor Cardoso
- Leonardo Gualtieri
title: Equilibrium configurations of fluids and their stability in higher dimensions
---
Introduction
============
The shape and stability of liquid drops has been the subject of intense study for more than 200 years, with exciting revivals occurring more or less periodically. In the early days of the topic, the hope was that centimeter-sized drops held together by surface tension could be good models for giant liquid masses held together by self-gravitation. Liquid drops, which could be controlled in the laboratory, should model self-gravitating stars where gravity would replace surface tension. A great deal of what is now known about (rotating) liquid drops is due to the work of J. Plateau [@plateau] who, despite suffering from blindness throughout a good part of his scientific life, conducted many rigorous experimental investigations on the subject (we refer the reader to [@plateau] for details). Starting with a non-rotating drop of spherical shape, Plateau and collaborators progressively increased the rate of rotation (angular velocity, assuming the drop rotates approximately as a rigid body). Plateau found that as the rotation rate increased, the drop progressed through a sequence of shapes which evolved from axisymmetric for slow rotation, to ellipsoidal and two-lobed and finally toroidal at very large rotation. These experimental works have been repeated with basically the same outcome, even in gravity-free environments [@rotatingdropexperiments] (see also [@rotatingdropexperiments2]). Theoretical analysis of Plateau’s experiments were started by Poincaré [@poincaresurfacetension], who proved that there are two very different possible solutions to the equilibrium problem: one is axisymmetric like disks or torii and the other is comprised of asymmetric shapes with two lobes. The analysis was completed by Chandrasekhar [@chandra], improved upon by Brown and Scriven [@brownscriven] (see also [@brownscrivenprl; @heine]) and is summarized in Figure \[fig:brown\], taken from [@brownscriven].
![The bifurcation and stability diagram for rotating liquid drops with surface tension (self-gravity is neglected). Shapes are stable along solid branches, unstable along broken ones. The plot depicts dimensionless angular velocity as function of dimensionless angular momentum. The dimensionless angular velocity is $\Omega=\sqrt{\frac{\rho
\hat{\Omega}\hat{R}}{8T}}$, where all hatted quantities are dimensionfull, $\hat{R}$ is the radius of a sphere with the same volume as that of the drop, $\rho$ and $T$ are the density and surface tension of the liquid. Likewise, the dimensionless angular momentum J is defined as $J=\hat{J}/4T \hat{R}^2\left (8T\hat{R}^2
\rho\right )$. Although not shown in the plot, the axisymmetric curve bends back to lower angular velocities after it has passed the four-lobed family neutral point. Note also that the maximum allowed angular momentum for [*stable*]{} shapes is denoted by $J_{II}$ in the diagram. Taken from [@brownscriven].[]{data-label="fig:brown"}](Fig1.eps){width="10" height="8"}
The Figure shows the (dimensionless, as defined by Brown and Scriven, see caption in Fig. \[fig:brown\]) angular velocity as a function of the angular momentum of the drop, and it displays the stability and bifurcation points of the evolution. Let’s focus first on the evolution of axisymmetric figures. Starting from zero angular momentum, where the shape is a spherical surface, one increases the angular momentum. The resulting equilibrium shape evolves from a sphere through oblate shapes to biconcave shapes (see also [@chandra]). There are no (axisymmetric) equilibrium figures with $\Omega>\Omega_{\rm max}\sim0.753$. The axisymmetric family does not terminate at $\Omega=\Omega_{\rm max}$, instead it bends back to lower angular velocity, and there are now [*two*]{} axisymmetric figures of equilibrium. The steepest descending branches are torii [@rotatingdropexperiments2; @gulliver].
Stability calculations show that all shapes of the axisymmetric family are stable against axisymmetric perturbations, however a real drop is also perturbed by non-axisymmetric modes. The stability calculations show that the drop becomes unstable at $\Omega =\Omega_{II}\sim 0.5599$ [@brownscriven], where the axisymmetric family is neutrally stable to a two-lobed perturbation. This point marks the bifurcation of a two-lobed family (therefore a non-axisymmetric family). All axisymmetric shapes with $\Omega>\Omega_{II}$ are unstable against two-lobed perturbations, but there are additional points of neutral stability, which are also bifurcation points: the three- and four-lobed perturbations are shown in the diagram \[fig:brown\] but there are others [@heine]. The first, most important non-axisymmetric surface, the two-lobed family, branches off at $\Omega=\Omega_{II}$, and is initially (near the bifurcation point) stable. In Figure \[fig:twolobes\] we show some typical members of that sequence (see [@phdheine], where this figure was taken from, for further details).
![This plot shows a sequence of members of the two-lobed family of drop shapes for several values of angular velocity. The two-lobed drops seem to approach a limiting surface consisting of two spheres touching each other in one point as $\Omega$ approaches 0. The first shape, with $\Omega=0.5599$, corresponds to the first (two-lobed) bifurcation point in Fig. \[fig:brown\]. For $\Omega=0.2885$ for instance, the figure is peanut shaped, and we can situate the corresponding family in Fig. \[fig:brown\]. Taken from [@phdheine].[]{data-label="fig:twolobes"}](Fig2.eps){width="8" height="9"}
It turns out that the evolution of self-gravitating fluids, i.e., stars, follows a very similar diagram, as shown in Figure \[fig:bifgrav\]. Finding the equilibrium configurations of self-gravitating rotating homogeneous fluids is an extremely important problem, which has met intense activity ever since the first discussions by Newton and others (see [@history] and also the excellent and concise introduction in [@chandrabookellipsoidal]), which was revisited some 50 years ago by Chandrasekhar and co-workers [@chandrabookellipsoidal]. In three spatial dimensions, the most basic figures of equilibrium of rotating bodies are the MacLaurin and Jacobian sequences [^1]. The MacLaurin sequence is a sequence of oblate spheroids along which the eccentricity of meridional sections increases from zero to one. Jacobi showed that for each value of $\Omega^2$, the square of the angular velocity of rotation, there are actually two permissible figures of equilibrium. The other is a tri-axial ellipsoid, belonging to the Jacobi sequence, which branches off from the MacLaurin sequence. This is the first of several points of bifurcation which distinguish the permissible sequences of figures of equilibrium of uniformly rotating homogeneous masses. If one starts from a non-rotating homogeneous fluid ellipsoid in the MacLaurin sequence and increases the angular momentum of the body, one will first come across the Jacobi branch, where the ellipsoid becomes unstable with regard to the first non-axisymmetric perturbation [@chandrabookellipsoidal] (which are the self-gravitating analogues of the shapes in Figure \[fig:twolobes\]). Continuing, still along the MacLaurin sequence and increasing angular momentum, one next comes to the bifurcation point of other non-axisymmetric figures: the “triangle”, the “square” and “ammonite” sequences [@chandrabookellipsoidal; @eriguchi]. One finally encounters the first axisymmetric sequence. The bodies of this sequence pinch together at the centre and eventually form Dyson’s anchor-rings, or torus-like configuration [@dyson; @bardeen; @wong; @eriguchisugimoto; @ansorg]. It is apparent that the bifurcation diagram for self-gravitating fluids, Figure \[fig:bifgrav\] is very similar to the one for non-gravitating fluids with surface tension, shown in Figure \[fig:brown\]. Thus, whenever the treatment of self-gravitation proves too difficult, the surface tension model is a good guide.
![The bifurcation diagram for rotating self-gravitating, incompressible homogeneous stars. Here we show the angular velocity as function of the angular momentum for several possible families. The data and notation are those of Eriguchi and Hachisu [@eriguchi]. The line $A-A'$ is the MacLaurin sequence, $B-B'$ the Jacobi sequence, $C-C'$ triangle sequence, $D-D'$ the square sequence, $E-E'$ ammonite sequence, $F-F'$ the one-ring sequence, $G-G'$ the two-ring sequence, $H-H'$ the pear-shaped sequence, $I-I'$ the dumb-bell sequence, $J-J'$ the Darwin sequence and finally $K-K'$ is the two point-mass sequence. Note the similarities with Fig. \[fig:brown\]. See [@eriguchi] for further details. []{data-label="fig:bifgrav"}](Fig3.eps){width="8" height="6"}
A realistic star is neither homogeneous nor incompressible and it certainly isn’t held together by surface tension. However, it seems logical to build our understanding using simpler models. We have gone a long way with simpler related models: another well-known example, where the fluid drop model proved extremely useful, is Bohr and Wheeler’s [@bohrwheeler] proposal to describe nuclear fission as the rupture of a rotating charged liquid drop, where now the surface tension plays the role of nuclear forces. A review with a detailed and unified discussion of charged nuclei, astronomical bodies and fluid drops can be found in [@swiatecki] (a concise and complete account of the history of this subject, along with some good quality figures can be also found in [@phdheine]). Liquid drops can model self-gravitating systems. Can they model extremely compact, highly relativistic objects like black holes? The answer, quite surprisingly, seems to be yes [@cardosodiasglrp]. Regarding the event horizon as a kind of fluid membrane is a position adopted in the past [@membraneparadigm]. A simple way to see that this is a natural viewpoint, is to take the first law of black hole mechanics [@BardCartHawk] which describes how a black hole (we will take for simplicity uncharged, static objects), characterized by a mass $M$, horizon area $A$, and temperature $T=1/(32 \pi M)$, evolves when we throw an infinitesimal amount of matter into it: $$dM=TdA \,.\label{firstlaw}$$ We now know that the correct interpretation of this law is that black holes radiate, and therefore can be assigned a temperature [@bekenstein; @hawking]. The second law of black hole mechanics $dA\geq0$ is then just the second law of thermodynamics. But we can also argue that Equation (\[firstlaw\]) can be looked at as a law for fluids, with $T$ being an effective surface tension [@booktension]. The first works in black hole mechanics actually considered $T$ as a surface tension (see the work by Smarr [@smarr] and references therein), which is rather intuitive: the potential energy for fluids, associated with the storage of energy at the surface, is indeed proportional to the area. Later, Thorne and co-workers [@membraneparadigm] developed the “membrane paradigm” for black holes, envisioning the event horizon as a kind of membrane with well-defined mechanical, electrical and magnetic properties. Not only is this a simple picture of a black hole, it is also useful for calculations and understanding what black holes are really like. This analogy was put on a firmer ground by Parikh and Wilczek [@wilczek], who provided an action formulation of this membrane picture. There are other instances where a membrane behavior seems evident: Eardley and Giddings [@eardley], studying high-energy black hole collisions found a soap bubble-like law for the process, while many modern interpretations of black hole entropy and gravity “freeze” the degrees of freedom in a lower dimensional space, in what is known as holography [@holo]. In [@kovtun1] it was shown that a “membrane” approach works surprisingly well, yielding precisely the same results as the AdS/CFT correspondence. There have also been attempts to work the other way around: computing liquid surface tension from the (analog) black hole entropy [@surfacetbh]. As a last example of a successful use of the membrane picture, Cardoso and Dias [@cardosodiasglrp] have recently been able to mimic many of the aspects of a gravitational instability by using a simple fluid analogy. We will comment on this work later on.
Having said this, it strikes one as very odd that the study of simple fluid models has not been extended to higher dimensions. Extra dimensions, besides the four usual ones, seem to be needed for consistency in all or many of candidate theories of unification. We have in mind not only string theory [@zwie] where the extra dimensions are compactified to a very small scale, but also recent theories with large [@add] or warped extra dimensions [@rs]. Therefore, an understanding of higher dimensional gravitating objects is imperative. However, even though there are several known exact solutions to Einstein equations in more than four dimensions describing black objects (with event horizons), nothing or very little is known about the simpler Newtonian fluid equilibrium configurations, with or without self-gravity.
For instance, higher dimensional black hole solutions with spherical topology have long been known [@tangherlini; @myersperry], which generalize the four-dimensional Kerr-solution to higher dimensions. In four spacetime dimensions one can show that the topology of a black hole must be a sphere, but that’s no longer the case in higher dimensions. A simple example is to consider extra flat dimensions, for instance ${\rm
Schwarzschild} \times R$, which is still a solution to the vacuum Einstein equations. These extended solutions can then have different topologies, like a cylinder, and are called back strings [@horowitzstrominger]. Yet another example is the five-dimensional black ring [@Emparan:2001wn; @Elvang:2003mj], having the topology of a torus.
It is curious to note that we are in possession of several exact solutions to the full Einstein equations and yet nothing is known about the simpler case of self-gravitating Newtonian fluids or even fluids without gravity but with surface tension. To be specific, we found a solution as “strange” as a black ring in five dimensions, yet we do not know how a rotating, self-gravitating Newtonian fluid behaves as a function of the rotation rate, for higher dimensional spacetimes. These simpler objects should be much easier to study, while carrying precious information about their more compact cousins with event horizons. There are several topics that could and should be addressed within this research line:
\(i) Do the fluid equations admit toroidal-like configurations, which may collapse to black rings for very high densities? Are these stable?
\(ii) Ultra-spinning black holes in spacetime dimensions larger than four are conjectured to be unstable [@emparanmyers; @cardosodiasglrp] for large enough rotation rates. If the four-dimensional results for self-gravitating fluids or for fluids with surface tension carry over, this is to be expected. As we saw in the evolution diagrams, axisymmetric configurations are unstable for rotation rates larger than a given threshold, decaying to a two-lobe, asymmetric configuration. This suggests that the end-state of ultra-spinning black holes is a two lobe, asymmetric state like those in Figure \[fig:twolobes\]. This possibility was already put forward in [@emparanmyers], but the arguments presented here make this a much stronger possibility. These shapes would then presumably spin-down due to radiation emission. Can we use simple fluids with surface tension or with self-gravity to model this process? Can we estimate the rotation rate at which instability occurs?
\(iii) The black strings and p-branes are unstable against a gravitational sector of perturbations. Can we simulate this instability with fluids? A first step was taken in [@cardosodiasglrp], where it was shown that a simple fluid with surface tension can account for most of the properties of this instability. Can we add self-gravity to the model? Can we include rotational and charge effects?
\(iv) Can we build the evolution and bifurcation diagrams for fluids with surface tension (or with self-gravity) in higher dimensions? Can we deduce other solutions to the general relativistic equations using these simpler models? Can we infer other properties of already known solutions?
These are just some of the topics that could be addressed with simple Newtonian fluid models, but the list could easily be made much longer. The goal of the present work is to pave the way to some of these possible developments, but especially to show that these simple models do yield relevant information. We will attack the problem in its two versions: fluids held together by surface tension and by self-gravity.
Plan of the paper {#plan-of-the-paper .unnumbered}
=================
This paper is divided in two parts, Part I in which we deal with fluids held by surface tension and Part II where we deal with fluids held together by self-gravity.
We start with Section \[sec:symm\], where we review the meaning of angular momentum in higher dimensions and also the symmetries of a rotating body. We also specify the two most important cases (case I and II) we shall deal with throughout the paper. In Section \[sec:hydroeqs\] we lay down the basic equations of hydrostatics and hydrodynamics in higher dimensions and in Section \[sec:perturbations\] we write the formalism necessary to analyse small deviations for equilibrium. The formalism just laid down is then used in the next sections. In Section \[sec:eqshapesurface\] we study the equilibrium figure of rotating drops. Oscillation modes of non-rotating spherical drops are derived in Appendix \[app:surfacetensiondrop\] and constitutes a generalization to arbitrary number of dimensions of the well-known result in three spatial dimensions. In Section \[sec:toroidalsurface\] we discuss equilibrium torus-like configurations of rotating fluids held together by surface tension, and show that these share common features with their vacuum general-relativistic counterpart, the black rings. In Section \[sec:stabilitysurface\] we study the stability of the equilibrium figures discussed in Section \[app:surfacetensiondrop\] against small perturbations. We compute neutral modes of oscillation and the onset of instability which allows one to derive the bifurcation diagram for the case of one non-vanishing angular momentum. We finish this part with Section \[sec:cylindersurface\] by reviewing the Rayleigh-Plateau instability of a thin long cylinder in higher dimensions [@cardosodiasglrp], which could be a good model for the Gregory-Laflamme instability. We discuss possible effects of rotation and magnetic fields.
In the second part we deal with self-gravitating fluids in higher dimensions. We start in Section \[sec:cylindergravity\] by considering an infinite cylinder of self-gravitating homogeneous fluid. We show that this cylinder is unstable against a classical instability, here termed Dyson-Chandrasekhar-Fermi (DCF) instability, which we demonstrate to be the exact analogue of the Gregory-Laflamme [@gl] instability of black strings. We discuss implications of this result for rotating and charged fluids and black strings. We then discuss in Section \[sec:torusselfgravitating\] toroidal equilibrium configurations of self-gravitating fluid (often termed Dyson rings) and we compare them to exact general relativistic toroidal-like configurations, the black ring solution. Finally, in Section \[sec:MacLaurin\] we write down the basic equations that generalize the MacLaurin and Jacobi sequence to higher dimensions, equations which could be used in future numerical work to reconstruct the whole bifurcation diagram. Oscillation modes of non-rotating globes are derived in Appendix \[app:gravitysphere\] and constitute a generalization to arbitrary number of dimensions of the well-known result by Kelvin in three spatial dimensions. We close in Section \[sec:conclusions\] with a discussion of the results, implications and suggestions for future work.
This work is by no means complete. Many important topics are left for future work, we discuss some of these in the concluding remarks.
Notation {#notation .unnumbered}
========
There are many different symbols used throughout the text. For clarity we find it useful to make a short list of the most often used symbols.
- $D \rightarrow$ total number of spatial dimensions.
- $x^i=(x^a,x^A) \rightarrow$ coordinates of a rotating body. We consider two particular cases: in case I there is a single rotation plane, and $a=1,2$; $A=3,...,D$; in case II, $D$ is odd and there are $\frac{D-1}{2}$ rotation planes. In this case $a=1,...D-1$; $A=D$.
- $C_D \rightarrow$ the area of the unit $(D-1)$-sphere, $C_D=\frac{2\pi^{D/2}}{\Gamma\left(\frac{D}{2}\right)}$.
- $\varpi \rightarrow$ radial coordinate in cylindrical coordinates.
- $\omega\rightarrow$ Fourier transform variable. The time dependence of any field is written as $e^{\omega t}$. In the case of a stable oscillation, it is related to the frequency by $\omega^2 =-\sigma^2$. Note that pure imaginary values of $\omega$ mean a stable oscillation.
- $\Omega\rightarrow$ angular velocity.
- ${\cal B}({\bf x})\rightarrow$ Gravitational potential.
- $G \rightarrow$ Newton’s constant in $D-$dimensions, defined by writing the gravitational potential as $${\cal B}({\bf x})\equiv G
\int_V\frac{\rho(\vec x')}{|\vec x-\vec x'|^{D-2}}d^Dx'\,.$$ This is perhaps not the most common definition. A popular definition of Newton’s constant is the one adopted by Myers and Perry [@myersperry] and the relation between the two is derived in Appendix \[app:newton\].
- $T\rightarrow$ magnitude of surface tension, defined by $P=Tn^i_{,i} $ where $P$ is the pressure in the interior of the body (we take the exterior pressure to be zero). We take ${\bf n}$ to be the outward normal to the surface the body.
- $P\rightarrow$ Pressure.
- $V\rightarrow$ Volume.
- $\rho\rightarrow$ density.
- $\lambda \rightarrow$ wavelength of a given perturbation.
- $k \rightarrow$ wavenumber $\equiv 2\pi/\lambda$.
- $I^{ij}\equiv \int_V \rho x^ix^jd^Dx \rightarrow$ Moment of inertia.
\[sec:symm\]Symmetries and geometrical description of a multi-dimensional rotating body
=======================================================================================
Let us consider a stationary rotating object in $D$-dimensional Euclidean space, with coordinates $x^i=(x^1,\dots,x^D)$. The rotation group in $D$ dimensions, $SO(D)$, has $D(D-1)/2$ generators $J^{ij}=J^{[ij]}$. Each generator $J^{ij}$ corresponds to an $SO(2)$ rotation, on the plane $x^i-x^j$.
In a suitable coordinate frame $J=(J^{ij})$ can be brought to the standard form (see for instance [@myersperry]) $$J=\left(\begin{array}{ccc}
\begin{array}{cc} 0 & J_1 \\ -J_1 & 0 \\ \end{array} & & \\
& \begin{array}{cc} 0 & J_2 \\ -J_2 & 0 \\ \end{array} & \\
& & \ddots \\ \end{array} \right)\,.$$ This means that in this frame the rotation is a composition of a rotation in the plane $x^1-x^2$, a rotation in the plane $x^3-x^4$, and so on, without mixings. The rotation is then characterized by $[D/2]$ generators (where $[\cdot]$ denotes the integer part), which are the generators of the Cartan subalgebra of $SO(D)$ (namely, the maximal abelian subalgebra of $SO(D)$). In other words, in $D$ dimensions there are $${\cal N}=\left[\frac{D}{2}\right]$$ independent angular momenta.
We will consider two particular cases:
- Only one angular momentum is non-vanishing: $J_1\neq 0$, $J_{i>1}=0$.
- All angular momenta are non-vanishing, and they are equal: $J_1=J_2=\dots=J_{\cal N}$; in this case we also assume $D$ to be odd.
These will be referred to as case I and II respectively throughout the whole article.
Case I. One non-vanishing angular momentum
------------------------------------------
In this case we are selecting a one-dimensional rotation subgroup $$SO(2)\subset SO(D)\,.$$ This means selecting a single rotation plane, $x^1\!-\!x^2$. We assume our object has the maximal symmetry compatible with this rotation: $$G=SO(2)\times SO(D-2)\subset SO(D)\label{sym}\,.$$ If we decompose the coordinates under $G$ as $$x^i=(x^a,x^A);~~~~a=1,2;~A=3,\dots,D\,,\label{coordinatesa}$$ and we define $$\begin{aligned}
\varpi&\equiv&\sqrt{(x^1)^2+(x^2)^2}\nonumber\,,\\
z&\equiv&\sqrt{(x^3)^2+\dots+(x^D)^2}\,,\label{deftoz1} \end{aligned}$$ the surface of our object depends only on $\varpi$ and $z$, and its equation can be written as $$z=f(\varpi) \,.$$ i.e. $$\Xi \equiv z-f(\varpi)=0\,,$$ therefore the surface does not depend on the angle in the $x^1\!-\!x^2$ plane, and it does not depend on the $D-3$ angles among $x^3,\dots,x^D$. We also define $$\phi\equiv\frac{df}{d\varpi}\,.$$ Notice that it must be $\phi=-\infty$ when $z=0$ for the surface to be smooth.
Being $\varpi_{,a}=\frac{x^a}{\varpi}$, the normal to the surface is $$\Sigma_{,i}=(\Sigma_{,a},\Sigma_{,A})=\left(-\phi\frac{x^a}{\varpi},
\frac{x^A}{z}\right)$$ with modulus $\delta^{ij}\Sigma_{,i}\Sigma_{,j}=1+\phi^2$, so the normal unit vector is $$n^i=\left(-\phi\frac{x^a}{\varpi\sqrt{1+\phi^2}},
\frac{x^A}{z\sqrt{1+\phi^2}}\right)\,.$$ Since the pressure is related to surface tension via $P=T n^i_{,i}$, we will need a general relation to compute $n^i_{,i}$. Let us then compute $n^i_{,i}=n^a_{,a}+n^A_{,A}$ for this case I. We have $$\begin{aligned}
n^a_{,a}&=&-\frac{\phi'}{(1+\phi^2)^{3/2}}
-\frac{\phi}{\varpi\sqrt{1+\phi^2}}=-\frac{1}{\varpi}\frac{d}{d\varpi}
\frac{\varpi\phi}{\sqrt{1+\phi^2}}\nonumber\\
n^A_{,A}&=&\frac{1}{\sqrt{1+\phi^2}}\frac{D-3}{z}\,. \label{divna} \end{aligned}$$
Case II. All angular momenta coincident, and $D$ odd.
-----------------------------------------------------
In this case the rotation group is broken to its Cartan subalgebra $$(SO(2))^{\frac{D-1}{2}}\subset SO(D)\,.$$ The rotation planes are $$x^1-x^2,~x^3-x^4,~\dots,~x^{D-2}-x^{D-1} \,,$$ and the $x^D$ axis is left unchanged by the rotation[^2]. The decomposition of the coordinates is now $$x^i=(x^a,x^A);~~~a=1,\dots,D-1\,;~~~A=D \,.\label{coordinatesb}$$ We define $$\begin{aligned}
\varpi&\equiv&\sqrt{(x^1)^2+(x^2)^2+\dots+(x^{D-1})^2}\nonumber\,,\\
z&\equiv&x^D\,. \end{aligned}$$ The surface does not depend on the $D-2$ angles among the $x^a$. We also define $$\phi\equiv\frac{df}{d\varpi}\,.$$ Notice that it must be $\phi=-\infty$ when $z=0$ for the surface to be smooth.
Being $\varpi_{,a}=\frac{x^a}{\varpi}$, the normal to the surface is $$\Sigma_{,i}=\left(-\phi\frac{x^a}{\varpi},0\right)\,,$$ with modulus $\delta^{ij}\Sigma_{,i}\Sigma_{,j}=1+\phi^2$, so the normal unit vector is $$n^i=\left(-\phi\frac{x^a}{\varpi\sqrt{1+\phi^2}},0\right)\,.$$ Finally, $$\begin{aligned}
n^i_{,i}&=&n^a_{,a}=-\frac{1}{\varpi^{D-2}}\frac{d}{d\varpi}\left
(\frac{\varpi^{D-2}\phi}{\sqrt{1+\phi^2}}\right)\,.\label{divnb} \end{aligned}$$
\[sec:hydroeqs\]Hydrodynamical and hydrostatical equations
==========================================================
Let ${\bf X}=(X^i)$ the position vector of a fluid element in the inertial frame, and ${\bf x}=(x^i)$ its position vector in the rotating frame, i.e. $${\bf X}=R{\bf x}$$ where $R=R_I$ or $R_{II}$ according to the case under consideration and $$R_I=\left(\begin{array}{cc|ccc}
\cos\Omega t & \sin\Omega t &&& \\
-\sin\Omega t & \cos\Omega t &&& \\
\hline
& & 1 && \\
& & & \dots & \\
& & & & 1 \\
\end{array}\right)\,;
~~~~~~~~ R_{II}=\left(\begin{array}{cccc}
\begin{array}{cc} \cos\Omega t & \sin\Omega t \\
-\sin\Omega t & \cos\Omega t \\ \end{array} & & & \\ %\hline
& \ddots & \\ %\hline
& & \begin{array}{cc} \cos\Omega t & \sin\Omega t \\
-\sin\Omega t & \cos\Omega t \\ \end{array} & \\ %\hline
& & & 1 \\
\end{array}\right)$$ are $D$-dimensional rotation matrices in the cases I, II respectively. We introduce the velocity ${\bf \cal V}$ and acceleration ${\bf \cal A}$ by $$\begin{aligned}
\dot{\bf X}&=&R{\bf \cal V}\nonumber\\
\ddot{\bf X}&=&R{\bf \cal A} \end{aligned}$$ where an overdot denotes time differentiation.
The hydrodynamical equation, in the inertial frame, is $$\ddot X^i=-\frac{1}{\rho}\frac{\partial P}{\partial X^i}\label{eqmotion}$$ (with $P$ pressure and $\rho$ constant density of the fluid) which, applying $R^{-1}$, gives $${\cal A}^i=-\frac{1}{\rho}\frac{\partial}{\partial x^i}P\,.$$ Let us compute ${\bf \cal A}$. We have $${\bf \cal V}=R^T\dot R{\bf x}+\dot{\bf x}$$ and then $${\bf \cal A}=R^T\dot R{\bf \cal V}+\dot{\bf \cal V}= (R^T\dot R)^2{\bf x}+2R^T\dot R \dot{\bf x}+\ddot{\bf
x}\,.$$ We have that the only non-vanishing components of $R^T\dot R$ are $$(R^T\dot R)^{ab}=-\Omega\epsilon^{ab}$$ where $\epsilon^{ab}=\epsilon^{[ab]}$, and $\epsilon^{12}=1$ in the case I, $\epsilon^{12}=\epsilon^{34}=\dots=\epsilon^{(D-2)(D-1)}=1$ in the case II.
Therefore, in both cases $((R^T\dot R)^2)^{ab}=\Omega^2\delta^{ab}$, and $$\begin{aligned}
\ddot x^a&=&-\frac{1}{\rho}\frac{\partial P}{\partial x^a}+\Omega^2
x^a+2\Omega\epsilon^{ab}\dot x^b \label{eq11}\\
\ddot x^A&=&-\frac{1}{\rho}\frac{\partial P}{\partial x^A}\,.\label{hydyn} \end{aligned}$$ The second term in the right hand side of (\[eq11\]) is the centrifugal force and the third term is the generalized Coriolis force. At equilibrium $\dot x^i=\ddot x^i=0$, therefore $$\begin{aligned}
\frac{\partial P}{\partial x^a}&=&\rho\Omega^2 x^a\nonumber\,,\\
\frac{\partial P}{\partial x^A}&=&0 \,.\end{aligned}$$ thus $$P=P_0+\frac{1}{2}\rho\Omega^2\varpi^2\,. \label{press}$$ This formula is valid both in case I and in case II, i.e., when there is one non-vanishing angular momentum and when all of them coincide (and $D$ is odd).
\[sec:perturbations\]The virial theorem, surface-energy tensor and small perturbations
======================================================================================
We would ultimately wish to assess stability of rotating drops, in order to generalize Fig. \[fig:brown\] to higher dimensions. This calculation would also provide some hints as to whether or not one should expect to see instabilities in black objects in higher dimensions, as we stressed in the Introduction. To attack this problem we will follow Chandrasekhar’s [@chandra] approach, to which we refer for further details. Alternative (numerical) approaches, perhaps more complete but more complex also, are described by Brown and Scriven [@brownscriven] and also Heine [@heine].
The virial theorem in the inertial frame
----------------------------------------
Consider a fluid of volume $V$ with uniform density $\rho$, bounded by a surface $S$. The external pressure, for simplicity, is taken to be zero. Due to the existence of surface tension, the pressure $P$ immediately adjacent to $S$, on the interior, is given in terms of the surface tension magnitude $T$ by $$P=Tn^i_{,i} \label{eqdrop}\,.$$ Multiplying (\[eqmotion\]) by $X^j$ and integrating over the volume $V$ of the fluid we get $$\frac{d}{dt}\int_V \rho \dot{X}^iX^j d^DX=2{\cal I}^{ij}-\int_V X^j \frac{\partial P}{\partial X^i} d^DX
\,,\label{st1}$$ where ${\cal I}^{ij}=\frac{1}{2}\int_V \rho \dot{X}^i \dot{X}^j d^D X$ is the inertia tensor. Integrating by parts the second term on the R.H.S. of (\[st1\]) we have $$\int_V X^j \frac{\partial P}{\partial X^i} d^DX=\int_S X^jPdS^i -\delta^{ij}\int_V P d^DX\,,\label{st2}$$ where $dS^i=n^idS$ and $dS$ is the $(D-1)$-dimensional surface element on $S$. Using (\[eqdrop\]) we further have $$\int_V X^j \frac{\partial P}{\partial X^i} d^DX= T\int_S X^jn^i_{,i} dS^i-\delta^{ij} \int _V P
d^DX\,.\label{st3}$$ Introducing this result in (\[st1\]) one gets the virial theorem $$\frac{d}{dt}\int_V \rho \dot{X}^iX^j d^DX=2{\cal I}^{ij}+ {\mathfrak S}^{ij}+\delta^{ij}\Pi \,,$$ where $$\begin{aligned}
\Pi&=&\int_V P d^DX\,,\\
{\mathfrak S}^{ij}&=&-T\int_S X^j n^k_{,k}dS^i\,. \end{aligned}$$ The quantity ${\mathfrak S}^{ij}$ is the surface-energy tensor [@chandra]. Conservation of angular momentum guarantees that $\mathfrak{S}^{ij}$ is symmetric. Indeed, $$\frac{d}{dt}\int_V\rho(\dot X^iX^j-\dot X^jX^i)d^D X=\mathfrak{S}^{ij}-\mathfrak{S}^{ji}=0\,.$$
The virial theorem in the rotating frame
----------------------------------------
In a rotating frame of reference the equations of motion are written down in (\[hydyn\]). Multiplying those equations by $x^j$, integrating over $V$ and using the previous sequence of transformations we get $$\begin{aligned}
\frac{d}{dt}\int_V \rho \dot{x}^a x^j d^Dx&=&2{\cal I}^{aj}+{\mathfrak
S}^{aj}+\delta^{aj}\Pi+\Omega^2I^{aj}
+2\Omega \epsilon^{ac}\int_V \rho \dot{x}^{c}x^{j}d^Dx \,,\label{v1}\\
\frac{d}{dt}\int_V \rho \dot{x}^A x^j d^Dx&=&2{\cal I}^{Aj}+
{\mathfrak S}^{Aj}+\delta^{Aj}\Pi \,,\label{v2}\end{aligned}$$ where $$I^{aj}=\int_V \rho x^{a}x^{j}d^Dx\,.$$ The tensor $\mathfrak{S}^{ij}$ is symmetric in the rotating frame, too, as can be shown by transforming it with the rotation matrix $R$.
We remind the reader that according to our conventions the indices $i,j$ can take any values between $1$ and $D$, but the indices $a,\,b$ and $A,\,B$ are restricted according to the situation we consider (case I or II), described in Equations (\[coordinatesa\]) and (\[coordinatesb\]). In the following we always consider the rotating reference frame.
\[virial\] Virial equations for small perturbations
---------------------------------------------------
Having solved for the equilibrium equations and found the shape for a given rotating drop (this problem will be explicitly worked out in the next Section \[sec:eqshapesurface\]), we ask whether this is a stable equilibrium configuration or not. To answer this we follow the usual procedure to assess stability: we perturb the equilibrium configuration. Suppose then that a drop in a state of equilibrium is slightly disturbed. All background quantities $Q_0$ are then changed by an amount $\delta Q_0$. The motions are described by a Lagrangian displacement of the form $${\bf \xi}(x)e^{\omega t}\,,$$ where $\omega$ is to be determined from the problem. In the case of constant density, $\xi$ is divergence free. Define $$\Gamma^{i;j}=\int_V \rho \xi^ix^jd^Dx\,,$$ where the semi-colon indicates that a moment with respect to coordinate space is being evaluated. Define also $$\Gamma^{ij}\equiv \Gamma^{i;j}+ \Gamma^{j;i}\,.$$ Then the perturbation equations read [@chandra] $$\begin{aligned}
\omega^2\Gamma^{a;j}-2\omega\Omega \epsilon^{ac}\Gamma^{c;j} &=&\delta {\mathfrak S}^{aj}+
\Omega^2\Gamma^{aj}+\delta^{aj}\delta \Pi \,,\label{eqvir1}\\
\omega^2\Gamma^{A;j}&=&\delta {\mathfrak S}^{Aj}+\delta^{Aj} \delta \Pi \,.\label{eqvir2} \end{aligned}$$ The surface integrals $\delta\mathfrak S^{ij}$ have the form $$\begin{aligned}
\delta\mathfrak S^{ij}&=&-T\delta\int_Sx^jn^k_{~,k}dS^i
=-T\left[\int_S\xi^jn^k_{~,k}dS^i+\int_Sx^j\delta(n^k_{~,k})dS^i
+\int_Sx^jn^k_{~,k}\delta(dS^i)\right]\nonumber\\
&=&-T\left[\int_S\xi^jn^k_{~,k}dS^i+\int_Sx^j\delta(n^k_{~,k})dS^i
-\int_Sx^jn^k_{~,k}\frac{\partial \xi^l}{\partial x^i}dS^l\right]\label{eqkS}\\\end{aligned}$$ where we have used the general result (see for instance [@chandra]) $$\delta(dS^i)=\xi^j_{~,j}dS^i-\frac{\partial \xi^j}{\partial x^i}dS^j$$ and $\xi^j_{~,j}=0$. Furthermore, we have (see [@chandra]) $$\begin{aligned}
\delta(n^i_{~,i})=(\delta n^i)_{~,i}-\frac{\partial n^i}{\partial x^k}
\frac{\partial \xi^k}{\partial x^i}\nonumber\\
\delta n^i=\left(n^kn^l\frac{\partial \xi^k}{\partial x^l}\right)n^i -n^j
\frac{\partial \xi^j}{\partial x^i}\,.\label{expdiv}\end{aligned}$$
Equations (\[eqvir1\]), (\[eqvir2\]) can be decomposed as $$\begin{aligned}
\omega^2\Gamma^{a;b}-2\omega\Omega \epsilon^{ac}\Gamma^{c;b}&=&\delta {\mathfrak S}^{ab}+
\Omega^2\Gamma^{ab}+\delta^{ab}\delta \Pi \,,\label{even1}\\
\omega^2\Gamma^{A;B}&=&\delta {\mathfrak S}^{AB}+
\delta^{AB}\delta\Pi \,,\label{even2}\\
\omega^2\Gamma^{a;A}-2\omega\Omega \epsilon^{ac}\Gamma^{c;A}&=&\delta {\mathfrak S}^{aA}+
\Omega^2\Gamma^{aA} \,,\label{odd1}\\
\omega^2\Gamma^{A;a}&=&\delta {\mathfrak S}^{Aa} \label{odd2}\,.\end{aligned}$$ These are the basic equations we will use later on to perform a stability analysis of rotating drops in higher dimensions. To complete the calculations we first need to compute the equilibrium figures of rotating drops and we this in the following.
\[sec:eqshapesurface\]Equilibrium shape of a rotating drop
==========================================================
\[sec:eqcasea\] Case I. One non-vanishing angular momentum
----------------------------------------------------------
From equations (\[divna\]), (\[press\]), (\[eqdrop\]) we find $$P_0+\frac{1}{2}\rho\Omega^2\varpi^2=T\left(-\frac{1}{\varpi}\frac{d}{d\varpi}
\frac{\varpi\phi}{\sqrt{1+\phi^2}}
+\frac{1}{\sqrt{1+\phi^2}}\frac{D-3}{z}\right)\,,\label{equationcasea}$$ thus $$\frac{1}{2}P_0\varpi^2+\frac{1}{8}\rho\Omega^2\varpi^4=
-T\frac{\varpi\phi}{\sqrt{1+\phi^2}}+(D-3)TI+C\label{eq0}$$ where $C$ is a constant, and $I$ is the integral $$I(\varpi)=\int_0^\varpi{\frac{d\varpi'\varpi'}{f(\varpi')\sqrt{1+ \left(\frac{df}{d\varpi'}
\right)^2}}}\,.\label{I}$$ If the drop encloses the origin (which we suppose for the time being), then $\varpi=0$ must belong to the surface; being $I(0)=0$, this means that $C=0$ and equation (\[eq0\]) reduces to $$\frac{1}{2}P_0\varpi^2+\frac{1}{8}\rho\Omega^2\varpi^4=
-T\frac{\varpi\phi}{\sqrt{1+\phi^2}}+(D-3)TI\,.\label{eq}$$ When $D=3$ this equation can be solved analytically [@chandra], but when $D>3$ this is not possible, because equation (\[eq\]) cannot be expressed in the form $\frac{df}{d\varpi}=F(\varpi)$.
We thus have to numerically integrate equation (\[equationcasea\]), which we re-write as $$\frac{P_0}{T}+\frac{1}{2T}\rho\Omega^2\varpi^2=\left(-\frac{1}{\varpi}\frac{d}{d\varpi}
\frac{\varpi\phi}{\sqrt{1+\phi^2}}
+\frac{1}{\sqrt{1+\phi^2}}\frac{D-3}{z}\right)\,,\label{equationcasea2}$$ We can re-scale coordinates $\varpi \rightarrow a\varpi$ and $z
\rightarrow az$, where $a$ will henceforth be the equatorial radius of the equilibrium figure. In this way all quantities are dimensionless and equation (\[equationcasea2\]) is written as $$\frac{P_0a}{T}+4\Sigma_a\varpi^2=-\frac{1}{\varpi}\frac{d}{d\varpi}
\frac{\varpi\phi}{\sqrt{1+\phi^2}}
+\frac{1}{\sqrt{1+\phi^2}}\frac{D-3}{z}\,,\label{equationcasea3}$$ with $\Sigma_a \equiv\frac{\rho \Omega^2a^3}{8T}$.
This equation can be integrated numerically, starting at $\varpi=0$ where $z=z(0)$ and $z'(0)\equiv \phi=0$. In this sense we are solving an eigenvalue equation: given $\Sigma_a$, $P_0a/T$ and $z(0)$ are determined by requiring that at the equator $r=a$ we have $z(a)=0$ and $z'(a)=-\infty$. The problem is highly non-linear, so the numerical procedure involved varying [*both*]{} $P_0a/T$ and $z(0)$ simultaneously over a rectangular grid, in order to determine those values for which the outer boundary conditions ($z(a)=0$ and $z'(a)=-\infty$) were satisfied. The accuracy gets worse for higher $\Sigma$ and $D$ but one usually has a precision of about $5\%$ in our numerical results.
![Typical (axi-symmetric) equilibrium figures of rotating drops for case I (one non-vanishing angular momentum), in $D=4$. The (blue) dashed line corresponds to a spherical surface with $\Sigma_a=0$. The (magenta) dotted line corresponds to $\Sigma_a=0.46$ and the (black) solid line corresponds to the limit $\Sigma_a=0.86$. In all three cases the equatorial radius is unity.[]{data-label="fig:eqfigsa"}](Fig4.eps){width="8" height="8"}
Numerical results for $D=4$ are shown in Figure \[fig:eqfigsa\] for some values of $\Sigma_s$, which can be compared to the $D=3$ results presented by Chandrasekhar in [@chandra]. For $\Sigma_a=0$, the non-rotating case, we start by verifying numerically that the sphere is a solution, as it should, with $P_0a/T=D-1$. The equilibrium figures are very similar and indeed the similarity seems to continue for larger $D$. The three different equilibrium shapes shown in Figure \[fig:eqfigsa\] all enclose the origin. As in the $D=3$ case (see [@chandra]), we find that for $\Sigma_a$ larger than a certain critical value, the drop no longer enclosed the origin, giving rise to toroidal shapes. The condition to be satisfied at the critical $\Sigma$ is that $z=0$ at $r=0$. For $D=3$ the critical value of $\Sigma_a$ is $\Sigma_a\sim 2.32911$. For $D>3$ it is harder to determine the critical parameter, since there is no integral expression for the shape. Nevertheless we determine $\Sigma_{\rm crit}\sim 2.3$ for $D=4,5$ with a $5\%$ uncertainty. An open important problem is to determine if a critical point exists for all $D$. If not, this might mean that the bifurcation curve \[fig:brown\] does not bend down, and therefore that the angular momentum is unbounded, which might be a good analogue for what happens with ultra-rotating [@myersperry] black holes. This may require more accurate numerical integrations, since we cannot trust our method for $D>5$.
Case II. All angular momenta coincident, and $D$ odd.
-----------------------------------------------------
In case II we will find that the results are very similar to those for $D=3$. From equations (\[divnb\]), (\[press\]), (\[eqdrop\]) we find $$P_0+\frac{1}{2}\rho\Omega^2\varpi^2=-\frac{T}{\varpi^{D-2}}
\frac{d} {d\varpi}\left
(\frac{\varpi^{D-2}\phi}{\sqrt{1+\phi^2}}\right )\,,$$ which is immediately integrated to give $$P_0\frac{\varpi^{D-1}}{D-1}+\frac{\rho
\Omega^2\varpi^{D+1}}{2(D+1)}
=-\frac{T\varpi^{D-2} \phi}{\sqrt{1+\phi^2}}+C\,, \label{eq1}$$ where $C$ is a constant.
Let’s first suppose the drop encloses the origin. In this case, there are points where $\varpi=0$ and therefore we must have $C=0$. Equation (\[eq1\]) can then be written as $$P_0\frac{\varpi}{D-1}+\frac{\rho \Omega^2\varpi^3}{2(D+1)}
=-\frac{T\phi}{\sqrt{1+\phi^2}}\,. \label{eq2}$$ At the equator, where $\varpi=a$ we must have $\phi \rightarrow
-\infty$ and then (\[eq1\]) implies that $$\frac{P_0a}{(D-1)T}=1-\Sigma_b\,,\quad \Sigma_b \equiv
\frac{\rho \Omega^2a^3}{2(D+1)T}\,.\label{sigmadef}$$ Equations (\[eq2\])-(\[sigmadef\]) are the same as equations (A4)-(A5) by Chandrasekhar [@chandra]. So, figures of equilibrium enclosing the origin do not depend on the dimensionality of spacetime, if measured in terms of the dimensionless number $\Sigma_b$ (by this we mean of course that $z(\varpi,\Sigma_b)$ is independent of $D$; anyway, the relation between $\Sigma_b$ and $\Omega$ depends on $D$, thus $z(\varpi,\Omega)$ depends on $D$). By measuring $\varpi$ and $\phi$ in units of $a$, we get $$\phi=\frac{d f}{d\varpi}=-\frac{\varpi\left (1- \Sigma_b+\Sigma_b
\varpi^2\right )}{\sqrt{1-\varpi^2\left
(1-\Sigma_b+\Sigma_b \varpi^2\right )^2}}\,,$$ which can be written as $$z=f(\varpi)=\int_{\varpi}^{1}\frac{w\left (1-\Sigma_b+\Sigma_b w^2
\right )}{\sqrt{1-w^2\left
(1-\Sigma_b+\Sigma_b w^2\right )^2}}d w\,.$$ Transforming to a new variable $\theta$ defined through $$\varpi^2=1-A\tan^2({\theta/2})\,,\quad A\equiv
\frac{\sqrt{1+2\Sigma_b}}{\Sigma_b}\,,$$ the integral representation can be simplified to $$z=\frac{1}{2\Sigma_b\sqrt{A}}\int_0^{\phi}\left(1+A\Sigma_b-
\frac{2A\Sigma_b}{1+\cos{\theta}}\right
)\frac{d\theta}{\sqrt{1-\chi^2\sin^2\theta}}\,, \label{zsol}$$ where $$\chi^2=\frac{1}{2}\left [1+\frac{1}{A}\left ( \frac{1}{2}+
\frac{1}{\Sigma_b} \right )\right ]\,.$$ Here, $\phi$ is related to the value of $\varpi$ at the boundary by $$\varpi=\sqrt{1-A\tan^2{\phi/2}}$$ The solution (\[zsol\]) can be expressed in terms of elliptic integrals $$F(\chi,\phi)=\int_0^{\phi}\frac{d\theta}{\sqrt{1-\chi^2\sin^2\theta}}
\,,\quad E(\chi,\phi)=\int_0^{\varpi}
d\theta \sqrt{1-\chi^2\sin^2\theta}\,,$$ as $$z=\frac{1}{2\Sigma_b \sqrt{A}}\left [\left (1-\Sigma_b A\right
)F(\chi,\phi) + 2A\Sigma_b \left (E(\chi,\phi) -\frac{\sin\phi
\sqrt{1-\chi^2\sin^2\phi}}{1+\cos \phi}\right )\right ]
\label{shape}\,.$$ Some representative equilibrium solutions are displayed in Figure \[fig:eqfigs\], for some values of $\Sigma_b$. The shapes are very similar to their “case I” relatives. We show typical shapes for high values of $\Sigma$ (these can be compared with the ones presented in Figure 1 of Chandrasekhar [@chandra]).
We have been dealing with drops that enclose the origin, and found that a closed expression for the shape is given by (\[shape\]). When does this sequence end? A necessary and sufficient condition for the drop to enclose the origin is that $z=0$ and $\xi=0$ are simultaneously satisfied. Now, the condition $\xi=0$ implies that $\phi=\phi_{\rm max}=2\arctan A^{-1/2}$. Substituting this value in (\[shape\]) we get $$\Sigma_{\rm max}=2.32911\,.$$ For values of $\Sigma_b$ larger than the above, the figure no longer encloses the origin.
![Typical equilibrum figures of rotating drops. The (blue) dashed line corresponds to a spherical surface with $\Sigma_b=0$. The (magenta) dotted line corresponds to $\Sigma_b=2$ and the (black) solid line corresponds to the limit $\Sigma_b=2.329$. In all three cases the equatorial radius was chosen to be unity.[]{data-label="fig:eqfigs"}](Fig5.eps){width="8" height="10"}
\[sec:toroidalsurface\]Equilibrium toroidal shapes
==================================================
In the general case of figures which do not enclose the origin (i.e.,$\varpi=0$ is not attained), we have to use equation (\[equationcasea\]) and (\[eq1\]) for cases I and II respectively. Here we shall focus on case II, equation (\[eq1\]), since it is easily manageable. In three spatial dimensions, solutions to equation (\[eq1\]) include ellipsoids, disks, biconcave and torus-like (or “wheels” in the terminology of [@rotatingdropexperiments2]) forms [@heine; @rotatingdropexperiments2]. We now wish to consider if one can use these toroidal fluid configuration to model strong gravity configurations with the same topology. In particular, we consider a torus with a large radius $R$ and small thickness $r_+$ (the reason being that it is in this regime that one really expects a “pure” torus).
We will first look for torus-like solutions which start smoothly (i.e. $\phi=0$) at some distance $R_1$. We will call these “type A” toroidal solutions. The outer radius will be denoted $R_2$, where $\phi=-\infty$. In this case, we can easily determine the constants $p_0\,,C$ and write (\[eq1\]) as $$P_0\frac{\varpi^{D-1}-R_1^{d-1}}{T(D-1)}+\frac{\rho \Omega^2 \left
(\varpi^{D+1}-R_1^{D+1}\right
)}{2T(D+1)}=-\frac{\varpi^{D-2}\phi}{\sqrt{1+\phi^2}}\,, \label{eq122}$$ where $$P_0\frac{R_2^{D-1}-R_1^{D-1}}{T(D-1)}=R_2^{D-2}-\frac{\rho \Omega^2 \left
(R_2^{D+1}-R_1^{D+1}\right)}{2T(D+1)}\,. \label{eq1222}$$ We have numerically integrated equation (\[eq1222\]) for several values of the parameters, and we find the law $$R_2 \left (R_2-R_1 \right )^2(D+1)\Lambda\simeq 3.57\,,$$ which is independent of dimensionality $D$. Here $$\Lambda\equiv \frac{\rho \Omega^2}{2T(D+1)}\,.\label{lambdadef}$$ The meridional cross-section for a typical type A torus is displayed in Fig. \[fig:smoothtorus\].
![Typical “smooth” Type A torus-like equilibrium shape. We have taken $D=4$ and $\Lambda=3.5,\,R_1=99.9549,\,R_2=100$.[]{data-label="fig:smoothtorus"}](Fig6.eps){width="8.5" height="6"}
As shown by Gulliver [@gulliver] and Smith and Ross [@smithross] there are several possible families of toroidal fluids (some of which have been experimentally observed already, see [@rotatingdropexperiments2]). We have just described one of those families, that starts smoothly at $R_1$.
We can find a second family, the type B toroidal shape, which is more similar to the gravitational Dyson rings which will be described in Section \[sec:torusselfgravitating\]. What we now look for is a family which starts “abruptly” at $R_1$. Our numerical experiments seem to suggest a very rich behavior for the members of this family. For very small $\Lambda$, we find the following approximate scaling law, $$r_+ R^2\Lambda \sim \frac{1.96}{D+1}\,,\quad R\equiv R_2\sim
R_1\,.\label{typeIItorus}$$ and the meridional cross-section of a typical member of this family is shown in Figure \[fig:torus\]. Such a family seems to have been discussed already by Appell [@appell].
![Typical Type B torus-like equilibrium shape. We have taken $D=4$ and $\Lambda=10^{-7},\,R_1=2\times 10^4,\,R_2=2.00000099\times 10^4$. We chose $\frac{P_0}{T(D-1)}=1$.[]{data-label="fig:torus"}](Fig7.eps){width="8.5" height="6"}
Let’s pause a moment on this scaling law. If, following the suggestion in [@cardosodiasglrp], we try to model gravitating black bodies with fluids held together by surface tension, the first law suggests [@cardosodiasglrp] that the surface tension is the equivalent of the Hawking temperature (up to a multiplicative constant). Now, the temperature of a black hole scales as $1/r$, where $r$ is some typical radius. For large “black rings” the relevant radius is the thickness $r_+$ of the ring. Thus we set $T \sim
1/r_+$. Using (\[lambdadef\]) and (\[typeIItorus\]) we get $$R^2\Omega^2\frac{M}{R r_+^{D-3}}\sim 1\,,$$ where we replaced $\rho=M/V$.
On the other hand, the mass per unit length of such a “black torus” should be given by the mass of a $D-1$ Tangherlini black hole [@tangherlini] (i.e., a higher dimensional non-rotating black hole): $$(D-2)C_DG M \sim 4\pi(D-2)C_{D-2}R\,r_+^{D-3}\,. \label{massblackobjects0}$$ Using the “constitutive” relation (\[massblackobjects0\]) for black objects and $J\sim \Omega M R$ we get finally $$\frac{J^2}{M^2R^2} \sim 1\,. \label{surfacering}$$ Fortunately, there is an exact solution of Einstein’s equations describing a general relativistic black ring, in five dimensions ($D=4$) [@Emparan:2001wn]: the black ring solution by Emparan and Reall [@Emparan:2001wn]. The black ring solution is described by two parameters $\nu$ and $R$. They are characterized by having mass and angular momentum equal to [@Emparan:2001wn; @Elvang:2003mj]: $$\begin{aligned}
M&=&\frac{3\pi
R^2}{4G_L}\frac{\bar\lambda(1+\bar\lambda)}{1+\nu}\,\\ J&=&\frac{\pi
R^3}{2G_L}\frac{\sqrt{\bar\lambda \nu}(1
+\bar\lambda)^{5/2}}{(1+\nu)^2}\,. \end{aligned}$$ Here, the parameter $\bar\lambda=2\nu/(1+\nu^2)$ for black rings. Large rings can be obtained by setting $\nu \rightarrow 0$, in which case $R$ can be looked at as the radius of the ring. For very small $\nu$ we then have $$\frac{J}{MR}=\frac{\sqrt{2}}{3} \label{grring}\,,$$ This is consistent with relation (\[surfacering\]) obtained in the model of a fluid held together by surface tension. The case $D=3$ seems to be special for in three dimensions there are no black strings (all black holes must be topological spherical in this case), whereas the surface tension model seems to allow for rings;
\[sec:stabilitysurface\]Stability analysis of a rotating drop in case I. one non-vanishing angular momentum
===========================================================================================================
To complete our discussion on drops held by surface tension, we consider the generalization to $D$-spatial dimensions of Figure \[fig:brown\], the bifurcation and stability diagram. Here we restrict ourselves to case I, i.e. only one non-vanishing angular momentum, corresponding to rotation in the plane $x^1-x^2$, and we will necessarily leave many topics for future research. For instance, all asymmetric figures will be left out of this exercise. We will follow the approach of Chandrasekhar [@chandra], based on the formalism of virial tensors. The special case in which the rotation is zero, i.e., a spherical drop, is considered in Appendix \[app:surfacetensiondrop\] where we compute the oscillation modes of a $D$-dimensional incompressible fluid sphere held by surface tension.
We have found the basic equations describing the equilibrium shape of rotating drops in the previous Sections. We would now like to investigate if these shapes are stable. The formalism to do this was laid down in Section \[virial\], and equations (\[even1\])-(\[odd2\]) are the most general set of equations describing perturbed rotating drops.
From now on we will restrict our analysis to a special kind of perturbations, termed “toroidal” modes by Chandrasekhar [@chandra]. The reason is that these modes are expected to play a role in the bifurcation and stability analysis. Future work should include an analysis of the other modes (termed “transverse-shear” and “pulsation modes”). We refer the reader to the work by Chandrasekhar [@chandra] for further details on the formalism. The toroidal modes are defined by $$\begin{aligned}
\xi^1&=&\alpha x^1+\beta x^2\nonumber\\
\xi^2&=&\beta x^1-\alpha x^2\nonumber\\
\xi^A&=&0\label{tor}\end{aligned}$$ with $\alpha,\beta$ constant parameters.
As shown in the Appendix \[toroidal\], if a perturbation has the form (\[tor\]), the only non-vanishing virial integrals are $$\begin{aligned}
\Gamma^{1;1}&=&-\Gamma^{2;2}=\pi\rho\alpha\frac{C_{D-2}}{D-2}
\int_0^a\varpi^3\left[f(\varpi)\right]^{D-2}d\varpi
\label{vir11}\\
\Gamma^{1;2}&=&\Gamma^{2;1}=\pi\rho\beta\frac{C_{D-2}}{D-2}
\int_0^a\varpi^3\left[f(\varpi)\right]^{D-2}d\varpi
\label{vir12}\end{aligned}$$ where $\varpi=a$ is the intersection of the surface of the drop with the $x^1-x^2$ plane. These quantities satisfy the equations (\[even1\]), which have the form $$\begin{aligned}
\omega^2\Gamma^{1;1}-2\omega\Omega \Gamma^{2;1} &=&\delta {\mathfrak{S}}^{11}+
\Omega^2\Gamma^{11}+\delta \Pi \,,\label{1e}\\
%
\omega^2\Gamma^{2;2}+2\omega\Omega\Gamma^{1;2}&=& \delta {\mathfrak{S}}^{22}+
\Omega^2\Gamma^{22}+\delta \Pi \,,\label{2e}\\
%
\omega^2\Gamma^{1;2}-2\omega\Omega \Gamma^{2;1}&=& \delta {\mathfrak{S}}^{12}+
\Omega^2\Gamma^{12} \,,\label{3e}\\
%
\omega^2\Gamma^{2;1}+2\omega\Omega\Gamma^{1;2}&=& \delta {\mathfrak{S}}^{21}+ \Omega^2\Gamma^{21}\label{4e}\end{aligned}$$ and, combined, become $$\begin{aligned}
\frac{1}{2}\omega^2\left(\Gamma^{11}-\Gamma^{22}\right) -2\omega\Omega \Gamma^{12}&=&\delta
{\mathfrak{S}}^{11}- \delta {\mathfrak{S}}^{22}+
\Omega^2 \left(\Gamma^{11}-\Gamma^{22}\right)\,,\label{eqv1}\\
%
\omega^2\Gamma^{12}+\omega\Omega\left(\Gamma^{11}-\Gamma^{22}\right) &=&2\delta {\mathfrak{S}}^{12}+
2\Omega^2\Gamma^{12} \,.\label{eqv2}
%\end{aligned}$$ In order to solve equations (\[eqv1\]), (\[eqv2\]), where the virial integrals are given by (\[vir11\]), (\[vir12\]), we have to compute the surface integrals $\delta\mathfrak S^{12}$ and $\delta\mathfrak S^{11}-\delta\mathfrak S^{22}$.
In the case of perturbations of the form (\[tor\]), they are (see Appendix \[toroidal\]) $$\begin{aligned}
\delta\mathfrak S^{12}&=& -\frac{T}{2}\beta\pi C_{D-2}\int_0^a
d\varpi\phi\varpi^4Q(\varpi)[f(\varpi)]^{D-3}\\
\delta\mathfrak S^{11}-\delta\mathfrak S^{22} &=&-T\alpha\pi C_{D-2}
\int_0^a d\varpi\phi\varpi^4Q(\varpi)
[f(\varpi)]^{D-3}\,.\end{aligned}$$ where $$Q=\frac{1}{\varpi ^2}\left (\frac{\phi'(\phi^2-2)}{(1+\phi^2)^{5/2}}
+\phi\frac{2+3\phi^2}{\varpi(1+\phi^2)^{3/2}}-\frac{D-3}{z}
\frac{\phi^2}{(1+\phi^2)^{3/2}} \right )\,. \label{QQ}$$ Therefore equations (\[eqv1\]), (\[eqv2\]) are equivalent to $$\begin{aligned}
\omega^2\alpha-2\omega\Omega\beta&=&-2{\mathfrak I}\alpha+
2\Omega^2\alpha\nonumber\\
\omega^2\beta+2\omega\Omega\alpha&=&-2{\mathfrak I}\beta+ 2\Omega^2 \beta\label{systv}\end{aligned}$$ where $${\mathfrak{I}}\equiv\frac{T(D-2)}{4\rho}
\frac{\int_0^a\phi\varpi^4Q(\varpi) [f(\varpi)]^{D-3}d\varpi}{
\int_0^a\varpi^3\left[f(\varpi)\right]^{D-2} d\varpi}\,. \label{jota}$$ Setting $\omega^2=-\sigma^2$, the algebraic system (\[systv\]) admits solutions $$\sigma=\Omega\pm\sqrt{2{\mathfrak I}-\Omega^2}\,.\label{freq}$$ A neutral mode of oscillation ($\sigma=0$) occurs at $\Omega^2={\mathfrak I}$, but (in the absence of dissipative mechanisms as stressed in [@chandra]) the drop is still stable in the vicinity of this point. If $$\Omega^2\le2{\mathfrak I}\,, \label{instcond}$$ then $\sigma$ is real and the perturbation is stable; if the rotation frequency of the drop is larger, then the drop is unstable against toroidal perturbations. To ascertain stability or instability, we thus have to use the results of section \[sec:eqcasea\] for the equilibrium figures of rotating drops and insert the numerically generated shapes $z=z(\varpi)$ in (\[jota\]) and (\[freq\]).
If we measure $\varpi$ and $z$ in units of $a$ and $\Omega$ in units of $\sqrt{\frac{8T}{\rho a^3}}$ we get $$\sigma=\sqrt{\Sigma_a} \pm\sqrt{2J-\Sigma_a}\,,\label{sigma}$$ where the dimensionless quantity $J$ is $$J\equiv\frac{(D-2)}{32} \frac{\int_0^a\phi\varpi^4Q(\varpi)
[f(\varpi)]^{D-3}d\varpi}{
\int_0^a\varpi^3\left[f(\varpi)\right]^{D-2} d\varpi}\,.$$ The results of solving (\[sigma\]) are displayed in Fig. \[fig:frediagram\] for $D=4$, which is a first step towards reconstructing the whole bifurcation diagram. In Fig. \[fig:frediagram\], $\sigma_1$ is (for real solutions) the largest of the two roots in (\[sigma\]) , i.e., as long as $2J>\Sigma_a$. For $2J<\Sigma_a$, we display only one root (the other root is the complex conjugate of this) by plotting the real and imaginary components of $\sigma$.
In three dimensions ($D=3$) it was found by Chandrasekhar [@chandra] that rotating drops exhibit a neutral mode of oscillation at $\Sigma=0.4587$. He also found that rotating drops become unstable for $\Sigma>0.84$. We find that for general $D$ the results are similar. For instance, in $D=4$ there is a neutral mode of oscillation at $\Sigma_a=0.46$, and above the critical $\Sigma_a=0.86$ the drops are unstable.
![The dependence on the quantity $\Sigma$ of the characteristic frequencies of the toroidal modes of a rotating four-dimensional drop in $D=4$. The diagram is very similar to the three dimensional one ($D=3$) presented in Figure 3 of Chandrasekhar [@chandra]. As a check on our numerical code, we verify that in the non-rotating limit our numerical results agree well with the analytical prediction $\sigma=\sqrt{D+1}/2$.[]{data-label="fig:frediagram"}](Fig8.eps){width="8" height="6"}
The shapes of the drops at these values of $\Sigma$ are displayed in Figure \[fig:eqfigsa\]. Results for higher $D$ are very similar. Our numerical method only allowed to probe dimensions up to $D=6$, for larger $D$ we loose accuracy and the results cannot be trusted.
\[sec:cylindersurface\] Stability of fluid cylinders held by surface tension: the Rayleigh-Plateau instability
==============================================================================================================
To end our discussion on fluids held by surface tension and also as a warm up exercise for the up-coming Sections, we will review the results of [@cardosodiasglrp], describing a simple capillary instability - the Rayleigh-Plateau instability [@rayleigh; @tomotika] - as a good model for the Gregory-Laflamme instability. Most calculations were omitted in [@cardosodiasglrp] so we present the full detailed computations here.
\[subsec:rp\] Rayleigh-Plateau instability in higher dimensions
---------------------------------------------------------------
We take an infinite non-rotating cylinder of radius $R_0$ held together by surface tension (therefore without self-gravity). Neglecting the external pressure, the unperturbed configuration is worked out easily to yield an internal pressure $P_0=(D-2)T/R_0$, where $D$ is the total number of spatial dimensions. We perturb slightly this configuration, according to $$r=R_0+\epsilon e^{\omega t+ikz}\,. \label{pertsurface}$$ The equations governing small perturbations are $$\begin{aligned}
P_0+\delta P&=&T{\rm div}\vec{n}\,,\\
\frac{\partial \bf u}{\partial t}&=&-{\rm grad}
\Upsilon \label{gradupssurface}\,,\\
{\rm div} {\bf u}&=&0\,,\label{divergenlesssurface}\\
\Upsilon &\equiv& \frac{\delta P}{\rho}\label{defupssurface}\,, \end{aligned}$$ where $P_0$ is the pressure in the undisturbed cylinder and ${\bf u}$ is the fluid velocity. For instance, $u_r$ is the radial component of the velocity vector. Equation (\[gradupssurface\]) yields $$\begin{aligned}
u_r&=&-\frac{1}{\omega}\frac{d}{dr}\Upsilon\,,\label{ursurface}\\
u_z&=&\frac{-ik}{\omega}\Upsilon\,.\label{uzsurface} \end{aligned}$$ Equation (\[divergenlesssurface\]) implies that $$\frac{du_r}{dr}+\frac{D-2}{r}u_r+\frac{du_z}{dz}=0 \,.$$ Substituting in this the relations (\[ursurface\]) and (\[uzsurface\]) we get $$\frac{d^2\Upsilon}{dr^2}+\frac{D-2}{r}\frac{d\Upsilon}{dr}-k^2
\Upsilon =0\,,$$ with (well-behaved at the origin) solution $$\Upsilon=C r^{3/2-D/2}I_{\alpha}(kr)\,, \label{Upsilonsurface}$$ where $\alpha=(D-3)/2$. By relation (\[ursurface\]) we get $$u_r=-C\frac{r^{3/2-D/2}}{\omega}\left ( \frac{3-D}{2r}
I_{\alpha}(kr) +kI'_{\alpha}(kr)\right ) \,,$$ where $I'_{\alpha}(x)$ stands for the derivative of the Bessel function with respect to the argument $x$. Now, for this velocity to be compatible with the radial perturbation (\[pertsurface\]) one must clearly have $$\epsilon \,e^{\omega t+ikz}=-C\frac{R_0^{3/2-D/2}}{\omega^2} \left
( \frac{3-D}{2R_0} I_{\alpha}(kR_0)+kI'_{\alpha}(kR_0)\right )
\,.\label{epsilonsurface}$$ With the displacement (\[pertsurface\]) we have to first order in $\epsilon$ $$\frac{\delta P}{\rho}=-\frac{T}{R_0^2\rho}\left
(D-2-(kR_0)^2\right) \epsilon e^{\omega t+ikz}\,.\label{deltaP}$$ Using (\[Upsilonsurface\]) and (\[epsilonsurface\]) together with (\[defupssurface\]), (\[deltaP\]) we finally find $$R_0^2\omega^2=\frac{T}{\rho}\frac{\left ( \frac{3-D}{2R_0}
I_{\alpha}(kR_0)+ kI'_{\alpha}(kR_0)\right )}{I_{\alpha}}\left (
D-2-(kR_0)^2 \right ) \,, \label{srefacerp}$$ where $I'_{\alpha}(x)$ stands for the derivative of the Bessel function with respect to the argument $x$. This is the sought relation yielding the characteristic frequencies of an infinite cylinder of fluid held by surface tension. Formula (\[srefacerp\]) constitutes a generalization of a classic result by Rayleigh [@rayleigh], and it predicts that for $kR_0<\sqrt{D-2}$ the cylinder is unstable against this kind of “beaded” perturbations. In Figure \[fig:insttimesurface\] we show $\omega$ as function of wavenumber $k$ for several space dimensions $D$ (actually what is shown in the plots are dimensionless quantities $\omega R_0$ and $kR_0$).
![The Rayleigh-Plateau instability of a hyper-cylinder held together by surface tension, in several dimensions. Here the effective density were chosen to match those of a higher-dimensional non-rotating black hole. The instability gets stronger as the spacetime dimensionality increases, and the critical wavenumber (for which $\omega=0$) increases with $D$.[]{data-label="fig:insttimesurface"}](Fig9.eps){width="8" height="6"}
In Section \[sec:cylindergravity\] we will discuss this instability in more detail, when we compare it to its self-gravitating counter-part.
Rayleigh-Plateau and Gregory-Laflamme
-------------------------------------
In General Relativity, there is an interesting class of black objects (objects with an event horizon) without spherical topology, the black branes [@horowitzstrominger]. A broad class of these objects are unstable against gravity, in a mechanism known as Gregory-Laflamme (GL) instability [@gl]. This is a beaded instability that makes any small perturbation with wavelength $\lambda$ of the order of, or larger than, the radius of the cylinder $R_0$ grow exponentially with time. It is a very robust long-wavelength instability. In [@cardosodiasglrp] it was proposed to mimic the GL instability with the Rayleigh-Plateau instability, based on the suggestive appearance of the second law of black hole mechanics which endows the horizon with an effective surface tension. It was found that all aspects of the GL instability could be accounted for by the Rayleigh-Plateau instability in fluid cylinders. In some sense, the effective surface tension can mimic all gravitational aspects of black strings. If this is indeed correct, we should be able to predict the effects of rotation and charge on the GL instability, as discussed in the following.
### \[sec:extrapolationsurface\]Effects of magnetic fields
Even though not pursued here, a generalization for charged and/or rotating cylinders is of interest. In any case, we can lean on the four-dimensional results ($D=3$ spatial dimensions), for which the charged and rotating cases have been worked out [@chandrabookstability; @interfacial]. A magnetic field has a stabilizing effect [@chandrabookstability], which is in agreement with the fully relativistic results for the Gregory-Laflamme instability, where it has been shown that adding charge to the black strings stabilizes it [@gl]. There is good evidence that extremal charged black strings are marginally stable, and indeed in the capillary case (cylinder held by surface tension) it is possible to completely stabilize the cylinder if the magnetic field has a certain strength [@chandrabookstability] (dependent on the conductivity of the fluid), so this may simulate the marginal stability of extremal black strings very well [@gl]. Nevertheless an extension of the calculations presented here to the case of magnetically charged cylinders with surface tension is needed.
### \[sec:extrapolationsurface2\] Effects of rotation
The effects of rotation on the capillary Rayleigh-Plateau instability can be discussed more intuitively. The centrifugal force, scaling with $r$, contributes with a de-stabilizing effect since there is an increased pressure under a crest but a reduced pressure under a trough. So in principle, the threshold wavelength should be smaller [@interfacial], i.e., the instability should get stronger when rotation is added. It is natural to expect that rotation will lift the degeneracy of the modes and that unstable modes will in general be non axi-symmetric. What can we expect for the general relativistic GL instability once rotation is added? A very naive reasoning is that since extremal charged black strings are marginally stable then perhaps the same happens with rotation, and so rotation has a stabilizing effect. This is incorrect as a simple argument shows. Consider a rotating black string in $D+1$ spacetime dimensions (with again $D$ being the number of spatial dimensions) of the form $$g={\rm Kerr}^{D} \times {\bf R}\,,$$ where ${\rm Kerr}^{D}$ is the metric of higher dimensional Myers-Perry rotating black hole in $D$ spacetime dimensions, with only one angular momentum parameter (case I). If we focus on $D>4$ the angular momentum of these black holes is unbounded and one can show [@emparanmyers] that for very large angular momentum $${\rm Kerr}^D \sim S_{chw}^{D-2}\times {\bf R^2}\,,{\rm for \,\,large\,\,rotation}$$ where $S_{chw}^{D-2}$ is the metric of a higher dimensional Schwarzschild black hole, often referred to as Tangherlini [@tangherlini] black holes. We thus finally get $${\bf g}=S_{chw}^{D-2} \times {\bf R^3}\,,$$ This is the metric of a [*non-rotating*]{} black membrane in $D+1$ spacetime dimensions extended along the $3$ dimensions. But this is precisely the geometry considered by Gregory and Laflamme [@gl], and so it is unstable. While this argument says little about the strength of the instability as function of rotation, it does predict that rotating branes are still unstable. Putting this and the fluid membrane argument together we predict that rotation makes the GL instability stronger. One can also use an entropy argument, similar to the one in [@mu-inpark], to show that rotating black strings should be unstable. Again, an extension of the calculations presented in Section \[subsec:rp\] to the case of rotating cylinders with surface tension is needed. The $D=3$ case, studied in [@surfacetensionrotating], indicates that other effects may come into play with rotation. For instance, other instabilities seem to manifest themselves for high enough rotation, so this is a particularly exciting topic to consider.
\[sec:cylindergravity\]Stability of self-gravitating fluid cylinders: the Dyson-Chandrasekhar-Fermi instability
===============================================================================================================
It is only appropriate that we start the study of self-gravitating fluids by examining the (self-gravitating) counterpart of the cylinder studied in the last Section \[sec:cylindersurface\].
A simple consistent solution can be found in $D$-dimensions, which corresponds to a static self-gravitating infinite cylinder made of fluid with a density $\rho$. For the moment we will not consider rotational effects (these will be discussed later), and we will focus on incompressible fluids. The stability of such a solution for the case of 3 spatial dimensions was studied for the first time by Chandrasekhar and Fermi [@chandrafermi] (see also [@chandrabookstability]), who showed that such a solution is unstable against cylindrically symmetric perturbations (those for which the cross-section is still a circle, with a $z-$dependent radius), and stable against all others. If the radius of the cylinder is $R_0$, this “varicose” instability (in Chandrasekhar’s [@chandrabookstability] terminology) sets in for wavelengths $\lambda>2\pi
R/1.0668$.
A similar phenomenon had however been studied already in 1892 by Dyson [@dyson], who investigated the stability of self-gravitating rings of fluid. In his terminology, the ring is stable against fluted (the ring remains symmetrical about its axis but the cross-section is deformed) and twisted (the cross-section remains circular but the circular axis of the ring is deformed) perturbations. He also showed that sufficiently thin rings are unstable against a third kind of perturbation, so-called beaded perturbations (those for which the circular axis is undisturbed, but the cross-section is a circle of variable radius). But these correspond precisely to Chandrasekhar-Fermi type of perturbation (with the added bonus that Dyson effectively shows that the instability goes away if the length of the cylinder or ring is less than the critical wavelength), and thus we shall refer to this classical fluid instability as the Dyson-Chandrasekhar-Fermi (DCF) instability.
A self-gravitating, non-rotating $D$-dimensional fluid cylinder
---------------------------------------------------------------
We will now generalize the DCF instability for an arbitrary number of dimensions. Consider an infinite cylinder of incompressible fluid, with density $\rho$ and radius $R$. To fix conventions we consider spacetime with $1$ (time)$+1$ (axis of cylinder)$+D-1=D+1$ total number of spacetime dimensions. In the following the coordinates $r\,,z$ have their usual meaning in polar coordinates. The gravitational potential ${\cal B}({\bf x})$ inside and exterior to the cylinder satisfies respectively $$\begin{aligned}
\nabla^2{\cal B}({\bf x})^i&=&-(D-2)C_D G \rho\,,\label{vi}\\
\nabla^2{\cal B}({\bf x})^e&=&0\,. \label{ve} \end{aligned}$$ Now, in $D$ spatial dimensions with cylindrical symmetry we have $$\nabla^2=\frac{1}{r^{D-2}}\frac{d}{dr}\left (r^{D-2}\frac{d}{dr} \right
)+\frac{d^2}{dz^2}\,.\label{laplacian}$$ Since we consider a cylinder with constant radius (along the axis in the $z$-direction), the last term drops off and we get the following solution (we demand that both the potential and its derivative be continuous throughout) $$\begin{aligned}
{\cal B}({\bf x})^i_0&=& -\frac{(D-2)C_D G\rho}{2(D-1)}r^2\,,\label{Vi}\\
{\cal B}({\bf x})^e_0&=&-\frac{(D-2)C_D}{2} G \rho R_0^2 \log r/R_0-
\frac{D-2}{4}C_D G \rho R_0^2\,, \quad
D=3\,.\\
{\cal B}({\bf x})^e_0&=&-\frac{(D-2)C_D G \rho R_0^2}{2(D-1)}
\left ( \frac{D-1}{(D-3)}-
\frac{2R_0^{D-3}}{(D-3)r^{D-3}}\right )\,, \quad D>3\,. \end{aligned}$$ Hydrostatic equilibrium demands that the pressure $P_0$ satisfies $$\frac{d}{dr}\left (\frac{P_0}{\rho}-{\cal B}({\bf x})\right )=0 \,,$$ which yields, upon requiring $P_0(r=R_0)=0$, $$P_0=-\frac{(D-2)C_D G_L \rho ^2}{2(D-1)}(r^2-R_0^2)\,.$$
Axisymmetric perturbations
--------------------------
![Black strings and fluid cylinders are unstable to perturbations on the extended dimension, i.e., along the axis of the cylinder. Ripples propagating along this axis grow exponentially with time for wavelengths of order of the radius of the cylinder.[]{data-label="fig:inst"}](Fig10.eps){width="8" height="3"}
Suppose now the system is subjected to a perturbation (see Fig. \[fig:inst\]) such that the boundary is described by $$r=R_0+\epsilon e^{\omega t+ikz}\,. \label{pert}$$ The equations governing small perturbations are $$\begin{aligned}
\nabla^2 \delta {\cal B}({\bf x})&=&0\,,\\
\frac{\partial \bf u}{\partial t}&=&-{\rm grad} \Upsilon \label{gradups}\,,\\
{\rm div} {\bf u}&=&0\,,\label{divergenless}\\
\Upsilon &\equiv& \frac{\delta P}{\rho}-\delta {\cal B}({\bf x})^i
\label{defups}\,. \end{aligned}$$ Here ${\bf u}$ stands for the velocity of the perturbed fluid. Let’s solve for the first of these equations. When solving for the gravitational potential, the last term in (\[laplacian\]) cannot be neglected. We find $$\begin{aligned}
{\cal B}({\bf x})^e&=&{\cal B}({\bf x})^e_0+A\epsilon e^{\omega t+ikz}
r^{3/2-D/2}K_{\alpha}(kr) \,,\\
{\cal B}({\bf x})^i&=&{\cal B}({\bf x})^i_0+B\epsilon e^{\omega t+ikz}
r^{3/2-D/2}I_{\alpha}(kr)\,, \end{aligned}$$ where $$\alpha=\frac{D-3}{2}\,,$$ and $K_{\alpha}$ and $I_{\alpha}$ are Bessel functions of argument $\alpha$. The constants $A$ and $B$ are determined by requiring continuity of the potential and its derivative on the boundary. By keeping only dominant terms in $\epsilon$ we find $$\begin{aligned}
A&=&(D-2)C_D G \rho R_0 I_{\alpha}(kR_0)\,,\\
B&=&(D-2)C_D G \rho R_0 K_{\alpha}(kR_0)\,. \end{aligned}$$ Therefore we have $$\delta {\cal B}({\bf x})^i=\epsilon (D-2)C_D G \rho R_0
K_{\alpha}(kR_0) e^{\omega t+ikz}r^{3/2-D/2}I_{\alpha}(kr)\,.$$ We also have $u \sim e^{i\omega t+ikz}$, therefore Eq. (\[gradups\]) yields $$\begin{aligned}
u_r&=&-\frac{1}{\omega}\frac{d}{dr}\Upsilon\,,\label{ur}\\
u_z&=&\frac{-ik}{\omega}\Upsilon\,.\label{uz} \end{aligned}$$ (Here $u_r$, for instance, is the radial component of the velocity vector). Equation (\[divergenless\]) implies that $$\frac{du_r}{dr}+\frac{D-2}{r}u_r+\frac{du_z}{dz}=0 \,.$$ Substituting the relations (\[ur\]) and (\[uz\]) we get $$\frac{d^2\Upsilon}{dr^2}+\frac{D-2}{r}\frac{d\Upsilon}{dr}-k^2\Upsilon
=0\,,$$ with (well-behaved at the origin) solution $$\Upsilon=C r^{3/2-D/2}I_{\alpha}(kr)\,. \label{Upsilon}$$ By relation (\[ur\]) we get $$u_r=-C\frac{r^{3/2-D/2}}{\omega}\left ( \frac{3-D}{2r}
I_{\alpha}(kr) +kI'_{\alpha}(kr)\right ) \,,$$ where $I'_{\alpha}(x)$ stands for the derivative of the Bessel function with respect to the argument $x$. Now, for this velocity to be compatible with the radial perturbation (\[pert\]) one must clearly have $$\epsilon \,e^{\omega t+ikz}=-C\frac{R_0^{3/2-D/2}}{\omega^2}
\left ( \frac{3-D}{2R_0}
I_{\alpha}(kR_0)+kI'_{\alpha}(kR_0)\right ) \,.\label{epsilon}$$ Our problem is completely specified once we impose the boundary condition that the pressure must vanish at the boundary, $$\delta P+\epsilon \frac{\partial P_0}{\partial r}=0\,,
\Rightarrow \rho \Upsilon+\rho \delta {\cal B}({\bf
x})^i+\epsilon \frac{\partial P_0}{\partial r}=0\,.$$ This gives us finally $$\frac{\omega^2}{(D-2)C_D G\rho}=\frac{(3-D)I_{\alpha}(kR_0)+
kR_0I'_{\alpha} (kR_0)}{2I_{\alpha}(kR_0)}\left
(I_{\alpha}(kR_0)K_{\alpha}(kR_0)-\frac{1}{D-1}\right )\,.
\label{dcfinstability}$$
![The DCF instability of a self-gravitating hyper-cylinder in several dimensions. Here the effective densities were chosen to match those of a higher-dimensional non-rotating black hole. The instability gets stronger as the spacetime dimensionality increases, and the critical wavenumber (for which $\omega=0$) increases with $D$.[]{data-label="fig:insttime"}](Fig11.eps){width="8" height="6"}
In Fig. \[fig:insttime\] we plot the dispersion relation $\omega(k)$ for several values of the dimensionality $D$, using Eq. (\[dcfinstability\]). When $D$ increases the maximum $\omega$ grows which means that the instability gets stronger with space dimensionality. The threshold wavenumber $k_c$, defined as the wavenumber for which $\omega=0$, also increases with $D$. In Table \[tab:thresholdmode\] we present the threshold wavenumber for selected values of $D$.
$D$ 4 5 6 7 8 9 49 99
---------------------- ------- ------ ------ ------ ------ ------ ------ ------
[DCF]{} 1.41 1.70 1.96 2.20 2.41 2.61 6.85 9.85
[Rayleigh-Plateau]{} 1.41 1.73 2.00 2.24 2.45 2.66 6.78 9.80
[Gregory-Laflamme]{} 0.876 1.27 1.58 1.85 2.09 2.30 6.72 9.75
-2mm
For large $D$ we can obtain an analytical expression for the threshold wavenumber. In fact we can expand the Bessel functions as [@stegun]: $$\begin{aligned}
I_{\alpha}(kR_0)&\sim& (kR_0/2)^{\alpha}\left (\frac{1}{\alpha!}+
\frac{(kR_0)^2}{4(\alpha+1)!}\right )\,,\\
K_{\alpha}(kR_0)&\sim& \frac{1}{2}(kR_0/2)^{-\alpha}\left (
(\alpha-1)!-\frac{(\alpha-2)!(kR_0)^2}{4}\right )\,. \end{aligned}$$ We get $$K_{\alpha}(kR_0)I_{\alpha}(kR_0)-\frac{1}{D-1} \sim
\frac{2}{D^2}-\frac{(kR_0)^2}{4\alpha^3}\,\Rightarrow\,\, kR_0 \sim
\sqrt{D}\,.$$ A comparison between these results and the ones in Section \[sec:cylindersurface\] is instructive. The Figs. \[fig:insttimesurface\] and \[fig:insttime\] show a remarkably similar behavior for both these instabilities. The threshold wavenumber (for which $\omega=0$), shown in Table \[tab:thresholdmode\] is also very similar to the corresponding one in both the Gregory-Laflamme and the Rayleigh-Plateau instability. Finally the $D \rightarrow \infty$ limit of the three instabilities is the same, $kR_0 \sim \sqrt{D}$, in the sense that the threshold wavenumber coincides in this limit.
DCF and Gregory-Laflamme
------------------------
In [@cardosodiasglrp], the authors proposed to mimic the GL instability with the Rayleigh-Plateau instability, based on the suggestive appearance of the second law of black hole mechanics which endows the horizon with an effective surface tension. What we now propose is that, while the fluid membrane analogy is extremely powerful (and perhaps more trustworthy when it comes to mimic black objects), the Newtonian classical counterpart of the GL instability is the DCF instability. The GL instability should be the limit of the DCF instability when the density is so large that the cylinder is in fact a black string. Not only are the details of the DCF instability very similar to the GL instability (for instance, compare Figure \[fig:insttime\] with Figure 1 of [@gl]) but for large $D$ we get exact agreement for the threshold wavenumber. In Table \[tab:thresholdmode\] we compare the threshold wavenumber for the three different types of instabilities (DCF, Rayleigh-Plateau and GL). The agreement is indeed very good, thus providing a convincing evidence for our conjectured analogy between DCF/Rayleigh-Plateau instabilities for fluids and GL instability for black strings.
It has been shown by Chandrasekhar and Fermi [@chandrafermi] (see also Chandrasekhar [@chandrabookstability]) that the DCF instability is rooted in the sign change of potential energy for these varicose or beaded perturbations. But this is precisely what happens for the capillary of Rayleigh-Plateau instability (see Section \[sec:cylindersurface\] and discussion in [@cardosodiasglrp]), which explains why both of them yield so similar results. This is another pleasant indication that these instabilities are related. Based on this agreement we can discuss what to expect for rotating and charged cylinders and discuss the significance of the results for the GL instability.
\[sec:extrapolation\]Effects of rotation and magnetic fields
------------------------------------------------------------
Even though not pursued here, a generalization for charged and/or rotating cylinders is of great interest. In any case, we can lean on the four-dimensional results ($D=3$ spatial dimensions), for which the charged [@chandrabookstability; @chandrafermi; @shivamoggi] and rotating cases [@karnik; @robe; @luyten] have been worked out.
It has been shown that a magnetic field has a [*stabilizing*]{} effect, reducing the threshold wavenumber, i.e., increasing the minimum wavelength at which instability occurs. This is in agreement with the fully relativistic results for the Gregory-Laflamme instability, where it has been shown that adding charge to the black string stabilizes it [@gl]. There is good evidence that the extremal case is marginally stable, and this could be perhaps the only real signature of an event horizon [^3], because in the horizonless, Newtonian case, “no magnetic field, however strong, can stabilize the cylinder for disturbances of all wavelengths: the gravitational instability of long waves will persist” [@chandrabookstability]. As we discussed in Section \[sec:cylindersurface\], a capillary model may be more accurate to describe extremal objects.
Focusing now on the effect of rotation, it has been shown [@karnik] (in three spatial dimensions) that rotation has a [*destabilizing*]{} effect. This agrees with the phenomenological arguments in [@cardosodiasglrp] for cylinders held by surface tension (see also [@interfacial]) of the capillary model. It also agrees with the arguments in Section \[sec:cylindersurface\] suggesting that rotating black strings are unstable and thus a solid conjecture is that the Gregory-Laflamme instability of black strings gets stronger if one adds rotation. Stronger here means the threshold wavenumber decreases (the threshold wavelength increases). An interesting extension of the results presented here would be to consider a $D$-dimensional rotating cylinder of fluid.
\[sec:torusselfgravitating\] Self-gravitating torus and black rings
===================================================================
Toroidal shapes are thought to be common configurations in the universe. In some of them these rings or tori are associated with a massive central object, such as Saturn’s rings, whereas in others the central object plays almost no role at all (see [@wong] and references therein). Toroidal configurations in full general relativity are also important (see [@ansorgringblackhole]), and they sometimes appear in the form of temporary black toroids, having an event horizon, in simulations of collapse of rotating matter [@shapiro].
The problem of a self-gravitating torus in Newtonian theory was investigated by many authors, the first studies belonging to Poincaré, Kowalewsky [@history] and Dyson [@dyson]. Dyson [@dyson] considers a self-gravitating ring of radius $R$ and thickness $2r_+$ (i.e. the cross-section of the ring is a circle with radius $r_+$). For small $r_+/R$ he finds a power series solution (the leading term had already been derived by Poincaré and Kowalewsky [@history]) to the equilibrium problem as $$\frac{\Omega^2}{\pi G_L\rho}= \frac{r_+^2}{R^2}\left ( \log\left
(\frac{8R}{r_+}\right )-\frac{5}{4} \right
)-\frac{r_+^4}{8R^4}\left (\log\left (\frac{8R}{r_+} \right )-
\frac{5}{12}\right )+{\cal O}(r_+/R)^5 \,,\quad D=3\,.$$ To generalize this result to arbitrary dimensions we first observe that the gravitational potential of a very large ring of radius $R$ can be obtained by using as gaussian surface a torus. We get, for $D>3$,
$${\cal B}({\bf x}) \sim -\frac{(D-2)C_DGM}{4\pi (D-3)C_{D-2}R
r_+^{D-3}}\,, \quad D>3\,,\label{Vring}$$
where we expressed everything in terms of the physically more significant quantity $M$, the total mass of the ring. For rings with a large radius $R$ the hydrostatic equilibrium condition can be written as [@dyson; @chandrabookellipsoidal] $\Omega^2R^2 \sim {\cal B}({\bf x})$. This gives us $$\Omega^2R^2 \sim \frac{(D-2)C_D M}{4\pi(D-3)C_{D-2}R \,r_+^{D-3}}
\label{equilibriumrings}$$ It seems plausible that this formula also describes general relativistic objects, as long as one uses an “equation of state” to deduce $r_+$ in terms of the other quantities. This is because we are considering large rings, without really appealing to the internal structure of the object. For large rings, the dominating force determining equilibrium should be well-described by Newtonian forces, even though general relativistic effects become important “locally” and therefore it is the latter that determine $r_+(R,M)$. To plug-in the information that we are dealing with a black object, we use the relation between $r_+$, $M$ and $R$. For a black ring, the mass per length should be given, to a good accuracy, by the mass of a $D-1$ Tangherlini black hole [@tangherlini]: $$(D-2)C_DG M \sim
4\pi(D-2)C_{D-2}R\,r_+^{D-3}\,. \label{massblackobjects}$$ Using (\[equilibriumrings\]) and re-expressing everything in terms of the angular momentum $J\sim MR^2\Omega$ we get $$\frac{J^2}{M^2R^2}\sim 1\,. \label{newtring}$$ We can compare this prediction against the exact general relativistic black ring solution [@Emparan:2001wn] in five spacetime dimensions ($D=4$). For large black rings, Eq. (\[grring\]), $\frac{J}{MR}=\frac{\sqrt{2}}{3}$, in very good agreement with the Newtonian prediction (\[newtring\]). Moreover these arguments also lead to conjecture the existence of black rings in higher dimensions. Again the case $D=3$ seems special for there are Dyson rings in this case (actually that’s the case where they were found), albeit with a peculiar logarithmic term. As we remarked at the end of Section \[sec:toroidalsurface\] there are no black rings in $D=3$ so this is an instance where the analogy breaks down. To end this section, we note that another accurate and simple model was put forward by Hovdebo and Myers [@Hovdebo:2006jy] to describe black rings in quasi-Newtonian terms, starting from the black string solution. Their conclusions are basically the same as ours, with the exception that instead of working with Newtonian forces, they extremize the entropy of the solution.
\[sec:MacLaurin\] Self-gravitating, rotating spheroids: the MacLaurin sequence
==============================================================================
To complete our discussion of self-gravitating objects, we consider the generalization to $D$-spatial dimensions of the simplest rotating objects: the MacLaurin sequence. Here we restrict ourselves to case of one non-vanishing angular momentum (case $I$ in the notation of Section \[sec:symm\]), corresponding to rotation in the plane $x^1-x^2$; the coordinates are then decomposed as (\[coordinatesa\]) $$x^i=(x^a,x^A)~;~~~a=1,2~;~~A=3,\dots,D\,.$$ We will follow the approach of Chandrasekhar [@chandrabookellipsoidal], based on the formalism of virial tensors. The special case in which the rotation is zero (i.e., a self-gravitating globe), is considered in Appendix \[app:gravitysphere\], where we compute the oscillation modes of a $D$-dimensional incompressible fluid sphere.
The MacLaurin sequence in three dimensions
------------------------------------------
Here we briefly review the classical calculation of the MacLaurin sequence due to Chandrasekhar. We refer to [@chandrabookellipsoidal] for further details.
Let us consider a matter distribution, inside the volume $V$, with density $\rho(\vec x)$. Its gravitational potential is $${\cal B}(\vec x)\equiv G\int_V\frac{\rho(\vec x')}{|\vec x-\vec
x'|}d^3x'\,.$$ We define the tensor $${\cal B}_{ij}(\vec x)\equiv G\int_V\rho(\vec x')\frac{(x^i-x^{\prime
i})(x^j-x^{\prime j}) }{|\vec x-\vec
x'|^3}d^3x'\,,$$ which satisfies ${\cal B}_{ii}={\cal B}$. We also consider the potential energy $$I\!{\cal B}\equiv-\frac{1}{2}\int_V\rho(\vec x){\cal B}(\vec x)d^3x\,,$$ and the tensor $$I\!{\cal B}_{ij}\equiv-\frac{1}{2}\int_V\rho(\vec x){\cal B}_{ij}(\vec
x)d^3x\,.$$ The following relation holds: $$I\!{\cal B}_{ij}=\int_V\rho\frac{\partial{\cal B}}{\partial
x^i}x^jd^3x\,.\label{relxpto}$$ Indeed, $$\begin{aligned}
&&\int_V\rho\frac{\partial{\cal B}}{\partial x^i}x^jd^3x=
-G\int_Vd^3x\int_Vd^3x'\rho(\vec x)\rho(\vec x') \frac{x^i-x^{\prime
i}}{|\vec x-\vec x'|^3}x^j= G\int_Vd^3x'\int_Vd^3x\rho(\vec
x')\rho(\vec x) \frac{x^i-x^{\prime i}}{|\vec x-\vec x'|^3}x^{\prime
j}\nonumber\\ &&=-\frac{1}{2}G\int_Vd^3x\int_Vd^3x'\rho(\vec x)\rho(\vec x')
\frac{(x^i-x^{\prime i})(x^j-x^{\prime j})}{|\vec x-\vec
x'|^3}\,.\end{aligned}$$ A consequence of this formula is that $I\!{\cal B}_{ij}$ is symmetric in its indices.
We then have the equations of motion $$\begin{aligned}
\ddot x^a&=&-\frac{\partial P}{\partial x^a}+\Omega^2\rho x^a+2
\rho\Omega\epsilon^{ab}\dot
x^b+\rho\frac{\partial {\cal B}}{\partial x^a}\nonumber\,,\\
\ddot z&=&-\frac{\partial P}{\partial z}+\rho\frac{\partial{\cal B}}{\partial z} \,,\end{aligned}$$ where $z=x^A=x^3$.
At equilibrium, $$\begin{aligned}
\Omega^2\rho x^a
-\frac{\partial P}{\partial x^a}+\rho\frac{\partial {\cal B}}{\partial x^a}&=&0\nonumber\,,\\
-\frac{\partial P}{\partial z}+\rho\frac{\partial{\cal B}}{\partial z}&=&0\,.\label{eqeq} \end{aligned}$$ We compute the virial integrals by multiplying equations (\[eqeq\]) by $x^j$ and integrating in the volume domain $V$. The resulting integrals are: $$\begin{aligned}
\Omega^2\int_Vd^3x\rho x^ax^j&=&\Omega^2I_{aj}\nonumber\,,\\
-\int_V d^3x\frac{\partial P}{\partial x^i}x^j&=&\delta^{ij}\int_V d^3xP
\equiv\delta_{ij}
\Pi\nonumber\,,\\
\int_Vd^3x\rho\frac{\partial{\cal B}}{\partial x^i}x^j&=&I\!{\cal B}_{ij}\,. \end{aligned}$$ Thus we get virial equations $$\begin{aligned}
\Omega^2I_{ab}+I\!{\cal B}_{ab}&=&-\Pi\delta_{ab}\nonumber\,,\\
\Omega^2I_{a3}+I\!{\cal B}_{a3}&=&0\nonumber\,,\\
I\!{\cal B}_{3a}&=&0~~~~~\hbox{(therefore $I\!{\cal B}_{a3}=I_{a3}=0$)}\nonumber\,,\\
I\!{\cal B}_{33}&=&-\Pi\,. \end{aligned}$$ The inertia tensor at equilibrium is then, in general, $$I_{ij}=\left(\begin{array}{ccc} I_{11} & I_{12} & 0 \\
I_{21} & I_{22} & 0 \\ 0 & 0 & I_{33} \\ \end{array} \right)\,.$$ We note that equilibrium does not require necessarily $I_{11}=I_{22}$.
From the virial equations we get $$I\!{\cal B}_{11}+\Omega^2I_{11}=I\!{\cal B}_{22}+\Omega^2I_{22}=
I\!{\cal B}_{33}\label{vir2}\,.$$ There are two possible solutions of equations (\[vir2\]). The first is the simplest one, with $$I_{11}=I_{22}\,,~~~~~I\!{\cal B}_{11}=I\!{\cal B}_{22}\,,$$ namely, with axial symmetry; the rotation frequency is given by $$\Omega^2=\frac{I\!{\cal B}_{33}-I\!{\cal B}_{11}}{I_{11}}\,.$$ and can be easily computed; this solution corresponds to the MacLaurin sequence of oblate axially symmetric ellipsoids. The second, more complicate, has $I_{11}\neq I_{22}$, and corresponds to the Jacobi sequence of tri-axial ellipsoids; this solution exists only if the total angular momentum exceeds a critical value.
The MacLaurin sequence in higher dimensions
-------------------------------------------
The generalization of the three dimensional formalism to higher dimensions is more or less straighforward. Let us consider a matter distribution in $D$ dimensions, inside the volume $V$, with density $\rho(\vec x)$. Again, we define the gravitational potential $${\cal B}(\vec x)\equiv G\int_V\frac{\rho(\vec x')}{|\vec x-\vec
x'|^{D-2}}
d^Dx'\,. \label{gravd}$$ and $${\cal B}_{ij}(\vec x)\equiv\int_V\rho(\vec x')\frac{(x^i-x^{\prime i}) (x^j-x^{\prime j}) }{|\vec x-\vec
x'|^D}d^Dx'\,,$$ which satisfies ${\cal B}_{ii}={\cal B}$. We also consider the potential energy $$I\!{\cal B}\equiv-\frac{1}{2}\int_V\rho(\vec x){\cal B}(\vec x)d^Dx\,.$$ and the tensor $$I\!{\cal B}_{ij}\equiv-\frac{1}{2}\int_V\rho(\vec x){\cal B}_{ij}(\vec
x)d^Dx\,.$$ We can check that the following relation, which is the extension of (\[relxpto\]), holds: $$I\!{\cal B}_{ij}=\frac{1}{D-2}\int_V\rho\frac{\partial{\cal
B}}{\partial x^i}x^jd^Dx\,.$$ Indeed, $$\begin{aligned}
&&\int_V\rho\frac{\partial{\cal B}}{\partial x^i}x^jd^Dx=
-(D-2)G\int_Vd^Dx\int_Vd^Dx'\rho(\vec x)\rho(\vec x')
\frac{x^i-x^{\prime i}}{|\vec x-\vec x'|^D}x^j=
(D-2)G\int_Vd^Dx'\int_Vd^Dx\rho(\vec x')\rho(\vec x)
\frac{x^i-x^{\prime i}}{|\vec x-\vec x'|^D}x^{\prime j}\nonumber\\
&&=-\frac{D-2}{2}G\int_Vd^Dx\int_Vd^Dx'\rho(\vec x)\rho(\vec x')
\frac{(x^i-x^{\prime i})(x^j-x^{\prime j})}{|\vec x-\vec x'|^D}\,.\end{aligned}$$ Thus $I\!{\cal B}_{ij}$ is symmetric in its indices, as it was in three dimensions.
The equations of motion yield $$\begin{aligned}
\ddot x^a&=&-\frac{\partial P}{\partial x^a}+\Omega^2\rho
x^a+2\rho\Omega\epsilon^{ab}
\dot x^b+\rho\frac{\partial {\cal B}}{\partial x^a}\nonumber\,,\\
\ddot x^A&=&-\frac{\partial P}{\partial x^A}+\rho\frac{\partial{\cal B}}{\partial x^A}\,.\end{aligned}$$ At equilibrium we get, $$\begin{aligned}
\Omega^2\rho x^a
-\frac{\partial P}{\partial x^a}+\rho\frac{\partial {\cal B}}{\partial x^a}&=&0\nonumber\,,\\
-\frac{\partial P}{\partial x^A}+\rho\frac{\partial{\cal B}}{\partial
x^A}&=&0\,.\label{eqeqd}\end{aligned}$$ We compute the virial integrals by multiplying equations (\[eqeqd\]) by $x^j$ and integrating in the volume domain $V$. The resulting integrals are: $$\begin{aligned}
\Omega^2\int_Vd^Dx\rho x^ax^j&=&\Omega^2I_{aj}\nonumber\,,\\
-\int_V d^Dx\frac{\partial P}{\partial x^i}x^j&=&\delta^{ij}\int_V d^DxP\equiv\delta_{ij}
\Pi\nonumber\,,\\
\int_Vd^Dx\rho\frac{\partial{\cal B}}{\partial x^i}x^j&=&(D-2)I\!{\cal B}_{ij}\,. \end{aligned}$$ Thus we get the virial equations $$\begin{aligned}
\Omega^2I_{ab}+(D-2)I\!{\cal B}_{ab}&=&-\Pi\delta_{ab}\nonumber\,,\\
\Omega^2I_{aA}+(D-2)I\!{\cal B}_{aA}&=&0\nonumber\,,\\
(D-2)I\!{\cal B}_{Aa}&=&0
~~~~~\hbox{(therefore $I\!{\cal B}_{aA}=I_{aA}=0$)}\nonumber\,,\\
(D-2)I\!{\cal B}_{AB}&=&-\Pi\delta_{AB}\,. \end{aligned}$$ The inertia tensor at equilibrium is then, in general, $$I_{ij}=\left(\begin{array}{cccc} I_{11} & I_{12} & 0 & \\
I_{21} & I_{22} & 0 & \\ 0 & 0 & I_{33} & \\
& & & \ddots \\ \end{array} \right)\,.$$ We note that equilibrium does not require necessarily $I_{11}=I_{22}$.
From the virial equations we get $$(D-2)I\!{\cal B}_{11}+\Omega^2I_{11}=(D-2)I\!{\cal
B}_{22}+\Omega^2I_{22}=
(D-2)I\!{\cal
B}_{33}=\dots=(D-2)I\!{\cal B}_{DD}\label{vir2d}\,.$$ The simplest solution of Eqs. (\[vir2d\]) is the one with axial symmetry: $$I_{11}=I_{22}\,,~~~~~I\!{\cal B}_{11}=I\!{\cal B}_{22}\,,,$$ which is the generalization of MacLaurin sequence; the rotation frequency is given by $$\Omega^2=(D-2)\frac{I\!{\cal B}_{33}-I\!{\cal B}_{11}}{I_{11}}\,,$$ and can be easily computed. But equations (\[vir2d\]) leave open the possibility to have other, non axisymmetric equilibrium configurations: the generalization of the Jacobi sequence. Their detailed study will have to wait for other, possible numerical, studies.
\[sec:conclusions\]Concluding remarks
=====================================
\[subsec:conclusions\] Equilibrium configurations in higher dimensions
----------------------------------------------------------------------
We have started the investigation of equilibrium figures of rotating fluids in higher dimensions. We used two appealingly simple models: one in which the fluid is held together by surface tension and the other in which self-gravity is the cohesive force. We have developed the basic tools to start an investigation of this important subject. Nevertheless, this work represents only a first step towards the understanding of the behavior of rotating fluids in higher dimensions. Many questions remain to be addressed. It is clear that the general behavior observed for $D=3$ still holds for higher $D$. It is natural to expect that (for instance for fluid drops), a non axisymmetric figure bifurcates from the point where the axisymmetric figure starts to be unstable, the point at which relation (\[instcond\]) saturates. What is the shape of this figure? Is the instability point really a bifurcation point? This would allow one to generalize Figure \[fig:twolobes\] to higher dimensions. Furthermore, our study must be extended to encompass more general perturbations and more general angular momentum configurations. Another pressing issue is to determine whether there is always a toroidal figure of equilibrium in general $D$ dimensions: could it be that for certain $D$ a drop [*always*]{} encloses the origin regardless of its rotation? If so, this could be an excellent model for ultra-rotating Myers-Perry black holes, for which the angular momentum is unbounded. Concerning self-gravitating spheroids, there are many more unanswered questions. The formalism to study self-gravitating fluid drops must be further developed to find the actual equilibrium shapes as well as their stability. This would lead to a generalization of the MacLaurin and Jacobi sequences to higher dimensions. It would also be exciting to study numerically self-gravitating toroidal configurations (the analogue of the Dyson rings).
\[subsec:fluidsblackholes\] Fluids and black objects
----------------------------------------------------
### \[subsubsec:mimic\] Mimicking event horizons with cohesive forces
The work presented here also allows one to better understand how gravity (even in the strong field regime) behaves in higher dimensions. Given the number of new solutions and instabilities discovered in this setting over the last few years, this is a particularly fascinating topic. A lesson to learn from the examples shown here (the Rayleigh-Plateau instability, the toroidal configurations, the DCF instability, the Dyson rings, etecetera) is that many attractive forces, like Newtonian gravity and surface tension, seem to mimic strong gravity situations very well. There is probably something more to it, when one shows that surface tension mimics [*so*]{} well the behavior of horizons, but this remains to be explained. A particularly promising proposal [@emparan], which gathers support from our work, is that the physics of fluids and black holes can be related semi-quantitatively for smooth enough objects. If the object is very inhomogeneous or is highly distorted, then the gravitational self-interaction between different parts of the object will be strong, and then one expects that the details of the (fully non-linear) dynamics of the interaction will matter. So, for very inhomogeneous objects, general relativity or fluid dynamics would yield very different results. This means that the fluid approach won’t, in principle, be useful for cases such as: the deeply non-linear evolution of the GL instability, when large inhomogeneities develop; the Myers-Perry black holes close to the extremal singularity (in the cases they admit a Kerr bound); thick black rings, and in general black rings in the region where non-uniqueness occurs. This is unfortunate since these are very intriguing and important regimes. Nevertheless, a plethora of important and smooth objects (like the ones studied here) should then be described by these very simple models, and could even be a starting point for more complex situations.
### \[subsec:bifdiagram\] The bifurcation diagram of black objects
In the Introduction, the suggestion is made that bifurcation diagram in Fig. \[fig:brown\] is in some sense universal, the qualitative behavior holding for most objects held together by some cohesive force. The natural generalization is to conjecture it may describe bifurcation diagrams of black objects. We would like to comment on the limitation of that generalization [^4].
[*Kerr bound*]{} The first point to address concerns the existence of a Kerr bound. If the bifurcation diagram Fig. \[fig:brown\] were indeed universal, the angular momentum of black holes would be bounded. In fact, the results in Section \[sec:eqshapesurface\] imply that there is a bound on $J$ in [*any*]{} of the cases I and II. We know however that the angular momentum of black holes in “case I” is unbounded (for $D>4$) [@myersperry; @emparanmyers], whereas in case II it is bounded [@myersperry] (see for instance Eq. 7 in [@Kunduri:2006qa]). Thus the surface tension model does not accurately describe ultra-spinning black holes, for the cases studied numerically ($D=3,4,5$). Accurate numerical results are needed for higher dimensions, as well as results for self-gravitating fluids. This would clear the issue.
[*Instabilities*]{} The second point concerns the instability of the solutions. The results available in the literature concerning black hole instability can be summarized as
\(i) For $D>4$, case I, there should be both axisymmetric and non-axisymmetric instabilities [@emparanmyers].
\(ii) For $D=4$ spatial dimensions, case I, the analysis of [@emparanmyers] does not apply; it was been suggested in [@Emparan:2001wn] that there could be an axisymmetric instability before the Kerr bound, ending in a black ring.
\(iii) For case II, the study in [@Kunduri:2006qa] suggests no instability.
The studies in [@emparanmyers] and [@Kunduri:2006qa] focused only on a special kind of perturbation, related to a traceless tensor mode. This seems to belong to the same representation of our “toroidal mode”, even if one relates to the metric and the other to the fluid.
In Section \[sec:stabilitysurface\] we found that a “case I” rotating drop is unstable. This agrees with the black hole picture (i) above. We were not able to extend the discussion to rotating drops in case II. So far both pictures (black hole and fluids) agree. The long-known discrepancy is the $D=3$ case: Kerr black holes in $D=3$ spatial dimensions are stable, whereas the fluid picture would predict instability. Another point of discrepancy is that the most general vacuum solution of Einstein’s equations in $3$ spatial dimensions is the Kerr family, with spherical topology, so black rings are absent in this case. We can thus argue that $D=3$ is a remarkable exception.
[*Non-axisymmetric figures of equilibrium*]{} The final point we would like to touch upon is the question of existence of non-axisymmetric figures of equilibrium. In the simple fluid models (see Figs. \[fig:brown\] and \[fig:bifgrav\]) there is a family of non-axisymmetric stationary figures; one of the main points of our paper was to show that this holds for fluids in general higher dimensions, too. We went on to find non-axisymmetric instability which we argued brings the drop to a non-axisymmetric stationary configuration (in the case of gravitating fluid, we proved that there exist stationary, non-axisymmetric configurations analogous to the Jacobi sequence). However, it is shown in [@Hollands:2006rj] that higher dimensional rotating black objects must be axisymmetric. Therefore stationary non-axisymmetric black solutions do not exist, and one would be led to think that the bifurcation diagram for fluids is completely misleading.
The explanation for the disagreement is very simple, and lies in the fact that one is expecting from the model something it was not built to do: a more realistic analogy would have to take emission of gravitational waves into account, never considered up to now. Since non-axisymmetric objects radiate gravitational waves, we would not expect to find stationary equilibrium solutions of that kind. Indeed, for three spatial dimensions this problem was considered by Chandrasekhar [@chandragw]. Chandrasekhar proved that when one takes emission of gravitational waves into account (all other “curved spacetime” effects being neglected), the Jacobi configuration is unstable and evolves in the direction of increasing angular velocity, approaching the bifurcation point. Furthermore, he proved that radiation reaction does not make the fluid unstable past the point of bifurcation. Thus the bifurcation diagram for black objects would be drawn by first deleting all the non-axisymmetric figures.
It is reasonable to think that this result, i.e. the non-existence of non-axisymmetric stationary fluid configurations in general relativity, can be generalized to higher dimensions. If this is the case, this would be the analogue of the no-go result of [@Hollands:2006rj] for higher dimensional rotating black holes.
\[subsec:worktodo\] Future work
-------------------------------
The two examples we have studied (surface tension and self-gravitating fluids) can therefore improve and strengthen our understanding of black objects. The fact that rotating fluid drops develop instabilities at finite (usually small) rotation seems to suggest that black holes cannot have large rotation. This was also put forward by Emparan and Myers [@emparanmyers] in connection with the Gregory-Laflamme instability. The examples shown here and the classical three-dimensional results discussed in the Introduction suggest that the final state of this instability is a non axi-symmetric figure. Some evidence was put forward already in [@emparanmyers] and this work adds some more evidence to that possibility. In fact, it might not even be possible to find matter configurations with high enough rotation to form these ultra-spinning black holes following collapse: we prove in Appendix \[app:poincareinequality\] that self-gravitating, rotating homogeneous fluids always have a bounded angular velocity (this is a generalization of a classic result by Poincaré). Future work along these lines (mimicking strong gravity situations with analogue models) could start by investigating the black string phase transition [@mu-inpark; @Gubser:2001ac; @Sorkin:2004qq] within either the surface tension or the self-gravity models. Another potentially interesting topic would be to investigate the collapse of toroidal fluid configurations to form black rings.
Acknowledgements {#acknowledgements .unnumbered}
================
We warmly thank Emanuele Berti, Marco Cavaglià, Óscar Dias, José Lemos and Mu-In Park for a careful and critical reading of the manuscript and for many useful comments. We are indebted to Roberto Emparan for very useful suggestions and specially for having enlightened us as to the significance of some of the results. We are also indebted to Dr. Claus-Justus Heine for very useful correspondence on this subject and for preparing Figure \[fig:twolobes\] for us. We thank Dr. Robert Brown for kind permission to use Figure \[fig:brown\] and Drs. Yoshiharu Eriguchi and Izumi Hachisu for kind permission to use Figure \[fig:bifgrav\]. VC acknowledges financial support from Fundação Calouste Gulbenkian through the Programa Gulbenkian de Estímulo à Investigação Científica.
\[app:surfacetensiondrop\] Oscillation modes of a spherical drop held together by surface tension
=================================================================================================
What are the oscillation modes of a spherical drop held together by surface tension? In three spatial dimensions the answer to this classical problem can be found for instance in [@lamb; @reid]. We will now generalize it to an arbitrary number of dimensions. Take a spherical drop with radius $R_0$. Define again $$\Upsilon \equiv \frac{\delta P}{\rho}\,,\label{upsilonsurface2}$$ where $\delta P$ is the change in pressure inside the drop (the outside pressure is taken to be zero), after it is perturbed by a displacement of the form $$r=R_0+\epsilon e^{\omega t}Y({\rm angles})\,, \label{pert2surface}$$ where $Y({\rm angles})$ are hyper-spherical harmonics, with eigenvalue $l(l+D-2)$. The equations governing small perturbations are $$\begin{aligned}
P_0+\delta P&=&T{\rm div}\vec{n}\label{46e} \,,\\ \frac{\partial
\bf u}{\partial t}&=&-{\rm grad} \Upsilon \label{gradups2surface}\,,\\
{\rm div} {\bf u}&=&0\,,\label{divergenless2surface}\\ \Upsilon
&\equiv& \frac{\delta P}{\rho} \label{defups2surface}\,. \end{aligned}$$ With the displacement (\[pert2surface\]), equation (\[46e\]) can be solved to yield (for details we refer the reader to Landau and Lifshitz [@reid]) near $r=R_0$ and to first order in $\epsilon$ $$\frac{\delta P}{\rho}=-\frac{T}{R_0^2\rho}(l-1)(l+D-1) \epsilon
e^{\omega t}Y({\rm angles})\,.\label{deltaP2surface}$$ Therefore $$\Upsilon =-\frac{T (l-1)(l+D-1)}{\rho R_0^{l+2}}r^l \epsilon
e^{\omega t}Y({\rm angles})\,.\label{aiai}$$ On the other hand, observing that $\nabla^2\left (\frac{\delta
P}{\rho}\right )=0$ we may immediately write $$\frac{\delta P}{\rho}=\epsilon F_p\frac{r^l}{R_0^l} Y({\rm
angles})e^{\omega t}\,,$$ with $F_p$ a constant to be determined. We thus have from the definition of $\Upsilon$ that $$\Upsilon =\frac{r^l}{R_0^l}\epsilon \, F_p\,Y({\rm
angles})e^{\omega t}\,.\label{upsxxx}$$ Equation (\[gradups2surface\]) then yields $$u_r=-\frac{1}{\omega}\frac{d}{dr}\Upsilon\,,\label{ur2surface}$$ Now, for this velocity to be compatible with the radial perturbation (\[pert2surface\]) one must have $$F_p=-\frac{R_0\omega^2}{l} \,. \label{epsilon2surface}$$ Inserting this in (\[upsxxx\]) and equating the resulting expression with (\[aiai\]) we get $$\omega^2=-\frac{T l(l-1)(l+D-1)}{\rho R_0^3} \,,$$ which is a generalization of the three-dimensional result found in [@lamb; @reid]. Note that the drop is always stable ($\omega$ is imaginary) and that $l=0$ and $1$ modes do not exist. The first corresponds to spherically symmetric pulsation of the drop, which violates the constant volume assumption and the second corresponds to a translation of the drop as whole.
\[app:gravitysphere\] Oscillation modes of a self-gravitating, non-rotating sphere
==================================================================================
Here we analyze the modes of a self-gravitating fluid sphere, in which case we can do without the virial formalism and find an exact analytical expression for the oscillation frequencies. We aim at generalizing the four-dimensional result by Kelvin and others [@lamb]. The analysis from Section \[sec:cylindergravity\] carries over with minimal changes.
The gravitational potential ${\cal B}({\bf x})$ inside and outside the globe satisfy respectively $$\begin{aligned}
\nabla^2{\cal B}({\bf x})^i&=&-(D-2)C_D G \rho\,,\label{visphere}\\
\nabla^2{\cal B}({\bf x})^e&=&0\,. \label{vesphere} \end{aligned}$$ In the unperturbed state we get $$\begin{aligned}
{\cal B}({\bf x})^i_0&=& -\frac{(D-2)C_D G\rho}{2D}\left (-
\frac{DR_0^2}{D-2}+r^2\right )\,,\label{Vspherei}\\
{\cal B}({\bf x})^e_0&=&\frac{C_D G\rho R_0^D}{D}
\frac{1}{r^{D-2}}\,.\end{aligned}$$ Hydrostatic equilibrium demands that the pressure $P_0$ satisfies $$\frac{d}{dr}\left (\frac{P_0}{\rho}-{\cal B}({\bf x})\right )=0
\,,$$ which yields, upon requiring $P_0(r=R_0)=0$, $$P_0=\frac{(D-2)C_D G \rho ^2}{2D}(R_0^2-r^2)\,.$$ Suppose now the system is subjected to a perturbation, of general form $$r=R_0+\epsilon e^{\omega t}Y({\rm angles})\,, \label{pert2}$$ where $Y({\rm angles})$ are hyper-spherical harmonics, with eigenvalue $l(l+D-2)$.
The equations governing small perturbations are $$\begin{aligned}
\nabla^2 \delta {\cal B}({\bf x})&=&0\,,\\ \frac{\partial \bf
u}{\partial t}&=&-{\rm grad} \Upsilon \label{gradups2}\,,\\ {\rm div}
{\bf u}&=&0\,,\label{divergenless2}\\ \Upsilon &\equiv& \frac{\delta
P}{\rho}-\delta {\cal B}({\bf x})^i
\label{defups2}\,. \end{aligned}$$ Here ${\bf u}$ stands for the velocity of the perturbed fluid. Let’s solve for the first of these equations. We find $$\begin{aligned}
{\cal B}({\bf x})^e&=&{\cal B}({\bf x})^e_0+\epsilon A r^{-l-D+2}
Y({\rm angles})e^{\omega t} \,,\\
{\cal B}({\bf x})^i&=&{\cal B}({\bf x})^i_0+\epsilon Br^{l}
Y({\rm angles})e^{\omega t}\,. \end{aligned}$$ The constants $A$ and $B$ are determined by requiring continuity of the potential and its derivative on the boundary. By keeping only dominant terms in $\epsilon$ we find $$\begin{aligned}
A&=&\frac{(D-2)C_D G \rho}{2l+D-2}R_0^{l+D-1}\,,\\
B&=&\frac{(D-2)C_D G \rho}{2l+D-2}R_0^{1-l}\,. \end{aligned}$$ Therefore we have $$\delta {\cal B}({\bf x})^i=\epsilon \frac{(D-2)C_D G \rho}{2l+
D-2}R_0^{1-l} r^{l}Y({\rm angles})e^{\omega t}\,.$$ Observing that $\nabla^2\left (\frac{\delta P}{\rho}\right )=0$ we may immediately write $$\frac{\delta P}{\rho}=\epsilon F_p\frac{r^l}{R_0^l} Y({\rm
angles})e^{\omega t}\,,$$ with $F_p$ a constant to be determined. We thus have from the definition of $\Upsilon$, equation (\[defups2\]), that $$\Upsilon \frac{r^l}{R_0^l}\epsilon \left (F_p-\frac{(D-2)C_D G
\rho R_0}{2l+D-2}\right )Y({\rm angles})e^{\omega t}\,.$$ Equation (\[gradups2\]) then yields
$$u_r=-\frac{1}{\omega}\frac{d}{dr}\Upsilon\,,\label{ur2}$$
Now, for this velocity to be compatible with the radial perturbation (\[pert2\]) one must have $$F_p=\frac{(D-2)C_D G \rho R_0}{2l+D-2}-\frac{R_0\omega^2}{l}
\,. \label{epsilon2}$$ Imposing a vanishing pressure on the surface of the globe we have finally $$\omega^2=-\frac{2(D-2)C_D G \rho}{D}\frac{l(l-1)}{2l+D-2}\,,
\label{oscglobe}$$ which is a generalization of Kelvin’s four-dimensional result [@lamb] to an arbitrary number of dimensions. As in the previous Section, the globe is always stable and $l=0$ and $l=1$ modes do not exist. Result (\[oscglobe\]) can be compared with the results for the modes of spherically symmetric black holes in higher dimensions [@modeshigherdim; @modeshigherdim2].
\[app:poincareinequality\] An upper bound for the rotation of an homogeneous fluid mass
=======================================================================================
In four dimensions there is an upper bound, derived by Poincaré [@poincaresurfacetension] for the maximum angular velocity of a rigidly rotating, homogeneous fluid mass in equilibrium. Consider a region $V$ in three dimensional space bounded by a surface $S$. Let $\frac{\partial}{\partial n_e}$ denote differentiation of a function across $S$ in the direction of the unit normal $n_e$ outward from $V$. The divergence theorem states that for any (sufficiently smooth) vector field $\vec{W}$ $$\int_S\vec{W}.\vec{n}\,d\sigma=\int_V {\rm div}\vec{W}d^3x\,.$$ If we choose $\vec{W}={\rm grad}{\cal B}$, where ${\cal B}$ is the gravitational potential, we get $$\int_S\frac{\partial {\cal B}}{\partial n_e}\,d\sigma=\int_V \nabla^2{\cal B}d^3x=-\int_V4\pi G \rho
d^3x\,,$$ where we used Poisson’s equation in the last equality.
A necessary (but not sufficient) condition for equilibrium is that the force at every point in $S$ be directed inward. But the total force is $F_r=\frac{\partial U}{\partial n_e}$, where $U$ is the modified potential $${\cal U}={\cal B}+\frac{1}{2}\Omega^2\left ((x^1)^2+(x^2)^2\right )$$ (we choose $x^3=z$ to be the axis of rotation). We then must require $$\frac{\partial{\cal U}}{\partial n_e}<0\,\,\Rightarrow \int_S \frac{\partial {\cal U}}{\partial
n_e}d\sigma=\int_V \nabla^2 U d^3x =\int_V\left(-4\pi G\rho+2\Omega^2\right)d^3x <0\,.$$ thus we have the Poincaré’s inequality [^5] $$\Omega^2<2\pi \rho G\,.$$ This derivation can be easily generalized to $D>3$ dimensional Euclidean space. If we assume the gravitational potential is (\[gravd\]) $${\cal B}({\bf x})=G\int d^Dx'\frac{\rho({\bf x'})}{|{\bf x}-
{\bf x'}|^{D-2}}\,,$$ then Poisson’s equation becomes $$\nabla^2{\cal B}=-(D-2)C_DG\rho \,,\label{glaplacian}$$ where $$C_D=\frac{2\pi^{D/2}}{\Gamma\left(\frac{D}{2}\right)} \,,
\label{areasphere}$$ and $C_Dr^{D-1}$ is the area of the $(D-1)$-sphere. Indeed, $$\int_V\nabla^2\left(\frac{1}{r^{D-2}}\right)d^Dx= \int_S
\frac{\partial}{\partial
n_e}\left(\frac{1}{r^{D-2}}\right)d\sigma =-(D-2)\int_S\frac{x^i}{r^D}
\frac{x^i}{r}d\sigma=-(D-2)C_D\,.$$ We now consider the two cases labeled by I and II, that is,
- one non-vanishing angular momentum;
- all angular momenta coincident, and $D$ odd.
The modified potential, in both cases, writes as $${\cal U}={\cal B}+\frac{1}{2}\Omega^2\sum_a(x^a)^2 \,,$$ where $a=1,2$ in case I, $a=1,\dots,D-1$ in case II. Its Laplacian is $$\begin{aligned}
\nabla^2 {\cal U}&=&-C_D\rho G+2\Omega^2~~~~~~~~~~~~\,\hbox{in case I}\nonumber\,,\\
\nabla^2 {\cal U}&=&-C_D\rho G+(D-1)\Omega^2~~~~~\hbox{in case II}\,, \end{aligned}$$ therefore Poincaré’s inequality becomes $$\begin{aligned}
\Omega^2&<&\frac{C_D}{2}\rho G~~~~~\hbox{in case I}\nonumber\,,\\
\Omega^2&<&\frac{C_D}{D-1}\rho G~~~~~\hbox{in case II}\,. \end{aligned}$$ For angular velocities larger than this threshold, the gravitational pull is insufficient to balance the centrifugal force, and there body is disrupted.
Toroidal perturbations of a rotating drop {#toroidal}
=========================================
Here we compute some useful quantities in the case of a rotating drop with one non-vanishing angular momentum, constant density $\rho$ and with toroidal perturbations (following the denomination of Chandrasekhar [@chandra]), i.e. $$\begin{aligned}
\xi^a&=&T^{ab}x^b\nonumber\,,\\
\xi^A&=&0\nonumber\,,\\
T^{ab}&=&\left(\begin{array}{cc}\alpha&\beta\\\beta&-\alpha\,,\\
\end{array}\right)\,,\label{tor1}\end{aligned}$$ with $\alpha,\beta$ constant parameters. We remind that, in the case with one non-vanishing angular momentum, the coordinates $\{x^i\}$ are decomposed as in (\[coordinatesa\]), (\[deftoz1\]) $$x^i=(x^a,x^A)~;~~~~a=1,2~;~~A=3,\dots,D$$ and we have defined $$\begin{aligned}
\varpi&=&\sqrt{(x^1)^2+(x^2)^2}\nonumber\,,\\
z&=&\sqrt{(x^3)^2+\dots+(x^D)^2}\,,\end{aligned}$$ so that $$\begin{aligned}
x^1&=&\varpi\cos\theta\nonumber\,,\\
x^2&=&\varpi\sin\theta\,.\end{aligned}$$ The $D$-dimensional volume integration element is $$d^Dx=\varpi z^{D-3}d\varpi d\theta dzd\Omega_{D-3}\,.$$ If we integrate on the angular variables, we get $$\int_{S^1\times S^{D-3}}d^Dx=2\pi C_{D-2} \varpi z^{D-3}d\varpi dz\,,$$ where $C_{D-2}$ is the area of a unit $(D-3)$-sphere (\[areasphere\]).
Let us now consider the $(D-1)$-dimensional surface $S$ of equation $z= f(\varpi)$. Since on $S$, $dz=\phi
d\varpi$, the metric element on $S$ is $$d\sigma^2=(1+\phi)^2d\varpi^2+\varpi^2d\theta^2+
[f(\varpi)]^2d\Omega_{D-3}^2\,.$$ The surface integration element is then $$dS=\sqrt{1+\phi^2}\varpi [f(\varpi)]^{D-3}
d\varpi d\theta d\Omega_{D-3}\,.$$ In particular, being $$n^a=-\frac{\phi x^a}{\sqrt{1+\phi^2}\varpi}\,,$$ we have $$dS^a=n^adS=-\phi x^a[f(\varpi)]^{D-3}d\varpi d\theta d\Omega_{D-3}\,.$$ After integration on the $(D-3)$-sphere we get, $$\int_{S^{D-3}}dS^a=-\phi x^a[f(\varpi)]^{D-3}C_{D-2}d\varpi d\theta\,.
\label{dSa}$$ Notice that in the $D=3$ case we have $dS^a=-2\phi x^ad\varpi d\theta$ where the factor $2$ accounts for the two hemispheres of the drop.
Virial integrals
----------------
We can easily compute the virial integrals $$\Gamma^{i;j}=\int_V\rho\xi^ix^jd^Dx\,.$$ From (\[tor1\]) it follows $\Gamma^{A;j}=0$; furthermore, $$\int_V x^ix^jd^Dx=0~~~~{\rm if}~i\neq j$$ because the drop is symmetric with respect to reflections to any axis, therefore $\Gamma^{a,A}=0$.
We have $$\begin{aligned}
\int_V(x^1)^2d^Dx&=&\left(\int_0^{2\pi}d\theta\cos^2\theta\right) C_{D-2}~
\int_0^ad\varpi\varpi^3~\int_0^{f(\varpi)}
dzz^{D-3}\nonumber\,,\\
&=&\pi\frac{C_{D-2}}{D-2}\int_0^ad\varpi\varpi^3 \left[f(\varpi)\right]^{D-2}=\int_V(x^2)^2d^Dx\,,\end{aligned}$$ where $\varpi=a$ is the intersection of the drop with the $x^1-x^2$ plane (namely, $f(a)=0$); therefore, $$\begin{aligned}
\Gamma^{1;1}&=&-\Gamma^{2;2}=\pi\rho\alpha\frac{C_{D-2}}{D-2}
\int_0^a\varpi^3\left[f(\varpi)\right]^{D-2}d\varpi\,,\\
\Gamma^{1;2}&=&\Gamma^{2;1}=\pi\rho\beta\frac{C_{D-2}}{D-2}
\int_0^a\varpi^3\left[f(\varpi)\right]^{D-2}d\varpi\,.\end{aligned}$$ We notice that these formulae are valid for every $D\ge 3$. When $D>3$, $C_{D-2}$ represents the area of the unit $(D-3)$-sphere; when $D=3$, the factor $C_1=2$ accounts for the two hemispheres composing the drop.
Surface energy integrals
------------------------
Here we want to compute the quantities $\delta\mathfrak S^{12}$ and $\delta\mathfrak S^{11}-\delta\mathfrak S^{22}$ where (see (\[eqkS\])) $$\delta\mathfrak S^{ab}=-T\left[\int_S\xi^bn^k_{~,k}dS^a+\int_Sx^b \delta(n^k_{~,k})dS^a
-\int_Sx^bn^k_{~,k}\frac{\partial \xi^c}{\partial x^a}dS^c\right]\,,\label{eqkS1}$$ and $$dS^a=-\phi x^a[f(\varpi)]^{D-3}C_{D-2}d\varpi d\theta\,.$$ Being $\xi^a=T^{ab}x^b$, we have that $$\begin{aligned}
&&\int d\theta\left[\xi^2x^1-x^2\frac{\partial\xi^c}{\partial x^1}x^c
\right]=\int_0^{2\pi} d\theta[T^{2a}x^ax^1-T^{1a}x^2x^a]=
\int_0^{2\pi} d\theta\left((x^1)^2-(x^2)^2\right)\beta=0\nonumber\,,\\
&&\nonumber\\
&&\int_0^{2\pi} d\theta\left[\xi^1x^1-\xi^2x^2-x^1\frac{\partial\xi^c}{\partial x^1}x^c +x^2\frac{\partial\xi^c}{\partial
x^2}x^c\right] =\int_0^{2\pi} d\theta\left[T^{1a}x^ax^1-T^{2a}x^ax^2
-x^1T^{c1}x^c+x^2T^{c2}x^c\right]=0\nonumber\,.\end{aligned}$$ therefore the first and third integrals in (\[eqkS1\]) cancel for the components $\delta\mathfrak S^{12}$ and $\delta\mathfrak S^{11}-\delta\mathfrak S^{22}$: $$\begin{aligned}
\delta\mathfrak S^{12}&=&-T\int_S\delta(n^k_{~,k})x^2dS^1\nonumber\,,\\
\delta\mathfrak S^{11}-\delta\mathfrak S^{22}
&=&-T\int_S\delta(n^k_{~,k})x^1dS^1-T\int_S\delta(n^k_{~,k})x^2dS^2\,.\end{aligned}$$ The perturbation of the divergence $\delta(n^i_{~,i})$ is given by (\[expdiv\]) $$\begin{aligned}
\delta(n^i_{~,i})&=&(\delta n^i)_{,i}-\frac{\partial n^i}{\partial x^k}
\frac{\partial \xi^k}{\partial x^i}=(\delta n^a)_{,a}-T^{ab}n^a_{~,b}
+(\delta n^A)_{,A}\nonumber\,,\\
\delta n^a&=&\left(n^kn^l\frac{\partial \xi^k}{\partial x^l}\right)n^a
-n^j\frac{\partial \xi^j}{\partial x^a}=[n^bn^cT^{bc}]n^a-T^{ab}n^b\nonumber\,,\\
\delta n^A&=&\left(n^kn^l\frac{\partial \xi^k}{\partial x^l}\right)n^A =[n^bn^cT^{bc}]n^A\,.\end{aligned}$$ with $$n^a=-\phi\frac{x^a}{\varpi\sqrt{1+\phi^2}}\,,~~~~~n^A=
\frac{x^A}{z\sqrt{1+\phi^2}}\,.$$ We have $$n^a_{~,b}=-\frac{\phi'}{(1+\phi^2)^{3/2}}\frac{x^ax^b}{\varpi^2}
+\frac{\phi}{\sqrt{1+\phi^2}}\frac{x^ax^b}{\varpi^3}-\delta^{ab} \frac{\phi}{\sqrt{1+\phi^2}\varpi}\,,$$ then we find $$\delta\left (n^i_{~,i}\right )=-\frac{x^ax^bT^{ab}}{\varpi^2}
\left[\frac{\phi'(\phi^2-2)}{(1+\phi^2)^{5/2}}+\phi\frac{2
+3\phi^2}{\varpi(1+\phi^2)^{3/2}}-\frac{D-3}{z}
\frac{\phi^2}{(1+\phi^2)^{3/2}} \right ]\,.$$ Therefore, $$\delta\left (n^i_{~,i}\right )=-\left [\alpha \left ((x^1)^2-(x^2)^2 \right )+2\beta x^1 x^2\right ]Q(\varpi)
=-\varpi^2\left(\alpha\cos 2\theta+\beta\sin 2\theta\right) Q(\varpi)\,, \label{deltadiv1}$$ where $$Q=\frac{1}{\varpi ^2}\left
(\frac{\phi'(\phi^2-2)}{(1+\phi^2)^{5/2}}+\phi\frac{2
+3\phi^2}{\varpi(1+\phi^2)^{3/2}}-\frac{D-3}{z}
\frac{\phi^2}{(1+\phi^2)^{3/2}} \right )\,.
\label{QQ1}$$ The surface integrals $\delta\mathfrak S^{12}$, $\delta\mathfrak
S^{11}-\delta\mathfrak S^{22}$ are then $$\begin{aligned}
\delta\mathfrak S^{12}&=&TC_{D-2}\int_0^a d\varpi\int_0^{2\pi}
d\theta\phi x^1x^2[f(\varpi)]^{D-3}\delta
(n^k_{~,k})\nonumber\,,\\
&=&-\frac{T}{2}\beta\pi C_{D-2}\int_0^a d\varpi
\phi\varpi^4Q(\varpi)[f(\varpi)]^{D-3}\,,\\
\delta\mathfrak S^{11}-\delta\mathfrak S^{22}
&=&TC_{D-2}\int_0^a d\varpi\int_0^{2\pi} d\theta
\phi\left[(x^1)^2-(x^2)^2\right][f(\varpi)]^{D-3}\delta(n^k_{~,k})
\nonumber\,,\\
&=&-T\alpha\pi C_{D-2}\int_0^a
d\varpi\phi\varpi^4Q(\varpi)[f(\varpi)]^{D-3}\,.\end{aligned}$$
\[app:newton\] Newton’s constant
================================
Here we would like to relate our adopted definition of Newton’s constant to the one adopted in the by now classical work by Myers and Perry [@myersperry]. They define a Newton’s constant $G_{MP}$ such that Einstein’s equations read $$R_{\mu \nu}-\frac{1}{2}g_{\mu \nu}R=8\pi G_{MP}T_{\mu \nu}\,.$$ We will now derive “Newton’s law” from this equation, in order to relate the different Newton constants. To derive Newton’s law $a=-\nabla {\cal B}({\bf x})$, we consider the Newtonian limit of Einstein’s equation (slowly moving particles, weak and static gravitational fields).
A test particle follows the geodesics of the spacetime $$\frac{d^2x^{\mu}}{d\tau ^2}+\Gamma^{\mu}_{\rho
\sigma}\frac{dx^{\rho}}{d\tau}\frac{dx^{\sigma}}{d\tau} =0\,.$$ Under the “moving slowly” condition $\frac{dx^i}{d\tau}\ll
\frac{dt}{d\tau}$ this is simplified to $$\frac{d^2x^{\mu}}{d\tau ^2}+\Gamma^{\mu}_{00}\left
(\frac{dt}{d\tau}\right )^2=0\,.$$ Due to the staticity assumption, $$\Gamma^{\mu}_{00}=\frac{1}{2}g^{\mu \lambda}\left
(\partial_0g_{\lambda 0}+\partial_0g_{0
\lambda}-\partial_{\lambda}g_{0 0}\right )=-\frac{1}{2}g^{\mu
\lambda}\partial_{\lambda}g_{0 0}\,,$$ so that the geodesic equation simplifies to $$\frac{d^2x^{\mu}}{d\tau ^2}-\frac{1}{2}g^{\mu
\lambda}\partial_{\lambda}g_{0 0}\left (\frac{dt}{d\tau}\right
)^2=0\,.$$ Lastly, the Newtonian limit implies weak gravitational fields, so we set $$g_{\mu \nu}=\eta_{\mu \nu}+h_{\mu \nu}\,,\quad |h_{\mu \nu}|\ll
1\,.$$ Thus we have (we can show that to first order $g^{\mu \nu}=\eta^{\mu
\nu}-h^{\mu \nu}$) $$\Gamma^{\mu}_{00}=-\frac{1}{2}\eta^{\mu
\lambda}\partial_{\lambda}h_{0 0}\,,$$ and therefore $$\frac{d^2x^{\mu}}{d\tau ^2}=\frac{1}{2}\eta^{\mu
\lambda}\partial_{\lambda}h_{0 0}\left (\frac{dt}{d\tau}\right
)^2\,.$$ Dividing the spatial components of this equation by $\left (\frac{dt}{d\tau}\right )^2$ we have finally $$\frac{d^2x^i}{dt^2}=\frac{1}{2}\partial_ih_{00} \,,$$ which allows one to identify $h_{00}=-2{\cal B}({\bf x})$. On the other hand, from Myers and Perry’s equation (1.5), we get $$\nabla^2 h_{00}=-16\pi\frac{D-2}{D-1}G_{MP}\label{gmp}\,.$$ Comparing our (\[glaplacian\]) with (\[gmp\]), we get the relation between our adopted $G$ and $G_{MP}$ as (there’s an extra minus sign that follows from the conventions adopted here for the relation force-potential) $$C_D G=\frac{8\pi}{D-1} G_{MP} \,.$$
[99]{} J. Plateau, [*Statique Expérimentale et Théorique des Liquides Soumis aux Seules Forces Moléculaires*]{}, (Paris, Gauthier-Villars, 1873).
C. P. Lee, A. V. Anilkumar, A. B. Hmelo and T. G. Wang, “Equilibrium of liquid drops under the effects of rotation and acoustic flattening: results from USML-2 experiments in space,” J. Fluid Mech. [**354**]{}, 43 (1998).
P. Aussillous and D. Quéré, “Liquid marbles,” Nature [**411**]{}, 925 (2001); idem, “Shapes of rolling liquid drops,” J. Fluid. Mech. [**512**]{}, 133 (2004); idem, “Non-stick droplets,” Chem. Eng. Technol. [**25**]{}, 9 (2002).
H. Poincaré, “Sur l’équilibre d’une masse fluide animée d’un mouvement de rotation,” Acta Mathematica [**7**]{}, 259 (1885).
S. Chandrasekhar, “The stability of a rotating liquid drop,” Proc. Roy. Soc. Ser. A [**286**]{}, 1 (1965).
R. A. Brown and L. E. Scriven, “The shape and stability of rotating liquid drops,” Proc. Roy. Soc. London Ser. A [**371**]{}, 331 (1980).
R. A. Brown and L. E. Scriven, “New class of asymmetric shapes of rotating liquid drops,” Phys. Rev. Lett. [**45**]{}, 180 (1980).
C. -J. Heine, “Computations of form and stability of rotating drops with finite elements,” IMA J. Numer. Anal. [**26**]{}, 723 (2006).
R. Gulliver, “Tori of prescribed mean curvature and the rotating drop,” in E.H.A. Gonzales, I. Tamanini, ed.: [*Variational methods for equilibrium problems of fluids*]{} (Meeting held in Trento, Italy, from June 20 to 25, 1983), Asterisque [**118**]{}, 167 (1984).
C.-J. Heine, [*Computations of form and stability of rotating drops with finite elements*]{}, PhD thesis, available at http://www.mathematik.uni-freiburg.de/IAM/homepages/claus/
A very complete account of the topic up to the late 19 century can be found in I. Todhunter, [*History of the mathematical theories of attraction and the figure of the earth*]{} (Dover, New York, 1962). Particular important works include H. Poincaré, [*Les méthodes nouvelles de la méchanique céleste*]{} (Gauthier-Villars, Paris, 1892, 1894, 1899); “Sur la stabilité de l’anneau de Saturne,” C. R. Acad. Sci. [**100**]{}, 346 (1885); idem, Bull. Astr. [**2**]{}, 109 (1885); ibidem [**2**]{}, 405 (1885); S. Kowalewsky, “Zusaetze und Bemerkungen zu Laplace’s Untersuchungen ueber die Gestalt der Saturnringe,” Astron. Nachr. [**111**]{}, 37 (1895); F. Tisserand, [*Traité de Méchanique Céleste*]{} (Gauthier-Villars, Paris, 1894).
S. Chandrasekhar, [*Ellipsoidal figures of equilibrium*]{} (Dover, New York, 1987); S. Chandrasekhar, [*Selected Papers, Volume 4*]{} (The University of Chicago Press, Chicago, 1989).
A. Balaguera-Antolinez, D. F. Mota and M. Nowakowski, “Ellipsoidal configurations in the de Sitter spacetime,” Class. Quantum Grav. [**23**]{}, 4497 (2006).
Y. Eriguchi and I. Hachisu, “New equilibrium sequences bifurcating from MacLaurin sequence,” Prog. Theor. Phys. [**67**]{}, 844 (1982); Y. Eriguchi, I. Hachisu and D. Sugimoto, “Dumb-bell-shape equilibria and mass-shedding pear-shape of selfgravitating incompressible fluid,” Prog. Theor. Phys. [**67**]{}, 1068 (1982);
J. M. Bardeen, “A reexamination of the post-newtonian MacLaurin spheroids,” Astrophys. J. [**167**]{}, 425 (1971).
C-Y Wong, “Toroidal figures of equilibrium,” Astrophys. J. [**190**]{}, 675 (1974).
Y. Eriguchi and D. Sugimoto, “Another equilibrium sequence of self-gravitating and rotating incompressible fluid,” Prog. Theor. Phys. [**65**]{}, 1870 (1981).
F. W. Dyson, “The potential of an anchor ring,” Phil. Trans. Roy. Soc. [**184**]{}, 43 (1893); F. W. Dyson, “The potential of an anchor ring. Part II,” Phil. Trans. Roy. Soc. [**184A**]{}, 1041 (1893).
M. Ansorg, A. Kleinwächter and R. Meinel, “Uniformly rotating asisymmetric fluid configurations bifurcating from highly flattened MacLaurin spheroids,” Mon. Not. Roy. Astron. Soc. [**339**]{}, 515 (2003);
N. Bohr and J. A. Wheeler, “The mechanism of nuclear fission,” Phys. Rev. [**56**]{}, 426 (1939). See also C. Y. Wong, “Toroidal and spherical bubble nuclei,” Annals of Physics (NY) [**77**]{}, 279 (1973).
W. J. Swiatecki, “The rotating charged or gravitating liquid drop and problems in nuclear physics and astronomy,” in [*Proceedings of the International Colloquium on Drops and Bubbles*]{} (ed. D. J. Collins, M. S. Plesset and M. M. Saffren, 1974), Jet Propulsion Laboratory, Pasadena California; S. Cohen, F. Plasil and W. J. Swiatecki, “Equilibrium configurations of rotating charged or gravitating liquid masses with surface tension. II.,” Annals of Physics (NY) [**82**]{}, 557 (1974).
V. Cardoso and O. J. C. Dias, “Rayleigh-Plateau and Gregory-Laflamme instabilities of black strings,” Phys. Rev. Lett. [**96**]{}, 181601 (2006).
K. S. Thorne, D. A. MacDonald and R. H. Price, [*Black Holes: The Membrane Paradigm*]{} (Yale University Press, 1986).
J. M. Bardeen, B. Carter and S. W. Hawking, “The four laws of black hole mechanics,” Commun. Math. Phys. [**31**]{}, 161 (1973).
J. D. Bekenstein, “Black holes and entropy,” Phys. Rev. D [**7**]{}, 2333 (1973).
S. W. Hawking, “Particle creation by black holes,” Commun. Math. Phys. [**43**]{}, 199 (1975).
R. Defay, I. Prigogine, A. Bellemans and D. H. Everett, [*Surface tension and adsorption*]{}, (Joun Wiley and Sons, New York, 1966).
L. Smarr, “Mass formula for Kerr black holes,” Phys. Rev. Lett. [**30**]{}, 71 (1973).
M. Parikh and F. Wilczek, “An action for black hole membranes,” Phys. Rev. D [**58**]{}, 064011 (1998).
D. M. Eardley and S. B. Giddings, “Classical black hole production in high-energy collisions,” Phys. Rev. D [**66**]{}, 044011 (2002).
G. ’t Hooft, “Dimensional Reduction in Quantum Gravity,” gr-qc/9310026; L. Susskind, “The World as a hologram,” J. Math. Phys. [**36**]{}, 6377 (1995); J. M. Maldacena, “The Large N limit of superconformal field theories and supergravity,” Adv. Theor. Math. Phys. [**2**]{}, 231 (1998); E. Witten, “Anti-de Sitter space and holography,” Adv. Theor. Math. Phys. [**2**]{}, 253 (1998); R. Bousso, “A covariant entropy conjecture,” JHEP [**9907**]{}, 004 (1999).
P. Kovtun, D. T. Son and A. O. Starinets, “Viscosity in strongly interacting quantum field theories from black hole physics,” Phys. Rev. Lett. [**94**]{}, 111601 (2005); idem, “Holography and hydrodynamics: diffusion on stretched horizons,” JHEP [**0310**]{}, 064 (2003).
D. J. E. Callaway, “Surface tension, hydrophobicity, and black holes: The entropic connection,” Phys. Rev. E [**53**]{}, 3738 (1996).
B. Zwiebach, in [*A First Course in String Theory*]{}, (Cambridge University Press, Cambridge, 2004).
N. Arkani-Hamed, S. Dimopoulos and G. R. Dvali, “The hierarchy problem and new dimensions at a millimeter,” Phys. Lett. B [**429**]{}, 263 (1998); I. Antoniadis, N. Arkani-Hamed, S. Dimopoulos and G. R. Dvali, “New dimensions at a millimeter to a Fermi and superstrings at a TeV,” Phys. Lett. B [**436**]{}, 257 (1998);\
N. Arkani-Hamed, S. Dimopoulos and G. R. Dvali, “Phenomenology, astrophysics and cosmology of theories with sub-millimeter dimensions and TeV scale quantum gravity,” Phys. Rev. D [**59**]{}, 086004 (1999).
L. Randall and R. Sundrum, “An alternative to compactification,” Phys. Rev. Lett. [**83**]{}, 4690 (1999); L. Randall and R. Sundrum, “A large mass hierarchy from a small extra dimension,” Phys. Rev. Lett. [**83**]{}, 3370 (1999).
F. R. Tangherlini, “Schwarzschild field in $n$ dimensions and the dimensionality of space problem,” Nuovo Cim. [**27**]{}, 636 (1963).
R. C. Myers and M. J. Perry, “Black Holes In Higher Dimensional Space-Times,” Annals Phys. [**172**]{}, 304 (1986).
G. T. Horowitz and A. Strominger, “Black strings and p-branes,” Nucl. Phys. B [**360**]{}, 197 (1991).
R. Emparan and H. S. Reall, “A rotating black ring in five dimensions,” Phys. Rev. Lett. [**88**]{} (2002) 101101.
H. Elvang and R. Emparan, “Black rings, supertubes, and a stringy resolution of black hole non-uniqueness,” JHEP [**0311**]{}, 035 (2003).
R. Emparan and R. C. Myers, “Instability of ultra-spinning black holes,” JHEP [**0309**]{}, 025 (2003).
J. Kunz, F. Navarro-Lerida and J. Viebahn, “Charged rotating black holes in odd dimensions,” Phys. Lett. B [**639**]{}, 362 (2006); see also A. N. Aliev, “Rotating black holes in higher dimensional Einstein-Maxwell gravity,” Phys. Rev. D [**74**]{}, 024011 (2006).
R. Gregory and R. Laflamme, “Black strings and p-branes are unstable,” Phys. Rev. Lett. [**70**]{}, 2837 (1993);“The instability of charged black strings and p-branes,” Nucl. Phys. B [**428**]{}, 399 (1994).
D. R. Smith and J. E. Ross, “Universal shapes and bifurcation for rotating incompressible fluid drops”, MAA [**1**]{}, 210 (1994).
P. Appell, [*Traité de mécanique rationnelle, Tome IV: Figures d’équilibre d’une masse liquide homogène en rotation*]{} (Gauthier-Villars, Paris, 1932).
L. Rayleigh, “On the capillary phenomena of jets”, Proc. Lond. Math. Soc. [**10**]{}, 4 (1878).
S. Tomotika, “On the instability of a cylindrical thread of a viscous liquid surrounded by another viscous fluid,” Proc. Roy. Soc. Lond. Ser. A [**150**]{}, 322 (1935).
S. Chandrasekhar, [*Hydrodynamic and hydromagnetic stability*]{} (Dover Publications, New York, 1981).
L. E. Johns and R. Narayanan, [*Interfacial instability*]{}, (Springer-Verlag, New York, 2002).
S. Chandrasekhar and E. Fermi, “Problems of gravitational stability in the presence of a magnetic field,” Astrophys. J. [**118**]{}, 116 (1953).
M-I Park, “The final state of black strings and p-branes and the Gregory–Laflamme instability,” Class. Quant. Grav. [**22**]{}, 2607 (2005).
R. Zahniser, “Instabilities of rotating jets,” B.Sc. thesis, MIT, unpublished (2004).
B. Kol and E. Sorkin, “On Black-Brane Instability In an Arbitrary Dimension,” Class. Quant. Grav. [**21**]{}, 4793 (2004).
M. Abramowitz and I. A. Stegun, in [*Handbook of Mathematical Functions*]{} (Dover Publications,Inc., New York 1965).
B. K. Shivamoggi, “Hydromagnetic gravitational instability of an infinite cylinder of a compressible fluid,” Astrophysics and Space Science [**86**]{}, 371 (1982).
A. M. Karnik, “Effects of rotation on the stability of a gravitating cylinder,” Astrophysics and Space Science [**83**]{}, 209 (1982).
H. Robe, “Équilibre et oscillations des cylindres polytropiques compressibles en rotation,” Annales d’Astrophysique [**31**]{}, 549 (1968).
P. J. Luyten, “Equilibrium and stability of rotating masses with a toroidal magnetic field,” Astrophysics and Space Science [**128**]{}, 289 (1986).
K. S. Thorne, “The Absolute Stability of Melvin’s Magnetic Universe,” Physical Review [**139**]{}, B244 (1965).
J. A. Wheeler, in [*Relativity, Groups and Topology*]{}, edited by B. and C. DeWitt (Gordon and Breach Publishers, New York, 1964).
See for instance M. Ansorg, A. Kleinwachter, R. Meinel, “Relativistic Dyson rings and their black hole limit,” Astrophys. J. [**582**]{}, L87 (2003); B. Zink, N. Stergioulas, I. Hawke, C. D. Ott, E. Schnetter and E. Mueller, “Formation of supermassive black holes through fragmentation of toroidal supermassive stars,” Phys. Rev. Lett. [**96**]{}, 161101 (2006).
S. L. Shapiro, S. A. Teukolsky and J. Winicour, “Toroidal black holes and topological censorship,” Phys. Rev. D [**52**]{}, 6982 (1995).
J. L. Hovdebo and R. C. Myers, “Black rings, boosted strings and Gregory-Laflamme,” Phys. Rev. D [**73**]{}, 084013 (2006).
R. Emparan, private communication.
S. S. Gubser, “On non-uniform black branes,” Class. Quant. Grav. [**19**]{}, 4825 (2002). E. Sorkin, “A critical dimension in the black-string phase transition,” Phys. Rev. Lett. [**93**]{}, 031601 (2004). H. K. Kunduri, J. Lucietti and H. S. Reall, “Gravitational perturbations of higher dimensional rotating black holes,” arXiv:hep-th/0606076. S. Hollands, A. Ishibashi and R. M. Wald, “A higher dimensional stationary rotating black hole must be axisymmetric,” arXiv:gr-qc/0605106. S. Chandrasekhar, “Solutions of two problems in the theory of gravitational radiation,” Phys. Rev. Lett.[**24**]{}, 611 (1970).
See for instance H. Lamb, “On the oscillations of a viscous spheroid”, Proc. London Math. Soc. [**13**]{}, 51 (1881); S. Chandrasekhar, “The oscillations of a viscous liquid globe”, ibid. [**9**]{}, 141 (1959).
W. H. Reid, “The oscillations of a viscous liquid drop”, Quart. Appl. Math. [**18**]{}, 86 (1960); L. D. Landau and E. M. Lifshitz, “Fluid Mechanics”, [Butterworth-Heinenann, Oxford, 1981]{}.
V. Cardoso, O. J. C. Dias and J. P. S. Lemos, “Gravitational radiation in D-dimensional spacetimes,” Phys. Rev.D [**67**]{}, 064026 (2003); V. Cardoso, J. P. S. Lemos and S. Yoshida, “Quasinormal modes of Schwarzschild black holes in four-dimensions and higher dimensions”, Phys. Rev. D [**69**]{}, 044004 (2004); idem, “Scalar gravitational perturbations and quasinormal modes in the five-dimensional Schwarzschild black hole”, JHEP [**0312**]{}, 041 (2003).
E. Berti, M. Cavaglia and L. Gualtieri, “Gravitational energy loss in high energy particle collisions: Ultrarelativistic plunge into a multidimensional black hole,” Phys.Rev. D [**69**]{}, 124011 (2004).
See for instance Y. Hagihara, [*Theories of Equilibrium Figures of a Rotating Homogeneous Fluid Mass*]{}, (NASA, 1970).
[^1]: We consider only Minkowski spacetimes. In the presence of cosmological constant some new features arise. For instance, the cosmological constant allows triaxial configurations of equilibrium rotating about the minor axis as solutions of the virial equations [@ellipsoidsdesitter]
[^2]: Black hole solutions with similar angular momentum configurations have been considered in [@kunz].
[^3]: The fact that extremal charged strings are marginally stable may not be directly related to horizons in themselves but rather to pure general relativistic effects. Thorne [@thornemelvin] has shown for instance that the Melvin universe in general relativity is absolutely stable. The Newtonian counterpart of the Melvin universe, on the other hand, is unstable [@wheelermelvin].
[^4]: We thank an anonymous referee for suggesting this discussion.
[^5]: A more sophisticated analysis by Crudelli [@crudelli] sharpened this inequality to half its value.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We give a very concise review of the group field theory formalism for non-perturbative quantum gravity, a higher dimensional generalisation of matrix models. We motivate it as a simplicial and local realisation of the idea of 3rd quantization of the gravitational field, equivalently of a quantum field theory of simplicial geometry, in which also the topology of space is fully dynamical. We highlight the basic structure of the formalism, and discuss briefly various models that are being studied, some recent results and the many open issues that future research should face. Finally, we point out the connections with other approaches to quantum gravity, such as loop quantum gravity, quantum Regge calculus and dynamical triangulations, and causal sets.'
author:
- '[**Daniele Oriti**]{}'
title: The group field theory approach to quantum gravity
---
Introduction and motivation
===========================
Group field theories (GFTs) [@iogft; @laurentgft] were developed at first as a generalisation of matrix models for 2d quantum gravity to 3 and 4 spacetime dimensions to produce a lattice formulation of topological theories. More recently, they have been developed further in the context of spin foam models for quantum gravity, as a tool to overcome the limitations of working with a fixed lattice in the non-topological case. In our opinion, however, their role in quantum gravity research goes much beyond this original scope, and GFTs should be seen as a fundamental formulation of quantum gravity and not just as an auxiliary tool. The bottom line of this perspective, here only tentatively outlined and still to be fully realised, hopefully, after much more work, can be summarised as follows: GFTs are quantum field theories [**of**]{} spacetime (as opposed to QFTs [*on*]{} spacetime), that describe the dynamics of both its topology and geometry in local, simplicial, covariant, algebraic terms, and that encompass most of the other approaches to non-perturbative quantum gravity, providing a link between them and at the same time going beyond the limitations of each. It will be clear from this brief account that we have just began to explore the structure of these models, and that very little is known about them; at the same time, we hope it would also be clear that what is known is enough for suggesting that in the GFT framework lies the potential for important developments towards a complete theory of quantum gravity.
The idea of defining a quantum field theory [*of*]{} geometry, i.e. a QFT on superspace (the space of 3-geometries) for given spatial topology, say $S^3$, was already explored in the past [@giddstro; @banks; @mcguigan]. The context was then a quantum cosmologyone, i.e. a global one. Such a theory would produce, in its perturbative expansion, a sum over different topologies each corresponding to a possible Feynman graph of the theory and to a possible interaction process for universesrepresented by the basic 3-sphere. The spatial topology change would be limited therefore to a changing number of disjoint copies of $S^3$. The field would represent a 2nd quantisation of the canonical wave function on superspace, here describing the one-particle sectorof the theory, turning it into an operator. The quantum amplitude for each Feynman graph, corresponding to a particular spacetime topology with $n$ boundary components, would be given by a path integral for gravity on the given topology, i.e. by a sum-over-histories quantisation of gravity with amplitude for each history obtained by exponentiating ($i$ times) the Einstein-Hilbert action for General Relativity. The difficulties in making mathematical sense of the continuum path integral itself are well-known, and it is a safe guess that the technical difficulties in turning this 3rd quantisation idea in a mathematically rigorous framework in the continuum are even more formidable. Also, such a cosmological setting (having a wave function and then a field for the whole universe) presents notorious interpretation problems, here only enhanced by treating also space topology as a variable) and lead to more than a few worries about the physical meaning of the whole setting. The general idea however is appealing in that such a framework would provide a natural mechanism for implementing topology change in quantum gravity and a prescription for weighting different spacetime topologies within a covariant sum-over-histories quantisation of gravity. In particular one could imagine that the interpretation issues, if not the technical difficulties, would be made easier if it was possible to implement the above ideas in a [**local**]{} framework, for example by generalising the superspace construction to [*open chunks of the universe*]{}, for example 3-balls, and then describing in a 3rd quantised language the interaction of these local pieces of the universe generating dynamically the whole universe and spacetime through their evolution. Again, however, the continuum setting seems to prevent a rigorous realisation of these ideas. By turning to a [**simplicial**]{} description of spacetime, the group field theory formalism gives a mathematically better defined realisation of these appealing ideas, providing a discrete sum-over-histories quantisation of gravity, and allowing an easier physical interpretation, since the picture of geometry and topology it is based on is intrinsically [*local*]{}.
The general formalism
=====================
Kinematics
----------
The geometry of a simplicial space (a triangulation) is fully characterised by a countable, if not finite, number of variables only, i.e. superspace becomes discrete, thus facilitating greatly the task of defining a field theory on it. Also, it is known that every closed D-dimensional simplicial complex can be obtained by gluing fundamental D-dimensional building blocks, each with the topology of a D-ball, along their boundaries (given by (D-1)-simplices). A [**local**]{}, and thus more physically sensible, realisation of the idea of a field theory on superspace is then possible, by considering first a wave function associated to each (D-1)-dimensional simplicial building block of space (if spacetime is D-dimensional), i.e. a functional of the geometry of one of them only, and then second quantising it, i.e. turning it into an operator. The quantum geometry of larger blocks of a spatial simplicial complex will be encoded in the tensor product of such wave functions/operators for the individual building blocks forming them. The issue is then: how to characterise the geometry of each simplicial building block, and thus of the full simplicial complex? in other words, which variables to choose to represent the arguments of our wave function/field? Here group field theories differ from other simplicial approaches to quantum gravity, and follow instead the path traced by loop quantum gravity, in that they seek to describe quantum geometry in terms of algebraic data, i.e. group and representation variables. This descends [@iogft; @review; @alexreview] from the classical description of gravity in terms of connection variables valued in the Lie algebra of the Lorentz group of the appropriate dimension, discretised to give elementary group valued parallel transports along paths in the (dual of the) simplicial complex, or equivalently in terms of Lie algebra-valued (D-2)-forms as in BF-like formulations of the same [@Peldan; @FKP], discretized to give the volumes of the (D-2)-dimensional cells of the simplicial complex, labelled by irreducible representations of the Lorentz group. The equivalence between these two sets of variables is given by the harmonic analysis on the group manifold that expresses their conjugate nature. More concretely, the field is taken to be a $\mathbb{C}$-valued function of D group elements, for a generic group $G$, one for each of the D boundary (D-2)-faces of the (D-1)-simplex the field corresponds to: $$\phi(g_1,g_2,...,g_D):
G^{\otimes D}\rightarrow \mathbb{C}.$$ The order of the arguments in the field corresponds to a choice of orientation for the (D-1)-simplex it represents; therefore it is natural to impose the field to be invariant under even permutations of its arguments (that do not change the orientation) and to turn into its own complex conjugate under odd permutations; this choice ensures that only orientable complexes are generated in the Feynman expansion of the field theory [@DP-P]. Other symmetry properties can also be considered [@DP-F-K-R]. The closure of the D (D-2)-faces to form a (D-1)-simplex is expressed algebraically by the invariance of the field under diagonal action of the group $G$ on the D arguments of the field: $\phi(g_1,...,g_D)=\phi(g_1g,...,g_Dg)$, which is also imposed [@BB; @carlomike]. This is also the simplicial counterpart of the Lorentz gauge invariance of first order gravity actions. As anticipated, the field can be expanded in modes using harmonic analysis on the group: $$\phi(g_i)=\sum_{J_i,\Lambda,k_i}\phi^{J_i\Lambda}_{k_i}\prod_iD^{J_i}_{k_il_i}(g_i)C^{J_1..J_4\Lambda}_{l_1..l_4},$$ with the $J$’s labelling representations of $G$, the $k$’s vector indices in the representation spaces, and the $C$’s being intertwiners of the group $G$, an orthonormal basis of which is labelled by an extra parameter $\Lambda$. That this decomposition is possible is not guaranteed in general but it is in fact true for the models being developed up to now, all based on the Lorentz group or on extensions of it. Group variables represent thus the GFT analogue of configuration space, while the representation parameters label the corresponding momentum space. Geometrically, the group variables, as said, represent parallel transport of a connection along elementary paths dual to the (D-2)-faces, while the representations $J$ can be put in correspondence with the volumes of the same (D-2)-faces, the details of this correspondence depending on the specific model [@review; @alexreview; @thesis]. The 1st quantisation of a geometric (D-1)-simplex in terms of these variables was performed in great detail in the 3 and 4-dimensional case in [@BB], but a similar analysis is lacking in higher dimensions. A simplicial space built out of $N$ such (D-1)-simplices is then described by the tensor product of $N$ such wave functions, at the 1st quantised level, with suitable constraints implementing their gluing, i.e. the fact that some of their (D-2)-faces are identified, at the algebraic level. For example, a state describing two (D-1)-simplices glued along one common (D-2)-face would be represented by: $\phi^{J_1 J_2..J_D\Lambda}_{k_1 k_2...k_D} \phi^{\tilde{J}_1
J_2...\tilde{J}_D\tilde{\Lambda}}_{\tilde{k}_1 k_2...\tilde{k}_D}$, where the gluing is along the face labelled by the representation $J_2$, and effected by the contraction of the corresponding vector indices (of course, states corresponding to disjoint (D-1)-simplices are also allowed). The corresponding state in configuration variables is: $\int dg_2
\phi(g_1,g_2,...,g_D)\phi(\tilde{g}_1,g_2,...,\tilde{g}_D)$. We see that states of the theory are then labelled, in momentum space, by [*spin networks*]{} of the group $G$ (see chapters by Thiemann and Perez). The second quantisation of the theory is obtained by promoting these wave functions to operators, as said, and the field theory is then specified by a choice of field action and by the definition of the partition function of the corresponding quantum theory. The partition function is then expressed perturbatively in terms of Feynman diagrams, as we are going to discuss. This implicitly assumes a description of the dynamics in terms of creation and annihilation of (D-1)-simplices, whose interaction generates a (discrete) spacetime as a particular interaction process (Feynman diagram) [@DP-F-K-R]. This picture has not been worked out in detail yet; in other words, no clear Fock structure on the space of states has been constructed, and no explicit creation/annihilation operators, acting on a suitably third quantised vacuum, resulting from the classical Hamiltonian structure of a group field theory, have been defined. Work on this is in progress [@ahmed-io-jimmy] (see also [@mikovic]).
Dynamics
--------
On the basis of the above kinematical structure, spacetime is viewed as emerging from the interaction of fundamental building blocks of space, and inherits therefore their simplicial characterisation, in that it is represented by a D-dimensional simplicial complex. More precisely, any such simplicial spacetime is understood as a particular interaction process among (D-1)-simplices, described in term of Feynman diagrams in typical quantum field-theoretic perturbative expansion. With this in mind, it is easy to understand the choice of classical field action in group field theories. This action, in configuration space, has the general QFT structure:
$$\begin{aligned}
S_D(\phi, \lambda)= \frac{1}{2}\left(\prod_{i=1}^D\int
dg_id\tilde{g}_i\right)
\phi(g_i)\mathcal{K}(g_i\tilde{g}_i^{-1})\phi(\tilde{g}_i)
+
\frac{\lambda}{(D+1)!}\left(\prod_{i\neq j =1}^{D+1}\int dg_{ij}\right)
\phi(g_{1j})...\phi(g_{D+1 j})\,\mathcal{V}(g_{ij}g_{ji}^{-1}), \label{eq:action}\end{aligned}$$
where the choice of kinetic and interaction functions $\mathcal{K}$ and $\mathcal{V}$ define the specific model considered. Of course, the same action can be written in momentum space after harmonic decomposition on the group manifold. The combinatorial structure of the dynamics is evident from the pairing of the arguments of these two functions and from their degree in field variables: the interaction term describes the interaction of D+1 (D-1)-simplices to form a D-simplex by gluing along their (D-2)-faces (arguments of the fields), that are pairwise linked by the interaction vertex. The nature of this interaction is specified by the choice of function $\mathcal{V}$. The (quadratic) kinetic term involves two fields each representing a given (D-1)-simplex seen from one of the two D-simplices (interaction vertices) sharing it, so that the choice of kinetic functions $\mathcal{K}$ specifies how the information and therefore the geometric degrees of freedom corresponding to their D (D-2)-faces are propagated from one vertex of interaction (fundamental spacetime event) to another. We will detail this structure when describing the Feynman graphs and the quantum dynamics of the theory. Before going on to that, let us discuss briefly the classical dynamics of GFTs. What we have is an almost ordinary field theory, in that we can rely on a fixed background metric structure, given by the invariant Killing-Cartan metric, and the usual splitting between kinetic (quadratic) and interaction (higher order) term in the action, that will later allow for a straightforward perturbative expansion. However, the action is also [*non-local*]{} in that the arguments of the D+1 fields in the interaction term are not all simultaneously identified, but only pairwise; while this is certainly a complication with respect to usual field theories in Minkowski space, on the other hand this pairwise identification implies that, even if the interaction is of order D+1, in terms of number of fields involved, it is still quadratic in terms of the individual arguments of the fields, which is in some sense a simplification of the usual situation (e.g. with respect to renormalisation issues). The classical equations of motion following from the above action are (schematically): $$\prod_j\int d\tilde{g}_{1j}\mathcal{K}(g_{1j}\tilde{g}_{1j}^{-1})\phi(\tilde{g}_{1j}) +
\frac{\lambda}{D+1}\left(\prod_{j=1}^{D+1}\int dg_{2j}..dg_{D+1 j}\right)
\phi(g_{2j})...\phi(g_{D+1 j})\,\mathcal{V}(\{g_{ij}\})=0$$ and are therefore, assuming $\mathcal{K}$ to be a generic differential operator, of integral-differential type. No detailed analysis of these equations in any specific GFT model nor of their solutions has been carried out to date, but work on this is in progress [@etera-laurent-aristide]. The role and importance of these equation from the point of view of the GFT per se are obvious: they define the classical dynamics of the field theory, they would allow the identification of classical background configurations around which to expand in a semi-classical perturbation expansion, etc. However, what is their meaning from the point of view of quantum gravity, in light of the geometric interpretation of the field itself as representing a (D-1)-simplex and of the GFT as a [*local simplicial 3rd quantisationof gravity*]{}? The answer is simple if striking: just as the Klein-Gordon equation represents at the same time the classical equation of motion of a (free) scalar field theory and the full quantum dynamics for the corresponding 1st quantised (free) theory, in the same way these classical GFT equations encode fully the quantum dynamics of the underlying (simplicial) canonical quantum gravity theory, i.e. the one in which quantum gravity wave functions are constructed, as for example in loop quantum gravity, by imposing the quantum version of the classical Hamiltonian and diffeomorphism constraints. This means that solving the above equations amounts to identifying non-trivial quantum gravity wave functions satisfying [**all**]{} the quantum gravity constraints, an important and still unachieved goal of canonical quantum gravity. We will come back to this canonical interpretation of GFTs after having introduced the perturbative expansion of their partition function in Feynman diagrams.
Another issue that still needs a careful investigation is that of the classical symmetries of the above action and of the resulting equations. It should be obvious, given the previous discussion, that the symmetries should correspond to the full set of symmetries of the simplicial theory one is quantising. Some of these symmetries, holding regardless of the specific choice of kinetic and interaction operators, are the above-mentioned closuresymmetry imposed on each field: $\phi(g_i)=\phi(g_i g)$, $\forall g\in G$, encoded in the symmetry property of the kinetic and vertex operators: $\mathcal{K}(g_i\tilde{g}_i^{-1})=\mathcal{K}(g
g_i\tilde{g}_i^{-1}g')$, $\mathcal{V}(g_{ij}\tilde{g}_{ji}^{-1})=\mathcal{V}(g_i
g_{ij}\tilde{g}_{ji}^{-1}g_j)$, and the global symmetry of the action under : $\phi(g_i)\rightarrow \phi(g g_i) \,\,\forall g\in G$. The presence of additional symmetries depends on the specific choice of kinetic and vertex operators, as said, as they would corresponds to specific symmetries of the classical discrete theories being quantised. The identification of such GFT analogues of the classical symmetries of specific discrete theories on spacetime is no easy task, and it is known already in the simpler example of GFT formulations of BF theories that there exist characteristic symmetries of the BF theories, i.e. so-called translation or topological symmetries, that can be correctly identified at the level of the GFT Feynman amplitudes, as identities between amplitudes for different graphs, but do not correspond to the above obvious symmetries of the GFT action [@laurentdiffeo].
Let us now turn to the quantum dynamics of group field theories. Most of the work up to now has focused on the perturbative aspects of this quantum dynamics, and basically the only thing that has been explored to some extent is the expansion in Feynman diagrams of the GFT partition function and the properties of the resulting Feynman amplitudes. This expansion is given by:
$$Z\,=\,\int
\mathcal{D}\phi\,e^{-S[\phi]}\,=\,\sum_{\Gamma}\,\frac{\lambda^N}{sym[\Gamma]}\,Z(\Gamma),$$ where $N$ is the number of interaction vertices in the Feynman graph $\Gamma$, $sym[\Gamma]$ is a symmetry factor for the graph and $Z(\Gamma)$ the corresponding Feynman amplitude. The Feynman amplitudes can be constructed easily after identification of the propagator, given by the inverse of the kinetic term in the action, and the vertex amplitude of the theory; each edge of the Feynman graph is made of $D$ strands, one for each argument of the field, and each one is then re-routed at the interaction vertex, with the combinatorial structure of an $D$-simplex, following the pairing of field arguments in the vertex operator. Diagrammatically:
{width="5.5cm" height="3cm"}
{width="5.5cm" height="4cm"}
Each strand in an edge of the Feynman graph goes through several vertices, coming back where it started, for closed Feynman graphs, and therefore identifies a 2-cell (for open graphs, it may end up on the boundary, identifying an open 2-cell). Each Feynman graph $\Gamma$ is then a collection of 2-cells (faces), edges and vertices, i.e. a 2-complex, that, because of the chosen combinatorics for the arguments of the field in the action, is topologically dual to a D-dimensional simplicial complex [@DP-F-K-R; @DP-P]. Clearly, the resulting complexes/triangulations can have arbitrary topology, each corresponding to a particular [*scattering process*]{} of the fundamental building blocks of space, i.e. (D-1)-simplices.
Even if every 2-complex arising as a GFT Feynman graph can be understood as topologically dual to a D-dimensional triangulation, this would not necessarily be a simplicial [*manifold*]{}. The data in the GFT Feynman graphs do not constrain the neighbourhoods of simplices of dimensions from (D-3) downwards to be spheres. This implies that in the general case, the resulting simplicial complex, obtained by gluing D-simplices along their (D-1)-faces, would correspond to a [ *pseudo-manifold*]{}, i.e. to a manifold with [*conical singularities*]{} [@DP-F-K-R; @DP-P; @cranematter]. A precise set of conditions under which the GFT Feynman graphs correspond to manifolds is identified and discussed at length in [@DP-P], both at the level of simplicial complexes and of the corresponding dual 2-complexes, in 2, 3 and 4 dimensions. All the relevant conditions can be checked algorithmically on any given Feynman graph. It is not clear, at present, whether and how one can construct suitably constrained GFT models satisfying these conditions, i.e. that would generate only manifold-like complexes in their Feynman expansion.
Each strand carries a field variable, i.e. a group element in configuration space or a representation label in momentum space. Therefore in momentum space each Feynman graph is given by a spin foam (a 2-complex with faces labelled by representation variables), and each Feynman amplitude (a complex function of the representation labels, obtained by contracting vertex amplitudes with propagator functions) by a spin foam model (see chapter by Perez):
$$Z(\Gamma)=\sum_{J_f}\prod_f A(J_f) \prod_e A_e(J_{f\mid e}) \prod_v A_v(J_{f\mid v}).$$ As in all spin foam models, the representation variables have a geometric interpretation (edge lengths, areas, etc) (see [@review; @alexreview]) and so each of these Feynman amplitudes corresponds to a definition of a sum-over-histories for discrete quantum gravity on the specific triangulation dual to the Feynman graph, although the quantum amplitudes for each geometric configuration are not necessarily given by the exponential of a discrete gravity action. For more on the quantum geometry behind spin foam models (geometric meaning of the variables, form and properties of the amplitudes, etc, we refer to the literature [@review; @alexreview; @causal].
One can show that the inverse is also true: any local spin foam model can be obtained from a GFT perturbative expansion [@carlomike; @laurentgft]. This is a particularly important result, since it implies that the GFT approach [*subsumes*]{} the spin foam in its perturbative aspects, while at the same time going beyond it, since there is of course much more in a QFT than its perturbative expansion. The sum over Feynman graphs gives then a sum over spin foams (histories of the spin networks on the boundary in any scattering process considered), and equivalently a sum over triangulations, augmented by a sum over algebraic data (group elements or representations) with a geometric interpretation, assigned to each triangulation. This perturbative expansion of the partition function also allows for a perturbative evaluation of expectation values of GFT observables, as in ordinary QFT. These are given [@laurentgft] by gauge invariant combinations of the basic field operators, that can be constructed in momentum space using spin networks according to the formula: $$O_{\Psi=(\gamma, j_e,i_v)}(\phi)=\left(\prod_{(ij)}dg_{ij}dg_{ji}\right) \Psi_{(\gamma, j_e,i_v)}(g_{ij}g_{ji}^{-1})\prod_i \phi(g_{ij}),$$ where $\Psi_{(\gamma, j_e,i_v)}(g)$ identifies a spin network functional (see chapter by Thiemann) for the spin network labelled by a graph $\gamma$ with representations $j_e$ associated to its edges and intertwiners $i_v$ associated to its vertices, and $g_{ij}$ are group elements associated to the edges $(ij)$ of $\gamma$ that meet at the vertex $i$.
In particular, the transition amplitude (probability amplitude for a certain scattering process) between certain boundary data represented by two spin networks, of arbitrary combinatorial complexity, can be expressed as the expectation value of the field operators having the same combinatorial structure of the two spin networks [@laurentgft].
$$\langle \Psi_1\mid\Psi_2\rangle = \int \mathcal{D}\phi\,O_{\Psi_1}\,O_{\Psi_2}\,e^{-S(\phi)} = \sum_{\Gamma/\partial\Gamma=\gamma_{\Psi_1}\cup\gamma_{\Psi_2}}\,\frac{\lambda^N}{sym[\Gamma]}\,Z(\Gamma)$$ where the sum involves only 2-complexes (spin foams) with boundary given by the two spin networks chosen. The above perturbative expansion involves therefore two very different types of sums: one is the sum over geometric data (group elements or representations of $G$) entering the definition of the Feynman amplitudes as the GFT analogue of the integral over momenta or positions of usual QFT; the other is the overall sum over Feynman graphs; the last sum includes a sum over all triangulations for a given topology and a sum over all topologies. Both sums are potentially divergent. First of all the naive definition of the Feynman amplitudes implies a certain degree of redundancy, depending on the details of the kinetic and vertex operators chosen, resulting from the symmetries of the defining GFT. This means that a proper gauge fixing of these symmetries, especially those whose group is non-compact, is needed to avoid the associated divergences [@laurentdiffeo]. Even after gauge fixing, the sum over geometric data has a potential divergence for every bubbleof the GFT Feynman diagram, i.e. for every closed surface of it identified by a collection of 2-cells. This is the GFT analogue of loop divergences of usual QFT. Of course, whether the GFT amplitudes are divergent or not depends on the specific model. For example, while the most natural definition of the group field theory for the Barrett-Crane spin foam model [@DP-F-K-R], presents indeed bubble divergences, a simple modification of it [@P-R], producing a different version of the same model, possesses [*finite*]{} Feynman amplitudes, i.e. it is [ *perturbatively finite*]{} without the need for any regularization. In general however a regularization and perturbative renormalisation procedure would be needed to make sense of the sums involved in the definition of a given GFT. No systematic study of renormalisation of GFTs has been carried out to date. The importance and need of this study is obvious, given that GFTs are presently [*defined*]{} by their perturbative expansion and that only a proof of renormalisability of such model would ensure that such perturbative expansions make physical sense. The sum over Feynman graphs, on the other hand, is most certainly divergent. This is not surprising both from the quantum gravity point of view and the QFT one. As said, the sum over Feynman graphs gives a sum over [*all*]{} triangulations for [*all*]{} topologies, each weighted by a (discrete) quantum gravity sum-over-histories. That such a sum can be defined constructively thanks to the simplicial and QFT setting is already quite an achievement, and to ask for it to be finite would be really too much! Also, from the strictly QFT perspective, i.e. leaving aside the quantum gravity interpretation of the theory and of its perturbative expansion, it is to be expected that the expansion in Feynman graphs of a QFT would produce at most an asymptotic series and not a convergent one. This is the case for all the interesting QFTs we know of. What makes the usual QFT perturbative expansion useful in spite of its divergence is the simple fact that it has a clear physical meaning, in particular the fact that one knows what it means to compute a transition amplitude up to a given order. In the GFT case this means, among other things, providing a clear physical interpretation to the coupling constant $\lambda$. This can be done, actually, in more than one way. First of all, defining $\alpha=\lambda^{\frac{1}{D-1}}$ and redefining $\tilde{\phi}=\alpha\phi$, we can recast the GFT action in the form $S_{\lambda}[\phi]=\frac{1}{\alpha^2}S_{\lambda=1}[\tilde{\phi}]$. One can then perform a loop expansion of the GFT partition function, that is an expansion in the parameter $\alpha$, instead of a perturbative expansion in the coupling constant. This gives, for a generic transition amplitude between two boundary states $\Psi_1$ and $\Psi_2$: $\langle\Psi_1\mid\Psi_2\rangle_\alpha=\frac{1}{\alpha^2}\sum_{i=0}^{\infty}\alpha^{2
i} \langle\Psi_1\mid\Psi_2\rangle_i$, where $\langle\Psi_1\mid\Psi_2\rangle_i$ is a sum over Feynman graphs with $i$ loops. The point here is to realise that adding a loop to a given Feynman diagram is equivalent [@laurentgft] to adding a handle to the simplicial complex dual to it. This means that the parameter $\alpha=\lambda^{\frac{1}{D-1}}$ governs the strength of topology changing processes in the GFT perturbative expansion. This interpretation can also be confirmed by analysing the Schwinger-Dyson equations for a generic GFT [@laurentgft]. A different perspective on the physical meaning of $\lambda$ is obtained by noticing that $\lambda$ weights somehow the sizeof the spacetimes emerging in the GFT perturbative expansion, assuming that the number of D-simplices is a measure of the D-volume of spacetime. Then one could define $\lambda =
e^{i\Lambda}$, interpret the Feynman amplitude as defining the exponential of an (effective) action for gravity $Z(\Gamma)=e^{iS(\Gamma)}$ and thus rewrite heuristically the GFT partition function as: $Z = \sum_\Gamma \frac{1}{sym(\Gamma)} e^{i (
S(\Gamma) + \Lambda V(\Gamma))}$. Now, assuming for $S(\Gamma)$ a form given by, say, the Regge action for pure gravity with no cosmological constant on a triangulation (dual to $\Gamma$) with fixed edge lengths, then $\Lambda$ would play the role of a cosmological constant contribution to the gravity action. Indeed this would be exactly the expression for a dynamical triangulations model [@renate] with $\Lambda$ being the bare cosmological constant. While this is just a very heuristic argument and there is no easy way to make it rigorous for a general GFT, it can in fact be made rigorous by considering a tensor model [@Ambjorn; @DP-P], a special case of the GFT formalism. This is obtained from the GFT for BF theories roughly by constraining the representations of $G$ used to be all equal to a given one with dimension $N$; we will give more details on both tensor models and GFTs for BF theories in the following. For tensor models, the Feynman expansion of the partition function is given by [@DP-P]: $$Z=\sum_{T}
\frac{1}{\textnormal{sym}(T)}\lambda^{n_D} N^{n_{D-2}}.$$ By redefining $\lambda=e^{-(\Lambda \,V_{D}+\frac{D(D+1)}{2}\frac{\arccos{1/D}}{16\pi
G}\,V_{D-2})}$ and $N=e^{\frac{V_{D-2})}{8G}}$, where $G$ is the bare Newton constant, $V_D$ is the volume of an equilateral D-simplex with edge lengths equal to $a$, and $V_{D-2}$ is the volume of an equilateral (D-2)-simplex, the expression above turns into that defining the dynamical triangulations approach [@renate], if the triangulations are restricted to those defining manifolds as opposed to pseudo-manifolds. Here, as anticipated, $\Lambda$ plays the role of the bare cosmological constant. As said, it is not obvious whether and how such interpretation of the GFT coupling constant $\lambda$ extends to the general GFT case, and more work on this is needed. However, we want to stress that, on the one hand, the two proposed interpretations for $\lambda$ are compatible with each other, and that finding a clear link between the two would mean linking the value of the bare, and then of the renormalised, cosmological constant to the presence of spatial topology change. This would be the realisation in a more rigorous context of an old idea [@banks; @coleman] that was indeed among the first motivations for developing a 3rd quantisationformalism for quantum gravity.
Before going on to discuss some specific GFT models, we would like to remark once more on the connection between GFT and canonical quantum gravity, and to point out to one important recent proposal concerning this relation. As already mentioned, the classical GFT equations of motion are interpreted as encoding the full quantum dynamics of the corresponding 1st quantised theory; this is a simplicial quantum gravity theory whose kinematical quantum states are labelled by D-valent spin networks for the group $G$. If the spin foams describing the histories of such states were to be restricted to those corresponding to triangulation of trivial spacetime topology, then a sum over such histories weighted by an appropriate amplitude would be a possible definition of the canonical inner product for a simplicial version of loop quantum gravity based on the group $G$, if the 2-point function so defined is positive semi-definite as it should. In other words such sum-over-histories would define implicitly a projection onto physical quantum states of gravity, thus fully encoding the dynamics of the theory usually given in terms of an Hamiltonian constraint operator (see chapters by Thiemann and Perez). Now, the GFT formalism allows a natural definition of such sum, that involves, as expected, only the classical information encoded in the GFT action. In fact, one could consider the restriction of the GFT perturbative expansion given above to [*tree level*]{}, for given boundary spin network observables,[@laurentgft]:
$$\langle \Psi_1\mid\Psi_2\rangle = \sum_{\Gamma_{\textnormal{tree}}/\partial\Gamma=\gamma_{\Psi_1}\cup\gamma_{\Psi_2}}\,\frac{\lambda^N}{sym[\Gamma]}\,Z(\Gamma).$$
This is the GFT definition of the canonical inner product, i.e. of the matrix elements of the projector operator [@carlomike], [*for Feynman amplitudes that are real and positive*]{}, as for example those of the BF or Barrett-Crane models. The definition is well posed, because at tree level every single amplitude $Z(\Gamma)$ is finite whatever the model considered due to the absence of infinite summation (unless it presents divergences at specific values of the configuration/momentum variables). Moreover, it possesses all the properties one expects from a canonical inner product: 1) it involves a sum over Feynman graphs, and therefore triangulations, with the cylindrical topology $S^{D-1}\times [0,1]$, for closed spin networks $\Psi_i$ associated with the two boundaries, as it is easy to verify; 2) it is real and positive, as said, but not strictly positive; it has a non-trivial kernel that can be shown [@laurentgft] to include all solutions of the classical GFT equations of motion, as expected. This means that the physical Hilbert space for canonical spin network states can be constructed, using the GNS construction, from the kinematical Hilbert space of all spin network states by modding out those states belonging to the kernel of the above inner product.
On the one hand this represent a concrete (and the first) proposal for completing the definition of a loop formulation of quantum gravity, that can now be put to test, so the possible end of a long quest started more than 15 years ago, and thus a proof of the usefulness of GFT ideas and techniques if only with respect to the canonical/Hamiltonian quantum gravity program. On the other hand, and, in our opinion, most importantly, it makes clear that the GFT formalism contains and therefore can achieve much more than any canonical quantum theory of gravity, given that the last is fully contained in the very small subset of the former corresponding to the classicallevel only. Within the GFT formalism, we can go beyond this investigating, for example, the full quantum structure, topology change, quantum gravity 2-point functions (in superspace) that are [**not**]{} the canonical inner product [@generalised], etc.
Some group field theory models
==============================
Pure gravity
------------
Let us now discuss some specific GFT models. The easiest example is the straightforward generalisation of matrix models for 2d quantum gravity to a GFT [@eac], given by the action: $$S[\phi]=\int_G dg_1dg_2 \frac{1}{2}\phi(g_1,g_2)\phi(g_1,g_2) +
\frac{\lambda}{3!}\int dg_1dg_2dg_3\phi(g_1,g_2)\phi(g_1,g_3)\phi(g_2,g_3)$$ where $G$ is a generic compact group, say $SU(2)$, and the symmetries mentioned above are imposed on the field $\phi$ implying, in this case: $\phi(g_1,g_2)=\tilde{\phi}(g_1g_2^{-1})$. The relation with matrix models is apparent in momentum space, expanding the field in representations $j$ of $G$ to give:
$$S[\phi]=\sum_j \textit{dim}(j) \left( \frac{1}{2}tr{(\tilde{\phi}_j^2)} +\frac{\lambda}{3!}tr(\tilde{\phi}_j^3)\right)$$
where the field modes $\tilde{\phi}_j$ are indeed matrices with dimension $\textit{dim}(j)$, so that the action is given by a sum of matrix models actions for increasing dimensions. Alternatively, one can see the above as the action for a single matrix model in which the dimension of the matrices has been turned from a parameter into a dynamical variable. The Feynman amplitudes for such a theory are given by $Z(\Gamma)=\sum_j \textit{dim}(j)^{2-2 g(\Gamma)}$ and shows that the GFT above gives a quantisation of BF theory (with gauge group $G$) on a closed triangulated surface, dual to $\Gamma$, of genus $g(\Gamma)$, augmented by a sum over all such surfaces defined as a Feynman expansion[@eac]. A quantisation of 2d gravity along the same lines would use an $U(1)$ connection (and therefore $G=U(1)$) and a restriction on the representations summed over to single one, together with additional data encoding bundle information in the GFT action, see [@iocarlosimone].
The extension to higher dimensions can proceed in two ways. The first attempt at this was done in [@Ambjorn] where the first tensor modelwas introduced, with action:
$$S[\phi]=\sum_{\alpha_{i}}\left(\frac{1}{2}\phi_{\alpha_1\alpha_2\alpha_3}\phi_{\alpha_1\alpha_2\alpha_3}
+\frac{\lambda}{4!}\phi_{\alpha_1\alpha_2\alpha_3}\phi_{\alpha_3\alpha_4\alpha_5}\phi_{\alpha_5\alpha_2\alpha_6}\phi_{\alpha_6\alpha_4\alpha_1}\right)
\label{eq:tensor},$$
for a $N\times N\times N$ tensor $\phi$.
This model generates 3-dimensional simplicial complexes in its Feynman expansion, that include [@Ambjorn; @DP-P] both manifold- and pseudo-manifold-like configurations. It is not however a GFT until group structure is added. This is done again as a straightforward generalisation of the above model for the 2d case. In fact the following choice of kinetic and vertex term in the generic action \[eq:action\]:
$$\mathcal{K}(g_i,\tilde{g}_i) = \int_G dg \prod_i
\delta(g_i\tilde{g}_i^{-1}g),\;\;\;\;\;\;\mathcal{V}(g_{ij},g_{ji}) =
\prod_i\int_G dg_i \prod_{i<j}\delta(g_i g_{ij}g_{ji}^{-1}g_j^{-1}),$$ where the integrals impose the gauge invariance under the action of $G$, gives the GFT quantisation of BF theories, for gauge group $G$, in any dimension [@boulatov; @ooguri]. In particular, in 3 dimensions, the choice of $G=SO(3)$ or $G=SO(2,1)$ provides a quantisation of 3D gravity in the Euclidean and Minkowskian signatures, respectively, and the so-called Ponzano-Regge spin foam model, as evaluation of the GFT Feynman amplitudes (see also [@iogft] for more details on this GFT, first proposed by Boulatov [@boulatov]), while the choice of the quantum group $SU(2)_q$ gives the Turaev-Viro topological invariant. For the $SU(2)$ case, the action is then: $$\begin{aligned}
S[\phi]=\prod_i \int_{SU(2)} \phi(g_1,g_2,g_3)\phi(g_1,g_2,g_3)+
\frac{\lambda}{4!}\prod_{i=1}^{6}\int_{SU(2)}dg_i \;\phi(g_1,g_2,g_3)\phi(g_3,g_4,g_5)\phi(g_5,g_2,g_6)\phi(g_6,g_4,g_1).
\;\;\;\;\end{aligned}$$
Lots is known about the last model (see chapter by Freidel), including the appropriate gauge fixing procedure of its Feynman amplitudes, the coupling of matter fields, etc. Here, we mention only one result that is of interest for the general issue of GFT renormalisation. This is the proof [@laurentborel] that a simple modification of the GFT above gives a model whose perturbative expansion is Borel summable. The modification amounts to adding another vertex term of the same order to the original one, to give:
$$\begin{aligned}
+\frac{\lambda}{4!}\prod_{i=1}^{6}\int_{SU(2)}dg_i\left[
\phi(g_1,g_2,g_3)\phi(g_3,g_4,g_5)\phi(g_5,g_2,g_6)\phi(g_6,g_4,g_1)+ \,\delta\, \phi(g_1,g_2,g_3)\phi(g_3,g_4,g_5)\phi(g_4,g_2,g_6)\phi(g_6,g_5,g_1)\right].\end{aligned}$$
The new term corresponds simply to a slightly different recoupling of the group/representation variables at each vertex of interaction, geometrically to the only other possible way of gluing 4 triangles to form a closed surface. This Borel-summability is interesting for more than one reason: 1) it shows that it [*is*]{} possible to control the sum over triangulations of all topologies appearing in the GFT perturbative expansion, despite the factorial growth in their number with the number of 3-simplices forming them; whether the sum is summable or not, then, depends clearly only on the amplitude weighting them; 2) even if the above modification of the Boulatov GFT model has no clear physical interpretation yet from the quantum gravity point of view, e.g. in terms of gravity coupled to some sort of matter, it is indeed a very mild modification, and most importantly one that one would expect to be forced upon us by renormalisation group-type of argument, that usually require us to include in the action of our field theory all possible terms that are compatible with the symmetries.
Still in the 3D case, it is easy to check that a GFT action with the same combinatorial structure of the Boulatov model, but for a real field defined on the homogeneous space $SO(3)/SO(2)\simeq S^2$ (in turn obtained choosing $G=SO(3)$ and then averaging over the subgroup $SO(2)$ in each argument of the field [@laurentborel]), and with the invariance under the global $SO(3)$ action in each field having been dropped, gives a generalisation of the tensor model \[eq:tensor\] analogous to the GFT generalisation of matrix models previously described. In fact the action for this model, in momentum space, is:
$$S[\phi]=\sum_{j_i,\alpha_{i}}\left(\frac{1}{2}\phi^{j_1j_2j_3}_{\alpha_1\alpha_2\alpha_3}\phi^{j_1j_2j_3}_{\alpha_1\alpha_2\alpha_3}
+\frac{\lambda}{4!}\sum_{j_i,\alpha_i}\phi^{j_1j_2j_3}_{\alpha_1\alpha_2\alpha_3}\phi^{j_3j_4j_5}_{\alpha_3\alpha_4\alpha_5}\phi^{j_5j_2j_6}_{\alpha_5\alpha_2\alpha_6}\phi^{j_6j_4j_1}_{\alpha_6\alpha_4\alpha_1}\right)$$ where the indices $\alpha_i$ run over a basis of vectors in the representation space $j_i$, and its partition function is: $Z=\sum_{\Gamma}\frac{(-\lambda)^{n_\Gamma}}{\textnormal{sym}(\Gamma)}\sum_{j_f}\prod_f
(2j_f+1)$, with $f$ being th faces of the 2-complex/Feynman graph, which is divergent and has to be regularised. There are three ways of doing it, and, by doing this, reducing the partition function to that of a tensor model: 1) simply dropping the sum over the representations $j_f$ by fixing them to equal a given $J$; 2) placing a cut-off on the sum by restricting $j_i<N$, obtaining $Z=\sum_{\Gamma}\frac{(-\lambda)^{n_v(\Gamma)}}{\textnormal{sym}(\Gamma)}[(N+1)^2]^{n_f(\Gamma)}$; 3) equivalently, but more elegantly, by defining the model not on $S^2$ but on the non-commutative 2-sphere $S^2_N$, which also carries a representation of $SU(2)$ but implies a bounded decomposition in spherical harmonics (labelled by $j<N$), thus giving the same result for the partition function. We recognise in the above result the partition function for the tensor model \[eq:tensor\] and for a dynamical triangulations model [@renate].
Let us now discuss the situation in the 4D case. Here GFT model building has followed the development of spin foam models for 4D quantum gravity. The guiding idea in the spin foam setting (see chapter by Perez) has been the fact that [ *classical*]{} gravity can be written [@Peldan; @FKP] as a constrained version of a BF theory for the Lorentz group; the construction of the appropriate spin foam model then required [@BC] first a description of simplicial geometry in terms of bivectors (discretion of the $B$ field of the BF theory) subject to the discrete analogue of the continuum constraints [@BC], and then finding the algebraic translation of these at the quantum level as constraints on the representations labelling the spin foam faces. The Barrett-Crane spin foam models obtained as a result of this amount roughly to a restriction of spin foam models for BF theories to involve so-called [*simple*]{} representations only of the Lorentz group ($SO(4)$ or $SO(3,1)$, or the corresponding double covering, depending on the signature wanted for spacetime) [@BC; @alexreview; @review]. The same restriction can be imposed at the GFT level, starting from the GFT describing 4D BF theory, by projecting down the arguments of the field from $G=SO(4)$ ($SO(3,1)$) to the homogeneous space $SO(4)/SO(3)\simeq S^3$ ($SO(3,1)/SO(3)$ or $SO(3,1)/SO(2,1)$ in the Lorentzian case), exploiting the fact that only simple representations of $G$ appear in the harmonic decomposition of functions on these spaces. The GFT action is then defined [@DP-F-K-R] (we deal explicitly only with the Riemannian case for simplicity) as:
$$\begin{aligned}
\lefteqn{S[\phi]=\frac{1}{2}\left(\prod_i \int_{SO(4)} dg_i\right)
P_gP_h\phi(g_1,g_2,g_3,g_4)P_gP_h\phi(g_1,g_2,g_3,g_4) +}\nonumber \\
&+&\frac{\lambda}{5!}\left(\prod_{i=1}^{10}\int_{SO(4)}dg_i\right)
\left[
P_gP_h\phi(g_1,g_2,g_3,g_4)P_gP_h\phi(g_4,g_5,g_6,g_7)P_gP_h\phi(g_7,g_8,g_3,g_9)\right.
\nonumber \\ &\left.\right.& \left.\hspace{4cm}P_gP_h\phi(g_9,g_5,g_2,g_{10})P_gP_h\phi(g_{10},g_8,g_6,g_1)\right]\;\;\end{aligned}$$
where the projection $P_h\phi(g_i)=\prod_i\int_{SO(3)}dh_i \,\phi(g_i
h_i)$ from the group to the homogeneous space imposes the wanted constraints on the representations, in momentum space, and the projection $P_g\phi(g_i)=\int_{SO(4)} dg \,\phi(g_i g)$ ensures that gauge invariance is maintained.
Different variations of this model can be constructed [@P-R; @thesis] by inserting the two projectors $P_h$ and $P_g$ in the action in different combinations. However, the symmetries of the resulting models remain the same, and the corresponding Feynman amplitudes have the structure:
$$Z(\Gamma)=\sum_{J_f}\prod_f \textnormal{dim}(J_f) \prod_e A_e(J_{f\mid e}) \prod_v \mathcal{V}_{BC}(J_{f\mid v}),$$
where $\textnormal{dim}(J_f)$ is the measure for the representation $J_f$, labelling the faces of the 2-complex/Feynman graph, entering the harmonic decomposition of the delta function on the group, and the function $\mathcal{V}_{BC}(J_{f\mid v})$, depending on the ten representations labelling the ten faces of $\Gamma$ incident to the same vertex $v$ is the so-called Barrett-Crane vertex [@BC; @review; @alexreview]; the exact form of the edge amplitude $A_e$, on the other hand, depend on the version of the specific variation of the above action being considered.
The above Feynman amplitudes, modulo the exact form of the edge amplitudes, can be obtained and justified in various ways, e.g. starting from a discretisation of classical BF theory and a subsequent imposition of the constraints [@review; @alexreview], and in particular there is a good consensus on the validity of the Barrett-Crane construction for the vertex amplitude. As a hint that at least [*some*]{} of the properties of gravity we want are correctly encoded in the above Feynman amplitudes, we cite the fact that there exists at least a sector of the geometric configurations summed over where an explicit connection with classical discrete gravity can be exhibited. In fact [@ruthjohn], for non-degenerate configurations (i.e. those corresponding to non-degenerate simplicial geometries) the asymptotic limit of the Barrett-Crane amplitude $\mathcal{V}_{BC}(J)$ is proportional to the cosine of the Regge action for simplicial gravity, i.e. the correct discretion of the Einstein-Hilbert action for continuum General Relativity. It is therefore very close to the form one would have expected for the amplitudes in a sum-over-histories formulation of quantum gravity on a given simplicial manifold.
All the above models (for BF theory and quantum gravity) in both 2, 3 and 4 dimensions share the following general properties: 1) their Feynman amplitudes are real; 2) they do not depend on the orientation chosen for the complexes, or to put it differently, no unique orientation for the (various elements of the) triangulation dual to any Feynman graph can be reconstructed from the amplitude associated to it; 3) in quantum gravity models, the asymptotic limit of the vertex amplitude gives (in the non-degenerate sector) the [*cosine*]{} of the Regge action instead of the exponential of it. These properties suggest the interpretation [@generalised] of the corresponding models as defining the quantum gravity analogue of the Hadamard 2-point function for a relativistic particle (anti-commutator of field operators in QFT), or, more appropriately in our field theory context, as defining a QFT picture for quantum gravity with Hadamard propagators used everywhere in the internal linesof the graphs, instead of the usual Feynman propagators. Also, these are the wanted properties if our aim is to define, through a GFT, the inner product of a canonical/Hamiltonian theory of quantum gravity based on spin networks, as we have discussed, i.e. we want to impose the dynamics encoded in a Hamiltonian constraint using covariant methods. In other words, the Feynman amplitudes of these GFT models would correspond not to a simplicial version of the path integral formalism for quantum gravity, but to the symmetrized version of the same over opposite spacetime orientations, that indeed gives a path integral definition of solutions of the Hamiltonian constraint operator of canonical quantum gravity [@halliwellhartle]. However, there are several reasons why one may want to go beyond this type of structure. We cite three of them. 1) From a field theoretic perspective applied to quantum gravity, i.e. from the point of view of a field theory on a simplicial superspace we are advocating here, the most natural object one would expect a GFT to define with its 2-point functions is not a canonical inner product, solution of the Hamiltonian constraint, but a Green function for it. This is what happens in ordinary QFT, for the free theory, and in the formal context of continuum 3rd quantisation for quantum gravity. In the last case, in fact, the Feynman amplitudes of the theory, in absence of spatial topology change, correspond to the usual path integral formulation of quantum gravity for given boundary wave functions, with amplitude given by the exponential of the GR action [@giddstro], which is a Green function for the Hamiltonian constraint, and not a solution of the same, in each of its arguments. 2) The orientation of the GFT 2-complexes can be given, for Lorentzian models, a [*causal*]{} interpretation [@causal; @feynman], if one also interprets the vertices of the same as fundamental spacetime events, and thus the orientation independence of the spin foam amplitudes associated to the usual models implies that these models define [*a-causal*]{} transition amplitudes for quantum gravity. Again, this is consistent with the field-theoretic interpretation of the same amplitudes given above and with the wish to construct the canonical inner product, but it also suggest that one should be able to construct other types of models defining [*causal*]{} quantum gravity transition amplitudes [@causal; @feynman] instead and thus different types of GFTs. 3) If there is, as we have seen, a clear interpretation from the perspective of canonical quantum gravity for the amplitudes defined by the GFT in the case of trivial spacetime topology (i.e. no spatial topology change), this occurs in a very limited subsector of the theory, and no clear meaning can be given in the same perspective to the GFT amplitudes for Feynman graphs beyond the tree level, when spatial topology change is present, and thus a canonical/Hamiltonian formalism is not well-defined. For all these reasons one would like to be have a more general class of GFT models that do depend on the orientation of the GFT Feynman graphs, that can be interpreted consistently as analogues of causal transition amplitudes of QFT, that are in more direct contact with usual path integral formulations of (simplicial) gravity, and that reduce to the above type of models (and corresponding amplitudes) when suitably restricted. This was achieved in [@generalised]. Here a generalised version of the GFT formalism was defined, for a field $\phi(g_i,s_i): (G\times\mathbb{R})^{\otimes
4}\rightarrow\mathbb{C}$, with action (in the 4d Riemannian case): $$\begin{aligned}
\lefteqn{S_{gen}(\{\lambda\})=\sum_{\mu,\alpha}\frac{1}{4}\prod_{i=1}^{4}\int
dg_i\,\int_\mathbb{R}ds_i \left\{
\phi^{-\mu\alpha}(g_i,s_i)\left[\prod_i\left(
-i\mu\alpha\partial_{s_i}+\nabla_{i}\right)\right]\phi^{\mu\alpha}(g_i,s_i)\right\} +} \nonumber \\ &+&\sum_{\mu}\sum_{\alpha_i}
\frac{\lambda_{\{\alpha_i,\mu\}}}{5!}\prod_{i\neq
j=1}^{5}\int_G dg_{ij}\int_\mathbb{R}ds_{ij}
\left\{ P_h\phi^{\mu\alpha_1}(g_{1j},s_{1j})
P_h\phi^{\mu\alpha_2}(g_{2j},s_{2j})
........ \right.
\nonumber \\
&& \left. ........P_h\phi^{\mu\alpha_5}(g_{5j},s_{5j})
\prod\,\theta(\alpha_i s_{ij}+ \alpha_j s_{ji})K\left(g_{ij},
g_{ji};\mu(\alpha_i s_{ij} + \alpha_j
s_{ji})\right)\right\},\;\;\;\;\end{aligned}$$ where: $g_i\in G=SO(4)$, $s_i\in\mathbb{R}$, $\mu=\pm 1$ and $\alpha_i = \pm 1$ are orientation data that allow to reconstruct the orientation of the Feynman graph from the [**complex**]{} amplitude associated to it, $\phi^{+}(g_i,s_i)=\phi(g_1,s_1;...,g_4,s_4)$ and $\phi^{-}(g_i,s_i)=\phi^{\dagger}(g_i,s_i)$, $P_h$ is the projector imposing invariance under the $SO(3)$ subgroup, $\nabla$ is the D’Alambertian operator on the group $G$, $\theta(s)$ is the step function and $K(g,s)$ is the evolution kernel for a scalar particle on the group manifold $G$ with evolution parameter $s$. The field is assumed invariant under the diagonal action of $G$, as in the models discussed above. The form of the kinetic operator and the use of the evolution kernels with proper time parameter $s$, plus the restriction on its range (through the theta functions), just as in the definition of the Feynman propagator of a relativistic particle, in the vertex term, impose a non-trivial dependence on the orientation data in fully covariant way. The resulting Feynman amplitudes [@generalised] have all the properties wanted, being complex and orientation-dependent, and have the natural interpretation as analogues of Feynman transition amplitudes for quantum gravity [@generalised; @feynman]. Also, when expressed in momentum space, but only for the variables $s_i$, whose conjugate variables are labelled suggestively as [*variable masses*]{} $m_i^2$, the amplitude for each vertex is given by: $$\begin{aligned}
\prod_{f_{ij}\mid v} \int_{0}^{+\infty}ds_f\,\theta(\alpha_i\alpha_j
s_f)\, K(\vartheta_{f_{ij}},\mu_v\alpha_{i}\alpha_{j}s_f)e^{-im_f^2s_f}=\prod_{f_{ij}\mid v}
\frac{1}{4\pi}\frac{1}{\sin\vartheta_{f_{ij}}}e^{i\mu_v
\sqrt{1-\mu\alpha_{i}\alpha_{j}m^2_{f_{ij}}}\vartheta_{f_{ij}}},\end{aligned}$$ with $\sqrt{1-\mu\alpha_{i}\alpha_{j}m^2_{f_{ij}}}$ interpreted as the area of the triangle dual to the face $f_{ij}$, and the angle $\vartheta_{f_{ij}}$ measuring the holonomy around the portion of the face dual to the same triangle inside the 4-simplex dual to $v$; thus this amplitude is [@generalised] of the form of the exponential of the Regge action in 1st order formalism, times an appropriate measure factor. It remains to be proven that this form still holds, in these variables, for the amplitude associated to the whole Feynman graph [@IoTimGFT; @GFTsimplicial]. Other models in the same class, i.e. with the same type of fields and the same symmetries, can be defined, and share similar properties [@IoTimGFT; @GFTsimplicial].
As desired, the generalised GFT model reduces to the usual ones (i.e. to the Barrett-Crane model, in this case) of one drops the $\theta$s from the vertex function in the action, or equivalently averages over different orientation data in the same term, [ *and*]{} goes to the [*static-ultra-local*]{} limit of the kinetic term, i.e. replaces the derivative operators with simple delta functions identifying pairwise the arguments of the two fields. What we have, then, is truly a generalisation of the usual formalism, that, moreover, gives GFTs a structure even more similar to ordinary QFTs, thanks to the presence of derivatives in the action, and may thus possibly simplify their future analysis and development.
Other types of GFTs have been constructed in the literature, ranging from a Boulatov-like model for 3d gravity based on the quantum group $DSU(2)$ [@kirill], with links to models of 3d quantum gravity coupled to matter, that we are going to discuss below, to a modified version [@ioetera] of the GFTs for the Barrett-Crane models, with a tunable extra coupling among the 4-simplices and a possible use in the renormalisation of spin foam models. For all these, and others [@grosse], we refer to the literature.
Matter coupling
---------------
We want instead to discuss now, briefly, the coupling of matter fields to quantum gravity in the group field theory context. This has been sketched in the 4-dimensional case in [@mikovicmatter], but has been described in full details only in the 3D case in [@iojimmy], with an earlier model, based on similar ideas but covering only the spinless case, having been introduced in [@iojimmylaurent]. The idea followed in this work is to define a GFT that would treat quantaof gravity and of matter on equal footing and generate, in perturbative expansion, a sum over Feynman diagrams of gravity, i.e. spin foam 2-complexes/triangulations, and Feynman diagrams of any given matter field theory, with the latter suitably topologically embedded in the former. The degrees of freedom of matter fields should also be suitably coupled to those of geometry to reproduce, at the level of the spin foam amplitudes, the correct dynamical interaction of gravity and matter. The spin foam amplitudes describing this coupling in 3d, were introduced and studied in [@PR1], and provided the key insight needed to devise an appropriate GFT reproducing them. The kinematical structure of the new GFT is given [@iojimmy] by [*two*]{} different types of fields: one is the usual $\phi$ of the Boulatov theory discussed above, and represents a triangle with no particle present, the other is a new field $\psi_s(g_1,g_2,g_3;u): SU(2)^{\otimes 4}\rightarrow \mathbb{C}$ , on which we impose a global $SU(2)$ symmetry under the simultaneous right shift of its 4 arguments, and represents a tetrahedron with a single particle of spin $s$ located at one of its vertices, whose degrees of freedom are encoded in the group variable $u$. The field $\psi_s$ can be partially expanded in modes as: $
\psi_s(g_1,g_2,g_3;u)=\sum_{I,n}\psi^{I}_{sn}(g_1,g_2,g_3)D^{I}_{sn}(u),
$ thus clarifying the interpretation of $s$ as the spin of the particle and of $I$ as its total angular momentum [@PR1]. In the language of spin networks, the first type of field corresponds to a simple pure gravity 3-valent spin network vertex, that will give closed spin network states when suitably contracted with other vertices through its open links; the second gives a 4-valent vertex with one extra open link labelled by the particle degrees of freedom; the contraction of such vertices with both pure gravity vertices and other 4-valent ones produces [*open*]{} spin networks, that are at the same time quantum gravity states and multi-particle states. The GFT should then describe dynamically the creation/annihilation of these two types of fundamental structures, producing both a (simplicial) spacetime and a matter Feynman graph for a particle species with spin $s$ embedded in it. The mass appears dynamically in the interaction with gravity, i.e. with the geometric degrees of freedom. For a single matter field of spin $s$ and in the simplest case of 3-valent interaction, the full GFT action doing the job is (in a shortened but intuitive notation):
$$\begin{aligned}
\lefteqn{S[\phi,\psi_s]=\frac{1}{2}\int
\prod^{3}_{i=1}dg_i\;\phi(g_i)\phi(g_i)+\frac{1}{2}\int\prod^{3}_{i=1} dg_i\, du\;
\psi_{s}(g_i;u)\psi_{s}(g_i;u)}\\
&+&\frac{\lambda}{4!}\int\prod^{6}_{i=1}dg_i\;\phi(g_1,g_2,g_3)\phi(g_3,g_5,g_4)\phi(g_4,g_2,g_6)\phi(g_6,g_5,g_1)\\
&+&\mu_2\int \prod^{6}_{i=1}dg_i\, du_a\;
\psi_{s}(g_i,g_3u_a^{-1}hu_a;u_a)\psi_{s}(g_4u_a^{-1}h^{-1}u_a,g_i;u_a)\phi(g_i)\phi(g_i)\\
&+&\mu_3\int \prod^{6}_{i=1}dg_i\, du_a\, du_b\, du_c\;
\psi_{s}(g_i,g_3u_a^{-1}hu_a;u_a)\psi_{s}(g_4u_b^{-1}h^{-1}u_b,g_i;u_b)\psi_{s}(g_i,g_2u_c^{-1}hu_c;u_c)\\
&\times&
\phi(g_6,g_5,g_1)\delta(u_a^{-1}hu_au_b^{-1}h^{-1}u_bu_c^{-1}hu_c)\sum_{I_\alpha, n_\alpha}
D^{I_a}_{sn_a}(u_a)D^{I_b}_{sn_b}(u_b)D^{I_c}_{sn_c}(u_c)C^{I_a\,I_b\,I_c}_{n_an_bn_c}\end{aligned}$$
where the third and fourth terms produce, in Feynman expansion, spacetime building blocks given by tetrahedra with a single particle propagating along two of its edges (third term) and by tetrahedra with three particles travelling along three of its edges and interacting at one of its vertices (fourth term). Here even the pure propagation of matter degrees of freedom has been encoded in one of the GFT interaction terms because it involves an interaction with the geometric degrees of freedom, but it could be as well encoded in one of the GFT kinetic terms. For a more extended discussion of this action and for a clarifying graphical representation of the various terms in it, and of the resulting Feynman graphs for gravity and matter, we refer to [@iojimmy]. As anticipated, the Feynman amplitudes of this GFT are given by the coupled Ponzano-Regge spin foam model for gravity coupled to spinning particles studied in [@PR1], so we do not need to reproduce them here. These can be understood as discrete path integrals for quantum gravity on a given spacetime triangulation, coupled to particles whose trajectories are characterised by the particle Feynman graph embedded in the same triangulation. Equivalently, they correspond to the evaluation in the pure gravity Ponzano-Regge spin foam model of appropriate particle observablescharacterised by the same Feynman graph [@PR1; @barrettfeynman]. On the other hand, we want to stress once more the GFT perspective on the issue of matter coupling in quantum gravity. This is, again, a purely field-theoretic perspective, but that of a field theory [**of**]{} spacetime with matter, and not on spacetime. Accordingly, as said, the field represents at once the basic building blocks [**of space and matter**]{} and the theory, as presently understood, builds up as Feynman graphs in a perturbative expansion both the histories of space (possible spacetimes), i.e. of both its geometry and topology, and of matter (possible particle evolutions/interactions). Another way in which matter degrees of freedom may be encoded in a GFT description can be found in [@eterawinston]
The extension of these results to 4D, in both the spin foam and GFT setting, can proceed in various directions. On the one hand, one can re-express the Feynman graph amplitudes for any matter field theory in flat space as a spin foam model for particles coupled to a topological theory [@aristidelaurent4d], and then try to build up a full quantum gravity spin foam model reproducing this in the appropriate topological limit. Having the explicit form of a such a candidate spin foam model, one could then use it as a guide for the construction of the corresponding GFT extending to 4d the model we just presented. On the other hand, the straightforward extension of the above GFT model to 4D gives a fully non-perturbative quantum description, in a discrete setting, of topological gravity (BF theory) coupled to topological strings, producing in perturbation expansion a sum over both simplicial spacetimes and discrete string worldsheets embedded in them [@iojimmystrings], and with kinematical and dynamical structures consistent with the canonical analysis of such system performed in [@alexjohn]. Yet another direction towards the understanding of quantum gravity coupled to matter in 4d is the development of GFT model for gravity coupled to Yang-Mills theories, following earlier results obtained in the spin foam setting [@iohendryk], that would produce, again in its perturbative expansion, a sum over all simplicial complexes with amplitudes given by the spin foam analogue of a path integral for gravity coupled to gauge theory, in the form of a lattice gauge theory on a random simplicial lattice with dynamical geometry [@iojimmyrobert]. We conclude this discussion of matter-gravity coupling in the GFT language by mentioning a rather speculative but highly intriguing idea put forward in [@cranematter]. The suggestion is that the GFTs for pure gravity of the type outlined above already contain all the right matter degrees of freedom, and therefore there is no need to modify them by explicitly coupling different types of data and extending the field content of the models: the idea is that the conical singularities arising in the GFT perturbative expansion, and characterising the triangulations not satisfying manifold-like conditions at their vertices [@DP-P], could be re-expressed as matter degrees of freedom, moreover possessing apparently the right type of (internal) structure for reproducing the matter content of the standard model [@cranematter].
Connections with other approaches
=================================
Where do group field theories stand with respect to other approaches to quantum gravity? We have already mentioned, albeit briefly, several links with other types of formalisms, and we would like to recapitulate them here. In doing so we sketch a (rather speculative at present) broader picture that sees GFTs as a [*generalised formalism for quantum gravity*]{}, in which other discrete and continuum approaches can be subsumed. We have already shown the way spin foam models arise universally as Feynman amplitudes for GFTs, so we don’t need to stress again how on the one hand the GFT approach can reproduce (at least in principle) any result obtained in the spin foam context, in its perturbative expansion around the GFT vacuum, while on the other hand providing much more than that in its, still unexplored, non-perturbative properties.
In the perspective on GFTs we proposed, they realise a [**local simplicial 3rd quantization of gravity**]{}, with its Feynman amplitudes interpreted as discrete path integrals for gravity on a given simplicial complex, and a sum over simplicial spacetimes of all topologies realised as a perturbative sum over Feynman graphs. The natural question is then what is the exact relationship with the more traditional path integral quantisations of simplicial gravity: quantum Regge calculus (see chapter by Williams) and dynamical triangulations (see chapter by Ambjorn et al.). Let us consider the first of the above. This uses a fixed triangulation of spacetime for defining its quantum amplitudes, and thus should be reproduced at the level of the GFT Feynman amplitudes for given Feynman graph. The geometric interpretation of the GFT variables (group representations, group elements, etc) appearing in the various models, is well understood [@causal; @review; @alexreview] and was discussed above. In general one would expect a generic spin foam model for gravity to give a path integral quantization of discrete gravity in 1st order formalism, i.e. treating on equal footing the (D-2)-volumes and the corresponding dihedral angles (equivalently, appropriate parallel transports of a Lorentz connection) as fundamental variables, as opposed to the 2nd order formulation of traditional Regge calculus in terms of edge lengths; but this may be consider a somewhat minor difference, even if its consequences deserve to be better understood. The main issue that needs to be solved in order to establish a clear link with the quantum Regge calculus approach has to do with the fact that the quantum amplitudes weighting histories of the gravitational field are given, in the latter approach, by the exponential of the Regge action for discrete gravity, and in the most studied spin foam models the connection between the quantum amplitudes and the Regge action is clear only in a particular regime and even there it is rather involved, as we have discussed. However, it seems plausible that the new generalised models of [@generalised], or a suitably modification of the same, can indeed give amplitudes with the same structure as in quantum Regge calculus, with a measure being uniquely determined by the choice of GFT action, thus making the connection with discrete gravity clear and at the same time subsuming the quantum Regge calculus approach within the GFT formalism, where it would arise for a specific type of models, in the perturbative sector only. The same type of amplitudes is needed also to establish a solid link with the dynamical triangulations approach, that in fact use again the exponential of the Regge action for weighting this time not the set of geometric data assigned to a single triangulation, but the combinatorial structure of the triangulation itself, which is treated as the only true dynamical variable within a sum over all possible triangulations of a given topology. Once the right type of amplitudes is obtained, the dynamical traingulations approach would once more arise as a subsector of the GFT formalism, if one could find the right way of trivialising the extra structure associated to each triangulation, thus dropping the now redundant sum over geometric data. Of course, more work would be needed then to impose the extra conditions (fixed slicing structure, absence of baby universe nucleation, etc) that seem to be needed in the modern version of the approach (see chapter by Ambjorn et al.) to have the right type of continuum limit. Work on this is in progress [@GFTsimplicial].
That a covariant path integral quantisation can be understood as being more general than the corresponding canonical/Hamiltonian one, and that this is even more true when a sum over topologies is implemented through some sort of 3rd quantisation, is well known. In the GFT framework, one expects then to be able to reproduce the setting and results of a canonical quantum gravity based on the same type of variables (group elements, interpreted as holonomies of a gravitational connection, or group representations, interpreted as (D-2)-volumes conjugate to them), and of quantum gravity states (spin networks): loop quantum gravity. How this can indeed be realised [@laurentgft], in a way that goes even beyond the present results of the LQG approach, has been shown above, when discussing the GFT definition of the canonical inner product. The main differences between the particular version of the LQG formalism that the GFT approach reproduces, and the traditional one (see chapter by Thiemann), are two: 1) the spin networks appearing as boundary states or observables in the GFT framework are inherently adapted to a simplicial context in that they are always D-valent in D spacetime dimensions, being dual to appropriate (D-1)-triangulations, while the spin networks arising in the [*continuum*]{} loop quantum gravity approach are of arbitrary valence; 2) the group used to label these states and their quantum evolution amplitudes in the GFT case is the Lorentz group of the corresponding dimension, which means, in dimension 4 and minkowskian signature, the non-compact group $SO(3,1)$, while LQG uses $SU(2)$ spin networks also in this context. The first of these differences is not so crucial, since on the one hand any higher valent spin network in LQG can be decomposed into lower-valent ones, and on the other hand any coarse graining procedure to be implemented to approximate simplicial (boundary) structures with continuum ones, and applied to the GFT boundary states or to GFT observables, would likely remove any restriction on the valence. The second difference is more troubleful, and establishing an explicit connection between the fully covariant GFT spin network structures and the $SU(2)$-based LQG ones is no easy task. However, lots of work has already been done on this issue [@eterasergei] (see chapter by Livine) and can be the starting point for 1) establishing a well-defined canonical formalism from the GFT structures first, and then 2) linking (more appropriately, reducing, probably through some sort of gauge fixing) this formalism to that of traditional LQG.
A fourth approach that can be linked to the GFT one is the causal set approach (see chapter by Henson). Recent work on spin foam models and GFT [@causal; @feynman; @generalised] has shown how the GFT Feynman amplitudes can be written in the form of models for causal evolution of spin networks [@fotinilee], by a correct implementation of causality requirements. A key step in doing this is the causal interpretation, in the Lorentzian context, of the GFT Feynman graph, this being [*directed graph*]{} i.e. a graph with directionsor arrows labelling its edges thus endowed with an orientation. In this interpretation, the vertices of the graph, i.e. the elementary GFT interactions, dual to D-simplices, are the fundamental spacetime events, and the links of the graph each connecting two such vertices, dual to (D-1)-simplices and corresponding to elementary propagation of degrees of freedom in the GFTs, represent the fundamental causal relations between spacetime events. A directed graph is very close to a causal set, from which it differs for just one, albeit important, property: it possibly includes closed loops of arrows. This, from the point of view of causal set theory, is a violation of causality, the microscopic discrete equivalent of a closed timelike loop in General Relativity, forbidden in the basic axioms defining the approach. No such restriction is imposed, a priori, on the corresponding GFT structures. There are several possible attitudes towards this issue from the GFT perspective: 1) it is possible that such configurations, even though they are present in the set of allowed configurations, are not relevant for the continuum approximation of them, i.e. they disappear or give a negligible contribution to the sum under the appropriate coarse graining procedure; 2) in the specific GFT models that will turn out to be of most interest for quantum gravity, Feynman graphs possessing such closed timelike loopsmay end up being assigned quantum amplitudes that strongly suppress them compared to other configurations; 3) one may be able to give a purely field-theoretic interpretation of such loops in the GFT context and then identify some sort of superselection rulesthat could eventually be imposed on the GFT perturbative expansion to suppress them; 4) finally, one may decide that there is no fundamental reason to ban such configuration from the admissible ones and rather find the way to interpret them physically and study their observable consequences. Only further work will tell. Finally, one more difference with the causal set framework is worth mentioning: due again to the simplicial setting in which they are realised, the GFT Feynman graphs have vertices of finite and fixed valence depending on the spacetime dimension, while the causal set vertices have no restriction on their valence. Once more, it is well possible that one has to welcome such restriction because it results in may be interpreted as one more sign of a fundamental spacetime discreteness, that may be attractive from both philosophical and physical reasons; at the same time, it is possible that such restriction on valence will be removed automatically and necessarily in the study of the continuum approximation of such discrete substratum for spacetime, by means of coarse graining procedures (as mentioned above regarding GFT kinematical states) or of renormalisation group arguments (e.g. inclusion of more types of interaction terms in the GFT action).
The GFT formalism is therefore able to encompass several other approaches to quantum gravity, each carrying its own set of ideas and techniques; strengthening the links with these other approaches will be, in our opinion, of great importance for both developing further the GFT framework itself, and also for making progress on the various open issues that such other approaches still have to face, thanks to their embeddinginto a different context, that naturally brings in a fresh new perspective on the same open issues.
Outlook
=======
Let us summarise. The group field theory approach aims to describe the dynamics of both spacetime geometry and topology down to the Planck scale, in a background independent and non-perturbative way (even if at present almost only the perturbation expansion around the complete vacuumis well understood), using a field-theoretic formalism. In essence, as discussed, a group field theory is a field theory over a group manifold, as for the mathematical formulation, and at the same time a field theory over a simplicial superspace (space of geometries), as for the physical interpretation. Thanks to the discrete description of geometry that the simplicial setting allows for, it corresponds to a [*local*]{} 3rd quantisation of gravity, in which the quantabeing created and annihilated in the quantum evolution are not universes, as in the traditional approach, but appropriately defined chunks of space.
What is particularly attractive, in our opinion, about this approach is the combination of [*orthodoxy*]{} in the mathematical language and technology used and of [*radicalness*]{} in the ideas that this language expresses. On the one hand, in fact, GFTs are almost ordinary field theories, defined on a group manifold with fixed metric and topology, and thus background dependent, speaking from a purely formal point of view, and using the ordinary language of fields and actions, of Feynman graphs, propagators and vertices of interaction, gauge symmetries, etc. This means that GFTs allow, at least in principle, to tackle any of the traditional questions in quantum gravity using techniques and ideas from QFT, thus making use of the vast body of knowledge and methods developed in a background dependent context that appeared for long time not directly applicable to quantum gravity research. On the other hand, the overall picture of spacetime and of gravity that this approach is based on, despite the traditional nature of the language used to express it, is definitely radical and suggests the following. There exist fundamental building blocks or atoms of space, which can be combined in all sorts of ways and can give rise to all sorts of geometry and topology of space. At the perturbative level spacetime is the (virtual) history of creation/annihilation of these fundamental atoms; it has no [*real*]{} existence, at least no more real existence in itself than each of the infinite possible interaction processes corresponding to individual Feynman graphs in any field theory; the interaction/evolution of these building blocks does not leave neither the geometry nor the topology of space fixed, but treats them on almost equal footing as dynamically evolving; the description of this evolution is necessarily background independent (from the point of view of spacetime) because spacetime itself is built from the bottom up and all of spacetime information has to be reconstructed from the information carried by the atomsand thus by the Feynman graphs. At the non-perturbative level, for what we can see given the present status of the subject, spacetime is simply not there, given that the non-perturbative properties of quantum gravity would be encoded necessarily either in the GFT action, and in the resulting equations of motion, or in the GFT partition function, and the related correlation functions, to be studied non-perturbatively, neither of which need any notion of spacetime to be defined or analysed (through instantonic calculations, or renormalisation group methods, or the like). Spacetime information is thus necessarily encoded in structures that do not use [*per se*]{} a notion of spacetime. Finally, there would be a [*fundamental discreteness*]{} of spacetime and a key role played by [*causality*]{}, in the pre-geometric sense of [*ordering*]{} related to the notion of orientation (so that it would probably be better to talk about pre-causality).
Many of these ideas had been of course put forward several times in the past, and occur in more than one other approach to quantum gravity, but the GFT formalism brings all of them together within a unique framework and, as said, expresses them in a rather conventional and powerful language; moreover, thanks to the possibility (yet to be realised) of subsuming many other approaches within the GFT one, all the ideas and techniques developed and the results obtained within these other approaches will maybe find a new role and application in the GFT setting.
Let us sketch also some examples of how traditional field theoretic methods can be used to tackle within a radically new perspective some crucial open issue in quantum gravity research. We have already mentioned several of these examples: the long-standing problem of solving the Hamiltonian constraint equation of canonical quantum gravity can be identified with the task of solving the classical GFT equations of motion, and how the other long-standing issue of defining a canonical inner product for quantum gravity states is turned into the task of analysing the tree level truncation of the (perturbative expansion of the) appropriate GFT. Also, the perturbation theory around such quantum gravity states would be governed, according to the above results, by the approximation of the GFT partition function around its classical solutions, and this suggests a new strategy for investigating the existence of gravitons (propagating degrees of freedom) in specific GFT/spin foam models. The most outstanding open issue that most of the discrete non-perturbative approaches to quantum gravity still face is however that of the continuum approximation. This problem has been formulated and tackled in a variety of ways, depending on the particular approach to quantum gravity under scrutiny. Obviously, given the role that formalisms like dynamical triangulations, quantum Regge calculus, causal sets or loop quantum gravity can play within the group field theory framework, all the various techniques developed for them can be adapted to the GFTs. However, the field theory language that is at the forefront of the GFT approach suggests once more new perspectives. Let us sketch them briefly. The continuum approximation issue can be seen as the search for the answer to two different types of question. a) What is the best procedure to approximate a discrete spacetime, e.g. a simplicial complex, with a continuum manifold, and to obtain some effective quantum amplitude for each geometric configuration from the underlying fundamental discrete model? In the context of spin foam models, the most directly linked with the GFT approach, this amounts to devising a background independent procedure for coarse grainingthe spin foam 2-complexes and the corresponding amplitudes [@robert; @fotini] to obtain a smooth approximation of the same. b) If a continuum spacetime or space are nothing else than some sort of condensateof fundamentally discrete objects, as in some emergent gravityapproaches (see chapter by Dreyer), and as suggested by condensed matter analog models of gravity [@volovik; @analog], what are these fundamental constituents? what are their properties? what kind of (necessarily background independent) model can describe them and the whole process of condensation? What are the effective hydrodynamic variables and what is their dynamics in this condensed or fluid phase? How does it compare to GR?
For what concerns the first (set of) question(s), the GFT approach offers a potentially decisive reinterpretation: since spin foam are nothing else than Feynman graphs of a GFT, and that spin foam models are nothing else than their corresponding Feynman amplitudes, the coarse graining of a spin foam model, either performed as outlined in [@robert] or with the techniques introduced in [@fotini], corresponds exactly to the [*perturbative renormalisation*]{} of the corresponding GFT. On the one hand this suggests to deal with the problem of continuum approximation of spin foams using all the perturbative and non-perturbative renormalisation group techniques from ordinary field theory adapted to the GFT case. On the other hand gives a further justification for the idea, proposed in [@fotini], that the Connes-Kreimer Hopf algebra of renormalisation developed for QFT could be the right type of formalism to use in such a quantum gravity context.
As for the second (set of) question(s), the GFT approach identifies uniquely the basic building blocks of a quantum space, those that could be responsible for the kind of condensationprocess or the transition to a fluid phase at the root of the emergence of a smooth spacetime in some approximation and physical regime, and gives a precise prescription for their classical and quantum dynamics, that can now be investigated. In particular, one could develop a statistical mechanics picture for the dynamics of the GFT atomsof space, and then the above idea of a condensationor in general of the possibility of an hydrodynamic description could be tested in specific GFT models, and in very concrete and precise terms, using once more usual techniques from statistical field theory (as applied for example in condensed matter systems). A more detailed discussion of the possible development of GFTs along these lines can be found in [@ioGFTemergence].
Whether any of the above ideas will be realised, and whether other, not yet imagined, possibilities for new developments will become manifest in the near future, only further work will tell. In our opinion, however, it is already clear that the GFT approach can be the right framework for asking the most fundamental questions about Quantum Gravity, and for finding answers to them.
[99]{} D. Oriti, in [*Mathematical and Physical Aspects of Quantum Gravity*]{}, B. Fauser, J. Tolksdorf and E. Zeidler, eds., Birkhaeuser, Basel (2006); gr-qc/0512103; L. Freidel, Int. J. Theor. Phys. 44, 1769 (2005), hep-th/0505016; S. Giddings, A. Strominger, Nucl. Phys. B 321, 481 (1989); T Banks, Nucl. Phys. B 309, 493 (1988); M. McGuigan, Phys. Rev. D 38, 3031 (1988); P. Peldan, Class. Quant. Grav. 11, 1087 (1994), gr-qc/9305011; J. C. Baez, J. W. Barrett, Adv. Theor. Math. Phys. 3, 815 (1999), gr-qc/9903060; L. Freidel, K. Krasnov, R. Puzio, Adv. Theor. Math. Phys. 3, 1289 (1999), hep-th/9901069; R. De Pietri, L. Freidel, K. Krasnov, C. Rovelli, Nucl. Phys. B 574, 785 (2000), hep-th/9907154; A. Mikovic, Class. Quant. Grav. 18, 2827 (2001), gr-qc/0102110; J. W. Barrett, L. Crane, Class. Quant. Grav. 17, 3101 (2000), gr-qc/9904025; D. Oriti, Rept. Prog. Phys. 64, 1489 (2001), gr-qc/0106091; A. Perez, Class. Quant. Grav. 20, R43 (2003), gr-qc/0301113; D. Oriti, J. Ryan, A. Youssef, [ *Hamiltonian and Fock sructure for generalised group field theories*]{}, in preparation; A. Baratin, L. Freidel, E. Livine, [*Solving spin foam models: instantons and group field theory*]{}, in preparation; M. Reisenberger, C. Rovelli, Class. Quant. Grav. 18, 121 (2001), gr-qc/0002095; J. Ambjorn, J. Jurkiewicz, R. Loll, Phys. Rev. D 72, 064014 (2005), hep-th/0505154; J. Ambjorn, B. Durhuus, T. Jonsson, Mod. Phys. Lett. A 6, 1133 (1991); S. Coleman, Nucl. Phys. B 310, 643 (1988); L. Freidel, D. Louapre, Nucl. Phys. B 662, 279 (2003), gr-qc/0212001; R. De Pietri, C. Petronio, J. Math. Phys. 41, 6671 (2000), gr-qc/0004045; A. Perez, C. Rovelli, Phys. Rev. D 63, 041501 (2001), gr-qc/0009021; A. Perez, C. Rovelli, Phys. Rev. D 64, 064002 (2001), gr-qc/0011037; E. Livine, A. Perez, C. Rovelli, Class. Quant. Grav. 20, 4425 (2003), gr-qc/0102051; D. Oriti, C. Rovelli, S. Speziale, Class. Quant. Grav. 22, 85 (2005), gr-qc/0406063; D. Boulatov, Mod. Phys. Lett. A 7, 1629 (1992), hep-th/9202074; H. Ooguri, Mod. Phys. Lett. A 7, 2799 (1992), hep-th/9205090; D. Oriti, Phys. Rev. D 73, 061502 (2006), gr-qc/0512069; L. Freidel, D. Louapre, Phys. Rev. D 68, 104004 (2003), hep-th/0211026; D. Oriti, [*Spin foam models of quantum spacetime*]{}, PhD thesis, University of Cambridge (2003), gr-qc/0311066; J. Halliwell, J. Hartle, Phys. Rev. D 43, 1170 (1991); J. W. Barrett, R. M. Williams, Adv. Theor. Math. Phys. 3, 209 (1999), gr-qc/9809032; E. Livine, D. Oriti, Nucl. Phys. B 663, 231 (2003), gr-qc/0210064; D. Oriti, Phys. Rev. Lett. 94, 111301 (2005), gr-qc/0410134; K. Krasnov, hep-th/0505174; E. Livine, D. Oriti, JHEP 0702, 092 (2007), gr-qc/0512002; H. Grosse, K. Schlesinger, Phys. Lett. B 528, 106 (2002), hep-th/0109080; A. Mikovic, Class. Quant. Grav. 19, 2335 (2002), hep-th/0108099; L. Freidel, D. Louapre, Class. Quant. Grav. 21, 5685 (2004), hep-th/0401076; D. Oriti, J. Ryan, Class. Quant. Grav. 23, 6543 (2006), gr-qc/0602010; L. Freidel, D. Oriti, J. Ryan, submitted for publication, gr-qc/0506067; W. Fairbairn, E. Livine, gr-qc/0702125; J. W. Barrett, Class. Quant. Grav. 23, 137 (2006), gr-qc/0502048; A. Baratin, L. Freidel, gr-qc/0604016; A. Baratin, L. Freidel, [*Hidden quantum gravity in 4d Feynman diagrams: emergence of spin foams*]{}, in preparation; D. Oriti, J. Ryan, [*Group field theory formulation of topological gravity coupled to strings*]{}, in preparation; J. Baez, A. Perez, gr-qc/0605087; R. Oeckl, D. Oriti, J. Ryan, [*A group field theory for quantum gravity coupled to gauge fields*]{}, in preparation; L. Crane, gr-qc/0110060; S. Alexander, L. Crane, M. D. Sheppeard, gr-qc/0306079; D. Oriti, [*Group field theory and simplicial gravity*]{}, in preparation; D. Oriti, T. Tlas, [*GFT and causal 3d spin foam models*]{}, in preparation; D. Oriti, H. Pfeiffer, Phys. Rev. D 66, 124010 (2002), gr-qc/0207041; S. Alexandrov, E. Livine, Phys. Rev. D 67, 044009 (2003), gr-qc/0209105; F. Markopoulou, L. Smolin, Nucl. Phys. B 508, 409 (1997), gr-qc/9702025; R. Oeckl, Nucl. Phys. B 657, 107 (2003), gr-qc/0212047; F. Markopoulou, Class. Quant. Grav. 20, 777 (2003), gr-qc/0203036; G. Volovik, Phys. Rept. 351, 195 (2001), gr-qc/0005091; C. Barcelo, S. Liberati, M. Visser, Living Rev. Rel. 8, 12 (2005), gr-qc/0505065; D. Oriti, hep-th/0612301.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Using the $\hbar$-expansion of the Green’s function of the Hartree-Fock-Bogoliubov equation, we extend the second-order Thomas-Fermi approximation to generalized superfluid Fermi systems by including the density-dependent effective mass and the spin-orbit potential. We first implement and examine the full correction terms over different energy intervals of the quasiparticle spectra in calculations of finite nuclei. Final applications of this generalized Thomas-Fermi method are intended for various inhomogeneous superfluid Fermi systems.'
author:
- 'J.C. Pei'
- Na Fei
- 'Y.N. Zhang'
- 'P. Schuck'
nocite: '[@*]'
title: ' Generalized Second-Order Thomas-Fermi Method for Superfluid Fermi Systems '
---
Introduction
============
Semiclassical treatments of quantum systems are always of broad interests. Examples are nuclei, metallic clusters, cold atomic gases, neutron star crusts, etc. This is particularly useful for large systems which challenge the capacities of supercomputers. In this context, the Thomas-Fermi approximation (or Local Density Approximation) has been extended to higher orders to include as much as possible quantum corrections. When pairing correlations are included, the extended Thomas-Fermi method for superconducting/superfluid systems becomes much more complicated than for normal fluid systems. In the past few years, only a very limited number of works concerning the formulism of the second-order superfluid Thomas-Fermi approach have been given in the literature [@schuck; @gross; @csordas] with either no or very limited practical calculations [@csordas]. So further studies and applications are very desirable. Furthermore, the density-dependent effective mass and the spin-orbit potential have not been considered yet, which are essential ingredients in, e.g., generalized nuclear density functionals such as the widely used Skyrme nuclear density functionals [@vautherin; @PGR]. Also in cold Fermi gases, the spin-orbit coupling [@wangpj] and the density-dependent effective mass [@bulgac08] are currently very interesting.
The Hartree-Fock-Bogoliubov (HFB) equations [@manybody], or Bogoliubov-de Gennes equations [@gennes], have been a framework for superfluid Fermi systems. The first $\hbar$-expansion of the HFB density matrix is based on the Wigner Kirkwood transformation of the Bloch propagator [@schuck], however, without practical applications. Later, the $\hbar$-expansion of the Green’s function of HFB solutions has been derived for superconducting systems [@gross]. More recently, a third paper based on the Green’s function method appeared with some applications to cold atomic systems [@csordas]. We also noticed that the coarse graining treatment in superfluid Fermi gases [@simonucci] has a connection to the second-order Thomas-Fermi method.
Our objective of the present work is to extend the formulism to include the density-dependent effective mass and the spin-orbit effect together with the second-order Thomas Fermi approximation for superfluid systems. Although such extensions [@brack] and even tensor interactions [@bartel] have been achieved long time ago for non-superfluid systems, they are not yet elaborated for superfluid systems. In this work, we first derive the full correction terms based on the $\hbar$-expansion of the Green’s function corresponding to the HFB equations. In addition, we implement these second-order correction terms in numerical examinations and compare the results with fully self-consistent Skyrme HFB calculations of nuclei where the density-dependent effective mass and the spin-orbit potential are included. In the past the Thomas-Fermi approximation with Gogny force has already been applied to superfluid nuclei in [@kucharek1; @kucharek2].
Recently, the adaptive multi-resolution 3D coordinate-space HFB method has been developed for complex superfluid systems but the computation of continuum states is still very costly [@Pei14]. The coordinate-space HFB calculations are very useful for describing weakly-bound systems and complex-shaped systems [@Pei14], in contrast to the conventional HFB approaches based on harmonic-oscillator basis expansions. A promising way to address large systems is to adopt the hybrid HFB calculations [@pei2011], i.e., using the Thomas-Fermi approximation for high-energy quasi-continuum states and the discretized HFB solutions for low-energy states. The hybrid HFB has first been adopted with the zeroth-order Thomas-Fermi method for cold atomic systems [@reidl; @xjliu]. In this respect, large superfluid systems such as the trapped cold atoms with $10^{5-6}$ particles [@partridge; @shin] and exotic neutron star crusts in large 3D cells [@mazy; @stone] are numerically very challenging. In such inhomogeneous systems, the violation of zeroth-order Thomas-Fermi approximation related to finite-size effects and complex spatial topologies could be non-negligible [@Sensarma].
Formulation
===========
The HFB equation in the coordinate-space representation takes the form [@Pei08]: $$\label{hfb1}
\left[
\begin{array}{cc}
h({\boldsymbol{r}})& \Delta({\boldsymbol{r}}) \\
\Delta^{*}({\boldsymbol{r}}) & -h({\boldsymbol{r}}) \\
\end{array}
\right]
\left[
\begin{array}{c}
U_{k}({\boldsymbol{r}}) \\
V_{k}({\boldsymbol{r}}) \\
\end{array}
\right]
= E_k
\left[
\begin{array}{c}
U_{k}({\boldsymbol{r}}) \\
V_{k}({\boldsymbol{r}}) \\
\end{array}
\right],$$ where $h=h_{\rm HF}-\lambda$; $h_{\rm HF}$ is the Hartree-Fock Hamiltonian; $\lambda$ is the Fermi energy (or chemical potential); $\Delta$ is the pairing potential; $U_k$ and $V_k$ are the upper and lower components of quasi-particle wave functions, respectively; $E_k$ is the quasi-particle energy.
The theoretical derivation of the second-order superfluid Thomas-Fermi approximation starts with the solution of the HFB equation using the Green’s function method (or Gorkov equation) [@gross; @csordas],
$$\left[ \omega \textbf{I}-
\Big(\begin{array}{cc}h({\boldsymbol{r}}) &\Delta({\boldsymbol{r}}) \vspace{3pt}\\
\Delta^{*}({\boldsymbol{r}}) & -h({\boldsymbol{r}})
\end{array}\Big)
\right]\textbf{G}({\boldsymbol{r}},{\boldsymbol{r}}',\omega)=\hbar \textbf{I}\delta({\boldsymbol{r}}-{\boldsymbol{r}}').$$
The Wigner transformation of the Green’s function form of the HFB equation can be written as: $$\displaystyle\Lambda [(\omega \textbf{I}-\textbf{H}(\textbf{R},\textbf{p})), \textbf{G}(\textbf{R}',\textbf{p}',\omega)]=\hbar \textbf{I},
\label{HFBe}$$ where ${\rm \textbf{H}}$ refers to the HFB Hamiltonian, and the operator $\Lambda$ can be expanded in powers of $\hbar$: $$\Lambda=\sum_{j=0}^{\infty}\hbar^j\Lambda_j=\sum_{j=0}^{\infty}\frac{\hbar^j}{j!}(\frac{-i}{2})^j\Big[(\nabla_{\textbf{R}}\cdot\nabla_{\textbf{p}'}
-\nabla_{\textbf{p}}\cdot\nabla_{\textbf{R}'})\Big]^j.
\label{operator}$$
Similar to non-superfluid systems, the spin-orbit potential can be included as a higher-order term in terms of $\hbar$ in the Hamiltonian [@grammaticos], $$\textbf{H}(\textbf{R},\textbf{p})=\textbf{H}_c(\textbf{R},\textbf{p})+\textbf{H}_{so}(\textbf{R},\textbf{p}),$$ in which ${\rm \textbf{H}}_c$ and ${\rm \textbf{H}}_{so}$ denote the central term and the spin-orbit term of the HFB Hamiltonian, respectively. The spin-orbit Hamiltonian is defined as $$\textbf{H}_{so}=
\Big(\begin{array}{cc} h_{so} & 0 \vspace{3pt}\\
0 & - h_{so}
\end{array}\Big)
, \hspace{8pt} h_{so} = -\frac{1}{\hbar}\textbf{w}({\boldsymbol{r}})\cdot \mathbf{\sigma} \times \mathbf{p}$$
The extended Thomas-Fermi method consists in expanding the Green’s function in Eq.(\[HFBe\]) in powers of $\hbar$ , i.e., $$\textbf{G}(\textbf{R},\textbf{p},\omega)=\hbar\sum_{n=0}^{\infty}\textbf{G}_n(\textbf{R},\textbf{p},\omega)\hbar^n .$$
The expansion of Eq.(\[HFBe\]) in terms of $\hbar$ to the second-order can be written as,
$$\begin{array}{l}
(\omega \textbf{I}-\textbf{H}_c)\textbf{G}_0=\textbf{I},
\end{array}$$
$$\begin{array}{ll}
(\omega \textbf{I}-\textbf{H}_c)\textbf{G}_1&+\Lambda_1[(\omega \textbf{I}-\textbf{H}_c), \textbf{G}_0] \vspace{5pt}\\
&+\Lambda_0[-\textbf{H}_{so}, \textbf{G}_0]=0,\\
\end{array}$$
$$\begin{array}{ll}
(\omega \textbf{I}-\textbf{H}_c)\textbf{G}_2&+\Lambda_2[(\omega \textbf{I}-\textbf{H}_c), \textbf{G}_0] \vspace{5pt}\\
&+ \Lambda_1[(\omega \textbf{I}-\textbf{H}_c), \textbf{G}_1] \vspace{5pt}\\
& +\Lambda_1[-\textbf{H}_{so}, \textbf{G}_0]\vspace{5pt}\\
& +\Lambda_0[-\textbf{H}_{so}, \textbf{G}_1]=0 .\\
\end{array}$$
\[HFBe2\]
The various expansion terms of $\textbf{G}$ can be obtained order by order. Compared to Ref.[@gross], we include the spin-orbit potential entailing more terms to appear in Eq.(\[HFBe2\]). First, $\textbf{G}_0$ corresponds to the zeroth-order solutions and does not change due to the spin-orbit potential. The matrix elements of $\textbf{G}_0$ are written as, $$\Big(\begin{array}{cc}\textbf{G}_0^{(\uparrow\uparrow)} & \textbf{G}_0^{(\uparrow\downarrow)} \vspace{3pt}\\
\textbf{G}_0^{(\downarrow\uparrow)} & \textbf{G}_0^{(\downarrow\downarrow)}
\end{array}\Big)=\frac{1}{D}
\Big(\begin{array}{cc}\omega+h & \Delta \vspace{3pt}\\
\Delta & \omega-h
\end{array}\Big),
\label{hfbe3}$$ where the $\textbf{G}_0^{(\uparrow\uparrow)}$ and $\textbf{G}_0^{(\uparrow\downarrow)}$ correspond to the normal density and pairing density, respectively. The abbreviation $D$ for the denominator in Eq.(\[hfbe3\]) is defined as $D=(\omega^{2}-h^{2}-\Delta^2)$.
$\textbf{G}_1$ changes due to the spin-orbit potential and is written as $$\begin{split}
&\textbf{G}_{1,\textbf{so}}^{(\uparrow\uparrow)}=\frac{1}{D^2} h_{so}[(\omega+h)^2-\Delta^2] \vspace{5pt} \\
&\textbf{G}_{1,\textbf{so}}^{(\uparrow\downarrow)} = \frac{1}{D^2}h_{so}\big( 2h\Delta \big)\\
\end{split}$$ Note that in this work we are treating systems in the case of real pairing potentials. The formulism with complex pairing potentials can be derived similarly but would be more complicated [@gross].
$\textbf{G}_2$ is quite complicated [@gross; @csordas] and the additional terms due to the spin-orbit potential are: $$\begin{split}
&\displaystyle\textbf{G}_{2,\textbf{so}}^{(\uparrow\uparrow)}=\displaystyle\frac{h_{so}^2}{D^3}\big[(\omega+h)^3-(\omega+3h)\Delta^2\big]
\vspace{5pt}\\\vspace{5pt}
&\displaystyle\textbf{G}_{2,\textbf{so}}^{(\uparrow\downarrow)}=\displaystyle\frac{h_{so}^2}{D^3}\big[(\omega^2-\Delta^2+3h^2)\Delta\big]
\end{split}$$ Note that terms involving derivatives of the spin-orbit potential have been omitted. Actually the terms in $\textbf{G}_2$ with $h_{so}$ do not contribute and only terms with $h_{so}^2$ do.
Next, the contributions to normal density and pairing density can be obtained by an appropriate integration of $\textbf{G}_2$ in the complex $\omega$ plane. For convenience of writing the expressions, we introduce a differential operator $\overleftrightarrow{\Lambda}$ as in Ref. [@schuck], $$\overleftrightarrow{\Lambda}=(\overleftarrow{\nabla}_{\textbf{R}}\overrightarrow{\nabla}_{\textbf{p}}-\overleftarrow{\nabla}_{\textbf{p}}\overrightarrow{\nabla}_{\textbf{R}}),
\label{poisson}$$ which is the operator in the Poisson bracket. Eq.(\[poisson\]) can be applied repeatedly and is related to Eq.(\[operator\]) without expansion coefficients. The 2$^{\rm nd}$-order normal density contribution terms are given as, $$\begin{array}{lcl}
\rho_2(\textbf{R},\textbf{p})&=&\eta_1(h\overleftrightarrow{\Lambda}^2h)+\eta_2(h\overleftrightarrow{\Lambda}^2\Delta)+\eta_3(\Delta\overleftrightarrow{\Lambda}^2\Delta) \vspace{5pt}\\
&& +\eta_4(h(\overleftrightarrow{\Lambda} h)^2)+\eta_5(h(\overleftrightarrow{\Lambda} h)(\overleftrightarrow{\Lambda} \Delta))\vspace{5pt}\\
&& + \eta_6(h(\overleftrightarrow{\Lambda} \Delta)^2)+\eta_7(\Delta(\overleftrightarrow{\Lambda} h)^2)\vspace{5pt}\\
&& + \eta_8 (\Delta(\overleftrightarrow{\Lambda} h)(\overleftrightarrow{\Lambda}\Delta))+\eta_9(\Delta(\overleftrightarrow{\Lambda}\Delta)^2)\vspace{5pt}\\
&& + \eta_{10} (h\overleftrightarrow{\Lambda}\Delta)^2+\displaystyle\frac{3h\Delta^2h_{so}^2}{4E^5}
\end{array}
\label{ndensity}$$ and the 2$^{\rm nd}$-order pairing density, $$\begin{array}{lcl}
\tilde{\rho}_2(\textbf{R},\textbf{p})&=&\theta_1(h\overleftrightarrow{\Lambda}^2h)+\theta_2(h\overleftrightarrow{\Lambda}^2\Delta)+\theta_3(\Delta\overleftrightarrow{\Lambda}^2\Delta)\vspace{5pt} \\
&& +\theta_4(h(\overleftrightarrow{\Lambda }h)^2)+\theta_5(h(\overleftrightarrow{\Lambda} h)(\overleftrightarrow{\Lambda} \Delta))\vspace{5pt}\\
&& + \theta_6(h(\overleftrightarrow{\Lambda} \Delta)^2)+\theta_7(\Delta(\overleftrightarrow{\Lambda} h)^2)\vspace{5pt}\\
&& + \theta_8 (\Delta(\overleftrightarrow{\Lambda}h)(\overleftrightarrow{\Lambda}\Delta))+\theta_9(\Delta(\overleftrightarrow{\Lambda}\Delta)^2)\vspace{5pt}\\
&& + \theta_{10} (h\overleftrightarrow{\Lambda}\Delta)^2+\displaystyle\frac{(2h^2-\Delta^2)\Delta h_{so}^2}{4E^5}
\end{array}
\label{pdensity}$$ In Eq.(\[ndensity\]) and Eq.(\[pdensity\]), the coefficients $\eta_i$ and $\theta_i$ have been given in Refs. [@schuck; @gross; @csordas], although they have different conventions. We refer to the Appendix for explicit expressions of these terms. Note that some terms vanish with the application of operators $\Lambda_j$, since we suppose a momentum-independent pairing potential. The additional contributions due to the spin-orbit potential are given in Eqs.(\[ndensity\]) and (\[pdensity\]) explicitly, in which $E$ denotes the HFB quasiparticle energies and takes the form as $$E=\sqrt{
(h_{\rm HF}(\mathbf{R},\mathbf{p})-\lambda)^2 +\Delta(\mathbf{R},\mathbf{p})^2} .$$
Up to now, we have not explicitly considered the density-dependent effective mass. The density-dependent effective mass is involved in the central part of the single-particle Hamiltonian, $$h_c=\frac{1}{2}f({\boldsymbol{r}})p^2+U({\boldsymbol{r}})~~~~~\mbox{with}~~~m^{*}({\boldsymbol{r}})=1/f({\boldsymbol{r}}).$$ In fact, the expressions of the $2^{\rm nd}$-order Thomas-Fermi do not need to be modified. However, three additional expansion terms appear compared to the cases without effective mass:
$$h\overleftrightarrow{\Lambda}^2h=\displaystyle fp^2\nabla^2f+2f\nabla^2U-\frac{2}{3}p^2(\nabla f)^2\\$$
$$\begin{array}{lcl}
h(\overleftrightarrow{\Lambda}h)^2&=&\displaystyle-\frac{1}{12}fp^4(\nabla f)^2+f(\nabla U)^2+\frac{1}{6}f^2p^4\nabla^2f\vspace{5pt}\\
&&\displaystyle+\frac{1}{3}f^2p^2\nabla^2U+\frac{1}{3}fp^2(\nabla f)(\nabla U) \\
\end{array}\\$$
$$h(\overleftrightarrow{\Lambda} h)(\overleftrightarrow{\Lambda} \Delta)=\displaystyle\frac{1}{6}fp^2(\nabla f)(\nabla \Delta)+f(\nabla U)(\nabla \Delta)$$
Finally, the normal density (and pairing density) in real space can be obtained by integrals over quasiparticle energies $E$ or single-particle energies $\varepsilon$ up to a certain cutoff $$\begin{array}{lcl}
\rho_2({\boldsymbol{r}})&=&\displaystyle\int\frac{\rho_2(\textbf{R},\textbf{p})}{(2\pi \hbar)^3} d^3p \vspace{5pt}\\
&=&\displaystyle\int_{E0}^{Ec}\frac{\rho_2(\textbf{R},\textbf{p})}{(2\pi \hbar)^3} \frac{pE}{f\sqrt{E^2-\Delta^2}}dE \vspace{5pt}\\
&=&\displaystyle\int_{\varepsilon 0}^{\varepsilon c}\frac{\rho_2(\textbf{R},\textbf{p})}{(2\pi \hbar)^3} \frac{p}{f}d\varepsilon
\end{array},$$ where $(E_0, E_c)$ or $(\varepsilon_0, \varepsilon_c)$ defines an energy interval in which the Thomas-Fermi approximation is used.
The 2$^{\rm nd}$-order kinetic density correction in real spaces is obtained as, $$\tau_2({\boldsymbol{r}})= \displaystyle \frac{1}{4}\nabla^2\rho +\int_{E0}^{Ec}\frac{\rho_2(\textbf{R},\textbf{p})}{(2\pi \hbar)^3} \frac{p^3E}{f\sqrt{E^2-\Delta^2}}dE$$
The 2$^{\rm nd}$-order spin-orbit density correction is given as: $$\textbf{J}_2({\boldsymbol{r}})= \displaystyle \int\frac{\rho(\textbf{R},\textbf{p})}{(2\pi \hbar)^3} (\textbf{p}\times {\bf{\sigma}}) d^3p$$ Due to the spin degeneracy, various correction densities should be multiplied by a factor of 2. Note that there is no zeroth-order spin-orbit density. Since the liner term of $(\textbf{p}\times {\bf{\sigma}})$ doesn’t contribute in the integration, the final expression of $\textbf{J}_2({\boldsymbol{r}})$ can be obtained by considering only the $\textbf{G}_{1,\textbf{so}}$’s contribution, $$\textbf{J}_2({\boldsymbol{r}})= \frac{\textbf{w}({\boldsymbol{r}})}{(\pi \hbar)^2}\displaystyle \int_{E0}^{Ec} \frac{-\Delta^2}{3E^3} \frac{p^3E}{f\sqrt{E^2-\Delta^2}}dE.$$
The limit of zero pairing gap
=============================
In the limit of the zero pairing gap, the derived generalized 2$^{\rm nd}$-order Thomas-Fermi approximation for the normal state should be recovered. For the normal density and pairing density, this has been demonstrated in Ref. [@gross]. In this work, we have to examine the additional terms of spin-orbit contributions and the spin-orbit density $\textbf{J}({\boldsymbol{r}})$ in the limit of zero pairing gap. For the spin-orbit contribution to normal density: $$\lim_{\Delta \to 0} \displaystyle\frac{3h\Delta^2h_{so}^2}{4E^5} = -\frac{1}{2}h_{so}^2\delta'(h).$$ The resulting density expression is equivalent to the 2$^{\rm nd}$-order Thomas-Fermi approximation of the non-superfluid state [@grammaticos]. Obviously, the resulting pairing density due to the spin-orbit contribution becomes zero in the limit of the zero pairing gap. Similarly, the spin-orbit density in the limit can be obtained, $$\begin{array}{ccl}
\lim\limits_{\Delta \to 0} \textbf{J}_2({\boldsymbol{r}}) &= &\displaystyle-2\frac{\textbf{w}({\boldsymbol{r}})}{(\pi \hbar)^2} \displaystyle \int_{\varepsilon0}^{\varepsilon c} \frac{\delta(h)}{3} \frac{p^3}{f}d\varepsilon \vspace{5pt} \\
&=&\displaystyle -2\rho({\boldsymbol{r}})\frac{\textbf{w}({\boldsymbol{r}})}{f({\boldsymbol{r}})}\\
\end{array}$$ which also agrees with the expression of the normal state [@grammaticos].
Implementation and calculations
===============================
![(Color online) The 0$^{\rm th}$-order (dotted line) and 2$^{\rm nd}$-order Thomas-Fermi contributions (0$^{\rm th}$-order plus 2$^{\rm nd}$-order ) to the neutron pairing density from 55 to 65 MeV, as well as the corresponding coordinate-space HFB solutions (as labeled by ‘Box’). The 2$^{\rm nd}$-order Thomas-Fermi contribution without explicitly considering the effective mass is also shown, as labeled by ‘2TF-(w/o f)’. []{data-label="fig-pdensity"}](pdensity.eps "fig:"){width="48.00000%"}\
![(Color online) The 2$^{\rm nd}$-order Thomas-Fermi terms corresponding to the pairing density in Fig.\[fig-pdensity\]. The terms denoted by $\theta_i$ as listed in Eq.(\[pdensity\]). The spin-orbit contribution (dashed line) and the summed correction (dotted line) are shown. []{data-label="fig-pdensity-t"}](pdensity-terms.eps "fig:"){width="48.00000%"}\
![(Color online) Similar to Fig.\[fig-pdensity\], but for the 0$^{\rm th}$-order and 2$^{\rm nd}$-order Thomas-Fermi contributions to the neutron normal density. The normal density from self-consistent coordinate-space HFB solutions is labeled by ‘Box’. []{data-label="fig-density"}](density.eps "fig:"){width="48.00000%"}\
![(Color online) Similar to Fig.\[fig-pdensity-t\], but for the 2$^{\rm nd}$-order contribution terms to the normal density as listed in Eq.(\[ndensity\]). []{data-label="fig-density-t"}](density-terms.eps "fig:"){width="48.00000%"}\
![(Color online) Similar to Fig.\[fig-pdensity\], but for the 0$^{\rm th}$-order and 2$^{\rm nd}$-order Thomas-Fermi contributions to the neutron kinetic density. []{data-label="fig-kdensity"}](kdensity.eps "fig:"){width="48.00000%"}\
![(Color online) Similar to Fig.\[fig-pdensity\], but for the 2$^{\rm nd}$-order Thomas-Fermi contribution (0$^{\rm th}$-order doesn’t contribute) to the divergence of the neutron spin-orbit density $\nabla\cdot \textbf{J}$. []{data-label="fig-sodensity"}](sodensity.eps "fig:"){width="48.00000%"}\
![(Color online) (a) The 2$^{\rm nd}$-order Thomas-Fermi terms to the neutron normal density with an energy cutoff from 25 MeV to 65 MeV; (b) the 2$^{\rm nd}$-order terms to the pairing density; (c) the 2$^{\rm nd}$-order terms to the pairing density with a pairing strength 10 times larger. []{data-label="fig-density-25"}](densities-25.eps "fig:"){width="45.00000%"}\
 plotted on a logarithmic scale to show surface distributions, compared to the coordinate-space HFB solutions. []{data-label="fig-psur"}](psurface.eps "fig:"){width="48.00000%"}\
To demonstrate applications of the 2$^{\rm nd}$-order Thomas-Fermi corrections, we perform self-consistent calculations of a finite deformed nucleus $^{^{238}}$U. For the particle-hole interaction channel, the often used Skyrme force SLy4 [@sly4] is adopted. For the pairing channel, the volume pairing, i.e., a delta interaction for the pairing force is adopted with a reasonable pairing strength of 200 MeV fm$^{-3}$. The Skyrme density functional contains the density-dependent effective mass and the spin-orbit potential, providing an ideal testing ground for the generalized 2$^{\rm nd}$-order Thomas-Fermi approximation.
Non-self-consistent Calculations
--------------------------------
The nuclear HFB solutions have deep-hole states [@pei2011] which correspond to deep-bound states in Hartree-Fock+BCS solutions. In the HFB approach, deep-hole states are narrow quasiparticle resonances with large quasiparticle energies $E_k$ and large occupation numbers $v_k^2$. These states contain important shell effects and can not be well described by the Thomas-Fermi approximation, since the latter does not account for shell effects. Therefore, we first aim to compare the contributions of various partial densities corresponding to high quasiparticle energies. Our final goal is to combine the HFB solutions at low-energy quasiparticle states and 2$^{\rm nd}$-order corrections in the quasiparticle high-energy region. In such a hybrid way, the computational costs for large superfluid Fermi systems can be remarkably reduced. Indeed, the high-energy states behave like quasi-continuum states and are distinctly different from low-energy states. One can suppose that the former can be well approximated by the Thomas-Fermi expressions.
First, self-consistent HFB calculations are performed for $^{^{238}}$U, then the zeroth-order and second-order Thomas-Fermi approximations are investigated by performing one iteration with densities from full HFB solutions. The HFB calculations are performed in the cylindrical coordinate space with the HFB-AX solver [@Pei08] that uses B-spline techniques. For calculations employing large box sizes and small lattice spacings, the discretized continuum spectra would be very dense, providing good resolutions. The pairing densities $\tilde{\rho}(r,z)$ are plotted along the diagonal coordinate $R=\sqrt{r^2+z^2}$$|_{(r=z)}$ rather than the axes, to avoid numerical errors at boundaries. We compare various pairing densities in Fig.\[fig-pdensity\]. In this work, our discussions are restricted to neutrons since pairing is absent in protons in $^{^{238}}$U. The pairing densities of the full self-consistent HFB calculation are summed for quasiparticle energies ranging from 55 to 65 MeV. This is the high energy window for which we want to study the Thomas-Fermi approximation, in discarding the influence of deep-hole states. The upper cutoff is taken as 65 MeV in all our results. The converged HFB densities are then inserted into the Thomas-Fermi expressions for one iteration and the zeroth and second order Thomas-Fermi densities are also summed (integrated) over the same energy interval. In this way the quantities shown in Fig.\[fig-pdensity\] are perturbative. It can be seen that the second-order contribution is very small but can slightly improve the pairing density distribution. Indeed, as shown in Eq.(\[ndensity\]) and Eq.(\[pdensity\]), the second-order corrections involve terms of order $1/E^5$ and $1/E^7$, which are suppressed in the high-energy region. With the derivative terms of effective mass, the second-order corrections are only slightly modified in the high energy region.
In Fig.\[fig-pdensity-t\], the second-order contributions from different terms in Eq. (\[pdensity\]) are displayed corresponding to Fig.\[fig-pdensity\]. As we can see, the spin-orbit effect has non-negligible contributions. The term of $h\overleftrightarrow{\Lambda}^2\Delta$ is dominating in the second-order pairing density. In the coarse graining treatment of Bogoliubov-de Gennes equations, only this term has been included in the pairing equations [@simonucci]. Based on our full analysis, we see that such a coarse graining treatment is reasonable, in particular at the surface. The terms of $h(\overleftrightarrow{\Lambda} h)^2$ and $h\overleftrightarrow{\Lambda}^2h$ are non-negligible but the two almost cancel each other in the high energy region.
Similar to Fig.\[fig-pdensity\], the contributions to normal densities are displayed in Fig.\[fig-density\]. Compared to Fig.\[fig-pdensity\], the high-energy contributions to the normal density are much smaller (by 2 orders of magnitude) than to the pairing density. The second-order correction can improve the density distribution at the surface. However, the second-order correction leads to enhanced oscillations in the inner part of the density distribution, due to non-self-consistent calculations. In Fig.\[fig-density-t\], we see that the terms of $\Delta(\overleftrightarrow{\Lambda}h)^2$ and $h\overleftrightarrow{\Lambda}^2 \Delta$ are dominating but the two almost cancel each other. The terms of $h\overleftrightarrow{\Lambda}^2h$ and $h(\overleftrightarrow{\Lambda}h)^2$ are also canceling each other. The term of $h(\overleftrightarrow{\Lambda}\Delta)^2$ is negligible. In contrast to the pairing density, it is hard to say which term plays a major role in the normal density.
In Fig.\[fig-kdensity\], the corrections to kinetic densities are displayed. Again, the second-order correction improves the description at surfaces compared to the zero-order correction. In addition, the second-order correction also improves the kinetic density in the inner part. In Fig.\[fig-sodensity\], the second-order correction to the divergence of the spin-orbit density, $\nabla\cdot \textbf{J}$, is shown. Note that there is no zero-order correction to the spin-orbit density. The obtained $\nabla\cdot \textbf{J}$ roughly agrees with the HFB solutions. It is important to obtain the spin-orbit density to do self-consistent semiclassical calculations.
To study the cutoff dependence, the second-order Thomas-Fermi correction terms to the densities with an enlarged energy interval from 25 MeV to 65 MeV are shown in Fig.\[fig-density-25\]. In Fig.\[fig-density-25\](a), the corrections to normal densities are displayed. Compared to Fig.\[fig-density-t\], the resulting densities are increased by a factor of 20. It is obvious that the terms of $h\overleftrightarrow{\Lambda}^2h$ and $h(\overleftrightarrow{\Lambda}h)^2$ do not cancel each other anymore. The same holds for the terms of $\Delta(\overleftrightarrow{\Lambda}h)^2$ and $h\overleftrightarrow{\Lambda}^2 \Delta$. The term $h(\overleftrightarrow{\Lambda}\Delta)^2$ is still negligible. In Fig.\[fig-density-25\](b), the corrections to pairing densities are shown. Compared to Fig.\[fig-pdensity-t\], the resulting pairing densities are increased by a factor of 10. In this case, we see that the total correction do not follow the term $h\overleftrightarrow{\Lambda}^2\Delta$ anymore which was dominating in the high energy window. The terms of $h(\overleftrightarrow{\Lambda} h)^2$ and $h\overleftrightarrow{\Lambda}^2h$ do not cancel each other. Or, in other words, the coarse graining treatment is not applicable in the low-energy region. The terms of second-order have to be fully taken into account for low-energy quasiparticle states.
In fact, the pairing in nuclei is rather weak, compared to cold atomic systems, for instance, close to the unitary limit. Therefore it is instructive to study for what happens if we increase artificially the pairing. In Fig.\[fig-density-25\](c), the resulting corrections to the pairing densities are obtained with a 10-times larger pairing strength compared to Fig.\[fig-density-25\](b). In this case, the term of $\Delta(\overleftrightarrow{\Lambda} h)^2$ acquires a role which is absent in Fig.\[fig-density-25\](b).
In Fig.\[fig-psur\], the pairing density contribution corresponding to Fig.\[fig-density-25\](c) are shown on a log scale. It can be seen that the second-order correction agrees exactly with the asympotics of coordinate-space HFB solutions at large distances. While the zeroth-order correction underestimates the pairing density at the nuclear surface by 10%. These examples demonstrate clearly the advantages of the full second-order Thomas-Fermi approximation for quasiparticle energy intervals with a lower edge reaching into the low energy domain (e.g., 25 MeV).
Hybrid HFB Calculations
-----------------------
![(Color online) The hybrid HFB calculations of neutron pairing density in $^{238}$U with 0$^{\rm th}$-order and 2$^{\rm nd}$-order Thomas-Fermi approximations within an quasiparticle energy interval from 25 MeV to 65 MeV, compared to the HFB-AX calculations. []{data-label="fig-hpairing"}](hpairing.eps "fig:"){width="48.00000%"}\
![(Color online) The hybrid HFB calculations of binding energies of $^{238}$U with 0$^{\rm th}$-order and 2$^{\rm nd}$-order Thomas-Fermi approximations at different cutoffs, compared to the HFB-AX calculations. []{data-label="fig-cutoff"}](cutoff.eps "fig:"){width="48.00000%"}\
One of our main motivations to implement the hybrid HFB calculations with the second-order Thomas-Fermi approximation is to reduce the computing costs for large systems. In our previous work, we have tested the hybrid HFB strategy with the zeroth-order Thomas-Fermi approximation [@pei2011]. In the hybrid strategy, the high energy deep-hole states and continuum states are separately treated. In the present work, the high-energy continuum is treated in the second-order Thomas-Fermi approximation. The deep-hole states, which are narrow quasiparticle resonances, can be described separately. Firstly, we diagonalize the single-particle Hamiltonian to obtain the deep-bound Hartree-Fock wavefunctions and single-particle energies.
$$(h-\lambda)v_i^{HF}({\boldsymbol{r}})= \varepsilon_i v_i^{HF}({\boldsymbol{r}}) \vspace{5pt}$$
$$(h-\lambda+\varepsilon_i)u_i^{HF}({\boldsymbol{r}})=\Delta({\boldsymbol{r}}) v_i^{HF}({\boldsymbol{r}})
\label{HF2}$$
\[H-HFB\]
Next, one natural way is to renormalize the deep-hole wavefunctions with the BCS approximation, as in Ref. [@pei2011]. We solve the Eq.(\[HF2\]) to describe the asympotics of scattering components in stead of BCS. Another more elaborate way is to solve the HFB equation perturbatively, as described in Ref. [@bulgac].
In the hybrid strategy, we introduce an energy interval (or window) of quasiparticle energies from a lower cutoff to 65 MeV. Below the lower cutoff, the box discretization is used to solve the HFB equation using the HFB-AX solver. Within the window, the zeroth-order or second-order Thomas-Fermi approximation for the continuum and the BCS approximation for deep-hole states are used. The quantal effects are taken into account by the standard HFB solutions below lower cutoffs and the deep-hole states. The hybrid HFB calculations are implemented self-consistently to keep the conservation of particle numbers. With iterations in terms of various densities, i.e., the normal density $\rho({\boldsymbol{r}})$, the kinetic density $\tau({\boldsymbol{r}})$, the spin-orbit density $\textbf{J}({\boldsymbol{r}})$ (does not appear in the zeroth-order), the pairing density $\tilde{\rho}({\boldsymbol{r}})$, the self-consistent hybrid Skyrme HFB calculations can be realized. The quantal and semiclassical are coupled via the summed densities. In our test calculations of $^{238}$U, we use the SLy4 force and the volume pairing interaction, as we used in the non-self-consistent calculations. In Fig.\[fig-hpairing\], the total pairing density profiles from hybrid calculations are shown. It can be seen that the agreements between hybrid and full HFB calculations are quite good.\
Concerning the treatment of the deep hole states, we should note that the BCS underestimates the pairing correlation compared to the HFB approach, due to the absent of continuum coupling. For example, to reproduce the neutron pairing gap of 1.245 MeV in $^{120}$Sn with SLy4 and the volume pairing, the pairing strengthes in BCS and HFB have to be adjusted to 283 and 187 MeV fm$^{-3}$, respectively. Hence we slightly increase the pairing in the BCS treatment of deep-hole states with one global factor. As a result the average pairing gaps at all cutoffs are close to the full HFB result. For systems such as trapped Fermi gases, there are no deep-hole states and the hybrid calculations will be simpler.
The results of hybrid HFB with different lower cutoffs of the energy windows are displayed in Fig.\[fig-cutoff\]. In Fig.\[fig-cutoff\], the deviations in total binding energies between the hybrid HFB and the coordinate-space HFB solutions increase as the lower cutoff decreases. Due to the self-consistency, the second-order approximation is obviously better than the zero-order Thomas-Fermi, in particular at low-energy cutoffs. As we can see, with the cutoff energy at 25 MeV, the deviation is $\thicksim$0.5 MeV over the total binding energy of 1790.52 MeV, which is quite satisfactory. Note that the deviations are not only from the approximate treatment of continuum but also from the treatment of deep-hole states. In the low-energy region, quasiparticle resonances can acquire considerable widths, leading to ambiguous contaminations. In addition, the numerical accuracy of derivatives is important in the second-order calculations. It will be improved by the multi-wavelet techniques [@Pei14] in the future. The full 3D coordinate-space HFB calculations are very expensive even with the efficient multi-wavelet techniques [@Pei14]. With the lower cutoff of 25 MeV, the number of eigen-functions to be solved would be reduced by half. On the other hand the extra numerical cost to include the second order Thomas-Fermi term with respect to take only the zeroth-order, is almost negligible. In this case, the resulting computational cost can be reduced at least by one order of magnitude and the desired accuracy is still retained.
Summary
=======
In summary, we extended the second-order Thomas-Fermi approximation of the Hartree-Fock-Bogoliubov solutions for superfluid systems by including the effective mass and the spin-orbit potential, which are essential for full calculations. In particular, the spin-orbit contribution is important because it is absent in the zeroth-order Thomas-Fermi approximation. The expressions of the new terms have been checked in the zero-pairing limit. The full second-order terms have been examined numerically in a perturbative manner in comparison with self-consistent coordinate-space HFB solutions. In general, the second-order corrections can improve various density distributions at surfaces compared to the zeroth-order corrections. The significance of including full superfluid second-order corrections definitely increases as the lower edge of the energy interval in which the Thomas-Fermi approximation acts goes down. Among the second-order correction terms, the pairing density contribution is relatively the most important one. Furthermore, we performed fully self-consistent hybrid HFB calculations with the second-order Thomas-Fermi for the continuum and satisfactory results have been obtained. This will be particulary useful for the 3D coordinate-space HFB calculations which are computationally very expensive. The second-order superfluid Thomas-Fermi method can be further extended by including the temperature dependence and rotational or vector fields. More interesting applications of the generalized second-order Thomas-Fermi approximation are intended for large complex inhomogeneous superfluid systems such as cold atomic condensates and neutron star crusts, in which the pairing fields are very large compared to finite nuclei.
Useful discussions with W. Nazarewicz, F.R. Xu, G. Fann, N. Hinohara, R. Id Betan, Y. Shi are gratefully acknowledged. This work was supported by the National Natural Science Foundation of China under Grants No.11375016, 11522538, 11235001, and by the Research Fund for the Doctoral Program of Higher Education of China (Grant No. 20130001110001); and the Open Project Program of State Key Laboratory of Theoretical Physics, Institute of Theoretical Physics, Chinese Academy of Sciences, China (Grant No. Y4KF041CJ1). We also acknowledge that computations in this work were performed in the Tianhe-1A supercomputer located in the Chinese National Supercomputer Center in Tianjin.
Second-order Expansion Terms
============================
The expansion coefficients of the second-order superfluid Thomas-Fermi approximation to the normal density in Eq. (\[ndensity\]) are listed here: $$\begin{split}
&\eta_1= -\frac{3h\Delta^2}{32E^5}\\
&\eta_2= \frac{(2h^2-\Delta^2)\Delta}{16E^5}\\
&\eta_3= \frac{2h\Delta^2-h^3}{32E^5}\\
&\eta_4= \frac{4h^2\Delta^2-\Delta^4}{16E^7} \\
&\eta_5= -\frac{(2h^2-3\Delta^2)h\Delta}{8E^7}\\
&\eta_6= -\frac{5h^2\Delta^2}{16E^7}\\
&\eta_7= -\frac{2h\Delta^3-3h^3\Delta}{16E^7}\\
&\eta_8= -\frac{3h^2\Delta^2-h^4-\Delta^4}{8E^7}\\
&\eta_9= -\frac{2h\Delta^3-3h\Delta^3}{16E^7}\\
&\eta_{10}= \frac{h^3+h\Delta^2}{16E^7}\\
\end{split}$$
The expansion coefficients of the second-order superfluid Thomas-Fermi approximation to the pairing density in Eq. (\[pdensity\]): $$\begin{split}
&\theta_1= -\frac{(\Delta^2-2h^2)\Delta}{32E^5}\\
&\theta_2= -\frac{h(h^2-2\Delta^2)}{16E^5}\\
&\theta_3= -\frac{3h^2\Delta}{32E^5}\\
&\theta_4= \frac{(3\Delta^2-2h^2)h\Delta}{16E^7}\\
&\theta_5= \frac{h^4+\Delta^4-3h^2\Delta^2}{8E^7}\\
&\theta_6= -\frac{(2\Delta^2-3h^2)h\Delta}{16E^7}\\
&\theta_7= -\frac{5h^2\Delta^2}{16E^7}\\
&\theta_8= -\frac{(2\Delta^2-3h^2)h\Delta}{32E^7}\\
&\theta_9= -\frac{h^4-4h^2\Delta^2}{16E^7}\\
&\theta_{10}= \frac{\Delta}{16E^5}\\
\end{split}$$
The non-zero derivative terms employed in Eq. (\[ndensity\]) and Eq. (\[pdensity\]):
$$h\overleftrightarrow{\Lambda}^2\Delta = \displaystyle f \nabla^2\Delta$$
$$h\overleftrightarrow{\Lambda}^2h=\displaystyle fp^2\nabla^2f+2f\nabla^2U-\frac{2}{3}p^2(\nabla f)^2\\$$
$$(\Delta\overleftrightarrow{\Lambda}h)^2 = \displaystyle\frac{1}{3} f^2p^2(\nabla\Delta)^2$$
$$\begin{array}{lcl}
h(\overleftrightarrow{\Lambda}h)^2&=&\displaystyle-\frac{1}{12}fp^4(\nabla f)^2+f(\nabla U)^2+\frac{1}{6}f^2p^4\nabla^2f\vspace{5pt}\\
&&\displaystyle+\frac{1}{3}f^2p^2\nabla^2U+\frac{1}{3}fp^2(\nabla f)(\nabla U) \\
\end{array}\\$$
$$h(\overleftrightarrow{\Lambda} h)(\overleftrightarrow{\Lambda} \Delta)=\displaystyle\frac{1}{6}fp^2(\nabla f)(\nabla \Delta)+f(\nabla U)(\nabla \Delta)$$
$$\Delta(\overleftrightarrow{\Lambda}h)^2 = \displaystyle \frac{1}{3}f^2p^2(\nabla^2\Delta)$$
$$h(\overleftrightarrow{\Lambda}\Delta)^2 = \displaystyle f (\nabla\Delta)^2$$
[999]{}
K. Taruishi and P. Schuck, Z. Phys. A 342, 397(1992).
C. A. Ullrich and E. K. U. Gross, Aust. J. Phys. 49, 103(1996).
A. Csordás, O. Almásy and P. Szépfalusy, Phys. Rev. A 82, 063609(2010).
D. Vautherin, D.M. Brink, Phys. Rev. C 5, 626(1972).
M. Bender, P.-H. Heenen, P.-G. Reinhard, Rev. Mod. Phys. [**75**]{},121(2003).
P.J. Wang, Z.-Q. Yu, Z.K. Fu, J. Miao, L.H. Huang, S.J. Chai, H. Zhai, and J. Zhang, Phys. Rev. Lett. 109, 095301 (2012).
A. Bulgac and M. McNeil Forbes, Phys. Rev. Lett. 101, 215301 (2008).
P. Ring and P. Schuck, *The Nuclear Many-Body Problem*, Springer-Verlag, 1980.
P.G. de Gennes, *Superconductivity of Metals and Alloys* (Benjamin, New York, 1966).
S. Simonucci and G.C. Strinati, Phys. Rev. B 89, 054511(2014).
M. Brack, B.K. Jennings and Y.H. Chu, Phys. Lett. B 65, 1(1976).
J. Bartel, K. Bencheikh, and J. Meyer, Phys. Rev. C 77, 024311(2008).
H. Kucharek, P. Ring, P.Schuck, R. Bengtsson, M. Girod, Phys Lett. B 216, 249(1989).
H. Kucharek, P. Ring, P. Schuck, Z. Phys. A 334,119(1989).
J.C. Pei, G. I. Fann, R. J. Harrison, W. Nazarewicz, Yue Shi, and S. Thornton, Phys. Rev. C **90**, 024317(2014).
J.C. Pei, A.T. Kruppa, and W. Nazarewicz, Phys. Rev. C 84, 024311(2011).
J. Reidl, A. Csordás, R. Graham, and P. Szépfalusy, Phys. Rev. A 59, 3816(1999).
X.J. Liu, H. Hu, and P.D. Drummond, Phys. Rev. A 76, 043605(2007); Phys. Rev. A 78, 023601(2008).
G.B. Partridge et al., Science 311, 503(2006).
Y. Shin et al., Phys. Rev. Lett. 97, 030401(2006).
J.G. Cao, D. Yang, Z.Y. Ma, N. Van Giai, Chin. Phys. Lett. 25, 73(2008).
H. Pais and J.R. Stone, Phys. Rev. Lett. 109, 151101(2012).
R. Sensarma, W. Schneider, R.B. Diener, M. Randeria, arXiv:0706.1741,2007.
B. Grammaticos and A. Voros, Ann. Phys. 123: 359(1979); *ibid*. 129, 153(1980).
.
A. Bulgac, arXiv:nucl-th/9907088.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Simple conformal loop ensembles (CLE) are a class of random collection of simple non-intersecting loops that are of particular interest in the study of conformally invariant systems. Among other things related to these CLEs, we prove the invariance in distribution of their nested “full-plane” versions under the inversion $z \mapsto 1/z$.'
author:
- Antti Kemppainen
- Wendelin Werner
date: University of Helsinki and ETH Zürich
title: The nested simple conformal loop ensembles in the Riemann sphere
---
Introduction
============
In [@Sh; @ShW], a one-dimensional natural class of random collections of simple loops in simply connected domains called Conformal Loop Ensembles has been defined and studied. We refer to the introduction of [@ShW] for a detailed account of the motivations that lead to their study. There are two basically equivalent (i.e. defining one enables to define the other one) versions of these simple CLEs, depending on whether one allows loops to be nested (i.e. one loop can surround another loop) or not. Let us recall various definitions and basic features of the latter (i.e. the non-nested) ones:
Such a CLE is defined in a simply connected planar domain $D$ (with $D \not= {{{\mathbb C}}}$) and it is a random countable collection $\Gamma = \{ \gamma_j, j \in J\}$ of simple loops that are all contained in $D$, that are disjoint (no two loops intersect) and non-nested (no loop in this collection surrounds another loop in this collection). Furthermore, the law of this random collection of loops is invariant under any conformal transformation from $D$ onto itself, and the image of $\Gamma$ under any given conformal map from $D$ onto some other domain $D'$ is a CLE in $D'$. The laws of CLEs can be characterized by an additional condition, called “Markovian exploration” that is described and discussed in [@ShW].
Alternatively, see also [@ShW], one can view CLEs as the collections of outer boundaries of outermost clusters in Poissonian collections of Brownian loops in $D$. Roughly speaking, one considers a Poissonian collection of Brownian loops in $D$. As opposed to the previous CLE loops, the Brownian loops are not simple, and they are allowed to overlap and intersect (and they often do, since they are sampled in a Poissonian – basically independent – way). Then one looks at the connected components of the unions of all these loops (i.e. one hooks up intersecting Brownian loops into clusters). It turns out that when the intensity of this Poisson collection of Brownian loops is not large, then there are several (in fact infinitely many) such clusters. Then, one only keeps the outer boundaries of these clusters (that turn out to be simple loops) and finally only keeps the outermost ones (as some clusters can surround others), one obtains a random collection of non-nested simple loops in $D$. It turns out that it is a CLE, and that this procedure (letting the intensity of the Brownian loops vary) does in fact construct all possible CLE laws.
A third description description relies on Oded Schramm’s SLE processes [@Sch]. It turns out that the loops in a CLE are very closely related to SLE$_\kappa$ curves, where the parameter $\kappa$ lies in the interval $(8/3, 4]$ (there is one CLE law for each such $\kappa$, this is called the CLE$_\kappa$), see again [@ShW]. This relation will be also useful in the present paper, as it is the one that exhibits some inside-outside symmetry property of the law of the loops. The precise SLE-based construction of the CLEs goes via SLE-based exploration tree (as explained in [@Sh]) or via a Poisson point process of SLE bubbles (see [@ShW]).
Finally, there is also a close and important relation between CLEs and the Gaussian Free Field (see e.g. [@MS1; @MS2; @MS3; @MS4] and the references therein) that we will briefly mention below, but, as opposed to the previous descriptions, we will not build on it in the present paper.
It is noteworthy to stress that these definitions of CLE all a priori take place in simply connected domains with boundary.
These loop models are of interest, in particular because they are the conjectural scaling limits of various discrete lattice models. For instance, the loops of the CLE should be the scaling limit of the outermost interfaces in various models from statistical physics (such as the critical Ising model) where some particular boundary conditions are imposed on the boundary of the (lattice-approximation of the) domain $D$. Loosely speaking, the boundary of the domain is therefore playing itself the role of an interface i.e. of another loop. This leads to the very natural definition of the [*nested CLE*]{} in the domain $D$ which is defined from a simple non-nested CLEs in an iterative i.i.d. fashion (like for a tree-like structure): Sample first a non-nested CLE in $D$, then sample independent CLEs in the inside each of this first generation CLE loops and so on. This defines, for each $\kappa$ in $(8/3,4]$ and each domain $D$, a [*nested CLE$_\kappa$*]{}. This is again a conformally invariant collection of disjoint loops in $D$ as before, but where each given point $z$ in $D$ is now typically surrounded by infinitely many nested loops. Conversely, if we are given a nested CLE sample in a simply connected domain, one just has to take its outermost loops to get a (non-nested) CLE sample. These nested CLEs are conjecturally the scaling limits of the joint laws of all the interfaces, including all the nested generations, of a wide class of two-dimensional models from statistical physics, such for instance as the $O(n)$ models.
One of our goals in the present paper is to study some properties of the natural version of these nested CLE defined in the entire plane. As we shall see in the first part of the present paper (this construction has been independently also written up in [@MWW]), the definition of the full-plane generalization of nested CLE is not a difficult task (building for instance on the Brownian loop-soup approach to the simple CLEs): One considers the limit when $R \to \infty$ of a nested CLE$_\kappa$ defined in the disc of radius $R$ around the origin, and shows that for any fixed $r$, the law of the picture restricted to this disc of radius $r$ converges as $R \to \infty$. More precisely, one can show that it is possible to couple the nested CLEs in two very large discs of radius $R$ and $R'$ such that with a very large probability $p$, they coincide inside the disc of radius $r$ (i.e. $p$ tends to $1$ when $R, R'$ go to infinity). Note that by scale-invariance, this procedure is equivalent to looking at the picture of a nested CLE in the unit disc and zooming at the law of the picture in the neighborhood of the origin. It is then easy to see from this construction that the law of this “full-plane” family of nested random loops is translation-invariant and scale-invariant.
However, with this definition, one property of this full-plane CLE turns out to be not obvious to establish, namely its invariance (in distribution) under the inversion $z \mapsto 1/z$. Indeed, in the nesting procedure, there is a definite inside-outside asymmetry in the definition of CLEs. One always starts from the boundary of a simply connected domain, and discovers the loops “from their outside” (i.e. the point at infinity in the Riemann sphere plays a very special role in the construction).
On the other hand, invariance of the full-plane CLEs under inversion is a property that is expected to hold. Indeed:
- The discrete $O(n)$ models that are conjectured to prove to these CLEs have a full-plane version, for which one expects such an inside-outside symmetry. In the particular case of the Ising model (which is the $O(1)$ model) that is known to be conformally invariant in the scaling limit (see [@CS; @CD]) and should therefore correspond to CLE$_3$, there is a full-plane version of the discrete critical Ising model that should in principle be invariant under $z \mapsto 1/z$ in the scaling limit as well.
- In the case where $\kappa=4$, the nested CLE$_4$ can be viewed as level (or jump) lines of the Gaussian Free Field, and it is possible (though we will not do this in the present paper) to define the full-plane CLE$_4$ in terms of a full-plane version of the Gaussian Free Field (which is then defined up to an additive constant, so a little care is needed to justify this – in particular, additional randomness is needed in order to define the nested CLE$_4$ from this full-plane GFF), and to see that the obtained CLE$_4$ is indeed invariant under $z \mapsto 1/z$, using the strong connection between CLE$_4$ and the GFF (in particular, the fact that CLE$_4$ is a deterministic function of the GFF when defined in a simply connected domain) derived in [@SchSh; @Dub].
While the previous provable direct connections of the full-plane CLE$_3$ and CLE$_4$ to the Ising model and the Gaussian Free Field respectively indicate quite direct roadmaps towards establishing their invariance under inversion (the CLE$_4$ case is actually quite easy), it is not immediate to adapt those ideas to the case of the other CLE$_\kappa$’s for $\kappa \in (8/3,4]$ (note for instance that the coupling between other CLEs and the GFF [@MS1; @MS3] involves additional randomness that does not seem to behave so nicely with respect to inversion).
One of our two main goals in this paper is to establish the following result:
\[mainthm\] For any $\kappa \in (8/3, 4]$, the law of the nested CLE$_\kappa$ in the full plane (as described above) is invariant under $z \mapsto 1/z$ (and therefore under any Möbius transformation of the Riemann sphere).
Since the law of nested CLE on the Riemann sphere is fully Möbius invariant and hence the law doesn’t depend on the choice of the root point, it makes sense to call it the *Conformal Loop Ensemble of the Riemann sphere* with parameter $8/3 < \kappa \leq 4$ and denote it by CLE$_\kappa(\hat{{{{\mathbb C}}}})$. One way to characterize this family of CLE’s is that they are random collection of loops such that the loops are nested, pair-wise disjoint and simple and that they have the following *restriction property*: if $A \subset \hat{{{{\mathbb C}}}}$ is a closed subset of the Riemann sphere with simply connected complement and if $z_0 \in \hat{{{{\mathbb C}}}} \setminus A$, then define the set $\tilde{A}$ to be the union of $A$ and all the loops that intersect $A$ together with their interiors — as seen from $z_0$, i.e. $z_0$ lies outside of these loops. Then the property, which we call restriction property of CLE$_\kappa(\hat{{{{\mathbb C}}}})$, is that the restriction of the CLE$_\kappa(\hat{{{{\mathbb C}}}})$ to the loops that stay in $U = {{{\mathbb C}}}\setminus \tilde{A}$ is the nested CLE$_\kappa$ in $U$.
One motivation for the present work comes from the fact that, as indicated for instance by the papers of Benjamin Doyon [@Do], it is possible to use such nested CLEs in order to provide explicit probabilistic constructions and interpretations of various basic concepts in Conformal Field Theory (such as the bulk stress-energy tensor). The paper [@Do] for instance builds on some assumptions/axioms about nested CLEs, that we prove in the present paper.
An instrumental idea in the present paper will be to use a “full-plane” version of a variant of the Brownian loop soup, where one only keeps the outer boundary of each Brownian loop instead of the whole Brownian loop. It turns out (this fact had been established in [@W]) that this soup of overlapping simple loops is invariant under $z \mapsto 1/z$, and that (as opposed to the Brownian loop soup itself) it creates more than one cluster of loops when the intensity of the soup is subcritical. We will refer to this loup soup as the SLE$_{8/3}$ loop soup. This is a random full-plane structure that is indirectly related to CLE, even though it is not the nested CLE itself.
Actually, *the other main purpose of the present paper* will be to derive and highlight properties of this particular full-plane structure that we think is interesting on its own right. We shall for instance see that outer boundaries of such clusters and inner boundaries are described by exactly the same intensity measure. More precisely, if one considers a full-plane SLE$_{8/3}$ loop soup, one can construct its clusters, and define those clusters $K_j$ of loops that surround the origin. Each one has an outside boundary $\gamma_j^e$ and an inside boundary $\gamma_j^i$. Then, one can define the intensity measures $\nu^i$ and $\nu^e$ by setting for each measurable set $A$ of simple loops (As the sigma algebra, we use always use the usual sigma algebra of events of staying in annular regions, see Section 3 of [@W].) $$\nu^i ( A) = {{\mathbf E}}( \# \{ j \, : \, \gamma_j^i \in A\} ) \hbox { and }Ê \nu^e ( A) = {{\mathbf E}}( \# \{ j \, : \, \gamma_j^e \in A\} ) .$$ Similarly, for a full-plane CLE (with the $\kappa$ corresponding to the intensity of the loop soup), one can define the intensity measure ${\nu^{\rm cle}}$ of the loops that surround the origin. Then,
\[secondthm\] For some constant $\alpha=\alpha(\kappa)$, one has $\nu^i = \nu^e = \alpha \times {\nu^{\rm cle}}$.
In fact, the proof will go as follows (even if we will not present the arguments in that order): One first directly proves that $\nu^i = \nu^e$ (which will be the core of our proofs), and then deduce Theorem \[secondthm\] from it using the inversion invariance of the SLE$_{8/3}$ loop soup, and then finally deduce Theorem \[mainthm\] from Theorem \[secondthm\].
This paper will be structured as follows. First, we will recall the basic properties of the SLE$_{8/3}$ loop soup, and deduce from it the definition and some first properties of the full-plane CLE$_\kappa$’s. Then, we will build on some aspects of the exploration procedure described in [@ShW] to define CLEs using SLE$_\kappa$ loops, and use some sample properties of SLE paths in order to derive Theorem \[secondthm\].
Chains of loops and clusters from the SLE$_{8/3}$ loop soup
===========================================================
Loop soups of Brownian loops and of SLE$_{8/3}$ loops
-----------------------------------------------------
The Brownian loop soup in ${{{\mathbb C}}}$ with intensity $c$ is a Poisson point process in the plane with intensity $c \mu$, where $\mu$ is the Brownian loop-measure defined in [@LW]. A sample of the Brownian loop soup in $D$ can be obtained from a sample of the Brownian loop soup in the entire plane, by just keeping those loops that fully stay in $D$. More precisely, if $\beta= \{\beta_j , j \in J\}$ is a Brownian loop soup in the plane with intensity $c$ and if $J_D = \{ j \in J \, : \, \beta_j \subset D \}$, then $\beta^D= \{\beta_j, j \in J_D\}$ is a sample of the Brownian loop soup with intensity $c$. It is shown in [@ShW] that when $c \le 1$, the Brownian loop-soup clusters in $D$ are disjoint, and that their outermost boundaries form a sample of a CLE$_\kappa$ (where $\kappa$ depends on $c$). [**For the rest of the present paper, the value of $c \in (0,1]$ (and the corresponding $\kappa (c) \in (8/3, 4]$) will remain fixed, and we will omit them (we will just write CLE instead of CLE$_\kappa$ and loop soup instead of loop soup with intensity $c$).**]{}
If one considers the full-plane Brownian loop soup, then because of the too many large Brownian loops (and the fact that infinitely many large Brownian loops in the loop soup do almost surely intersect the unit circle), it is easy to see that there exists almost surely only one dense cluster of loops. The Brownian loop soup does therefore not seem so well-adapted to define a full-plane structure.
The following observations will however be useful: Firstly, when $D$ is simply connected, define for each Brownian loop $\beta_j$ for $j \in J_D$, its outer boudary $\eta_j$ (the outer boundary of a Brownian loop is almost surely a simple loop, see [@W] and the references therein). Then, consider the outer boundaries of outermost clusters of loops defined by the family of simple loops $\eta_D= \{\eta_j, j \in J_D\}$ (instead of $\beta_D$). Clearly, this defines the very same collection of non-nested simple loops as the outer boundaries of outermost clusters of $\beta_D$, and it is therefore a CLE. Secondly, it is shown in [@W] that the family $\eta = \{\eta_j, j \in J\}$ is a Poisson point process of SLE$_{8/3}$ loops, and that this random family is invariant (in law) under any Möbius transformation of the Riemann sphere (in particular under $z \mapsto 1/z$). This yields a non-trivial “inside-outside” symmetry of Brownian loop boundaries (the proof in [@W] is based on the fact that this outer boundary can be described in SLE$_{8/3}$ terms). Hence, we see that the CLE$_\kappa$ is also the collection of outer boundaries of outermost cluster of loops of an SLE$_{8/3}$ loop soup in $D$ (that can itself be viewed as the restriction of a full-plane SLE$_{8/3}$ loop soup to those loops that stay in $D$).
Finally, we note that the outer boundary $\eta_j$ of the Brownian loop $\beta_j$ is clearly much “sparser” than $\beta_j$ itself (the inside is empty...). As indicated in [@W], it turns out that if one considers the soup $\eta$ of SLE$_{8/3}$ loops in the entire plane, then (for $c \le 1$) the clusters will almost surely all be finite and disjoint. Here is a brief justification of this fact:
- Note first that if we restrict ourselves to a (subcritical i.e., $c \le 1$) loop soup $\eta_{{\mathbb U}}$ in the unit disc ${{\mathbb U}}$, then the outer boundaries of outermost loop-soup clusters do form a CLE$_\kappa$. Therefore, the outermost cluster-boundary $\gamma$ (in the CLE in ${{\mathbb U}}$) that surrounds the origin is almost surely at positive distance of the unit circle. Hence, for some positive ${\epsilon}$, $${{\mathbf P}}( d ( \gamma , \partial {{\mathbb U}}) > {\epsilon}) \ge 1/2 .$$ Let us call $A_1$ this event $\{d ( \gamma , \partial {{\mathbb U}}) > {\epsilon}\}$.
- The total mass (for the SLE$_{8/3}$ loop measure defined in [@W]) of the set of loops that intersect both $\partial U$ and $(1-{\epsilon}) \partial {{\mathbb U}}$ is finite. This can be derived in various ways. One simple justification uses the description of this measure as outer boundaries of (scaling limits) of percolation clusters (see [@W]), and the fact that the expected number of critical percolation clusters that intersect both $R \partial {{\mathbb U}}$ and $(1-{\epsilon})R \partial {{\mathbb U}}$ is finite and bounded independently of $R$ (this is just the Russo-Seymour-Welsh estimate) and therefore also in the $R \to \infty$ limit. Alternatively, one can do a simple SLE$_{8/3}$ computation. Hence, if we perform a full-plane SLE$_{8/3}$ loop soup, then with positive probability, no loop in the soup with intersect both $\partial U$ and $(1-{\epsilon}) \partial U$. Let us call $A_2$ this event.
- The events $A_1$ and $A_2$ are independent ($A_1$ is measurable with respect to the set of loops in the loop soup that stay in ${{\mathbb U}}$ and $A_2$ is measurable with respect to the set of loops in the loop soup that intersect $\partial {{\mathbb U}}$). Hence, the probability that $A_1$ and $A_2$ hold simultaneously is strictly positive. This implies that with positive probability, there exists a cluster of loops in the full-plane SLE$_{8/3}$ loop soup, that surrounds the origin and is contained entirely in the unit disc.
- It follows immediately (via a simple $0-1$ law argument, because the event that there exists an unbounded loop-soup cluster does not depend on the set of loops that are contained in $R {{\mathbb U}}$ for any $R$, and is therefore also independent of the loop soup itself) that almost surely, all clusters in this soup are bounded (if not, the distance between the origin and the closest infinite cluster is scale-invariant and positive).
- The fact that the clusters are almost surely all disjoint can be derived in a rather similar way (just notice that if two different full-plane loop-clusters had a positive probability to be at zero distance from each other, then the same would be true for two CLE loops in the unit disc, with positive probability, and we know that this is almost surely not the case.
To sum things up: For any given $c \le 1$, the full-plane SLE$_{8/3}$ loop soup defines a random collection of clusters $(K_i, i \in I)$ that is [**invariant in distribution under any Möbius transformation of the Riemann sphere**]{} (including the inversion $z \mapsto 1/z$), and the boundaries (inner and outer boundaries) of these clusters are closely related to SLE$_\kappa$ paths for $\kappa = \kappa (c) \in (8/3, 4]$.
Markov chains of nested clusters and of nested loops
----------------------------------------------------
We are now going to pick two points on the Riemann sphere, namely the origin and infinity (but by conformal invariance, this choice is not restrictive), and we are going to focus only on those clusters that disconnect one from the other i.e. that surround the origin. Recall that almost surely, both the origin and infinity are not part of a cluster (and scale-invariance shows that there exist almost surely a countable family of loop-soup clusters that disconnect $0$ from infinity). We can order those clusters that disconnects infinity from the origin “from outside to inside”. We will denote this collection by $(K_j, j \in J)$ where $J \subset I$ is now a decreasing bijective image of ${{\mathbb Z}}$ (each $j$ in $J$ has therefore a successor denoted by $j+1$).
The boundaries of the complement of each $K_j$ consists of countably many loops, two of which (corresponding to the connected components $O_j^i$ and $O_j^e$ of ${{{\mathbb C}}}\setminus K_j$ that respectively contain the origin and infinity) surround the origin. We will call these boundaries $\gamma_j^i$ and $\gamma_j^e$. One therefore has a nested discrete sequence of loops, when $j$ in $J$, then $$\gamma_{j}^e \succ \gamma_j^i \succ \gamma_{j+1}^e$$ where $\gamma \succ \gamma'$ means that $\gamma$ surrounds $\gamma'$ (we however allow here the possibility that $\gamma$ intersects $\gamma'$ – indeed, for small $c$, it happens that for a positive fraction of the $j$’s, the inner and outer boundaries $\gamma_j^i$ and $\gamma_j^e$ of $K_j$ do intersect).
The scale invariance of the loop soup, as well as the fact that the expected number of clusters that surround the origin and have diameter between $1$ and $2$ say is finite, shows immediately that we can define three infinite measures $\nu$, $\nu^i$ and $\nu^e$ that correspond to the intensity measure of the families $(K_j)$, $(\gamma_j^i)$ and $(\gamma_j^e)$ respectively. In other words, for any measurable family $L$ of loops (see e.g. [@W] for details on the $\sigma$-field that one can use), $$\nu^i (L) = {{\mathbf E}}\bigg( \sum_{j \in J} {{\mathbbm{1}}}_{\gamma_j^i \in L} \bigg)$$ (and the analogous definition for the measure $\nu^e$ on outer loops and for the measure $\nu$ on clusters of loops, defined on an appropriately chosen $\sigma$-field). Clearly, these measures are scale-invariant i.e. for any set $L$ of loops and any positive $\lambda$, $$\nu^i (L) = \nu^i ( \{ \gamma \, : \, \lambda \gamma \in L \}).$$ Additionally, these measures have the following inversion relations which play an important role later.
\[prop: measures under z mapsto 1/z\] The measure $\nu$ is invariant under $z \mapsto 1/z$ and the image of the measure $\nu^i$ under $z \mapsto 1/z$ is $\nu^e$.
The claim follows from the fact that the full-plane SLE$_{8/3}$ loop soup is invariant under inversion.
Let us define three Markov kernels that are heuristically correspond to the mapping $K_{j} \mapsto K_{j+1}$, $\gamma^i_{j} \mapsto \gamma^i_{j+1}$ and $\gamma_{j}^i \mapsto \gamma_{j+1}^e$. Note that in the definition of these chains, we always explore from outside to inside and from one cluster to the next one.
More rigorously, for any simply connected domain $D \not= {{{\mathbb C}}}$ that contains the origin, sample an SLE$_{8/3}$ loop soup in $D$ and denote by ${\cal L}^{K}_D$ (respectively ${\cal L}^{i}_D$ and ${\cal L}^e_D$) the law of the outermost cluster that surrounds the boundary (resp. the inner boundary of this outermost cluster and the outer boundary of the outermost cluster). When $A$ is a compact set that surrounds the origin, denote (whenever it exists) by $D(A)$ the connected component of the complement of $K$ that contains the origin. Then, the kernels are defined as follows: Consider $Q^{\to K} (A, \cdot) := {\cal L}^K_{D(A)} ( \cdot)$ and similarly $Q^{\to i} (A, \cdot) := {\cal L}^i_{D(A)} (\cdot)$ and $Q^{\to e} (A, \cdot) := {\cal L}^e_{D(A)}(\cdot)$.
Let us now consider a full-plane SLE$_{8/3}$ loop soup, and take two different simply connected domains $D$ and $D'$ that contain the origin. For $D$ (respectively $D'$), we restrict the full-plane loop soup to the set of loops that are contained in $D$ (resp. $D'$). Hence, we have now three families of nested clusters that surround the origin:
- The clusters $(K_j, j \in J)$ of the full-plane loop soup.
- The clusters $(K_n^D, n \ge 1)$ and $(K_{n'}^{D'}, n' \ge 1)$ corresponding to the loop soups in $D$ and $D'$ respectively. These two sequences can ve viewed as Markov chains with kernel $Q^{ \to K}$ started from $\partial D$ and $\partial D'$ respectively. Similarly, their inner boundaries are Markov chains with kernel $Q^{\to i}$.
The properties of the full-plane SLE$_{8/3}$ loop soup imply immediately that almost surely, there exists $j_0$, $n_0$ and $n_0'$ so that for all $n \ge 0$, $$K_{j_0 + n } = K_{n_0 +n }^D = K_{n_0' + n }^{D'}.
\label {coupl}$$ Indeed, almost surely, for small enough $\epsilon$, no loop in the loop soup does intersect both the circle of radius $\epsilon$ around the origin and ${{{\mathbb C}}}\setminus D$ or ${{{\mathbb C}}}\setminus D'$, which implies that the “very small” clusters that surround the origin are the same in all three pictures.
Shapes of clusters and of loops
-------------------------------
Note that one can decompose the information provided by a loop $\gamma$ (or a set $K$) that surrounds the origin into two parts (we now detail this in the case of the loop):
- Its “size”, for instance via the log-conformal radius $\rho (\gamma)$ of its interior, seen from the origin (such that the Riemann mapping $\Phi_\gamma$ from the unit disc onto the interior of $\gamma$ such that $\Phi_\gamma(0) = 0$ and $\Phi_\gamma' (0) \in (0, \infty)$ satisfies $\Phi_\gamma' (0) = \exp ( \rho (\gamma))$.
- Its “shape” $S(\gamma)$ i.e. its equivalence class under the equivalence relation $$\gamma \sim \gamma'
\Longleftrightarrow \hbox {There exists some positive $\lambda$ for which $\gamma = \lambda \gamma'$}.$$
For a shape $S$ and a value $\rho$, we define $\gamma ( \rho , S)$ to be the only loop with shape $S$ and log-conformal radius $\rho$. The scale-invariance of $\nu^i$ implies immediately that there exists a constant $a_i$ and a probability measure $P^i$ on the set of shapes so that $\nu^i$ is the image of the product measure $a_i d\rho \otimes P^i$ under the mapping $(\rho, S) \mapsto \gamma (\rho, S)$.
The same of course holds for $\nu^e$, which defines a constant $a_e$ and a probability measure $P^e$, and for $\nu$ that defines a constant $a_K$ and a probability measure $P^K$. We can note that for any $R$, for a full-plane SLE$_{8/3}$ loop-soup sample, the number of inside cluster boundaries and the number of exterior cluster boundaries that are included in the annulus between the circles of radius $1$ and $R$ can differ only by at most $1$ from the number of interior cluster boundaries in this annulus (because the loops $\gamma_j^i$ and $\gamma_j^e$ are alternatively nested). It follows (letting $R \to \infty$ and looking at the expected number of such respective loops) that $a_e=a_i=a_K$ (and we will denote this constant by $a$).
We can note that the three kernels $Q^{\to i}$, $Q^{\to K}$, $Q^{\to e}$ induce kernels $\tilde Q^{ \to i}$, $\tilde Q^{\to K}$ and $\tilde Q^{ \to e}$ on the set of shapes (because the former kernels are “scale-invariant”). The coupling property (\[coupl\]) implies immediately that $P^K$ and $P^i$ are the unique stationary distributions for $\tilde Q^{\to i}$ and $\tilde Q^{\to K}$. It follows that $\nu^i$ and $\nu$ are (up to a multiplicative constant) the only scale-invariant measures that are invariant under $Q^{\to i}$ and $Q^{\to K}$ respectively.
Full-plane CLE and reversibility
================================
In the next two subsections, we will describe the construction of the full-plane CLEs. This has been independently and in parallel written up also in [@MWW], where it is used for another purpose.
Markov chain of nested CLE loops and its properties
---------------------------------------------------
In order to construct and study the nested CLEs, we will focus on the kernel $Q^{\to e }$ instead of $Q^{\to i}$. Let us first collect some preliminary simple facts:
1. Let us consider first a loop soup in the unit disc. We know a priori that the log-conformal radius of $\gamma^e_{1}$ is not likely to be very small: For instance, for any positive $x_0$, there exists $c >0$ so that for all $x$, $$\label {lem1}
{{\mathbf P}}\left( \rho (\gamma^e_1) \le -x-x_0 \right) \le e^{-cx_0} \, {{\mathbf P}}\left( \rho ( \gamma_1^e) \le -x \right).$$ Indeed, if $\rho (\gamma^e_1) \le -x-x_0$, then on the one hand, the annulus $\{ z \, : \, e^{-x_0} < |z| < 1 \}$ does not contain an SLE$_{8/3}$ loop in the loop soup (which is an event of probability strictly smaller than one), and on the other hand, if we restrict the loop soup to the disc $e^{-x_0} {{\mathbb U}}$, the outermost loop-soup cluster boundary $\tilde \gamma_1^e$ that surrounds the origin satisfies $ \rho ( \tilde \gamma_1^e ) \le \rho (\gamma_1^e) \le -x-x_0$. But these two events are independent, and the laws of $ \rho ( \tilde \gamma_1^e ) + x_0$ and of $\rho (\gamma_1^e)$ are identical by scale-invariance, so that (\[lem1\]) follows.
2. Consider a sequence $(\xi_n, n \ge 1)$ of i.i.d. positive random variables such that for some $x_0$ and $c >0$, and for all $x$, ${{\mathbf P}}( \xi_1 > x + x_0 ) \le e^{-cx_0} \, {{\mathbf P}}( \xi_1 > x)$. Define $S_n = \xi_1+ \ldots + \xi _n$ and for all $y >0$, the overshoot at level $y$ i.e. $O(y) = \min \{ S_n - y \, : \, n \ge 1 \hbox { and } S_n > y \}$. Then, for all $M$ that is a multiple of $x_0$, $$\label {lem2}
{{\mathbf P}}( O(y) \ge M ) \le e^{-cM}.$$ Indeed, if we suppose that $M$ is a multiple of $x_0$, then $$\begin{aligned}
{{\mathbf P}}( O( y) \ge M )
& = \sum_{n \ge 0} {{\mathbf P}}( S_n < y \hbox { and } \xi_{n+1} \ge M +y - S_n ) \\
& = \sum_{n \ge 0} {{\mathbf E}}\left[ \,
{{\mathbbm{1}}}_{S_n < y} \, {{\mathbf P}}( \xi_{n+1} \ge M +y - S_n \,|\, \sigma(\xi_1, \ldots, \xi_n) ) \,\right] \\
& \leq \sum_{n \ge 0} e^{-cM} \, {{\mathbf E}}\left[ \,
{{\mathbbm{1}}}_{S_n < y} \, {{\mathbf P}}( \xi_{n+1} \ge y - S_n \,|\, \sigma(\xi_1, \ldots, \xi_n) ) \,\right] \\
& = e^{-cM} \sum_{n \ge 0} {{\mathbf P}}( S_n < y \le S_{n+1}) = e^{-cM}
\end{aligned}$$ where $\sigma(\xi_1, \ldots, \xi_n) $ is the sigma algebra generated by the random variables $\xi_k$, $k = 1,2,\ldots, n$. Note that (\[lem2\]) shows that for a large enough given $M$, $P( O(y) \le M ) \ge 1/2$ (independently of $y$).
3. Let us briefly recall how to define a nested CLE in the simply connected domain $D$ (with $D \not= {{{\mathbb C}}}$). We first sample a simple CLE, that defines a countable collection of disjoint and non-nested loops in $D$. For each $z \in D$, it is almost surely surrounded by a loop denoted by $\gamma_1(z)$ in this CLE (note of course that while for each given $z$, $ \gamma_1 (z)$ almost surely exists, there exists a random fractal set with zero Lebesgue measure of points that are surrounded by no loop). In particular, if the origin is in the domain $D$, then the loop $\gamma_1 (0)$ is distributed like the loop $\gamma_1^e$.
Then, once this first-layer CLE is defined, we repeat (conditionally on this first generation of loops) the same experiment independently inside each of these countable many loops. For each given $z$, this defines almost surely a second-layer loop $\gamma_2 (z)$ that surrounds $z$. We then repeat this procedure indefinitely. Hence, for any fixed $z$, we get almost surely a sequence of nested loops $(\gamma_n (z), n \ge 1)$.
Let us now suppose that $0 \in D$. Clearly, if we focus only the loops that surround the origin $(\gamma_1, \gamma_2, \ldots ) := (\gamma_1 (0), \gamma_2 (0), \ldots)$, we get a Markov chain with kernel $Q^{\to e}$. We can now define the random variables $\xi_1 = \rho (D) - \rho (\gamma_1)$ and for all $j \ge 2$, $\xi_j = \rho (\gamma_{j-1}) - \rho(\gamma_j)$ corresponding to the successive jumps of the log-conformal radii. These are i.i.d. positive random variables, and combining the previous two items, we see that there exists $M$ and $c$ such that, for all $v < \rho (D)$, if $j_0$ is the first $j$ for which $\rho ( \gamma_j) < v$, then $${{\mathbf P}}( \rho (\gamma_{j_0}) \le v - M ) \le e^{-cM}.
\label {lem3prime}$$ In particular, for some given $M$, $${{\mathbf P}}( \rho (\gamma_{j_0}) \le v - M ) \le 1/2.
\label {lem3}$$
4. Let us now consider two bounded simply connected domains $D$ and $D'$ that surround the origin, and try to couple the (non-nested) CLEs in these two domains in such a way that the [*first*]{} loops $\gamma_1$ and $\gamma_1'$ that surround the origin coincide. We will assume in this paragraph that the log-conformal radii of $D$ and $D'$ are not too different i.e. that $$| \rho (D) - \rho (D') | \le M$$ (where $M$ is chosen as in (\[lem3\])).
We consider a realization of the SLE$_{8/3}$ loop soup in the full-plane, and then restrict them to $D$ and $D'$ respectively. This defines a coupling of the two loops $\gamma_1^e$ and ${\gamma_1^e}'$. Then, for this coupling, there exists a positive constant $u$ that does not depend on $D$ and $D'$ so that $${{\mathbf P}}( \gamma_1^e = {\gamma_1^e}' ) > u
\label {lem4}$$ Let us now briefly indicate how to prove this fact: Clearly, we can assume that $\rho (D') \ge \rho (D)$ (otherwise, just swap the role of $D$ and $D'$), and because of scale-invariance, we can assume, without loss of generality that $\rho (D) = 0$ (i.e. that there exists a conformal map $\Phi$ from $D$ onto ${{\mathbb U}}$ such that $\Phi (0)=0$ and $\Phi'(0)=e^0 = 1$). By Koebe’s $1/4$ Theorem, this implies that ${{\mathbb U}}\setminus 4 \subset D$ and similarly (because $\rho (D') \ge 0$), that ${{\mathbb U}}/ 4 \subset D'$.
Let us consider a full-plane loop soup of SLE$_{8/3}$ loops. Let us first restrict this loop soup to the disc ${{\mathbb U}}/8$, and define the event that there exists an outer-boundary of a cluster in this loop soup such that its outer boundary does fully surround ${{\mathbb U}}/16$ and such that the cluster itself does intersect ${{\mathbb U}}/16$. If such a cluster exists, then it is clearly unique – we denote it by $K$. Note that at this point, we have not required that no other cluster of the loop soup in ${{\mathbb U}}/8$ surrounds $K$.
Considerations from [@ShW] show that such a $K$ indeed exists with a positive probability $u_0$. Furthermore, we can discover this event “from the inside” by exploring all loop-clusters of the loop soup that do intersect the disc ${{\mathbb U}}/ 16$. Hence, for any simply connected domain $V$, the event that $K\subset V$ is independent from the loops in the full-plane loop soup that do not intersect $V$. It therefore follows easily that (on the event where $K$ exists) the conditional law of the loop soup outside of $K$ (given $K$) is just a SLE$_{8/3}$ loop soup restricted to the outer complement of $K$.
On the other hand, we also know (for instance from [@ShW]) that with positive probability (that is bounded from below independently of the shape of $D$), the outermost cluster $K_1$ in the CLE in $D$ is a subset of ${{\mathbb U}}/16$. It therefore follows that, conditionally on $K$, if we then sample the loops in $D$ that lie outside of $K$, with a conditional probability that is bounded uniformly away from $0$ (i.e. uniformly larger than some $u_1$), we do not create another cluster of loops that surrounds $K$. Hence, the conditional probability that $K_1 = K$ is greater than $u_0u_1$.
The same holds for $K_1'$ (using this time the fact that $0 \le \rho (D') \le M_0$), and (when one first conditions on $K$), the events $K_1 = K$ and $K_1' = K$ are positively correlated (they are both decreasing events of the loop soup outside of $K$). Hence, we conclude that the conditional probability that $K = K_1 = K_1'$ is bounded away from $0$ uniformly, from which (\[lem4\]) follows.
With these results in hand, we can now construct a coupling between nested CLEs between any two given simply connected domains $D$ and $D'$ that surround the origin, in such a way that they coincide in the neighborhood of the origin:
We will first only focus on the two sequences of loops that surround the origin $(\gamma_1, \gamma_2, \ldots)$ and $(\gamma_1', \gamma_2', \ldots)$ that we will construct from the outside to the inside in a “Markovian way”, and we will couple them in such a way that for some $n_0$ and $n_0'$, $\gamma_{n_0} = \gamma_{n_0'}'$. Then, we will choose the two nested CLEs in such a way that they coincide within this loop $\gamma_{n_0}$.
Suppose for instance that $\rho (D) \ge \rho (D')$ (the other case is treated symmetrically) – note however that we do not assume here that $\rho (D)$ and $\rho (D')$ are close. So, our first step is to try to discover some loops $\gamma_m$ and $\gamma_{m'}'$ in the two nested CLEs that have a rather close log-conformal radius.
We therefore first construct $\gamma_2, \gamma_3, \ldots$ (using the Markov chain $Q^{\to e}$) until $\gamma_{m_1}$, where $$m_1 = \min \{ m \ge 1 \, : \, \rho (\gamma_m) < \rho (D')\} .$$ Two cases arise:
- Case 1: $\rho (\gamma_{m_1}) \ge \rho (D') - M$. By (\[lem3\]), we know that this happens with probability at least $1/2$. In this case, these two sets have close enough conformal radii (so that we will then be able to couple $\gamma_{m_1 + 1}$ with $\gamma_1'$ so that they coincide with probability at least $u$) and we stop.
- Case 2: $\rho (\gamma_{m_1}) < \rho (D') - M$. Then, we start constructing the loops $\gamma_1', \ldots$ until we find a loop $\gamma_{m_1'}'$ such that $\rho (\gamma_{m_1'}') < \rho (\gamma_{m_1})$. Again, either the difference between the conformal radii is in fact smaller than $M$, and we stop. Otherwise, we start exploring the loops $\gamma_{m_1+1}, \ldots$ until we find $\gamma_{m_2}$ with $\rho (\gamma_{m_2}) < \rho (\gamma_{m_1'}')$, and so on. At each step, the probability that we stop is at least $1/2$, so that this procedure necessarily ends after a finite number of iterations.
In this way, we almost surely find $\gamma_m$ and $\gamma_{m'}'$ so that $|\rho (\gamma_m) - \rho (\gamma_{m'}')| \le M$. Furthermore, we have not yet explored/constructed the loops inside these two loops. Hence, we can now use (\[lem4\]) to couple $\gamma_{m+1}$ with $\gamma_{m'+1}'$ so that they are equal with probability at least $u$.
On the part of the probability space where the coupling did not succeed, we start the whole procedure again by continuing to construct loops inwards from these two loops $\gamma_{m+1}$ and $\gamma_{m'+1}'$. Again, since this coupling succeeds at each iteration with a probability at least $u$, we finally conclude that almost surely, using this construction, we will eventually find $\bar m$ and $\bar m'$ so that $\gamma_{\bar m} = \gamma_{{\bar m}'}'$.
A final observation is that (because of (\[lem3prime\])), for this construction $${{\mathbf P}}\left( \rho (\gamma_{\bar m}) \,<\, \min (\rho (D), \rho (D')) - x \right) \to 0$$ as $x \to \infty$, uniformly with respect to all choices of $D$ and $D'$.
Hence, we have obtained the following result.
\[prop: coupling of nested CLEs\] For any $D$ and $D'$, it is possible to couple the nested CLEs in $D$ and in $D'$ in such a way that almost surely:
- There almost surely exists an $n_0$ such that the two nested CLEs coincide inside the loop $\gamma_{n_0}$.
- Furthermore, $ {{\mathbf P}}( n_0 \ge j )$ tends to zero as $j$ tends to infinity, uniformly over all possible sets $D$ and $D'$ with $\rho (D') \ge \rho (D)$.
Definition of the full-plane CLE
--------------------------------
Proposition \[prop: coupling of nested CLEs\] enables us to define and state a few properties of the full-plane CLE.
- The law of the part of the nested loop soup in the disc $n {{\mathbb U}}$ that is contained in a finite ball of radius $r>0$ does converge when $n \to \infty$ to a limit.
- For any sequence of domains $D_n$ with $\rho (D_n) \to \infty$, the law of the part of the nested loop soup in the disc $n {{\mathbb U}}$ that is contained in a finite ball of radius $r>0$ does converge when $n \to \infty$ to the same limit.
The first statement is just obtained by noting that the previous proposition shows that it is possible to couple the nested CLE in $n {{\mathbb U}}$ with the nested CLE in $n'{{\mathbb U}}$, so that with a large probability (that tends to $1$ when $n, n' \to \infty$) they coincide inside the disc of radius $r$. The second just follows from the coupling between the CLEs in $D_n$ and $n {{\mathbb U}}$.
The above convergence enables to define the *full-plane nested CLE* (started from $\infty$) to be the law on nested loops in the entire plane that coincide with the limit inside each disc of radius $r$. Let $b \in {{{\mathbb C}}}$. We define the *full-plane nested CLE chain to $b$* (from $\infty$) to the restriction of the full-plane nested CLE to those loops that surround $b$. We use the notation CLE$(a \to {{{\mathbb C}}})$ and CLE$(a \to b)$ for the full-plane nested CLE started from $a$ and the corresponding chain to $b$, which are defined for other $a$ than $\infty$ by a Möbius transformation.
For this last definition, we used the fact that the full-plane CLE is scale-invariant and translation invariant in distribution in the following sense:
- If $\psi_\lambda$ is the scaling in the plane by a factor $\lambda>0$, then the law of CLE$(\infty \to 0)$ is invariant under $\psi_\lambda$. Also CLE$(\infty \to {{{\mathbb C}}})$ is invariant under $\psi_\lambda$.
- If $\phi_b$ is the translation in the plane by a complex number $b$, then the law of CLE$(\infty \to b)$ is the image of of the law CLE$(\infty \to 0)$ under $\phi_b$. The entire collection CLE$(\infty \to {{{\mathbb C}}})$ is in fact invariant under $\phi_b$.
This property follows from coupling the CLEs in $n{{\mathbb U}}$ with $n \lambda {{\mathbb U}}+ b$ (for a given $\lambda>0$ and $b \in {{{\mathbb C}}}$).
In the nested CLE in a domain $D$, the chains of loops to distinct points $\{b_1,b_2,\ldots,b_n\}$ are coupled so that the chains to $b_i$ and $b_j$ are the same until the loops of $b_i$ don’t any more surround $b_j$ and vice versa, after which the chains are conditionally independent. This shows that the full-plane CLEs from $\infty$ to any of the points $\{b_1,b_2,\ldots,b_n\}$ can be coupled to have the same property. Let us denote the restriction of CLE$(a \to {{{\mathbb C}}})$ to those loops that disconnect $a$ and a point in $\{b_1,b_2,\ldots,b_n\}$ by CLE$(a \to \{b_1,b_2,\ldots,b_n\})$. In the rest of the paper, we will show that the law of CLE$(\infty \to {{{\mathbb C}}})$ is fully Möbius invariant and by that result we can define the *Riemann sphere nested CLE*, denoted by CLE$(\hat{{{{\mathbb C}}}})$, whose law doesn’t depend on the starting point. One way to formulate this is that CLE$(a \to \{b_1,b_2,\ldots,b_n\})$ and CLE$(b_1 \to \{a,b_2,\ldots,b_n\})$ have the same law, if we ignore the order of exploration of the loops.
We can define ${\nu^{\rm cle}}$ as the infinite intensity measure of CLE$(\infty \to 0)$, i.e. the set of loops that surround the origin in the full-plane CLE, and apply the same arguments that at the end of Section 2: The measure ${\nu^{\rm cle}}$ is scale-invariant, invariant under $Q^{\to e}$, and its shape probability measure $P^{\rm cle}$ is invariant under $\tilde Q^{\to e}$. The previous coupling result shows that $P^{\rm cle}$ is the unique invariant shape distribution under $\tilde Q^{\to e}$, from which one can deduce that (up to a multiplicative constant) ${\nu^{\rm cle}}$ is the unique scale-invariant measure that is invariant under $Q^{\to e}$.
Roadmap to reversibility of the full-plane CLE
----------------------------------------------
Let us now briefly sum up the measures on translation-invariant and scale-invariant random full-plane structures that we have defined at this point:
- The nested CLE in the entire plane. When one focuses at the loops surrounding the origin, it has an intensity measure ${\nu^{\rm cle}}$ that is, up to multiplicative constants, the only scale-invariant measure that is invariant under $Q^{ \to e}$.
- The full-plane SLE$_{8/3}$ loop soup. When looking at clusters and their boundaries that surround the origin, it defines intensity measures $\nu$, $\nu^i$ and $\nu^e$. The former two are (up to multiplicative constants) the only invariant measures under $Q^{\to K}$ and $Q^{\to i}$ that are scale-invariant. Furthermore $Q^{\to e} \nu^i = \nu^{e}$.
We recall that (as opposed to the nested CLE) we already know at this stage that the full-plane SLE$_{8/3}$ loop soup is invariant under inversion and that therefore the image of $\nu^i$ under $z \mapsto 1/z$ is $\nu^e$, see Proposition \[prop: measures under z mapsto 1/z\].
Our roadmap is now the following: In the next section, we are going to prove (building on the SLE$_\kappa$ description of CLEs and on various properties of SLE) the following fact:
\[nui=nue\] The two measures $\nu^e$ and $\nu^i$ are equal.
Let us now explain how Proposition \[nui=nue\] implies Theorem \[mainthm\] (we defer the proof of the proposition to the next section): First, note that Proposition \[nui=nue\] implies immediately that $Q^{\to e} \nu^e = Q^{ \to e} \nu^i = \nu^e$, so that $\nu^e$ equal to is a constant times ${\nu^{\rm cle}}$ (because it is invariant under $Q^{\to e}$). Let us rephrase this rather surprising fact as a corollory in order to stress it:
The kernels $Q^{\to e}$ and $Q^{\to i}$ have the same scale-invariant measures.
As we know that $\nu^e$ is the image of $\nu^i$ under $z \mapsto 1/z$, we can already conclude that ${\nu^{\rm cle}}$ is in fact invariant under inversion.
In order to prove Theorem \[mainthm\], it is sufficient to prove the invariance in distribution under the map $z \mapsto 1/z$ of the nested family CLE$(\infty \to 0)$ of loops $(\gamma_j, j \in J)$. Indeed, on each of the successive annuli (in between $\gamma_j$ and $\gamma_{j+1}$), the conditional distribution (given the sequence $(\gamma_j)$) of the other loops of the same “generation” as $\gamma_{j+1}$ in the nested CLE (that are surrounded by $\gamma_j$ but by no other loop in between them and $\gamma_j$) is given by the outermost boundaries of loop-soup clusters in the annulus between $\gamma_j$ and $\gamma_{j+1}$, conditioned to have no cluster that surrounds $\gamma_{j+1}$. This description of the conditional distribution is nicely invariant under inversion (because the loop soups are), and this proves readily that the law of the entire nested CLE is invariant under $z \mapsto 1/z$. Since we already have translation-invariance and scale-invariance, this implies Theorem \[mainthm\].
It now remains to prove that the law of the nested family CLE$(\infty \to 0)$ of loops $(\gamma_j, j \in J)$ is invariant under inversion. Before explaining this, let us first make a little side-remark: Let us define the successive concentric annuli $(A_j, j \in J)$ in the nested CLE sequence where $A_j$ denotes the annular region in between the loop $\gamma_j$ and its successor $\gamma_{j+1}$ (i.e. the next loop in the sequence, inside of $\gamma_j$). As before, one can also define the scale-invariant “intensity measure” on the set of annuli that we call $\nu^A$. The Markovian definition of the nested CLE sequence shows immediately that $\nu^A$ can be described from the product measure ${\nu^{\rm cle}}\otimes \tilde P$ as follows:
- $\tilde P$ is the law of the outside-most loop $\tilde \gamma$ that contains the origin in a CLE in the unit disc.
- Starting from a couple $(\gamma, \tilde \gamma)$, one defines the annulus $A$ that is between $\gamma$ on the one hand (that is therefore the outer loop of the annulus) and $\phi_\gamma (\tilde \gamma)$ where $\phi_\gamma$ is the conformal map from the unit disc onto the inside of $\gamma$ such that $\phi(0)=0$ and $\phi'(0) > 0$.
But, one can observe that almost surely, in a nested CLE sequence, only one annulus between successive loops (that we call $A(1)$) does contain the point $1$. Hence, $\nu^A$ restricted to those annuli that contain $1$ is a probability measure, and this probability measure $P^{A, 1}$ is the law of $A(1)$.
One can apply a similar construction to the full-plane SLE$_{8/3}$ loop soup, focusing on the annuli that are in between a loop $\gamma_j^i$ and the next outer boundary $\gamma_{j+1}^e$. It follows (using the fact that ${\nu^{\rm cle}}$ is equal to a constant times $\nu^e$) that, up to a multiplicative constant (corresponding to the probability that a given point is in between to such loops) the measure $\nu^A$ describes also the intensity measure of such annuli in the full-plane loop soup. We can now use the inversion-invariance of the full-plane loop soup and the fact that $\nu^i=\nu^e$, to conclude that that $\nu^A$ can also be constructed from inside-out as follows: Define the inner loop via ${\nu^{\rm cle}}$ and choose the outer loop by sampling a CLE in the outside of the inner loop, and take the innermost loop in this CLE. From this, it follows that $\nu^A$ is invariant under $z \mapsto 1/z$, and therefore $P^{A, 1}$ too.
In order to prove the inversion invariance of the law of the entire nested sequence $(\gamma_j, j \in J)$, we proceed in almost exactly the same way, except that we now focus on the joint law of the $2n_0$ loops “closest” to $1$ in the sequence: Let us index the loops by $1/2 + {{\mathbb Z}}$ in such a way that the point $1$ is in between the two successive loops $\gamma_{-1/2}$ and $\gamma_{1/2}$. Let us choose any integer $n_0 \ge 1$, and look at the random family consisting of the $2n_0 $ loops nearest to the point $1$ i.e. $$\Gamma^{n_0}:= (\gamma_{-n_0+1/2}, \ldots, \gamma_{-1/2}, \gamma_{1/2}, \ldots , \gamma_{n_0 -1/2}).$$
One way to describe the law of of $\Gamma^{n_0}$ is to start with the infinite measure on $2n_0+2$-tuples of loops obtained by defining the first one according to the infinite scale-invariant measure ${\nu^{\rm cle}}$ and then to use $2n_0 -1$ times the Markov kernel $Q^{\to e}$ in order to define its successors, and then to restrict this infinite measure to the set of $(2n_0 +2)$-tuples of loops such that $1$ is in between the two middle ones.
Exactly the same arguments as for $n_0=1$ then show that $\Gamma^{n_0}$ can alternatively be defined inside out, so that the law of $\Gamma^{n_0}$ — and therefore of the entire sequence, as this holds for all $n_0$ — is invariant under $z \mapsto 1/z$.
Remarks on the Markov chain of annular regions
----------------------------------------------
Note that the previous annuli measure $\nu^A$ is scale-invariant; we can therefore define its associated shape probability measure $P^A$. We will denote by $m (A)$ the unique $m <1$ such that $A$ can be mapped conformally onto $\{ z \, : \, m < |z| < 1 \}$.
We can note that with the description of $\nu^A$ via ${\nu^{\rm cle}}\otimes \tilde P$, the modulus $m(A)$ of the annulus is fully encoded by $\tilde \gamma$ (as it is the modulus of the part of the unit disc that is outside of $\tilde \gamma$). In particular, restricting $\nu^A$ to the set of annuli of a certain modulus (say for $m(A) \in (m_1, m_2)$), one obtains a scale-invariant measure on annuli described by the previous method from ${\nu^{\rm cle}}\otimes \tilde P^{m_1, m_2}$ (where $\tilde P^{m_1, m_2}$ means the probability $\tilde P$ restricted to those loops that define an annulus with modulus in $(m_1, m_2)$). In other words, the “marginal measure” on the outside of such annuli is just a constant $c(m_1, m_2)$ times ${\nu^{\rm cle}}$, and its shape probability is still $P^{\rm cle}$.
But our CLE symmetry result shows that it is also possible to view the nested CLE sample as being defined iteratively from inside to outside. Furthermore, the modulus of an annulus $A$ surrounding the origin and of $1/A$ are identical. Hence, it follows immediately that the marginal measure of the inside loop of the annulus (restricted to those with modulus in $(m_1, m_2)$ is also $c(m_1, m_2) {\nu^{\rm cle}}$ and that its shape probability measure is again $P^{\rm cle}$.
Hence, it follows that:
If we define the Markov kernel $Q^{\to e, (m_1, m_2)}$ just as $Q^{\to e}$ except that we restrict ourselves to those jumps that correspond to an annulus with modulus in $(m_1, m_2)$, then ${\nu^{\rm cle}}$ is again (up to a multiplicative constant) its unique invariant measure.
The following two extreme cases are of course particularly worth stressing:
\(a) When $m_1=0$ and $m_2$ gets very small, we see on the one hand by standard distortion estimates that the shapes of the inside loop and of $\tilde \gamma$ become closer and closer, and on the other hand, that shape of the inside loop is always described by the shape of $\nu$. Hence, this leads to the following description of the CLE shape distribution $P^{\rm cle}$:
Consider a CLE in the unit disc and let $\hat \gamma$ denote the outermost CLE loop that surrounds the origin, and let $m$ denote the modulus of the annulus between $\hat \gamma$ and the unit circle. Then, the law of the shape of $\hat \gamma$ conditioned on the event $\{ m < {\epsilon}\}$ does converge to the shape distribution $P^{\rm cle}$ as ${\epsilon}\to 0$.
Loosely speaking, the very small loops in a simple non-nested CLE describe the stationary shape $P^{\rm cle}$.
\(b) When $m_2=1$ and $m_1$ is very close to one, then when $m (A) > m_1$, the inside and the outside loop are (in some conformal way) conditioned to get very close to each other. Again, both the outer and the inner shape are always described by $P^{\rm cle}$. It is actually possible to make sense of the limiting kernel $Q^{\to e, (m_1, 1)}$ as $m_1 \to 1$. This gives a scale-invariant measure on “degenerate” annuli where the inside and outside loops intersect, and where the marginals of the shape measure for both the inside and the outside loops are described via $P^{\rm cle}$. In the case where $\kappa = 4$, this is very closely related to the conformally invariant growing mechanism described in [@WW] and to work in progress, such as e.g. [@SWW].
Proof of $\nu^i = \nu^e$
========================
Exploring (i.e. dynamically constructing) loop-soup clusters
------------------------------------------------------------
In the present subsection, we review some ideas and tools introduced in [@ShW] about simple CLEs, and discuss some consequences in the present setup.
In the sequel, we will say that a conformal transformation $\varphi$ from ${{\mathbb H}}$ onto a subset of the Riemann sphere defines a “marked domain” (as it gives information about the domain $\varphi ({{\mathbb H}})$ as well as the image of some marked points, say of $i$ and $0$).
In [@ShW] (see also [@WW]), it has been studied and explained how to construct a simple CLE in a simply connected domain $D$ from a Poisson point process of SLE bubbles. Let us briefly and somewhat informally recall this construction in the case where $D$ is the disc of radius $R$ around the origin. First, note that when one continuously moves from $-R$ along the segment $[-R, 0]$, one encounters loops of a CLE one after the other. The CLE property loosely speaking states that if one discovers the whole loop as soon as one bounces it, then the law of the loops in the remaining to be explored domain is still that of a CLE in that domain. This leads to the fact that the loops that one discovers can be viewed as arising from a Poisson point process of boundary bubbles (that turn out to be SLE$_\kappa$ bubbles)– see [@ShW] for details. And indeed, it is in fact possible to construct a CLE, when starting from a Poisson point process of such SLE$_\kappa$ bubbles. More precisely, one first defines the infinite measure $\mu$ on SLE$_\kappa$ loops in the upper half-plane that touch the real line only at the origin. This is the appropriately scaled limit when $\epsilon \to 0$ of the law of an SLE$_\kappa$ from $0$ to $\epsilon$ in the upper half-plane. Then, one considers a Poisson point process of such bubbles with intensity $\mu$ and from this Poisson point process, one can construct all the simple non-nested CLE loops that for instance intersect a circle $C$ that surrounds the origin (one just replaces the previous segment by the path $[-R, -r]$, where $-r \in C$ to which one then attaches the circle $C$ (along which one then moves continuously and discovers all the CLE loops that it intersects). Note also that if one discovers a loop that surrounds the circle $C$ on the way, then it is possible to continue the exploration in its inside if one is considering a nested CLE. In this way, one can constructs in fact all the loops that intersect $C$ in a [*nested*]{} CLE. See [@ShW] for details.
This procedure shows the existence of a measure $\rho^{C}_R$ on the set of marked domains so that the image of the measure $\rho^{C}_{R} \otimes \mu$ under the map $(\varphi, \beta) \mapsto \varphi (\beta)$, and restricted to those pairs for which $\varphi (\beta)$ intersects $C$, is equal exactly to the intensity measure of nested CLE loops in the disc of radius $R$, restricted to those that intersect $C$. In other words, for any set $A$ of loops that intersect $C$, if $\Gamma_R$ denotes a nested CLE in the disc of radius $R$, then $$\label {relation1}
{{\mathbf E}}\left( \sum_{\gamma \in \Gamma_R} {{\mathbbm{1}}}_{ \gamma \in A} \right) = (\rho^{C}_{R} \otimes \mu ) ( \{ ( \varphi, \beta) \, : \, \varphi (\beta) \in A \} ).$$ The previously described convergence and coupling arguments on nested CLEs when $R \to \infty$ in fact readily show that the previous statement also holds in the full-plane CLE setting (just continuing independently the exploration inside each of the discovered loops). More precisely:
There exists a measure $\rho^C$, so that if $\Gamma$ denotes a full-plane CLE, (\[relation1\]) holds when one replaces $(\Gamma_R, \rho^C_R)$ by $(\Gamma, \rho^C)$.
Let us now consider a full-plane SLE$_{8/3}$ loop soup instead and its clusters. One can then apply almost the same argument as in the nested CLE to obtain the following statement: Let $\Gamma_e$ denote the set of [*outer boundaries*]{} of clusters in this full-plane SLE$_{8/3}$ loop soup. Then:
There exists a measure $\tilde \rho_C$ on the set of marked domains, so that (\[relation1\]) holds when one replaces $(\Gamma_R, \rho^C_R)$ by $(\Gamma_e, \tilde \rho^C)$.
The two little modifications that are needed in order to justify this fact are:
- That one needs to replace the measure on SLE$_\kappa$ bubbles (i.e. boundary-touching loops) by a measure on “boundary-touching clusters”. The existence and construction of this measure is obtained in exactly the same way as the existence and construction of the CLE bubble measure in Sections 3 and 4 of [@ShW].
- That when one encounters a cluster that surrounds (or intersects) $C$, then one continues to explore inside all connected components of its complement that intersect the circle $C$ independently.
Recall that the full-plane SLE$_{8/3}$ loop soup is invariant under any Möbius transformation of the Riemann sphere. Hence, we can reformulate the previous property after applying the conformal transformation $z \mapsto 1/ (z-z_0)$. We therefore obtain, for each point $z_0$ in the plane with $z_0 \not=0$, and any (small) circle $C$ surounding $z_0$, a description of the measure on the set of those boundaries of SLE$_{8/3}$ clusters, that intersect $C$ and separate $z_0$ from the rest of the cluster (this corresponds to the fact that the previous description was describing the “outer boundaries” of the clusters, which are those that separate the cluster from infinity). In the sequel, we shall in fact in particular focus on those loops that do disconnect the origin from infinity (i.e. the $\gamma_j^e$ and $\gamma_j^i$ loops). Among those, the previous procedure describes/constructs:
- The loops $\gamma^e_j$ that do not surround $z_0$ and intersect $C$.
- The loops $\gamma^i_j$ that do surround $z_0$ and intersect $C$.
In all the remainder this section, when we will mention “the $\epsilon$-neighborhood of $z_0$” in the plane (for $z_0 \not= 0$), this will always mean the disc of radius $|z_0| \sinh \epsilon$ around $z_0 \times \cosh \epsilon$ (i.e. with diameter $[z_0 e^{-\epsilon}, z_0 e^{\epsilon}]$). In particular, we see that with this definition, (i) when $\epsilon$ is very small, the $\epsilon$-neighborhood of $1$ is quite close to the Euclidean $\epsilon$-neighborhood of $1$. Furthermore, (ii) for all $z_0 \not=1$, the $\epsilon$-neighborhood of $z_0$ is equal to the image under $z \mapsto z_0 z$ of the $\epsilon$-neighborhood of $1$, and (iii), the $\epsilon$-neighborhood of $z_0$ is invariant under the inversion $z \mapsto z_0^2/z$.
Similarly, we will denote $d(z, K)$ to be the largest $r$ such that $K$ remains disjoint of the $r$-neighborhood of $z$.
If we apply our previous analysis to the case where the circle $C$ is the boundary of the ${\epsilon}$-neighborhood of $z_0$, we therefore obtain the existence of a measure $\tilde \rho^{z_0, {\epsilon}}$ on marked domains (here the marked domain is simply connected in the Riemann sphere but not necessarily simply connected in ${{{\mathbb C}}}$) such that:
- The measure $\nu^e$ restricted to those loops that intersect $C$ and do not surround $z_0$, is equal to the image of the measure $\tilde \rho^{z_0, \epsilon} \otimes \mu$ under the mapping $(\varphi, \beta) \mapsto \varphi (\beta)$, when restricted to those loops that do separate $0$ from infinity, and do not surround $z_0$.
- The measure $\nu^i$ restricted to those loops that intersect $C$ and do surround $z_0$, is equal to the image of the measure $\tilde \rho^{z_0, \epsilon} \otimes \mu$ under the mapping $(\varphi, \beta) \mapsto \varphi (\beta)$, when restricted to those loops that do separate $0$ from infinity, and do surround $z_0$.
We can also note that inversion-invariance of the loop-soup picture shows that in each of these two statements, it is possible to replace $\tilde \rho^{z_0, \epsilon}$ by its image $\hat \rho^{z_0, \epsilon}$ under $z \mapsto z_0^2/z$ (it just corresponds to exploring/constructing the image of the loop-soup cluster boundaries under this map).
Our goal is now to build on these constructions in order to show that the measures $\nu^e$ and $\nu^i$ are very close when ${\epsilon}\to 0$. Let us denote by $V_{\epsilon}^e (\gamma)$ (respectively $V_{\epsilon}^i (\gamma)$) to be the the set of points that are in the $\epsilon$-neighborhood of a loop $\gamma$ [*and* ]{} lie outside of it (respectively inside of it). Clearly, for each loop $\gamma$, $$\int_{{{{\mathbb C}}}} {{\mathrm d}}^2 z \, {{\mathbbm{1}}}_{\{ z \in V_{\epsilon}^e (\gamma) \} } / | V_{\epsilon}^e(\gamma) | = 1 ,$$ (where $|V|$ denotes here the [*Euclidean*]{} area of $V$) so that it is possible to decompose the measure $\nu^e$ as follows: $$\begin{aligned}
\nu^e ( F(\gamma) )
&=& \nu^e \left( F(\gamma) \int_{{{{\mathbb C}}}} {{\mathrm d}}^2 z \, {{\mathbbm{1}}}_{\{ z \in V_{\epsilon}^e (\gamma) \} } / | V_{\epsilon}^e(\gamma) | \right)
\\
&=&
\int_{{{\mathbb C}}}{{\mathrm d}}^2 z \, (\tilde \rho^{z,{\epsilon}} \otimes \mu)^e \left( F( \gamma) / | V_{\epsilon}^e(\gamma) | \right),\end{aligned}$$ where $(\tilde \rho^{z,{\epsilon}} \otimes \mu)^e$ denotes the measure on $\gamma= \varphi (\beta)$ restricted to the configurations where one constructs $\gamma$ “from the outside” a loop surrounding the origin (i.e. $z$ lies on the outside of this loop). Rotation and scale-invariance shows that $$(\tilde \rho^{z,{\epsilon}} \otimes \mu)^e \left( F( \gamma) / | V_{\epsilon}^e(\gamma) | \right)
= (\tilde \rho^{1,{\epsilon}} \otimes \mu)^e \left( F( z\gamma) / ( |z|^2 | V_{\epsilon}^e(\gamma) |) \right)
.$$ We can now interchange again the order of integration, which leads to $$\label {nue}
\nu^e ( F (\gamma) ) = (\tilde \rho^{1, {\epsilon}} \otimes \mu)^e \left( {\tilde F (\gamma) } / {| V_{\epsilon}^e(\gamma) |} \right)$$ where $\tilde F (\gamma) = \int_{{{{\mathbb C}}}} {{\mathrm d}}^2 z \, F ( z \gamma)/ |z|^2
$ (provided that $F$ is chosen so that the above integrals all converge, for instance if it is bounded and its support is included in the set of loops that wind around the origin and stay in some fixed annulus $D(0, r_2) \setminus D(0, r_1)$).
Using inversion, we get the similar expression for $\nu^i$, $$\label {nui}
\nu^i ( F (\gamma) ) = (\tilde \rho^{1, {\epsilon}} \otimes \mu)^i \left( { \tilde F (\gamma) } / { | V_{\epsilon}^i(\gamma) | } \right),$$ where this time, the notation $(\tilde \rho^{1, {\epsilon}} \otimes \mu)^i$ means that we now consider the loops $\gamma = \varphi (\beta)$ that surround both the origin and $y_0$.
Let us stress that the two identities (\[nue\]) and (\[nui\]) hold for all ${\epsilon}$.
In order to explain what is going to follow in the rest of the paper, let us now briefly (in one paragraph) outline the rest of the proof: We want to prove that $\nu^e (F) = \nu^i (F)$ for a sufficiently wide class of functions $F$ and to deduce from this that $\nu^e = \nu^i$. In order to do so, we will show that in the limit when ${\epsilon}\to 0$, the two right-hand sides of (\[nue\]) and (\[nui\]) above behave similarly. For this, we will use the results and ideas of [@LR] on the Minkowski content of SLE paths, that loosely speaking will show that when $d = 1 + \kappa/8$, for $\nu^e$ and $\nu^i$ almost all loop $\gamma$, there exists a deterministic sequence ${\epsilon}_n $ that tends to $0$ such that (this will be Lemma \[minkovski\]), $$| V_{{\epsilon}_n}^i(\gamma) | \sim | V_{{\epsilon}_n}^e (\gamma) | \sim {\epsilon}_n^{2-d} L(\gamma),$$ as $n \to \infty$, where $L(\gamma)$ is a positive finite quantity related to the “natural” (i.e. geometric) time-parametrization of the SLE loop. We will rely on the one hand on this fact, and on the other hand, on the fact that when ${\epsilon}\to 0$, the measure ${\epsilon}^{d-2} (\tilde \rho^{1, {\epsilon}} \otimes \mu)^e$ converges to the same measure $\lambda$ on loops that pass through $1$ and separate the origin from infinity as ${\epsilon}^{d-2} (\tilde \rho^{1, {\epsilon}} \otimes \mu)^i$ does. Basically (we state the following fact as it may enlighten things, even though we will not explicitly prove it because it is not needed in our proof), one has an expression of the type $$\nu^i ( F (\gamma) ) = \lambda (\tilde F (\gamma) / L(\gamma)) = \nu^e (F (\gamma)).$$ Our proof will be based on a coupling argument that enables to compare the right-hand sides of (\[nue\]) and (\[nui\]). Another fact that it will be handy to use in the following steps is the reversibility of SLE$_\kappa$ paths for $\kappa \in (8/3, 4]$. There exist now several different proofs of this result first proved by Dapeng Zhan in [@Zhan], see for instance [@MS1; @WW2].
We now come back to our actual proof of the fact that $\nu^e=\nu^i$, and state the following lemma; we postpone its proof to the next and final subsection of the paper, as it involves somewhat different arguments (and results of [@LR] on the Minkowski content of chordal SLE paths).
\[minkovski\] There exists a sequence ${\epsilon}_n$ that tends to $0$, such that for $\nu^e$ almost every loop, there exists a finite positive $L(\gamma)$ such that $$\label {equiv}
\lim_{n \to \infty} {\epsilon}_n^{d-2} | V_{{\epsilon}_n}^e (\gamma) | =
\lim_{n \to \infty} {\epsilon}_n^{d-2} | V_{{\epsilon}_n}^i (\gamma) | = L(\gamma).$$
Note that this is equivalent to the fact that almost surely, for any exterior boundary of a cluster that surrounds the origin in a full-plane loop-soup sample, (\[equiv\]) holds. One could also derive the more general statement (taking the limit when ${\epsilon}\to 0$ instead of along some particular sequence ${\epsilon}_n$) but Lemma \[minkovski\] will be sufficient for our purpose.
Let us now explain how to use this lemma in order to conclude our proof. We will couple $(\tilde \rho^{z,{\epsilon}} \otimes \mu)^e$ with $(\hat \rho^{z,{\epsilon}} \otimes \mu)^i$. Note first that because of inversion-invariance of the full-plane SLE$_{8/3}$ loop-soup, the total masses of the two measures $(\tilde \rho^{z,{\epsilon}} \otimes \mu)^e$ and $(\hat \rho^{z,{\epsilon}} \otimes \mu)^i$ are identical. In fact, these masses decay like a constant times ${\epsilon}^{2-d}$ as ${\epsilon}\to 0$ (but for what follows, it will be enough to note that they are bounded by a constant time ${\epsilon}^{2-d}$).
Let us now describe how we define our coupling: First, for a choice of a marked domain by $\tilde \rho^{z, {\epsilon}}$ in our first measure, we consider the one obtained by the inversion $y \mapsto z^2 / y$ for the “sample” of $\hat \rho^{z, {\epsilon}}$. Hence, after mapping our marked domains onto the unit disc in such a way that $\infty$ and the origin are mapped onto two real symmetric points $-a$ and $a$ respectively, we have to compare/couple the following two measures on bubbles:
- The measure on SLE$_\kappa$ bubbles in the unit disc, rooted at some point $e^{i \theta}$ and restricted to the set of bubbles that surround $a$ and not $-a$
- The measure on SLE$_\kappa$ bubbles in the unit disc, rooted at the point $-e^{i \theta}$ and restricted to the set of bubbles that surround $-a$ and not $a$.
By symmetry, these two measures have again the same mass. The goal is now to couple two loops $\beta^1$ and $\beta^2$ (each defined under these SLE$_\kappa$ bubble measures) in such a way that when $a$ is small, for most realizations of $\beta^1$ and $\beta^2$, the two loops are in fact very similar in the neighborhood of the origin (in the disc of radius $\sqrt {a}$, say). This would indeed then imply that when mapped back onto the marked domain, the loops are very close, except in a small neighborhood of $z$ (and therefore very close everywhere).
Let us first sample progressively a part of $\beta_1$ starting from its root $e^{i \theta}$. One natural way to encode this exploration in the present setting is to always map the complement of the curve in the unit disc back to the unit disc, in such a way that the two point $-a$ and $a$ are mapped onto two symmetric real values $-a_t$ and $a_t$. This fixes the conformal transformation (note also that $a_t$ is increasing with time, which can enable to use $a_t$ to define a convenient time-parametrization). The tip of the curve is mapped onto some $e^{i \varphi_t}$ while the target (i.e. one of the images of $e^{i\theta}$) is mapped onto some $e^{i \theta_t}$ on the unit circle. We are interested in the time $T$ (when it exists), which is the first time at which $ e^{i\varphi_t} = - e^{i \theta_t}$. Then at this time, after mapping back the complement of the already discovered part of $\beta^1$ to the unit disk, the remaining to be discovered path is an SLE$_\kappa$ from the random point $b:= e^{i \varphi_T}$ to $-b$ that we restrict to the event that it disconnects $-a_T$ from $a_T$.
We shall see a little later (it will convenient to explain this in the next subsection, together with some other result proved there) that:
\[thealemma\] Consider the SLE bubble measure rooted at $e^{i \theta}$ and restricted to those that disconnect $-a$ from $a$. When $a \to 0$, the proportion of loops such that $T < \tau_{a^{3/4}}$ tends to one. Here $\tau_r$ is the hitting time of the circle around $r$ in the previous parametrization.
By symmetry, for the begining of the loop starting from $- e^{i \theta}$, we can use exactly the symmetric path (with respect to the origin), so that at the same time $T$, the configuration is exactly symmetric one (see Figure \[couplingidea\]). In both cases, modulo the conformal transformation corresponding to the curve up to time $T$, the remaining curve is just an SLE from $b$ to $-b$ in the unit disc, restricted to those configurations that separate $-a_T$ from $a_T$. Then, in our coupling we can use the very same SLE sample (i.e. using reversibility of the SLE) in both cases for this remaining SLE.
\[couplingidea\]
This therefore defines a coupling $\bar \rho^{1,{\epsilon}}$ of the two measures $(\tilde \rho^{z,{\epsilon}} \otimes \mu)^e$ and $(\hat \rho^{z,{\epsilon}} \otimes \mu)^i$, in such a way that $(\gamma^1, \gamma^2)$ are defined each under these two measures (with same total mass) and are very close (for a large proportion of their masses). Furthermore, with this coupling, the quantities of the type $V_{\epsilon}^i (\gamma)$ and $V_{\epsilon}^e (\gamma)$ are very close as well (when ${\epsilon}$ is small) for $\gamma^1$ and $\gamma^2$ (because except on a small piece very close to $z_0$, the two loops are the conformal images of the same SLE$_\kappa$ path under almost the same conformal transformation).
Let $$\overline L (\gamma) := \sup_{n} \epsilon_n^{d-2} \max ( | V_{\epsilon_n}^i (\gamma) |, |V_{\epsilon_n}^e (\gamma) | ).$$ We know that this quantity is finite for almost all loops that we are considering (because of Lemma \[equiv\]).
Let $R>0$, and choose $\Phi_R: [0, \infty) \to [0,1]$ to be some smooth function that is equal to 1 on $[2/R, \infty)$ and to $0$ on $[0, 1/R]$. We let $\tilde \Phi_R (\gamma) = \Phi_R (\overline L(\gamma))$.
We can then simply estimate $$ | \^e ( F () \_R () ) - \^i ( F () \_R () ) | \_n\^[d-2]{} |\^[1, \_n]{} ( | - | ) . $$ By dominated convergence — the integrand is bounded, the total mass remains bounded and the integrand gets close to zero except on a smaller and smaller portion of the space — we see that this tends to $0$ as $n \to \infty$.
Since this holds for all $F$ and $R$, we conclude that in fact $\nu^i = \nu^e$.
Minkowski content and symmetry
------------------------------
This final subsection will mostly devoted to the proof of Lemma \[minkovski\]. In fact, we will deduce it from the following similar lemma concerning chordal SLE paths (and not loops). Suppose that $J$ is non-negative continuous compactly supported function from $\overline {{\mathbb H}}$ to ${{{\mathbb {R}}}}$ that is equal to $0$ on a neighborhood of the origin, and that $\beta$ is a simple curve from $0$ to infinity in ${{\mathbb H}}\cup \{ 0\}$ (starting at $0$, and tending to infinity at the other end). One can then define $H^+ (\beta)$ and $H^-(\beta)$, the two connected components of ${{\mathbb H}}\setminus \beta$ that respectively lie to “its right” and to “its left”, and define $$v_{{\epsilon}, J} (\beta)= \int_{{{\mathbb H}}} {{\mathrm d}}^2 z \, J(z) {{\mathbbm{1}}}_{ d(z,\beta) < {\epsilon}}, \quad v_{{\epsilon}, J}^ + (\beta)
= \int_{H^ + ( \beta)} {{\mathrm d}}^2 z \, J( z) {{\mathbbm{1}}}_{d(z,\beta) < {\epsilon}}$$ and $v_{{\epsilon}, J}^- (\beta) = v_{{\epsilon}, J} (\beta) - v_{{\epsilon}, J}^+ ( \beta)$.
\[minkovski2\] Suppose that $\kappa \in (8/3, 4]$. There exists a sequence ${\epsilon}_n$ that tends to $0$, such that for all non-negative continuous compactly supported function $J$ defined on ${{\mathbb H}}$, one has for almost every chordal SLE$_\kappa$ path $\beta$, $
v_{{\epsilon}_n, J}^+ (\beta) \sim v_{{\epsilon}_n, J}^- (\beta)
$ as $n \to \infty$.
Before proving this lemma, let us first explain how one can deduce Lemma \[minkovski\] from it:
1. The arguments of [@LR], see analogous statement in Theorem 3.1 of [@LR] (preprint version), go through with basically no modification in order to prove that for all given $J$ as above, for $d= 1+ \kappa/8$, and for almost all SLE$_\kappa$ curve $\beta$, $$\lim_{{\epsilon}\to 0} {\epsilon}^ {d-2} v_{{\epsilon}, J} (\beta) = L_{J} (\beta),$$ where $L_J (\beta)$ is positive on the set of curves that pass through the support of $J$. The only difference with [@LR] is the presence of the weighting $J$ (but this is in fact also treated there in the context of the “covariant” properties of the Minkowski content). Combining this with Lemma \[minkovski2\] ensures that almost surely, $$v_{{\epsilon}_n, J}^+ (\beta) \sim v_{{\epsilon}_n, J}^- (\beta) \sim v_{{\epsilon}_n, J} (\beta)/2 \sim{\epsilon}_n^{2-d} L_J (\beta) /2$$ as $n \to \infty$.
2. Consider now instead of a chordal SLE path, an SLE bubble defined under the infinite measure $\mu$ (its description as an SLE excursion measure is for instance described in [@ShW]), and keep a (compactly supported) function $J$ as before. Let $2r$ denote the distance between the origin and the support of $J$. We can discover a first bit of the SLE bubble until it reaches the circle of radius $r$ around the origin (if it does so, which is the case for a set of bubbles with finite $\mu$-mass). If we renormalize the measure appropriately, we can say that “conditionally” on this first part, the law of the remaining-to-be-discovered part of the bubble is a chordal SLE in the slit upper half-plane joining the tip of the already-discovered part to $0$. If we map this onto the upper half-plane and the tip and $0$ respectively onto $0$ and $\infty$, we see readily that the result stated in Step 1 (applied to a random function $\tilde J$ that depends on $J$ and the first part of the bubble) actually yields that for any $J$ as before, for $\mu$-almost every bubble $\beta$, $$v_{{\epsilon}_n, J}^i (\beta) \sim v_{{\epsilon}_n, J}^e (\beta) \sim {\epsilon}_n^{2-d} L_J (\beta) /2$$ as $n \to \infty$, where the superscripts $i$ and $e$ now stand for the “interior” and the “exterior” of the bubble $\beta$.
3. We are now going to restrict ourselves to the study of loops $\gamma$ in the plane that surround the origin and stay confined in a given annulus $A=\{ z \, : \, r_1 < |z| < r_2 \}$. Let us define the functions $f^+ $ and $f^-$ in the plane such that $z \mapsto f^+(z) = f^- (-z)$ is equal to $1$ on $\{ z \, : \, \Re (z) > r_1 /2 \}$, to $0$ on $\{ z \, : \, \Re (z) < -r_1 /2 \}$ and to $1/2 + (\Re (z))/ r_1)$ in the medial strip $\{ \Re (z) \in [-r_1/2, r_1/2]\}$. Note that $f^+ + f^- = 1$. We claim that in order to prove Lemma \[minkovski\], it is sufficient to show that for $\nu^e$ almost all loop $\gamma$, $$\lim_{n \to \infty} {\epsilon}_n^{d-2} | V_{{\epsilon}_n, f^+}^e (\gamma) | = \lim_{n \to \infty} {\epsilon}_n^{d-2} | V_{{\epsilon}_n, f^+}^i (\gamma) | = L_{f^+} (\gamma)$$ as $n \to \infty$, where $L_{f^+}$ is positive on the set of loops under consideration (that wind around the small disc of radius $r_1$). Indeed, by symmetry, the same result is then true for $f^-$, and adding the two contributions (for $f^+$ and $f^-$), we get that $$\lim_{n \to \infty} {\epsilon}_n^{d-2} | V_{{\epsilon}_n}^e (\gamma) | =
\lim_{n \to \infty} {\epsilon}_n^{d-2} | V_{{\epsilon}_n}^i (\gamma) | = (L_{f^+} + L_{f^-})(\gamma).$$
4. In order to prove this statement with $f^+$, we can consider a full-plane SLE$_{8/3}$ picture, and discover all loop-soup clusters “from the outside”, by first discovering (at once) all clusters that intersect the circle of radius $r_2$, and then “moving inwards” along the segment $[-r_2, -r_1]$. In this way, we can represent the measure $\nu^e$ (restricted to the set of loops that wind around the origin in the annulus $A$) as the image of a measure $(\rho \otimes \mu)$ under a map $(\varphi, \beta) \mapsto \varphi (\beta)$ as before, where $\varphi (0)$ is always on the segment $[-r_2, -r_1]$ which is a point in the neighborhood of which $f^+$ vanishes. Applying the result of Step 2 to the function $$| (\varphi^{-1})' (\cdot) |^{2-d} f^+ (\varphi^{-1} ( \cdot )),$$ we then immediately get the statement that is needed in Step 3 (the fact that this function explodes in the neighborhood of some part of the real line does not matter, as the bubble almost surely remains at positive distance from those parts).
It then remains to explain how to adapt the ideas of [@LR] in order to prove Lemma \[minkovski2\], and to also explain how to derive Lemma \[thealemma\]. The results on the Minkowski content of chordal SLE in [@LR] are based on the one hand on the fact that when $z_0$ is a point in the upper half-plane, then one has a very good control on the probability that the SLE intersects the ${\epsilon}$-neighborhood of $z_0$: More precisely, if one maps the upper half-plane onto the unit disc by the conformal transformation $\phi$ such that $\phi (z_0)=0$ and $\phi (\infty) = -1$, then one can view the (chordal) SLE from $\phi(0)$ to $-1$ in the unit disc as a Loewner chain in the unit disc, viewed from the origin. If one re-parametrizes it via the log-conformal radius $u=u(t)$ seen from $0$, one gets easily that the argument $\theta_u$ of $\phi_t ( W_t)$ (where $\phi_t$ maps ${{\mathbb H}}\setminus \beta[0,t]$ onto ${{\mathbb U}}$ with $\phi_t(\infty) = -1$ and $\phi_t (z_0) =0$) evolves according to the SDE $$\label {theSDE}
{{\mathrm d}}\theta_u = \sqrt {\kappa} \,{{\mathrm d}}B_u + \frac {4-\kappa}{2} \tan \frac {\theta_u}{2} \,{{\mathrm d}}u$$ where $B$ is an standard one-dimensional Brownian motion. The log-conformal radius of ${{\mathbb H}}\setminus \gamma [0,\infty)$ seen from $z_0$ can therefore expressed as the life-time of the diffusion $\theta$ (the time it takes before hitting $-\pi$ or $\pi$).
One can note that in this framework, it is also easy to see on which side of $\gamma$ the point $z_0$ is. It will be in $H^+$ if and only if the diffusion $\theta$ hits $-\pi$ before $\pi$. As a consequence, this setup and the same arguments can be used to see that when ${\epsilon}\to 0$, $$\label {thirdeq}
{{\mathbf P}}( z_0 \in H^+ (\beta) \, | \, d( \beta, z_0) \le {\epsilon}|z_0|) \to 1/2.$$ One can for instance use a coupling argument, at the first time at which the previous diffusion $\theta$ hits $0$ (after which one couples $\theta$ with $-\theta$), and to notice that the probability that the probability that the diffusion spends a long time in $(0, \pi)$ is much smaller than the probability that it spends a long time in $(-\pi, \pi)$, regardless of the starting point of the diffusion. This shows that the convergence in (\[thirdeq\]) is in fact uniform with respect to $z_0$. In other words, there exists a function $f (v)$ that decreases to $0$ as $v$ decreases to $ 0$, such that, for all $z_0$ and all ${\epsilon}< |z_0| / 2$, $$\label {firsteq}
\left|{{\mathbf P}}( z_0 \in H^+ (\beta) \, | \, d(z_0, \beta) < {\epsilon}) - {{\mathbf P}}( z_0 \in H^- (\beta) \, | \, d(z_0, \beta) < {\epsilon}) \right| \le f({\epsilon}/ |z_0|).$$ From this, it follows in particular that (with the notations of Lemma \[minkovski2\]) $E(v_{{\epsilon}, J}^+ (\beta)) \sim E( v_{{\epsilon}, J}^- (\beta))$ as ${\epsilon}\to 0$.
The second main ingredient in the proofs of [@LR] is a control on the second moments of $v_{{\epsilon}, J}$ and their variation with respect to ${\epsilon}$ (i.e. of the variations of a smoothed out version with respect to ${\epsilon}$). One can summarize this type of result (that can be found in [@LR]) as follows: Define $T_{\epsilon}(z)$ as before as the hitting time of the disc of radius ${\epsilon}$ around $z$ by the chordal SLE$_\kappa$ $\beta$:
\[secondlemma\] For any compact subset $K$ of the upper half-plane, for each small $a >0$, there exists a positive $b=b(a)$ and a constant $c(K)$ such that for all $z, y$ in $K$ and all ${\epsilon}^{1-2a} < |z-y| $, $${{\mathbf P}}\left( T_{{\epsilon}^{1-a}} (z) < T_{\epsilon}(y) < T_{\epsilon}(z) < \infty \right)
\le c(K) \, {\epsilon}^{b} \, {{\mathbf P}}( T_{\epsilon}(y) < T_{\epsilon}(z) < \infty ).$$
We do not repeat the proof here, but it can be found in Sections 2.3 and 4.2 of the preprint version of [@LR].
This estimate means in particular, that if one conditions on $T_{\epsilon}(y) < T_{\epsilon}(z) < \infty$, then with a large conditional probability, at time $T_{\epsilon}(y)$, the SLE did not get too close to $z$ yet. But by the previous estimate applied to the SLE paths after this time, the conditional probabilities that it gets then close to $z$ and passes to its right is very close to the conditional probability that it gets very close to $z$ and passes to its left.
Let us be more specific: Let us define the event $E := \{ T_{\epsilon}(y) < \infty, \, T_{\epsilon}(y) < T_{{\epsilon}^{1-a}} (z) \}$. Note that $E \subset \{ T_{\epsilon}(y) < \infty \}$ and that $E$ is measurable with respect to $\beta_{[0, T_{\epsilon}(y)]}$. Clearly, $$\begin{aligned}
\lefteqn {
\left| {{\mathbf P}}\big( T_{\epsilon}(y) < T_{\epsilon}(z) < \infty, \; z \in H^+ ( \beta) \big)
- {{\mathbf P}}\big( T_{\epsilon}(y) < T_{\epsilon}(z) < \infty, \; z \in H^- ( \beta) \big) \right| } \\
&\leq
{{\mathbf P}}( T_{{\epsilon}^{1-a}} (z) < T_{\epsilon}(y) < T_{\epsilon}(z) < \infty ) \\
& \qquad
+ \left| {{\mathbf E}}\left[ {{\mathbbm{1}}}_{E}
\left( \;{{\mathbf P}}\left[ T_{\epsilon}(z) < \infty, \; z \in H^+ ( \beta) \,\Big|\, \beta_{[0, T_{\epsilon}(y)]} \right]
- {{\mathbf P}}\left[ T_{\epsilon}(z) < \infty, \; z \in H^- ( \beta) \,\Big|\, \beta_{[0, T_{\epsilon}(y)]} \right] \;
\right) \right] \right| .
\end{aligned}$$ But on the event, $E$, conditionally on $\beta_{[0, T_{\epsilon}(y)]}$, if one applies the conformal Markov property at the time $T_{\epsilon}(y)$, simple distortion estimates yield that the disc of radius ${\epsilon}$ around $z$ gets mapped to a shape that is very close to a disc around the image of $z_0$, and that the radius of this disc is much smaller than the distance of this image to the real line. Furthermore, all these estimates are uniform enough, so that we can use (\[firsteq\]) in order to conclude that in fact that the last line in the previously displayed equation is bounded by $$f_1 ({\epsilon}) C'(K) {{\mathbf P}}( T_{\epsilon}(y) < T_{\epsilon}(z) < \infty )$$ and therefore $$\begin{aligned}
&& \left| {{\mathbf P}}\left( T_{\epsilon}(y) < T_{\epsilon}(z) < \infty, \, z \in H^+ ( \beta) \right)
- {{\mathbf P}}\left( T_{\epsilon}(y) < T_{\epsilon}(z) < \infty, \, z \in H^- ( \beta) \right) \right| \\
&&\le f_2 ({\epsilon}) \, C'(K) \, {{\mathbf P}}\left( T_{\epsilon}(y) < T_{\epsilon}(z) < \infty \right)
\end{aligned}$$ for some functions $f_1$ and $f_2$ that tend to $0$ at $0$.
Using reversibility of SLE$_\kappa$ and applying the same reasoning to the curve $-1/\beta$, one can gets the similar bound $$\begin{aligned}
&& {
\left| {{\mathbf P}}\left( T_{\epsilon}(z) < T_{\epsilon}(y) < \infty, \, z \in H^+ ( \beta) \right)
- {{\mathbf P}}\left( T_{\epsilon}(z) < T_{\epsilon}(y) < \infty, \, z \in H^- ( \beta) \right) \right| } \\
&& \le f_2( {\epsilon}) \, C''(K) \, {{\mathbf P}}\left( T_{\epsilon}(z) < T_{\epsilon}(y) < \infty \right) .\end{aligned}$$ Hence, adding up the two previous bounds, we get that $$\begin{aligned}
&& {
\left| {{\mathbf P}}\left( T_{\epsilon}(y) < \infty, \, T_{\epsilon}(z) < \infty, \,{ z \in H^+ ( \beta) }\right)
- {{\mathbf P}}\left( T_{\epsilon}(y) < \infty, \, T_{\epsilon}(z) < \infty, \,{ z \in H^- ( \beta) }\right) \right| }
\\
&& \le f_2( {\epsilon}) \,C'''(K)\,
{{\mathbf P}}( T_{\epsilon}(y) < \infty, T_{\epsilon}(z) < \infty ) .
\end{aligned}$$ Integrating this inequality with respect to ${{\mathrm d}}^2 y \, {{\mathrm d}}^2 z \, J(y) \, J(z)$, one gets $${{\mathbf E}}\left( | v_{{\epsilon}, J}^+ (\beta) - v_{{\epsilon}, J}^- (\beta) | \times v_{{\epsilon}, J} ( \beta) \right)
\le f_2({\epsilon}) \, c(J) \, {{\mathbf E}}\left( (v_{{\epsilon}, J} (\beta))^2 \right).$$ But clearly, the left-hand side is larger than $${{\mathbf E}}\left( ( v_{{\epsilon}, J}^+ (\beta) - v_{{\epsilon}, J}^- (\beta))^2 \right)$$ so that Lemma \[minkovski2\] follows via the Borel-Cantelli Lemma (just choose ${\epsilon}_n$ such that $\sum f_2 ({\epsilon}_n) < \infty$). Here we used that ${{\mathbf E}}\left( (v_{{\epsilon}, J} (\beta))^2 \right)$ is uniformly bounded in ${\epsilon}$, see analogous statement in Theorem 3.1 of [@LR] (preprint version).
We now very briefly indicate how to prove Lemma \[thealemma\]. Let us come back to the description of chordal SLE via the SDE (\[theSDE\]) $${{\mathrm d}}\theta_u = \sqrt {\kappa} \,{{\mathrm d}}B_u + \frac {4-\kappa}{2} \tan \frac {\theta_u}{2} \, {{\mathrm d}}u$$ and describe briefly the outline of this rather standard argument.
1. Let $\sigma$ denote the first time at which the SLE curve generated by $\theta_u$ (in this $u$-parametrization) reaches the circle of radius $a^{1/2}$ around the origin (if it does so this time is actually deterministic and equal to $\log (1/a) /2$, and otherwise we let $\sigma= \infty$). A standard Harnack inequality type argument shows that, uniformly over the choices of the starting point $\hat \theta $ and the end-point $\tilde \theta$ in $(- \pi, \pi)$, $${{\mathbf P}}\left(
\theta [0, \sigma] \subset (-\pi, 3\pi/4) \hbox { or }
\theta [0, \sigma] \subset (-3\pi/4, \pi) \, \left| \,
\sigma < \infty, \, \theta_0=\hat \theta , \, \theta_\sigma=\tilde \theta \right. \right) \to 0$$ as $a \to 0$.
2. Since the previous statement is uniform in $\tilde \theta$, it follows also that $${{\mathbf P}}\left(
\theta [0, \sigma] \subset (-\pi, 3\pi/4) \hbox { or }
\theta [0, \sigma] \subset (-3\pi/4, \pi) \, \left| \,
\sigma' < \infty, \, \theta_0=\hat \theta , \, \theta_{\sigma'}=\tilde \theta \right. \right) \to 0$$ as $a \to 0$, where this time $\sigma'$ is the time at which the SLE reaches the circle of radius $2a$ around the origin, if it does so.
3. Next, we notice that there exists $c>0$ so that for all $a$ small enough, and uniformly with respect to $\hat \theta$, $${{\mathbf P}}\left( \theta_{\sigma'} \in (-\pi/2, \pi/2) \, \left| \,
\theta_0 = \hat \theta , \, \sigma' < \infty \right. \right) \ge c$$ and $${{\mathbf P}}\left( \gamma \hbox { disconnects } -a \hbox { from } a \, \left| \,
\sigma' < \infty, \, \theta_0=\hat \theta , \, \theta_{\sigma'} \in (-\pi/2, \pi/2) \right. \right) \ge c.$$ If we put the pieces together, it follows that uniformly with respect to the starting point $ \theta_0$, we have for all small $a$, $$c^2 \; {{\mathbf P}}( \sigma' < \infty ) \le {{\mathbf P}}( \gamma \hbox { disconnect } -a \hbox { from } a )
\le {{\mathbf P}}( \sigma' < \infty).$$
4. We combine these estimates to get $$\begin{aligned}
&{{\mathbf P}}\left(
\theta [0, \sigma] \subset (-\pi, 3\pi/4) \hbox { or }
\theta [0, \sigma] \subset (-3\pi/4, \pi) \, \left| \,
\gamma \hbox { disconnects } -a \hbox { from } a , \, \theta_0=\hat \theta
\right. \right) \\
&\leq \frac{1}{c^2}
{{\mathbf P}}\left(
\theta [0, \sigma] \subset (-\pi, 3\pi/4) \hbox { or }
\theta [0, \sigma] \subset (-3\pi/4, \pi) \, \left| \,
\sigma' < \infty , \, \theta_0=\hat \theta
\right. \right)
\to 0\end{aligned}$$ uniformly over the choices of the starting point $\hat \theta $. If we let the starting point $\theta_0$ converge to $\pm \pi$, the lemma follows.
[**Acknowledgements.**]{} We acknowledge support of the Academy of Finland, of the Agence Nationale pour la Recherche under the grant ANR-MAC2, of the Fondation Cino Del Duca and the Einstein Foundation, and the hospitality of Université Paris-Sud, of TU Berlin, the University of Cambridge and FIM Zürich, where part of this work has been elaborated in these past years. We thank Greg Lawler for inspiring discussions, and Jason Miller and Sam Watson, David Wilson and Hao Wu for their useful comments.
[10]{}
D. Chelkak, H. Duminil-Copin, C. Hongler, A. Kemppainen, and S. Smirnov. , 352(2):157–161, 2014.
D. Chelkak and S. Smirnov. . , 189(3):515–580, 2012.
B. Doyon. . , 103: 233–284, 2013.
J. Dubédat. . , 22(4):995–1054, 2009.
G. Lawler, and M.A. Rezaei. Minkowski content and natural parameterization for the Schramm-Loewner evolution, , to appear.
G. Lawler, O. Schramm, and W. Werner. . , 16(4):917–955, 2003.
G. F. Lawler, O. Schramm, and W. Werner. . , 187(2):275–308, 2001.
G. F. Lawler and W. Werner. . , 128(4):565–588, 2004.
G. F. Lawler and B. M. Werness. . , to appear.
J. Miller and S. Sheffield. Imaginary Geometry I. Interacting SLEs, preprint.
J. Miller and S. Sheffield. Imaginary Geometry II. Reversibility of SLE$_\kappa (\rho_1;\rho_2)$ for $\kappa \in (0,4)$, preprint
J. Miller and S. Sheffield. Imaginary Geometry III. Reversibility of SLE$_\kappa$ for $\kappa \in (4,8)$, preprint
J. Miller and S. Sheffield. Imaginary Geometry IV: interior rays, whole-plane reversibility, and space-filling trees, preprint
J. Miller, S.S. Watson and D.B. Wilson. Extreme nesting in the conformal loop ensemble, preprint, arxiv:1401.0217.
. Nacu and W. Werner. . , 83, 789–809, 2011.
O. Schramm . 118:221–288, 2000.
O. Schramm and S. Sheffield. . , 202: 21–137, 2009.
O. Schramm, S. Sheffield, and D. B. Wilson. . , 288(1):43–53, 2009.
S. Sheffield. . , 147(1):79–129, 2009.
S. Sheffield, S.S. Watson, H. Wu.
S. Sheffield and W. Werner. . , 176, 1827–1917, 2012.
W. Werner. . , 337(7):481–486, 2003.
W. Werner. . Ecole d’été de physique des Houches LXXXIII, Pages 57–99, 2006.
W. Werner. . , 21(1):137–169, 2008.
W. Werner and H. Wu, . 320: 637–661, 2013.
W. Werner and H. Wu, , 18, paper 36, 2013.
D. Zhan, . 36:1472-1494, 2008.
Department of Mathematics and Statistics
P.O. Box 68
00014 University of Helsinki, Finland
Department of Mathematics
ETH Zürich, Rämistr. 101
8092 Zürich, Switzerland
[email protected]
[email protected]
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
We present a measurement of the cross section for $Z$ boson production times the branching fraction to tau lepton pairs $\sigma (p\bar{p} \to
Z + X) \cdot$Br$(Z\rightarrow \tau^+ \tau^-)$ in $p \bar{p}$ collisions at $\sqrt{s} = 1.96$ TeV. The measurement is performed in the channel in which one tau lepton decays into a muon and neutrinos, and the other tau lepton decays hadronically or into an electron and neutrinos. The data sample corresponds to an integrated luminosity of 1.0 fb$^{-1}$ collected with the D0 detector at the Fermilab Tevatron Collider. The sample contains 1511 candidate events with an estimated 20% background from jets or muons misidentified as tau leptons. We obtain $\sigma \cdot \rm{Br}$ $= 240 \pm
8$(stat)$\pm$12(sys)$\pm$15(lum) pb, which is consistent with the standard model prediction.
date: 'August 8, 2008'
title: 'Measurement of $\sigma (p\bar{p}\to Z + X)\cdot{\rm Br}(Z\to\tau^+\tau^-)$ at $\sqrt{s} = 1.96$ TeV'
---
list\_of\_authors\_r2.tex
The resonant production of tau lepton pairs is as interesting for the study of standard model (SM) physics as the production of lighter lepton pairs. For new phenomena, especially for decays of particles coupled to mass, such as SM or supersymmetric Higgs bosons, the detection of resonant pairs of tau leptons becomes even more interesting. This is due to the fact that tau leptons are much heavier than the other leptons, increasing the chance that these new phenomena would be observed first in this channel. Unfortunately, the detection of tau leptons is far more difficult than that of muons or electrons.
A measurement of $\sigma(p\bar{p} \to Z + X) \cdot$Br($Z\rightarrow
\tau^+ \tau^-$) in $p \bar{p}$ collisions at $\sqrt{s} = 1.96$ TeV is described in this Letter. The analysis is based on an event sample containing a single isolated muon from a tau lepton decay and a tau candidate reconstructed as a narrow jet that could be produced by a tau lepton decaying either hadronically or into an electron and neutrinos. This measurement is of interest not only as a test of the SM prediction but also because any excess over the expected $\sigma
\cdot$Br could be an indication of a source other than $Z$ bosons for events containing tau lepton pairs, such as the Higgs boson [@other_source_of_taus]. The precision of this result is significantly improved compared to earlier publications [@p14ZtautauPRD; @CDFZtautau].
The analysis presented here is based on data collected between September 2002 and February 2006 by the D0 experiment, corresponding to an integrated luminosity of 1003 $\pm$ 62 pb$^{-1}$ [@newlumi].
The D0 detector [@run2det] is a general purpose, axially and forward-backward symetric detector, consisting of a central-tracking system located within a 2 T superconducting solenoidal magnet, surrounded by three liquid-argon/uranium calorimeters and a muon detector. The spatial coordinates of the D0 detector are defined using a righthanded Cartesian system with the origin in the center of the detector. The positive $z$-axis is the direction of the proton beam, the positive $y$-axis points upwards and the positive $x$-axis points out of the Tevatron ring. The azimuthal angle $\phi$ is measured with respect to the positive $x$ direction. Pseudorapidity is defined as $\eta = - \ln [\tan(\theta/2)]$, where the polar angle $\theta$ is measured with respect to the positive $z$ direction. The tracking system has coverage up to $\eta \approx 3$. The calorimeter consists of a central section (CC) covering $|\eta| \lesssim 1.1$ and two end calorimeters (EC) that extend coverage to $|\eta|\approx 4.2$, all housed in separate cryostats and segmented into cells of dimensions $0.1 \times 0.1$ in $\eta - \phi$ space [@run1det]. The muon system [@run2muon] provides a coverage up to $\eta \approx 2$ and is located outside the calorimeter; it consists of a layer of tracking detectors and scintillation trigger counters before 1.8 T iron toroids, followed by two similar layers after the toroids. Luminosity is measured using plastic scintillator arrays located in front of the EC cryostats, covering $2.7<|\eta|<4.4$. A three level trigger system is designed to select most interesting events based on preliminary information from the tracking, calorimetry, and muon systems, reducing the number of recorded events from the collision rate of $\approx$ 2 MHz to a rate of $\approx$ 50 Hz, which is written to tape.
The triggering strategy used in this analysis is based on the tau lepton which decays into $\mu\nu_\mu\nu_\tau$. A single muon trigger requiring hits in the muon system in combination with a high transverse momentum ($p_T$) track reconstructed in the central tracking system is required. The average trigger efficiency, ultimately parametrized as a function of $\phi$, $\eta$ and $z$ using a data sample of $Z \rightarrow \mu^+ \mu^-$ events, is ($52.3 \pm 1.4$)%. No dependence on the muon $p_T$ is observed above 15 GeV.
Most backgrounds as well as the efficiency of the selection for signal $Z \rightarrow \tau^+ \tau^-$ events are estimated using Monte Carlo (MC) simulations. All simulated samples are generated with [ pythia]{} [@pythia] using the CTEQ6.1L parton distribution function (PDF) set. Simulation of the D0 detector is done using [ geant3]{} [@GEANT]. Noise in the detector and the contributions from other simultaneous interactions are simulated by adding random untriggered data events to the MC simulation. These events were chosen such that the effective instantaneous luminosity distribution in MC is the same as in data. The code used for the reconstruction of simulated events is identical to the one used for data.
Corrections are applied to all MC events to obtain overall good agreement between the simulation and collider data. The momentum scale and resolution for muons in the MC are tuned to reproduce the $Z$ boson invariant mass distribution observed in data. Similarly, the jet energy resolution is tuned to match that observed in data for each region of the detector. The $p_T$ spectrum of the $Z$ boson for events generated with [pythia]{} has a different shape than that measured in data; therefore the $p_T$ of the $Z$ boson is reweighted to fit the direct measurement in data [@ZpTpaper]. Small differences in acceptance between data and simulation are corrected for by weighting the simulated $z$ position of the primary vertex in MC events to reproduce that observed in data.
Reconstruction efficiencies for muons and tracks are calculated both in data and MC using samples of $Z \rightarrow \mu^+ \mu^-$ events. Efficiency correction factors for MC events as a function of muon or track $\phi$, $\eta$ and $z$ are applied. The signal or background samples are normalized to the expected number of events evaluated using the luminosity of the data sample and the theoretical values of the next-to-next-to-leading order (NNLO) cross sections in the case of $Z$ boson production [@NNLO_Zcrosssection1; @NNLO_Zcrosssection2] or next-to-leading order (NLO) cross sections for all other processes where the NNLO calculation is not available. $W$ boson production is normalized from data.
In this analysis, muons are identified starting with their signature in the muon detector. The track reconstructed from hits in the muon layers is required to match a track from the central tracking detectors. The muon momentum is measured using only the central tracking detectors.
A tau candidate is a collection of (i) a calorimeter cluster reconstructed using a simple cone algorithm [@simple_cone], (ii) tracks associated with the calorimeter cluster of which at least one has $p_T>1.5$ GeV but with a total invariant mass less than $1.8$ GeV, and (iii) electromagnetic (EM) sub-clusters constructed from the cells in the EM section of the calorimeter. The size of the cone used for reconstruction of the calorimeter cluster is ${\cal R}=\sqrt{(\Delta
\phi)^2 + (\Delta \eta)^2}=0.5$, where $\Delta \phi$ is the difference in azimuthal angle, and $\Delta \eta$ the difference in pseudorapidity between the cone axis and each of the calorimeter towers. Isolation variables are calculated using a cone of ${\cal
R}=0.3$. The tracks associated with the tau candidate must also be contained within this ${\cal R}=0.3$ cone.
Tau candidates are classified as type 1, 2 or 3, depending on the numbers of tracks and EM sub-clusters they possess. Type 1 tau candidates have exactly one associated track and no EM sub-clusters, type 2 have one associated track and one or more EM sub-clusters, and type 3 have at least two associated tracks. These categories correspond roughly to pure one-prong decays, one-prong plus neutral pion decays as well as decay into electrons, and three-prong decays of the tau lepton.
Due to the large number of jets reconstructed as tau candidates, additional selection criteria must be applied in order to distinguish tau leptons from jets. Three neural networks (NN), one for each tau type, are trained using $Z \to \tau^+ \tau^-$ MC events as signal and events with a jet back-to-back with a non-isolated muon from data as background. The NNs use isolation variables based on tracks, hadronic and EM calorimeter clusters, as well as shower shape variables and correlation variables between calorimeter and tracks. Figure \[fig:NN\] shows the discrimination obtained using the NNs. Requiring that the NN output be larger than 0.9 results in a background rejection of almost a factor of 50 for all three tau types. This reduces the probability for a jet to be misidentified as a tau lepton to 1.1% for the sum of all types (from 52% without the NN output requirement), while maintaining a total efficiency of close to 70% for tau leptons which decay hadronically or to an electron and neutrinos. Electrons are treated as type 2 tau candidates, since the efficiency for them to be reconstructed as such and pass the NN output requirement is 98%. For a complete description of the neural networks and details on their performance see Ref. [@confprocCristina].
The variable chosen to best illustrate the $Z \to \tau^+\tau^-$ signal is the visible mass, given by: $${\rm Visible~Mass = \sqrt{( P_{\mu} + P_{\tau} + \Ptmiss )^2} },$$ where ${\rm P}_{\mu,\tau}=(E_{\mu,\tau},p_{\mu,\tau}^x,
p_{\mu,\tau}^y, p_{\mu,\tau} ^z)$ are the four-momentum vectors of the muon and the tau candidate, and $\Ptmiss=(\etmiss, \etmiss^x,
\etmiss^y, 0)$, with $\etmiss$ being the missing transverse energy in the event and $\etmiss^x, \etmiss^y$ being its projections along the $x$ and $y$ directions. The uncorrected missing transverse energy is defined as the vector equal in length and opposite in direction to the vectorial sum of transverse energies of the calorimeter cells. The transverse momenta of muons are subtracted from this vector, after corrections for the energy deposited by the muons in the calorimeter have been applied. When the tau candidate matches a reconstructed electron, the energy corrections derived for electrons are applied. For jets corresponding to tau candidates, the tau energy corrections described below are applied. Jet energy corrections applied to all other jets in the event are propagated to the missing $E_T$ calculation.
To compare the visible mass distributions of the tau pairs between data and MC, it is important to have the correct energy scale for the tau candidate. For type 1 tau candidates, the momentum of the track is used as the best estimate of the energy of the tau candidate when the tracking resolution is superior to the calorimeter energy resolution (up to calorimeter cluster energy of 70 GeV). For type 2 candidates matching electrons, the energy corrections derived for electrons are applied. For type 2 candidates not matching electrons and type 3 tau candidates, the energy is estimated using $$E^{\text{corr}} = \sum_{i} p_i^{\text{trk}} + E^{\text{cal}} -
\sum_{i} R(p_i^{\text{trk}},\eta)\cdot p_i^{\text{trk}},$$ where $p_i^{\text{trk}}$ is the momentum of track $i$ associated with the tau candidate, $E^{\text{cal}}$ is the energy deposited by the tau candidate in the calorimeter, and $R(p_i^{\text{trk}},\eta)$ represents the response of the calorimeter to the $\pi^\pm$ which produced track $i$ associated with the tau candidate, as a function of the energy and rapidity of the $\pi^\pm$. Typically, $0.6 <
R(p_i^{\text trk}, \eta) < 0.9$. As the resolution of the calorimeter is better than that of the tracking at calorimeter cluster energies higher than 70 GeV (type 1), 100 GeV (type 2), or 120 GeV (type 3), the energy of the calorimeter cluster is used in these cases, after applying $\eta$ and energy dependent corrections obtained from MC.
The default program in the D0 [geant]{} simulation for hadronic interactions, [geisha]{} [@Geisha], does not reproduce the charged pion response in data well. Therefore g[calor]{} [@gCALOR] is used for a more precise simulation of single charged pion interactions. The charged pion response obtained using these special simulations was found to be in reasonable agreement with preliminary data measurements in the central calorimeter [@thesis_Cristina]. The energy measurement for neutral particles, mostly important for type 2 taus, is dominated by electromagnetic showers in the calorimeter. The simulation of electromagnetic showers in [geant]{} is sufficiently accurate for the purpose of this measurement.
The preselection requires one isolated muon reconstructed within the pseudorapidity interval $ | \eta| < 1.6$. The transverse momentum of the muon as measured by the central tracking detectors must satisfy $p_T^\mu > 15$ GeV. No other muon matched to a central track with $p_T
> 10$ GeV is allowed in the event. The muon isolation requires the sum of energies of all cells situated in a hollow cone around the direction of the muon with $0.1<{\cal R}<0.4$, as well as the sum of all tracks in a cone of ${\cal R}<0.5$, excluding the muon track, to be less than 2.5 GeV.
The preselection further requires one tau candidate with $p_T >
15$ GeV, $|\eta| < 2$, scalar sum of the transverse momenta of all tracks associated with the tau candidate $>15$ GeV for types 1 and 3 and $> 5$ GeV for type 2 tau candidates, $NN > 0.3$, and no other muon matching the tau candidate. Type 3 tau candidates with two tracks are only considered if both tracks have the same charge. The tau candidate is required to have a charge with opposite sign to that of the muon. The distances in the $z$ direction at the track’s point of closest approach between the muon and the primary vertex, the tau candidate and the primary vertex, as well as the distance between the muon and the tau candidate must be less than 1 cm.
In total 8316 events pass these criteria. To reduce the $W$ + jets and the $Z \to \mu^+ \mu^-$ backgrounds, another selection criterion is used, based on a variable which gives an approximation of the $W$ boson mass, referred to as $m^*$: $$m^* = \sqrt{2 E_{\nu} E_{\mu} (1 - \cos \Delta \phi)},$$ where $E_{\nu} = \etmiss \cdot E_{\mu}/p_T^\mu$ is an approximation of the neutrino energy, and $\Delta \phi$ is the angle between $\etmiss$ and the muon in the transverse plane.
For the final selection, all the preselection criteria are applied. Additionally, the lower limit on the NN output for the tau candidates is raised to 0.9 for types 1 and 2, and to 0.95 for type 3 tau candidates. The final selection also requires $m^* < 20$ GeV. A total of 1511 events pass all the selection criteria.
The dominant remaining background arises from multijet processes, mainly from $ b \bar{b}$ events where the muon isolation requirement is met and one of the jets satisfies the tau candidate selection criteria. Another significant source of events with isolated muons and tau candidates is $W$ + jets production, where the $W$ boson decays to $\mu\nu$ and one of the jets is misidentified as a tau candidate. The $Z \rightarrow \mu^+ \mu^-$ background is reduced by the requirement that no other muon be found in the event, but a small number of events will be selected when one of the muons is not reconstructed. Small contributions are also expected from $W \rightarrow \tau \nu$ and $WW
\rightarrow l \nu l \nu$, as well as $t \bar t$ production. Contributions from $WZ$ and $ZZ$ events yield less than one event each after the final selection criteria and are therefore neglected. All backgrounds, except the multijet background, are estimated using MC simulations.
The multijet background is estimated using the data events that satisfy all requirements placed on the signal sample except that the muon and the tau candidate have the same sign charge. We call this the same-sign (SS) sample. To obtain the appropriate normalization for this background, a special data sample is selected, the “multijet sample,” containing events that pass all requirements placed on the signal sample except the isolation criteria and the cut on the tau NN output. Instead of the isolation requirement used for the signal events, the events in the multijet sample have the sum of energies of all calorimeter cells inside the hollow isolation cone in the range 2.5 to 10 GeV. The sum of all non-muon tracks $p_T$ within the track isolation cone is required to be in the same interval $2.5 - 10$ GeV. To avoid contamination from $Z \rightarrow \tau^+ \tau^-$ signal events, an upper limit on the tau NN output is placed at 0.8. To increase the statistics of this sample, the muon $p_T$ is required to be at least 10 GeV instead of 15 GeV. The multijet sample is expected to be completely dominated by multijet processes, but may also include events in which a $W$ decaying into a muon is produced in association with a jet. The $W + {\rm jets}$ contribution is reduced by requiring that the muon and the tau candidate are back to back ($|\phi_\mu -
\phi_\tau| > 2.5$). A slight excess of opposite sign (OS) over SS events is observed in the multijet sample. No significant dependence of the OS/SS ratio as a function of $p_T$ or NN output is observed for the three types of tau candidates in the multijet sample. Correction factors ($f_{\rm mj}^i$) of $1.13 \pm 0.03, 1.08 \pm 0.01$, and $1.06 \pm 0.01$ for tau types 1 to 3 are obtained, being used as discussed below to normalize the multijet background in the final signal sample.
The number of events in the SS sample is corrected for the contribution from $Z \rightarrow \mu^+ \mu^-$, $Z \rightarrow \tau^+
\tau^-$, and $W \rightarrow \tau \nu$ obtained from MC, multiplied by an additional correction factor which takes into account the difference between the charge misidentification rates in data and MC. Totals of 6 events for type 1, 16 events for type 2, and 18 events for type 3 tau candidates from $Z \rightarrow \mu^+ \mu^-$, $Z \rightarrow \tau^+\tau^-$, and $W \rightarrow \tau \nu$ are estimated to have a misidentified charge after all cuts and are subtracted from the number of SS events when the multijet background is calculated. The contribution from $W \rightarrow \mu \nu$ events is accounted for separately.
A part of the $W$ + jets background is already included in the SS sample. However, we expect a significant excess of OS events compared to the number of SS events due to the fact that a high percentage of $W + 1~\rm{jet}$ events comes from quark jets. The number of $W +
\rm{jets}$ events in data is estimated by selecting a sample that is expected to have a large contribution from $W$ boson processes and low or negligible contributions from $Z$ boson production. Such a $W$+jets enriched sample can be obtained by requiring an isolated muon with $p_T > 20$ GeV, a tau candidate with 0.3 $< NN < $ 0.8, $|\phi_\mu - \phi_\tau| < 2.7$, and $m^* > 40$ GeV. Mostly multijet and $W$+jets events contribute to this sample. The excess of OS events compared to SS events is given for the multijet background by $f_{\rm
mj}^i$ for tau type $i$. For the $W + {\rm jets}$ sample, similar factors ($f_W^i$) of 2.39 $\pm$ 1.01, 3.15 $\pm$ 1.17, and 1.6 $\pm$ 0.26 are estimated from data, in the sample with the cuts listed above, but requiring a tighter cut $m^* > 60$ GeV. Using this, we can calculate the number of $W + {\rm jets}$ events in the $W + {\rm
jets}$ enriched data sample by solving the following system of two equations for each tau type $i$: $$N_{W}^i + N_{\rm mj}^i = N_{OS}^i + N_{SS}^i$$ $$\frac{f_{W}^i - 1}{f_{W}^i +1}N_{W}^i +\frac{f_{\rm mj}^i -
1}{f_{\rm mj}^i +1} N_{\rm mj}^i = N_{OS}^i - N_{SS}^i$$ where $N_{W}^i$ is the number of $W + {\rm jets}$ events, $N_{\rm mj}^i$ is the number of multijet events and $N_{OS}^i$, $N_{SS}^i$ are the numbers of OS, respectively SS events in the $W + {\rm jets}$ enriched data sample. The ratios between the number of $W$ + jets events calculated in data by solving the above system of equations and the one expected from MC for each tau type are used as normalization factors for this background in the signal region. The uncertainty on $N_{W}^i$ from data is taken as a systematic uncertainty. The estimated number of $W$ + jets events in the signal sample, not including those in the SS sample, is $14 ~\pm$ 5 events.
Several distributions such as muon and tau candidate transverse momentum, pseudorapidity, and azimuthal angle, as well as $\etmiss$, $m^*$, and visible mass are compared between the data and the predicted sum of backgrounds and $Z \rightarrow\tau^+ \tau^-$ for the SM cross section and branching ratio. All these distributions show good agreement after each of the preselection, NN selection, and anti-$W$ requirement stages.
In Fig. \[fig:mvis\] the visible mass distribution for events which pass the final selection requirements is shown separately for each of the tau types, while Fig. \[fig:mvislog\] shows the same distribution for the sum of all types. Good agreement is observed between the data and the sum of the background SM processes and $Z
\rightarrow\tau^+\tau^-$ signal, normalized using the NNLO SM prediction [@NNLO_Zcrosssection1; @NNLO_Zcrosssection2].
Table \[tab:events\] shows the number of events expected for each tau type from each of the backgrounds, as well as from the $Z
\rightarrow \tau^+ \tau^-$ signal. It also shows the total numbers of expected background and signal events in comparison to the numbers of events observed in data for three levels of selection: preselection, preselection plus NN output requirement, and after all selection criteria are applied. Good agreement is observed between the predicted and observed numbers of events at each level of selection for all tau types.
We estimate that approximately 1.2% of all $Z \rightarrow \tau^+
\tau^-$ events have the wrong sign for either the muon or the tau candidate, therefore appearing as SS events. From the number of $Z
\rightarrow \tau^+ \tau^-$ events obtained by subtracting the estimated background from the number of events in the final sample, we calculate the number of $Z \rightarrow \tau^+ \tau^-$ events reconstructed as SS to be 17. This number is added to the number of events in the OS sample when calculating the $Z \rightarrow \tau^+
\tau^-$ cross section.
Reconstruction of a second track close to a first reconstructed track is found to be more efficient in MC than in data. A correction factor of 0.97 $\pm$ 0.028 is applied to simulated events containing type 3 tau candidates. This factor is obtained by comparing the ratios of type 3 tau candidates with two and three tracks in data and MC and taking into account that there are twice as many SS as OS combinations when one of the three tracks is lost.
---------------------------------------- ----- ----- ----- ----- ----- ----- ------ ---- ------ ----- ------ ----- ------ ----- ------ ----- ------ -----
Process
$Z /\gamma^*\rightarrow \tau^+ \tau^-$ 302 4 230 4 146 3 1469 9 1131 8 786 7 693 6 484 5 358 5
$Z/\gamma^* \rightarrow \mu^+ \mu^-$ 58 2 43 2 6.1 0.6 176 3 108 3 14.0 0.8 184 3 38 1 8.9 0.7
$WW$ 7.2 0.3 6.1 0.3 0.4 0.1 79 1 74 1 6.9 0.3 9.3 0.4 6.1 0.5 0.5 0.1
$t\bar{t}$ 2.7 0.3 2.0 0.3 0.2 0.1 33 1 28 1 2.4 0.3 29 1 4.2 0.4 0.5 0.1
$W \rightarrow \tau \nu$ 10 2 4 1 1.5 0.8 50 4 14.1 2.2 1.4 0.7 168 7 22.0 2.7 3.7 1.2
$W \rightarrow \mu \nu$ 127 11 42 5 2.1 0.9 470 18 116 9 6.7 1.9 1384 32 202 13 14.1 2.7
Multijet 208 15 46 8 25 5 584 25 123 12 61 8 2265 47 273 18 145 13
Predicted 715 18 373 11 181 7 2861 32 1594 18 878 12 4732 59 1029 23 531 15
Data
---------------------------------------- ----- ----- ----- ----- ----- ----- ------ ---- ------ ----- ------ ----- ------ ----- ------ ----- ------ -----
Systematic uncertainties on the multijet and $W + {\rm
jets}$ backgrounds are derived from the statistical uncertainties of the control samples used to estimate these backgrounds and from the systematic uncertainties on the correction factors used for their normalization.
The systematic uncertainty related to the tau energy measurement is estimated by scaling the charged pion response used for data by the largest difference found between the response measured in data and the response obtained using g[calor]{} (6%) and recalculating the acceptance applying all cuts. The value of this uncertainty is 1%.
NN systematic uncertainties are calculated using statistical ensembles of events in which each input variable is allowed to fluctuate within the difference observed between the distributions of that particular variable in data and MC. The RMS of the ratio of the number of events passing a certain NN cut to the number of events in the ensembles, called the ensemble cut ratio, is taken as a measure of the NN uncertainty. The estimated uncertainties are 4.3% for type 1, 2.0% for type 2, and 3.8% for type 3 tau candidates, which results in a total uncertainty of 2.7%.
The uncertainty due to the tau candidate track reconstruction efficiency is taken to be the same as the uncertainty on reconstructing muon tracks and is estimated using $Z \to \mu^+ \mu^-$ events to be 1.4%. The uncertainty on the correction factor due to differences between data and MC in tracking efficiency for type 3 taus is added in quadrature to this value, resulting in a total uncertainty related to the tau candidate tracks of 1.6%. The systematic uncertainties due to muon identification and muon track matching are determined to be 0.6% and 0.8%, respectively. The systematic uncertainty due to the charge misidentification is 1%. The uncertainty on the trigger efficiency is 2.7% and takes into account the bias related to the choice of the control sample, the variation due to possible background contamination, variations in time or due to changing luminosity, the choice of binning, and the choice of parameters for the efficiency, as well as the limited statistics. The uncertainty on the total integrated luminosity is 6.1% [@newlumi], with an additional systematic uncertainty of 1% related to the influence on the luminosity of applying the data quality criteria used to reject events with coherent calorimeter noise.
The PDF uncertainty of 2.0% is estimated using a NLO calculation [@Melnikov-Petriello] and the CTEQ6.1 error sets. This uncertainty is obtained from the variation in acceptance when these error sets are used, added in quadrature with the difference in acceptance when using the MRST2004 error sets at NLO and with the additional variation when going from NLO to NNLO with MRST2004. Table \[tab:syst\] summarizes all the systematic uncertainties.
Source Value
------------------------------------ -------
Tau energy scale 1.0 %
Tau identification 2.7 %
Tau track reconstruction 1.6 %
Multijet background 1.6 %
W $\rightarrow \mu \nu$ background 0.5 %
Trigger 2.7 %
Muon track match 0.8 %
Muon identification 0.6 %
Muon momentum resolution 0.4 %
Charge misidentification 1.0 %
MC statistics 0.6 %
PDF 2.0 %
Total (except luminosity) 5.2 %
Luminosity 6.2 %
: \[tab:syst\] Systematic uncertainties on the $\sigma(p
\bar{p} \rightarrow Z/\gamma^* + X) \cdot {\rm Br}(Z/\gamma^*
\rightarrow \tau^+ \tau^-)$ measurement.
The cross section times branching ratio for the process $p \bar{p}
\rightarrow Z/\gamma^*+X \rightarrow \tau^+ \tau^- + X$ is given by the number of signal events divided by the product of the total efficiency and the integrated luminosity. The number of signal events estimated from Table \[tab:events\], with the correction for signal events reconstructed as SS, is 1227. Since Table \[tab:events\] shows the estimated number of events in the $Z/\gamma^*$ mass range $15-500$ GeV, other corrections have to be made in order to compare the result of this analysis with theoretical cross sections. To limit the mass range to $60-130$ GeV, the number of events expected from the mass region $15-60$ GeV (7 events) as well as from the $130-500$ GeV mass region (26 events) are subtracted from the number of signal events in data. The total efficiency for $Z \rightarrow \tau^+
\tau^-$ events in the $60-130$ GeV mass region is $4.9 \times
10^{-3}$, which also includes the trigger efficiency of 52.3%. Finally, a factor of 0.98 [@098factor] is applied to estimate the $Z$ boson cross section as opposed to the $Z/\gamma^*$ cross section for this mass region.
Given the systematic uncertainties listed in Table \[tab:syst\] and an integrated luminosity of 1003 $\rm{pb}^{-1}$, we estimate $\sigma(p
\bar{p} \rightarrow Z + X) \cdot {\rm Br}(Z \rightarrow \tau^+
\tau^-)$ = 240 $ \pm$ 8 (stat) $\pm$ 12 (sys) $\pm$ 15 (lum) pb, which is in good agreement with the SM prediction of $251.9^{+5.0}_{-11.8}~\rm{pb}$ [@NNLO_Zcrosssection1; @NNLO_Zcrosssection2] that results from the NNLO calculation using the MRST2004 PDFs, as well as with the $241.6^{+3.6}_{-3.2}~\rm{pb}$ [@NNLO_Zcrosssection1; @cross_sect_CTEQ] value obtained at NNLO using the CTEQ6.1M PDF parametrization. This result is the most precise measurement of $\sigma(p \bar{p} \rightarrow Z + X) \cdot {\rm
Br}(Z \rightarrow \tau^+ \tau^-)$ to date, in good agreement with previous measurements of the $Z$ boson cross section times branching ratio to leptons at $\sqrt{s} = 1.96$ TeV [@p14ZtautauPRD; @CDFZtautau; @measured_crosssections].
acknowledgement\_paragraph\_r2.tex
[99]{} list\_of\_visitor\_addresses\_r2.tex D0 Collaboration, V.M. Abazov [*et al.*]{}, Phys. Rev. Lett. [**97**]{}, 121802 (2006); D0 Collaboration, V.M. Abazov [*et al.*]{}, arXiv:0805.2941 (2008), accepted by Phys. Rev. Lett. D0 Collaboration, V.M. Abazov [*et al.*]{}, Phys. Rev. D [**71**]{}, 072004 (2005); Erratum Phys. Rev. D [**77**]{}, 039901(E) (2008). CDF Collaboration, A. Abulencia [*et al.*]{}, Phys. Rev. D [**75**]{}, 092004 (2007). T. Andeen [*et. al.*]{}, FERMILAB-TM-2365 (2007). D0 Collaboration, V.M. Abazov [*et al.*]{}, Nucl. Instrum. Methods Phys. Res. A [**565**]{}, 463 (2006). D0 Collaboration, S. Abachi [*et al.*]{}, Nucl. Instrum. Methods Phys. Res. A [**338**]{}, 185 (1994). D0 Collaboration, V.M. Abazov [*et al.*]{}, Nucl. Instrum. Methods Phys. Res. A [**552**]{}, 372 (2005). T. Sjöstrand [*et al.*]{}, Comput. Phys. Commun. [$\bm {135}$]{}, 238 (2001). The version used was [6.323]{}. R. Brun and F. Carminati, CERN Program Library Long Writeup W5013, 1993 (unpublished). D0 Collaboration, V.M. Abazov [*et al.*]{}, Phys. Rev. Lett. [**100**]{}, 102002 (2008). R. Hamberg, W.L. van Neerven, and T. Matsuura, Nucl. Phys. [**B359**]{}, 343 (1991); Erratum ibid [ **B644**]{}, 403 (2002). An additional description of uncertainties can be found in arXiv:hep-ph/0308087 (2003). A.D. Martin, R.G. Roberts, W.J. Stirling, and R.S. Thorne, Phys. Lett. B [**604**]{}, 61 (2004). G. Blazey [*et al.*]{}, arXiv:hep-ex/0005012v2 (2000). C.F. Galea, Proc. “Physics at LHC 2006, Cracow, Poland,” Acta Physica Polon. [**B38**]{}, 769 (2007). H.C. Fesefeldt, University of Aachen Technical Report PITHA 85-02 (1985). C. Zeitnitz and A. Gabriel, “The GEANT-Calor Interface User’s Guide” ORNL, Oak Ridge (1996). C.F. Galea, PhD thesis, FERMILAB-THESIS-2008-22. J. Pumplin [*et al.*]{}, JHEP [**0207**]{}, 012 (2002);\
D. Stump [*et al.*]{}, JHEP [**0310**]{}, 046 (2003). K. Melnikov and F. Petriello, Phys. Rev.D [**74**]{}, 114017 (2006); C. Anastasiou, L.J. Dixon, K. Melnikov, and F. Petriello, Phys. Rev. D [**69**]{}, 094008 (2004). S. Frixione and B.R. Webber, JHEP [**06**]{}, 029 (2002). CDF Collaboration, D. Acosta [*et al.*]{}, Phys. Rev. Lett. [**94**]{}, 091803 (2005).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'I review three socio-economic models of economic opinions, urban segregation, and language change and show that the well known two-dimensional Ising model gives about the same results in each case.'
author:
- 'D. Stauffer'
title: 'Social applications of two-dimensional Ising models'
---
Introduction
============
Computer simulations of Ising models have been done for half a century, have been taught in courses on theoretical and computational physics, and are assumed to be known to the reader.[@landau] Here we discuss how simulations simulations can be applied to some problems of interest in the social sciences: urban segregation, language change, and economic opinions.
In the Ising model each site carries a spin which is up or down; neighboring spins “like” to be parallel, and the external field $h$ “wants” to orient the spins parallel to the field. At temperature $T$ the model is simulated with Boltzmann probabilities proportional to $\exp(-E/k_BT)$ by the Metropolis, Glauber, or the heat bath algorithms, or for constant magnetization $M$ using Kawasaki spin exchange.[@landau] The magnetization $M$ is the difference between the number of up and down spins and $E$ is the energy of the system. At low temperatures there is a spontaneous magnetization in zero field: Most of the spins are parallel to each other, either mostly up or mostly down. (For Kawasaki dynamics for which $M$ is fixed, the spins form two large domains.) At high temperatures the spins order into small clusters, without any long-range order. At very high temperatures in a field, the interaction between the spins becomes negligible. The spontaneous magnetization vanishes for $T \rightarrow T_c^-$, where $T_c \approx 2.27$ in units of the nearest-neighbor interaction energy. For $T$ slightly below $T_c$ in small lattices, $M$ switches irregularly between its positive and negative equilibrium values; for lower temperatures and a magnetization antiparallel to an external field, nucleation is possible and a small minority cluster parallel to the field can grow with a low probability to some critical size, and then grow rapidly to encompass the whole lattice. (In the literature this switching is called tunneling, but like nucleation it is not a quantum-mechanical effect.)
These basic algorithms and results can be taught in physics courses. We will see from the three following examples how these methods and results can be applied in social science.
Business confidence
===================
In Germany important business managers are regularly asked how they judge the future economy, and their opinions are summarized and published in the Ifo index in the form of positive, negative, and neutral opinions, as well as an overall average.[@ifo] Hohnisch[@hohnisch] noted that there is a tendency for either a strongly pessimistic or strongly optimistic opinion to form, similar to the Ising simulations shown in Fig. 1. Thus psychological interactions (“herding”) between the managers exist besides hard economic facts.
Apparently, the polled experts influence each other, like ferromagnetic spins. The high-tech bubble of the U.S. stock markets, which ended in spring 2000, might have been another example of psychological impact on economics: First numerous people thought that everything having to do with computers would make big profits, and then doubts spread. In 1971 Weidlich published a more complicated method of treating such binary opinions.[@weidlich]
Reference was followed by a study of the influence of temporally fluctuating external fields; too much such noise destroys the spontaneous magnetization.[@isshock] If there is good (bad) news, people may evolve to form an optimistic majority, but if good news too quickly changes into bad news and back, then people just ignore it and adopt random attitudes. We might think of curriculum reform in societies were the government sets the rules and changes them before the first students under the old rules have finished their studies.[@baumert]
Schelling’s segregation
=======================
In 1971 the 2005 Nobel economist Thomas Schelling published an Ising-like model[@schelling] to explain racial segregation in U.S. cities due to the personal preferences of the residents, without any outside pressures (for example, different housing costs and discrimination). As we will discuss, his model roughly corresponds to an Ising model at $T = 0$. It took 35 years before it was clearly realized that in contrast to reality (like the old Harlem in New York City) no large domains are formed in Schelling’s model,[@kirman] even though Schelling’s paper had become a classic in urban dynamics.[@fossett] Schelling’s model was corrected, for example, by Jones,[@jones] who introduced random dynamics: A small fraction of people move away and are replaced by people who are happy in their new residence. As a result large domains are formed as desired. This paper was widely ignored. The simple Ising model at low $T$ is a much earlier example of the self-organization of such domains.
Physicists were not aware of Ref. for nearly three decades,[@lls] but then simulated Ising and Potts models, with two and more groups for increasing values of $T$, as suggested by Weidlich. The temperature can be interpreted as a measure of the tolerance toward people of different groups. If $T$ increases fast enough, large domains can be avoided,[@ortmanns] with both two and more than two groups. (With more than two groups and conserved group membership as in Kawasaki kinetics, there is one domain for each group in low temperature equilibrium.)
At fixed $T$, the various complications by which the Schelling model differs from the simple Ising model are hardly relevant: usually $T = 0$ gives small clusters and $0 < T < T_c$ gives large domains.[@solomon] This lack of importance is also true for various versions that are intermediate between the Ising and the Schelling model.
More interesting than these intermediate cases is the feedback where each site has its own value of $T$. This value increases (decreases) if all neighbor spins have the same (opposite) orientation as the spin of interest; there also is a general tendency of $T$ to decrease, that is, for a person to forget his/her own tolerance. (Some drastic event shocks people into more tolerant behavior, and then they slowly forget.[@kantz].) In this way a stationary average ${<}T{>}$ self-organizes, with or without spontaneous magnetization depending on the forgetting rate. The usual way of assuming from the outset a social temperature $T$ is thus avoided by this self-organization, depending on the forgetting rate. Figure 2 shows how a change from a forgetting rate of 0.2 percent to 0.3 percent per iteration lowers only slightly the mean temperature, but creates a large spontaneous magnetization (which is analogous to urban segregation and large ghettos). This model suggests that life-long reinforcing of tolerance is needed to avoid ghetto formation.
Nettle’s language change
========================
Nettle[@nettle] simulated how one language (or language feature) which is spoken by everybody can be replaced by another language (feature) without any outside force or bias. His Fig. 1 is reminiscent of our Fig. 1: irregular switching between positive and negative spontaneous magnetizations. He found that the rate at which the majority switches language decays if the population becomes larger; in the Ising model the answer is already known: the rate decays exponentially with increasing linear dimension of the lattice.[@trapp] In the simulation of Ref. not everybody was equal: some speakers were more equal than others, as in social impact theory.[@latane] This complication made the model more realistic, but is not needed to get language change.
If we consider a very large lattice, the switching of Fig. 1 become exponentially rare. If we introduce a bias in favor of the new language (feature), the bias corresponds to an external field, and the previously mentioned nucleation process occurs, as has been well studied for Ising models and for languages.[@nettle; @wang]
Later work[@langchange] replaced a single Ising spin by 8 language features each of which can have 5 values as in the Potts model. The size dependence of the language change rate depends weakly on population size if people learn language features only from their neighbors, and depends strongly on the population size if people can learn from all other members of the population. More details are given in Refs. and . Because different simulations give different results for the size dependence of language change, much more work is needed to clarify that question of whether languages spoken by many people change more slowly than than those spoken by less people. In comparison, the Ising model has been simulated for nearly half a century. And also empirical work on change in real in contrast to simulated languages give conflicting results.[@nettle; @langchange]
Discussion
==========
The three examples have shown how the standard Ising model can be applied to socioeconomic questions. In two cases the original authors[@schelling; @nettle] were not aware of the Ising model and made their models more complicated than necessary to answer their questions. Of course, these complications may make the model more realistic and were simulated even in the first example for which the Ising model was known to the authors.[@hohnisch] Had the social scientists learned about Ising models, or had this author read more carefully the book[@lls] in which Schelling was cited by several physicists, progress could have been faster.
How can we use this material for teaching in physics classes? While teaching electricity and magnetism, I introduce ferromagnetism and the Ising model in connection with Maxwell’s equations in matter, where the magnetization gives the difference between the magnetic $B$ and $H$ fields and where the concept of a spontaneous magnetization can be introduced. Then the above examples can be presented. I used Schelling’s model when ethnic segregation was much in the news in Germany in the author’s home town.
Further details of the simulations
==================================
[*Hohnisch*]{}. The simulations in Ref. are more complicated than the Ising model and allow also for a neutral opinion, as in the Blume-Capel and Potts models.[@BC; @wu] In the Blume-Capel model there are three choices $+1$, 0, and $-1$ for each spin $S_i$, and a term proportional to $S_i^2$ is added to the usual interaction energy $\sum_{i,k}S_iS_k$. This additional term controls the fraction of neutral (zero) spins. In the Potts model, each spin has an integer value between 1 and $q$, and the interaction energy of two spins has a low value if the two spins are the same, and a high value if they are different. The special case of $q=2$ corresponds to the usual Ising model. For a fluctuating external field[@hohnisch] $B = \pm b$ applied to a fraction $p$ of the spins, there is a spontaneous magnetization if $b$ is less than some critical value $b_c$ but not above it. The value of $b_c$ decreases if $T \rightarrow T_c^-$ and/or if the influenced fraction $p$ increases. These fluctuating fields can be interpreted as external news that is good or bad.
[*Schelling*]{}. Schelling[@schelling] defines the neighborhood as the eight nearest and next-nearest neighbors on a square lattice. Each site is either empty, or occupied with someone from one of the two groups of people. People are defined as happy if at least half their neighbors are from their own group, and otherwise are unhappy. Unhappy people move to the closest empty residence where they are happy. Many variants were also studied by Schelling. Schelling’s and Nettle’s model may be regarded as (stochastic) cellular automata, but with sequential instead of the usual simultaneous updating.
[*Nettle*]{}. Nettle [@nettle] assumed that each site follows the majority of the neighbors, except that with some noise probability it does the opposite. Even for the case of everybody influencing everybody, and everybody being equally important, switching is possible, see Fig. 3(a). This simplification corresponds to a majority rule or voter model[@liggett] with noise.
In Fig. 3(a) we assumed that the spin $S_i = \pm 1$ to change as $S_i = {\rm sign}(\sum_k S_k)$, where $k$ runs over all sites and the sign function is replaced by a random choice if the spin sum gives zero. Then with some probability, taken here to be 39 percent, the spin is reversed (independently on whether it was reversed before). Similar results were obtained in Fig.3(b) on a $101 \times 101$ lattice with only 7.7 percent noise and influence only from the four nearest neighbors.
I thank S. Wichmann, M. Hohnisch, E. Holman, K. Müller, S. Pittnauer, C.Schulze, and S. Solomon for research collaborations in this field.
[99]{}
D. P. Landau and K. Binder, [*A Guide to Monte Carlo Simulations in Statistical Physics*]{} (Cambridge University Press, Cambridge, 2005), 2nd ed.
More information on the Ifo index (Information & Forschung at Munich University) can be found at [<www.ifo.de>](<www.ifo.de>).
M. Hohnisch, S. Pittnauer, S. Solomon, and D. Stauffer, “Socioeconomic interaction and swings in business confidence indicators,” Physica A [**345**]{}, 646–656 (2005).
D. Nettle, “Is the rate of linguistic change constant?,” Lingua [**18**]{}, 119–136 (1999).
C. Schulze, D. Stauffer, and S. Wichmann, “Birth, survival and death of languages by Monte Carlo simulation,” Comm. Comput. Phys., submitted.
W. Weidlich, [*Sociodynamics: A Systematic Approach to Mathematical Modelling in the Social Sciences*]{} (Gordon and Breach, London, 2000).
M. Hohnisch, D. Stauffer, and S. Pittnauer, “The impact of external events on the emergence of social herding of economic sentiment,” e-print arXiv:physics/0606237.
J. Baumert et al., “Teacher education in Northrhine-Westphalia” (in German), Ministerium für Innovation, Wissenschaft, Forschung und Technologie, Düsseldorf, April 2007. T. C. Schelling, “Dynamic models of segregation,” J. Math. Sociol. [**1**]{}, 143–186 (1971).
D. Vinkovic and A. Kirman, “A physical analogue of the Schelling model,” Proc. Natl. Acad. Sci. USA [**103**]{}, 19261–19265 (2006).
M. Fossett, “Ethnic preferences, social distance dynamics, and residential segregation: Theoretical explorations using simulation analysis,” J. Math. Sociol. [**30**]{}, 185–274 (2006).
F. L. Jones, “Simulation models of group segregation,” Aust. NZ. J. Sociol. [**21**]{}, 431–444 (1985).
H. Levy, M. Levy, and S. Solomon, [*Microscopic Simulation of Financial Markets*]{} (Academic Press, New York, 2000).
H. Meyer-Ortmanns, “Immigration, integration and ghetto formation,” Int. J. Mod. Phys. C [**14**]{}, 311–320 (2003); C. Schulze, “Potts-like model for ghetto formation in multi-cultural societies,” Int. J. Mod. Phys. C [**16**]{}, 351–356 (2003).
D.Stauffer and S. Solomon, “Ising, Schelling and self-organising segregation,” Eur. J. Phys. B, in press; K. Müller, C. Schulze, and D. Stauffer, “Inhomogeneous and self-organised temperature in Schelling-Ising model,” Max Born Symposium, Sept. 2007, arXiv:0706.2592.
E. G. Altmann, S. Hallerberg, and H. Kantz, “Reactions to extreme events: Moving threshold model,” Physica A [**364**]{}, 435–444 (2006).
B. Latané, “The psychology of social impact,” Am. Psychologist [**36**]{}, 343–365 (1981).
J. Ke, T. Gong, and W.S-Y. Wang, “Language change and social networks,” Fifth EVOLANG, Leipzig, March 31, 2004.
S. Wichmann, D. Stauffer, C. Schulze, and E. W. Holman, “Do language change rates depend on population size?,” e-print arXiv:0706.1842.
M. Blume, “Theory of first-order magnetic phase change in UO$_2$,” Phys. Rev. [**141**]{}, 517–525 (1966); H. W. Capel, “On the possibility of first-order phase transitions in Ising systems of triplet ions with zero-field splitting,” Physica [**32**]{}, 96–106 (1966).
F. Y. Wu, “The Potts model,” Rev. Mod. Phys. [**54**]{}, 235–268 (1982).
T. M. Liggett, [*Interacting Particle Systems*]{} (Springer, New York, 1985).
H. Meyer-Ortmanns and T. Trappenberg, “Surface tension from finite-volume vacuum tunneling in the 3D Ising model,” J. Stat. Phys. [**58**]{}, 185–198 (1990).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The emergence of local phases in a trapped two-component Fermi gas in an optical lattice is studied using quantum Monte Carlo simulations. We treat temperatures that are comparable or lower than those presently achievable in experiments and large enough systems that both magnetic and paired phases can be detected by inspection of the behavior of suitable short-range correlations. We use the latter to suggest the interaction strength and temperature range at which experimental observation of incipient magnetism and $d$-wave pairing are more likely and evaluate the relation between entropy and temperature in two-dimensional confined fermionic systems.'
author:
- Simone Chiesa
- 'Christopher N. Varney'
- Marcos Rigol
- 'Richard T. Scalettar'
title: 'Magnetism and pairing of two-dimensional trapped fermions'
---
Strong electron correlations in solids are believed to be at the origin of a remarkable host of phenomena that range from anomalous insulating states and magnetism to high temperature superconductivity [@bednorz86]. The Hubbard Hamiltonian [@hubbard63], a model which provides a simplified picture of the band structure and electron interactions, is thought to contain the necessary ingredients to describe such spectacular diversity. In the strongly interacting limit the half-filled model describes a Mott insulator and, on a two-dimensional (2D) square lattice, quantum Monte Carlo simulations [@hirsch89; @varney09] have convincingly established the existence of antiferromagnetism at $T=0$. This conjunction of insulating behavior and long-range antiferromagnetic order accurately describes the low temperature physics of the undoped parent compounds of the cuprate superconductors.
Despite intensive analytical and computational work, what happens as one dopes the antiferromagnet has remained controversial and many important questions remain unanswered [@scalapino06; @anderson07; @lee06]. A promising new route to study the model has recently emerged in the form of experiments with fermionic gases in optical lattices. The appeal of these experiments lies in the high degree of control over the different parameters of the system and the ensuing possibility of a close realization of a “pure” Hubbard Hamiltonian [@bloch08], free of the complexity characterizing solid state systems. Ground breaking achievements include loading an ideal quantum degenerate Fermi gas in a three-dimensional (3D) lattice [@kohl05] and the realization of the Mott metal-insulator transition in 3D [@jordens08; @schneider08].
Optical lattice experiments require the presence of a confining potential, usually harmonic, and are accordingly modeled by the Hamiltonian $$H\! = \!-t \!\sum_{\langle {i j} \rangle, \sigma}\!\!\left( c^\dag_{j
\sigma} c_{i \sigma}^{\phantom\dag} +\mathrm{H.c.}\right)\!
+ \sum_{i}\!\left( V_i n_{i} + U n_{ i \uparrow} n_{i \downarrow}\right),
\label{hubham}$$ with $V_i\equiv Vr_i^2$. The curvature $V$ causes the density to vary across the lattice, resulting in a situation at odds with the homogeneous system that one would like to emulate. This is a potential problem because numerical studies on related models [@white98] have indicated the existence of many competing phases with vastly different physical properties. Understanding the extent to which the delicate balance between these phases is affected by the presence of the trap is, therefore, of importance for assessing whether fermionic gases can be taken as accurate simulators of homogeneous models. Thanks to improvements in quantum Monte Carlo (QMC) codes [@vollhardt93; @georges96; @hettler98; @rubtsov05; @varney09], this and other issues can now be quantitatively investigated at temperatures that are comparable or below those of the latest optical lattice experiments [@jordens08; @schneider08].
Here we report results from state of the art determinant QMC (DQMC) simulations [@blankenbecler81] of the trapped 2D Fermi-Hubbard Hamiltonian. We found clear signatures of magnetic and pairing correlations at temperatures of the order of the magnetic scale $J=4t^2/U$, a perhaps surprising observation considering that the corresponding temperature range in strongly correlated solid state systems is well above $T_c$. We note, however, that the properties that we compute are local in character, and that recent advances in optical lattice experiments [@Gemelke09; @Bakr09; @Sherson10] have made possible the imaging of individual sites and short range correlations around them. Hence, our results suggest that the purity of optical lattices and the novel probes used in these experiments could allow the observation of local pairing signatures at higher temperatures than possible in a condensed matter environment.
![\[Fig1\] (a) Local density $n(i)$, (b) density fluctuations $\delta(i)$, (c) local staggered magnetization $M(i)$, and (d) $d$-wave pairing $\chi(i)$ are shown for $U =6t$, $T =
0.31t$, $\mu_0 = 3.0t$, and $V = 0.04t$. A Mott insulating domain is emerging in the density profile of panel (a), in the form of a half-filled ring 6-10 lattice spacings from the trap center. The density fluctuations are minimized in this region. The staggered magnetization and the $d$-wave pairing, however, show a pronounced maximum in the Mott domain.](fig1.eps){width="45.00000%"}
We analyze the properties of the trapped system in terms of the spatial dependence of the local density $n(i) = \langle n_i\rangle$, the density fluctuations $\delta(i) = \langle n_{i}^2 \rangle - \langle
n_{i}^{\phantom 2} \rangle ^2$, and the spin correlations $C_{xy}(i)=4\langle S^z_{i+(x,y)} S^z_i\rangle$. Pairing is analogously discussed by defining a local $d$-wave pair creation operator ${\Delta^d_i}^\dagger$ $$\begin{aligned}
\begin{split}
\label{fulld}
\Delta^\dagger_{ij} &= \frac{1}{\sqrt{2}}
\left(c^{\dagger}_{i\uparrow} c^{\dagger}_{j\downarrow} +
c^{\dagger}_{j\uparrow} c^{\dagger}_{i\downarrow}\right)\\
{\Delta^d_i}^\dagger &= \frac{1}{2} (\Delta^\dagger_{i,i+\hat{x}}
+ \Delta^\dagger_{i,i-\hat{x}} - \Delta^\dagger_{i,i+\hat{y}} -
\Delta^\dagger_{i,i-\hat{y}})
\end{split}\end{aligned}$$ and a corresponding correlation function $P_{xy}(i)=\langle
\Delta^{d}_{i+(x,y)} \Delta^{d\dagger}_{i}\rangle$. To isolate the effect of the pairing vertex [@white89], we focus on the connected correlation function in which suitable products of single particle propagators are subtracted from the expression in Eq. .
In Fig. \[Fig1\], we examine the spatial dependence of different properties at $U = 6\,t$, $V=0.04\,t$ and $T =0.31\, t$ for a system of 560 fermions. In this parameter regime, the average sign in DQMC is $0.3$ and decreases exponentially as $T$ is lowered. $T=0.31\, t$ is therefore very close to the bound that the sign problem [@loh90; @troyer05] imposes on the lowest achievable temperature. Despite this restriction, the emergence of a Mott insulator is clearly visible and depicted in Fig. \[Fig1\](a), which shows a density plateau region of commensurate filling, $n(i)\simeq
1$, and in Fig. \[Fig1\](b) through the formation of a pronounced minimum in the density fluctuations. This domain is also distinguished by enhanced antiferromagnetism \[Fig. \[Fig1\](c)\] and $d$-wave pairing \[Fig. \[Fig1\](d)\]. These last two properties are represented by the local order parameters $$M(i)=\sum_{x,y} (-1)^{x_i+y_i+x+y}C_{x,y}(i),\quad \chi(i)=\sum_{x,y}
P_{x,y}(i) ,$$ which are expected to diverge if long-range order develops around site $i$.
Given that the temperature $T= 0.31\,t$ is of the order of the antiferromagnetic exchange $J$, the sharp signal in $M(i)$ is clearly due to the formation (and partial ordering) of local moments in the Mott domain. The appearance of the peak in $\chi(i)$ in the same region is however surprising - one would expect the peak to occur away from the insulating phase. This shows that $M(i)$ or $\chi(i)$ can be accurate indicators of the appearance of local phases only at temperatures sufficiently low that the order parameter is dominated by long-range contributions. As Fig. \[Fig1\](d) demonstrates, this is not the case in our simulations nor in current experiments. However we shall argue that the temperature and interaction dependence of spin and pairing correlations, when examined at distances which are sensitive to the appropriate energy scales, can provide compelling evidence of incipient order [@supmat].
![\[Fig2\] Temperature dependence of the (a) density, (b) density fluctuations, (c) nearest-neighbor spin correlations $C_{10}(i)$, and (d) next-nearest-neighbor $d$-wave pairing $P_{11}(i)$ for $U=6t$. Lines are results within the local density approximation. The insets of (a) and (b) show the temperature dependence of the entropy and the double occupancy normalized to the number of particles, respectively, while the insets of (c) and (d) are the averages over the lattice of the local staggered magnetization and $d$-wave pairing.](fig2.eps){width="45.00000%"}
Using this argument, we proceed to determine the temperature and entropy scale that experimentalists need to achieve in trapped 2D lattices in order to observe the onset of antiferromagnetic and pairing correlations. Note that a related analysis has been recently carried out for the entropy scale of the [*half-filled*]{} homogeneous model [@paiva10].
As the temperature is lowered from $T=1.14\,t$ to $T=0.31\,t$ ($U=6\,t$), the density distribution \[Fig. \[Fig2\](a)\] is slightly compressed, the entropy per particle $S$ \[shown in Fig. \[Fig2\](a) inset\] decreases monotonically, and there is a marked increase in the total double occupancy in the system $D$ \[see inset in Fig. \[Fig2\](b)\], which is generated by the increase of the double occupancy at the trap center. We determine the entropy, $S =
\beta(\langle E \rangle-F)$, by first computing the Helmholtz free energy using constant-temperature coupling-constant integration over an identical set of traps containing approximately $560$ fermions and increasing $U$ [@supmat]. The Mott insulating region $n(i)\simeq 1$ is characterized by a deepening of the minimum in the density fluctuations $\delta(i)$ \[Fig. \[Fig2\](b)\] and by a three-fold increase in the nearest-neighbor (n.n.) spin-spin correlation $C_{10}(i)$ \[Fig. \[Fig2\](c)\]. As the latter may be the easiest way to observe the onset of antiferromagnetic correlations in experiments, Fig. \[Fig2\](c) makes clear that, although a weak signal may be visible at the highest $T$, a clear maximum in the Mott domain will only be apparent when $T<t$.
When analyzing pairing correlations it is important to realize that $P_{00}(i)$ and $P_{10}(i)$ contain large contributions from $C_{10}(i)$ and, consequently, carry little information on the actual pairing amplitude in the vicinity of a given site. This is analogous to the more familiar statement that local moment formation cannot be taken as indication of short range magnetic order. $P_{11}(i)$ \[Fig. \[Fig2\](d)\] is therefore the shortest-distance correlation that can be used to gain insights on the development of off-diagonal order; it is characterized by two maxima around the Mott region, sharply peaked at densities ($n=0.8$ and $n=1.2$) that lie in the center of the superconducting dome of the cuprate phase diagrams [@krockenberger08]. The analogous correlation for pairs in an extended $s$-wave state (see [@supmat]) is also significantly enhanced in a broad ring around $n=0.6$. Although the local character of $P_{11}$ and the high temperature do not allow a characterization of the dominant low-temperature pairing symmetry, our results suggest that $d$-wave and extended $s$-wave pairing may be detected as leading pairing symmetries in spatially distinct regions, in agreement with existing ground state calculations on related homogeneous models [@rigol09; @dagotto94].
![ Same quantities of Fig. \[Fig2\] but as a function of interaction strength $U/t$ at constant $T=0.50t$. The inset of panel (a) shows the entropy for temperatures $T / t = 0.67, 0.57,
0.50, 0.44, 0.40, 0.36, 0.33$, and $0.31$ (top to bottom).[]{data-label="Fig3"}](fig3.eps){width="45.00000%"}
Another important piece of information that will help experimentalists observe antiferromagnetism and pairing is the optimal value of the ratio $U/t$ at which those correlations are expected to be maximal. In order to address this issue, we first investigate the interaction dependence of the same physical properties represented in Fig. \[Fig2\] but at constant $T = 0.5\,t$ (Fig. \[Fig3\]), and then show that our conclusions are unaltered if it is $S$ to be held constant.
At intermediate coupling ($U=4\,t$), the smallness of the gap, the curvature of the trapping potential, and thermal fluctuations make it impossible for a flattening in the density profile to appear. Note, however, that signals of incipient insulating behavior are clear in the structure of density fluctuations \[Fig. \[Fig3\](b)\]. At large interaction strength ($U=10\,t,12\,t$), the existence of a Mott insulating domain can be directly inferred by inspection of the density profile: the large central plateau comprising about 300 sites is indicative of a phase with vanishing compressibility. Nearest-neighbor spin correlations acquire their largest value at $U=8\,t$. While the double occupancy is monotonically suppressed as $U$ increases, the staggered magnetization shows a similar behavior to $C_{10}(i)$, reaching a maximum when $U$ is of the order of the bandwidth. It is therefore clear that the optimal interaction strength must be near the bandwidth and that future experiments should focus their attention on this regime. Similarly $P_{11}(i)$ exhibits a sharp maximum at $n(i) = 0.8$ for $U \simeq 8\, t$ but contrary to what is observed for $C_{10}(i)$, there is no weakening when going from $U=8t$ to $U =12t$ and no dependence of the “optimal” density for pairing on $U$. The latter is reminiscent of the properties of cuprate superconductors where optimal doping is constant across the entire family of these materials.
One might wonder what would happen in the experimentally relevant case of a (quasi-)adiabatic evolution, i.e., a system at constant entropy, as a function of $U$. In the inset of Fig. 3(a), we show the evolution of the entropy per particle as a function of the interaction strength for various fixed temperatures. As $U$ is increased at constant $S$, we find that the system cools down and goes, for instance, from $T=0.5t$ at $U=2t$ to $T=0.31t$ at $U=10t$, while $S$ is held fixed at 0.5. Since this cooling is mild beyond $U=8t$, isothermal and adiabatic evolution in this parameter regime are essentially equivalent. For $U<8t$, an adiabatic increase of $U$ is accompanied by progressively lower temperatures and this, in turn, implies that $U=8t$ remains the optimal interaction strength to observe strong-correlation phenomena.
We finally touch on the extent to which properties in a trap are a faithful representation of those in a homogeneous system. To this aim we use DQMC simulations on homogeneous $8\times8$ clusters to estimate local properties as a function of chemical potential and plot the corresponding local density approximation (LDA) results in Figs. \[Fig2\] and \[Fig3\] as continuous lines. The agreement with the ab-initio results is excellent with the sole exception of the half-filled region at $T=0.31\,t$ \[Fig. \[Fig2\](c)\]. This discrepancy is a characteristic failure of the LDA when applied to the interface between coexisting phases [@rigol03Fa].
![\[Fig4\] Spin correlations $C_{xy}(i)$ \[for fixed site $i$ and variable separation $(x,y)$\] along the path illustrated in inset (a) around the trapped center. Inset (b) is an enlarged view of the longer correlations. $C_{xy}(i)$ exhibits the alternating sign characteristic of antiferromagnetism. The correlations decrease rapidly, but nevertheless extend out to several lattice spacings. The spin correlations are also shown in the LDA approximation for $d = 1$ and $d = 2$.](fig4.eps){width="45.00000%"}
A more dramatic failure of the LDA is expected when examining longer range correlations. Since in the fermionic Hubbard model the development of a certain type of order corresponds with a rather narrow density interval, different phases in a trap appear with a narrow-ring-like topology. If one disregards interfacial effects, these regions can be thought as one-dimensional homogeneous strips where the long range character of correlations is determined by the strip width. An example of such a situation is given by the development of long-range magnetic order in the half-filled annulus found in our simulations at $U=6t$. In particular, Fig. \[Fig4\] shows that 2D correlations, computed on a half-filled homogeneous cluster at the same temperature, represent a good approximation to the correlations in a trap only at short distances, overestimating the correct long-range value. Analogous results for a 1D chain fail at all distances and, in particular, decay too quickly at large separation. This behavior is generic and expected for the superfluid regions as well. When the system size increases, these quasi-1D regions grow both in width and length and their correlations ultimately cross over to a pure 2D character. This dimensionality effect must be taken into account when using finite size systems to infer the critical behavior.
In summary, we have addressed finite temperature properties of inhomogeneous Fermi-Hubbard systems using an ab-initio approach. Our findings of enhanced antiferromagnetic and pairing correlations just below the temperature scale $T \sim t$ thus open important perspectives for current experiments with ultracold fermions in optical lattices. While the lowest temperatures reported here are well above the $d$-wave coherence temperature, the enhanced signal in local properties is a promising sign: these results and the purity of the experimental optical lattice suggest that, in contrast to solid state systems, temperatures of the order of the hopping scale may be enough to observe clear local signature of [*both*]{} magnetic and pairing correlations. Tuning experiments to be in the regime where the onsite interactions are of the order the bandwidth ($U \sim 8 t$) provides the sharpest signal of the many-body effects. Our computation of the entropy $S$ indicates that adiabatic cooling occurs in the 2D Hubbard Hamiltonian with a position dependent (confining) potential as the interaction strength $U$ is increased via, [*e.g.*]{}, a Feshbach resonance. This allows our conclusions concerning the observability of spin and pairing order to be relevant to experiments at constant entropy.
This work was supported under ARO Award W911NF0710576 with funds from the DARPA OLE Program, by the DOE SciDAC program under grant DOE-DE-FC0206ER25793, by the US Office of Naval Research, and by the National Science Foundation under Grant No. OCI-0904597 and through TeraGrid resources provided by NICS under grant number TG-DMR100007. M.R. and R.T.S. thank the Aspen Center for Physics for hospitality. C.N.V. acknowledges NSF dissertation enhancement grant No. 0803230.
[28]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, in **, edited by (, ), p. .
, , , , ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, , , , ****, ().
, ****, ().
J. F. Sherson [*et al.*]{}, Nature (London) [**467**]{}, 68 (2010).
, ****, ().
, , ****, ().
, ****, ().
.
, , , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , , ****, ().
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: '[Quantum kicked rotor was recently realized in experiments with cold atomic gases and standing optical waves. As predicted, it exhibits dynamical localization in the momentum space. Here we consider the weak localization regime concentrating on the Ehrenfest time scale. The later accounts for the spread-time of a minimal wavepacket and is proportional to the logarithm of the Planck constant. We show that the onset of the dynamical localization is essentially delayed by four Ehrenfest times and give quantitative predictions suitable for an experimental verification. ]{}'
address: |
$^{1}$ Department of Physics, University of Minnesota, Minneapolis, MN 55455, USA\
$^{2}$ William I. Fine Theoretical Physics Institute, University of Minnesota, Minneapolis, MN 55455, USA\
$^{3}$ L. D. Landau Institute for Theoretical Physics, Moscow, 117940, Russia
author:
- 'C. Tian$^{1}$, A. Kamenev$^{1}$, and A. Larkin$^{1,2,3}$'
title: Weak dynamical localization in periodically kicked cold atomic gases
---
Unprecedented degree of control reached in experiments with ultra-cold atomic gases [@Chu] allows to investigate various fundamental quantum phenomena. A realization of quantum kicked rotor (QKR) is one such possibility that recently attracted a lot of attention [@Raizen95; @Ammann; @Raizen99; @Zhang04]. To this end cold atoms are placed in a spatially periodic potential $V_0\cos (2k_Lx)$ created by two counter-propagated optical beams. The potential is switched on periodically for a short time $\tau_p\ll T$, giving a kick to the atoms; here $T$ is a period of such kicks. The evolution of the atomic momenta distribution may be monitored after a certain number of kicks. If the gas is sufficiently dilute [@footnote1], one may model it with the single–particle Hamiltonian, that upon the proper rescaling takes the form [@footnote2] of the QKR: $${\hat H}={1\over 2}\,{\hat l}^{\,2}+ K \cos \theta \sum_n \delta
\left(t-n\right)\, .
\label{Hamiltonian}$$ Here $\theta \equiv 2k_Lx$ and time is measured in units of the kick period, $T$. The momentum operator is defined as $\hat
l=i{\mathchar'26\mkern-9muk}\, \partial_\theta$, where the dimensionless Planck constant is given by ${\mathchar'26\mkern-9muk}=8\hbar T k_L^2/(2m)$. Finally, the classical stochastic parameter is $K = {\mathchar'26\mkern-9muk}V_0\tau_p/\hbar \,$.
The classical kicked rotor is known to have the rich and complicated behavior [@Chirikov79]. In particular, for sufficiently large $K~(\gtrsim 5)$, it exhibits the chaotic diffusion in the space of angular momentum [@Chirikov79]. The latter is associated with the diffusive expansion of an initially sharp momenta distribution: $\delta\langle l^2(t)\rangle\equiv \langle
\left(l(t)-l(0)\right)^2\rangle=2D_{cl}t$ (dashed line on Fig. \[fig1\]). For sufficiently large $K$, the classical diffusion constant may be approximated by $K^2/4$ [@Chirikov79]. The higher order correction is an oscillatory function of the stochastic parameter, i.e., $D_{cl}(K)\approx {1\over 4}K^2(1-3J_2(K) +2J^2_2(K))$ [@Rec81; @KFA00]. It was realized a while ago [@CCFI79; @CIS81] that quantum interference destroys the diffusion in the long time limit and leads to localization: $\delta\langle l^2(t)\rangle\to \sim \xi^2$ at $t\gtrsim t_L\equiv D_{cl}/{\mathchar'26\mkern-9muk}^2= \xi^2/D_{cl}$, where the localization length is given by $\xi = D_{cl}/{\mathchar'26\mkern-9muk}$. For a large localization length $\xi\gg {\mathchar'26\mkern-9muk}$, there is a long crossover regime, $1<t<t_L$, between the classical diffusion and quantum localization.
It was suggested in Ref. [@FGP82], that the QKR may be mapped onto the one–dimensional Anderson localization with the long range disorder. The universal long–time behavior of the latter is described by the non–linear sigma–model [@AZ96], resulting in the standard weak–localization correction [@Altland93; @Basko]: $\delta\langle l^2(t)\rangle = 2D_{cl}t\,(1-0.75
\sqrt{t/t_L}\,)$ (dashed–dotted line on Fig. \[fig1\]). Notice, that the correction is linear in ${\mathchar'26\mkern-9muk}$ and non-analytic in time. In an apparent contradiction with this fact, explicit studies [@She87; @DP02] of the first few kicks show only renormalization of the diffusion constant starting from terms $\sim {\mathchar'26\mkern-9muk}^2$. The aim of this paper is to develop a quantitative description of the classical to quantum crossover for the QKR that, in particular, accounts for these conflicting observations.
It is known in various contexts [@LO68] that such crossover involves an additional time scale, $t_E$, called an Ehrenfest (or breaking) time. For a generic quantum mapping, it was first shown by Berman and Zaslavski [@BZ78; @Zaslavsky81], that quantum corrections become comparable to the classical limit at the time $t_E$. This is the time needed for a minimal quantum wavepacket, $\delta\theta_0\delta l_0\simeq {\mathchar'26\mkern-9muk}$, to spread uniformly over the angular direction. Due to the chaotic motion, trajectories diverge as $\delta\theta(t)=\delta\theta_0 e^{\lambda
t}$, where $\lambda$ is the classical Lyapunov exponent. For $K\gg 1$, $\lambda=\ln(K/2)$ [@Chirikov79]. Estimating $\delta l_0\approx K\delta\theta_0$ – a typical momenta dispersion after one kick, one finds for the Ehrenfest time: $$t_E={1\over \lambda}\ln \sqrt{K\over {\mathchar'26\mkern-9muk}}\, .
\label{ehrenfest}$$ It is widely believed [@CIS81; @BZ78; @Zaslavsky81; @Izrailev90] that this intermediate, $1<t_E < t_L$, time scale is indeed relevant for the quantum evolution of classically chaotic systems. This observation was put on the quantitative basis in Ref. [@AL96] in the context of localization caused by classical scatterers. In this paper we adopt methods of Ref. [@AL96] to the essentially different problem of the QKR.
In the leading order in ${\mathchar'26\mkern-9muk}$ we found for the momentum dispersion: $$\delta\langle l^2(t)\rangle =2D_{cl}t - \frac{8{\mathchar'26\mkern-9muk}\sqrt{D_{cl}}}{3\sqrt \pi}\,\theta (t- 4t_E) \left(t-
4t_E\right)^{3/2} ,
\label{result}$$ where $\theta(t)$ is the step function (long-dashed line on Fig. \[fig1\]). At relatively large time, $t_E \ll t <t_L$ our result approaches the standard weak–localization, mentioned above. However, corrections of the order of ${\mathchar'26\mkern-9muk}$ are absent for $t\leq 4t_E$. The delay is caused by the interference nature of the localization. Indeed, the first correction originates from the interference of two closed–loop counter–propagating trajectories, see Fig. \[fig2\]. It takes time about $t_E$ for classical trajectories passing through (almost) the same value of the momentum to diverge and take counter–propagating roots. As a result, the interference effects are practically absent at smaller times and show up only after $4t_E$.
One may show that the time interval $0\leq t\leq 4t_E$ is protected from higher order weak localization corrections as well. For example, in the second order weak localization correction there are two diagrams [@TL03] proportional to ${\mathchar'26\mkern-9muk}^2\, \theta(t-mt_E) (t-mt_E)^2$ with $m=6$ and $m=8$ correspondingly. This fact agrees with the perturbative studies of the QKR dynamics after a few kicks [@CCFI79; @CIS81; @Raizen95], where no localization effects were seen. (Though the classical diffusion coefficient, $D_{cl}$, is renormalized as an analytic function of ${\mathchar'26\mkern-9muk}^{\,2}$.) It is important to mention, however, that in reality there is no non–analyticity at the point $t=4t_E$ as may seem from Eq. (\[result\]). The small localization corrections, non–analytic in ${\mathchar'26\mkern-9muk}$, [*do*]{} exist for $t \lesssim 4t_E$. They are associated with the fluctuations of the Ehrenfest time. Since in the quantum mechanics the minimal separation between the trajectories is about $\delta\theta_0\sim \sqrt{{\mathchar'26\mkern-9muk}/K}$, it takes a finite time $(\sim t_E)$ for them to diverge. This time may fluctuate depending on initial conditions. The fluctuations are characterized by the time scale $\delta t_E =
\lambda_2t_E/\lambda^2$, where (cf. Eq. (\[ehrenfest\])) $\lambda_2=\langle\lambda^2\rangle - \langle
\lambda\rangle^{2}\approx 0.82\,$ for sufficiently large $K$ and the angular brackets denote averaging over the initial angle. The interference between rare trajectories, diverging faster than the typical ones, leads to quantum corrections at $t\lesssim 4t_E$ of the form: $$\delta\langle l^2(t)\rangle =2D_{cl}t -\frac{\Gamma
\left(\frac{5}{4}\right){\mathchar'26\mkern-9muk}}{3\pi/256 } \sqrt{D_{cl}}\, (\delta
t_E)^{3/2} f\left({4t_E-t\over \sqrt{\delta t_E}}\right)\, ,
\label{smalltime}$$ where $f(0)=1$ and $f(x)=6\sqrt {2\pi}/ \Gamma
\left(\frac{5}{4}\right) x^{-5/2}e^{-x^2/16}$ for $x\gg1$. Localization correction including effect of the Ehrenfest time fluctuations is depicted on Fig. \[fig1\] by a full line. It is rather close to the prediction of Eq. (\[result\]) (long-dashed line), however the singularity at $t=4t_E$ is rounded.
Having outlined our main results, we turn now to some details of the calculations. One starts from introducing the exact one period evolution operator as: $\hat U=\exp \left\{ i(K/{\mathchar'26\mkern-9muk})
\,\cos\hat\theta\right\}\exp \left\{i\hat l ^2/(2{\mathchar'26\mkern-9muk}) \right\}$. All physical quantities may be expressed in terms of the matrix elements of $\hat U^n$, where $n$ stays for the number of kicks (time). We shall be particularly interested in the four–point density–density correlator, defined as: $$\begin{aligned}
&& {\cal D}(l_+,l_-; l_+',l_-';\omega _+,\omega_-)
\label{QDiff}\\
&\equiv& \sum\limits_{n,n'=0}^\infty \langle l_+|e^{i\omega_+ n}
e^{ \frac{i\hat l ^2}{2{\mathchar'26\mkern-9muk}}}{\hat U}^{n}|l_+' \rangle
\overline{\langle l_-'|e^{i\omega_- n'}e^{\frac{i\hat l
^2}{2{\mathchar'26\mkern-9muk}}}{\hat U}^{n'}|l_- \rangle}\, , \nonumber\end{aligned}$$ where $|l_{\pm}\rangle$ denote momentum eigenstates. First we note that averaging over $\left(\omega_++\omega_-\right)/2$ leads to $n=n'$ [@AZ96]. Then performing the standard Wigner transform, one passes to the variables $l=(l_++l_-)/2$ and $\theta$ – the Fourier transform of $l_+-l_-$ (and similarly for the prime variables). Since the most unstable direction of the underlying classical dynamics is along the $\theta$ direction [@Chirikov79], any initially smooth distribution quickly relaxes in this direction. Thus, averaging over $\theta$ and $\theta'$ may be performed. The resulting correlator depends only on the relative momentum $l-l'$ and frequency $\omega \equiv \omega_+ - \omega_-$. Introducing finally angle $\varphi$ as the Fourier image of $l-l'$, one ends up with ${\cal D}={\cal D}(\varphi;\omega)$. Its classical limit, ${\cal D}_0(\varphi;\omega)$, may be found e.g. by using the diagrammatic technique of Ref. , where in the large $K$ limit it corresponds to the family of ladder diagrams. Alternatively, one may show from Eq. (\[QDiff\]) that ${\cal D}_0$ satisfies classical Liouville equation. Upon the proper regularization [@Rec81; @KFA00; @Zirnbauer99], that may be viewed as a coarse graining in the angular direction, one arrives to the classical diffusion propagator: $${\cal D}_0 (\varphi;\omega)= \left(-i\omega +D_{cl}
\varphi^2\right)^{-1}\, .
\label{DF}$$ This classical limit reflects the diffusion in the momentum space: $\delta\langle l^2(t)\rangle=2D_{cl} t$, with the diffusion coefficient, $D_{cl}(K)$, studied extensively in the literature [@Chirikov79; @Rec81; @KFA00].
The first quantum correction to Eq. (\[DF\]) is given by the one–loop weak–localization diagram, Fig. \[fig2\]a. It describes the interference of the two counter–propagating trajectories, passing through (almost) the same point in the momentum space, Fig. \[fig2\]b. In the Wigner representation such correction takes the form: $$\begin{aligned}
&&
\delta {\cal D} (l,\theta;l',\theta')=\int\int \frac{dl_0d\theta_0}{2\pi}\,
\frac{dl_1d\theta_1}{2\pi}\,\Big\{
{\cal C} (l_1,\theta _0;l_0,\theta _1)
\,
\nonumber \\
&\times&
\hat {\cal
X}(l_0,\theta_0;l_1,\theta_1)
\big[ {\cal D}_0(l,\theta;l_0,\theta_0)
\, {\cal
D}_0(l_1,\theta_1;l',\theta')\big] \Big\} \, ,
\label{qc1}\end{aligned}$$ where $\omega$ argument is omitted to shorten notations. The operator $\hat {\cal X}$ stays for the Hikami box [@Hikami81], which is given by $$\hat {\cal X} = -\exp\left\{-\frac{K^2(\delta\theta_0)^4}{4
{\mathchar'26\mkern-9muk}^2} +4i\frac{\delta l_0\delta\theta_0}{{\mathchar'26\mkern-9muk}}\right\}\,
D_{cl}\left( \nabla^{\,2}_{l_0}+\nabla^{\,2}_{l_1}\right)\, ,
\label{Hikamibox}$$ where $\delta\theta_0\equiv \theta_0+\theta_1$ and $\delta l_0
\equiv \left(l_0-l_1\right)/2$. It is clear from this expression that the quantum correction, Eq. (\[qc1\]), is non–zero as long as $\delta l_0\delta \theta_0\lesssim {\mathchar'26\mkern-9muk}\,$ and, therefore, it is proportional to ${\mathchar'26\mkern-9muk}$. The semiclassical Cooperon ${\cal C}
(l_0,\theta _0;l_1,\theta _1)$ gives the probability of return to (almost) the same momentum, $l_1\approx l_0$, at (almost) the opposite angle, $\theta_1\approx - \theta_0$. If these conditions were strict, such motion would be forbidden by the time–reversal symmetry. The quantum uncertainty makes it possible. It takes, however, a long time to magnify the initially small angular variation $\delta\theta_0\simeq \sqrt{{\mathchar'26\mkern-9muk}/K}$ (this estimate as well as $\delta l_0\simeq \sqrt{{\mathchar'26\mkern-9muk}K}$ follows directly from Eq. (\[Hikamibox\])) up to $\delta\theta_n\approx 1$, when the usual diffusion takes place.
To take this fact into account we divide the Cooperon trajectory onto two parts: the Ehrenfest region, where $\delta \theta_n \ll
1$ and the diffusive region with $\delta \theta_n \gtrsim 1$. We denote the corresponding propagators as $\cal W$ and ${\cal
C}_0$ and write in the time representation ${\cal C}(t)=\int dt'\,{\cal W}(t')\,{\cal C}_0(t-2t')$ (cf. Fig. \[fig2\]b), where, as we show below, $t'\approx t_E$. Notice that the diffusive part of the trajectory is shortened by $2t'$, leading to ${\cal
C}(\omega)={\cal W}(2\omega)\, {\cal C}_0(\omega)$. The diffusive Cooperon, ${\cal C}_0(l_0-l_1;\omega)$, has the same form as Eq. (\[DF\]) and thus ${\cal C}_0(0;\omega)\sim \int d\varphi (D_{cl}\varphi^2-i\omega)^{-1}$.
To evaluate propagator ${\cal W}(2\omega)$ in the Ehrenfest regime, we define ${\cal W}(z,n)$ as a probability to reach the deviation $\delta\theta_n\equiv e^z$ during $n$ kicks, starting from an initially small variation, $\delta\theta_0 \simeq
\sqrt{{\mathchar'26\mkern-9muk}/K}$. According to the classical equations (the standard map) $\theta_{n}=\theta_{n-1} +l_{n}\,$ and $l_{n}=l_{n-1}+K\sin\theta_{n-1}\,$, the variation evolves as $\delta\theta_{n}=\delta\theta_{n-1}(1+K\cos\theta_{n-1})+2\delta
l_{n-1}$. Since $\delta l_0\simeq K\delta\theta_0 $, in the leading order in $K\gg 1$ the evolution of $\delta\theta$ is given by $
\delta\theta_n \approx \delta\theta_0 \prod_{j=0}^{n-1}K\cos\theta_j$. Taking the logarithm of this expression, one obtains $${\cal W}(z;n) =\! \left\langle \delta\left(z-\ln |\delta
\theta_0|-\sum_{j=0}^{n-1}\ln
\left|K\cos\theta_j\right|\right)\right\rangle ,
\label{W}$$ where the averaging is performed over the initial distribution of $\delta\theta_0$ \[with the typical scale $\delta\theta_0\sim\sqrt {{\mathchar'26\mkern-9muk}/K}$, cf. Eq. (\[Hikamibox\])\] as well as over dynamics of the fast variable, that is $\delta l_n/\delta\theta_n$. For $K\gg 1$, one may treat $\cos\theta_j$ after successive kicks as independent random variables and employ the central limiting theorem to perform the averaging in Eq. (\[W\]). As a result, $${\cal W}(z;n) \approx \exp\left\{- \frac{(z-\ln \sqrt{{\mathchar'26\mkern-9muk}/K} -
n\lambda)^2}{2n\lambda_2}\right\}\, ,
\label{W1}$$ where the Lyapunov exponent $\lambda$ [@Chirikov79] and its dispersion $\lambda_2$ are defined as: $$\begin{aligned}
\lambda &\equiv& \left\langle\, \ln|K\cos\theta| \right\rangle
=\ln (K/2)\, ;
\label{lambda}\\
\lambda _2 &\equiv& \left\langle\, \ln^2|K\cos\theta|
\right\rangle -\lambda^2=\zeta(3)-\ln^2 2\approx 0.82\, ;
\nonumber\end{aligned}$$ the angular brackets imply integration over $\theta$. The Ehrenfest evolution crosses over to the usual diffusion at $\delta
\theta_n\approx 1$, meaning $z\approx 0$. Performing finally the Fourier transform as ${\cal W}(\omega)\equiv \sum_n e^{i\omega
n}{\cal W}(0,n)$ and employing the definition of the Ehrenfest time, Eq. (\[ehrenfest\]), and the fact that $\lambda_2\ll
\lambda$, one finds: $${\cal W}(2\omega)=\exp \left\{2i\omega t_E-\frac{2\omega ^2\lambda
_2 t_E}{\lambda ^2}\right\}\, .
\label{Gamma}$$
Due to the time–reversal invariance there is an exact symmetry between divergence and convergence of the two classical trajectories involved in the weak–localization correction. This symmetry is illustrated on Fig. \[fig2\]c. Therefore it takes an additional time $\sim t_E$ for the two distinct semiclassical diffusons to arrive to the point $l_0\approx l_1$ and $\theta_0\approx - \theta_1$, bringing, thus, another factor ${\cal W}(2\omega)$. In a slightly different language, one may define the Hikami box for a classically chaotic system [@AL96; @footnote3] as $2{\mathchar'26\mkern-9muk}{\cal W}^2(2\omega)
D_{cl}\nabla^2_l$, where one factor ${\cal W}(2\omega)$ comes from the two legs of the Cooperon, while another originates from the two diffusons. Finally, the quantum correction, Eq. (\[qc1\]), reduces to the renormalization of the diffusion coefficient in the classical propagator, Eq. (\[DF\]), as $D(\omega) = D_{cl}
+\delta D(\omega)$ with $$\delta D(\omega) =-{{\mathchar'26\mkern-9muk}D_{cl}\over \pi}\, {\cal W}^2(2\omega)\!
\int \!\! \frac{d \varphi}{-i\omega+D_{cl}\varphi^2}\,\, .
\label{DDclKW}$$ Equations (\[Gamma\]) and (\[DDclKW\]) constitute the main analytical results of this work. They describe quantitatively dynamical weak–localization of the QKR with the account for the Ehrenfest time phenomena. One may finally express the time evolution of the momentum dispersion in terms of the frequency–dependent diffusion coefficient. The exact relation reads as: $$\delta\langle l^2(t) \rangle =\!
\int\limits_{-\infty}^{\infty}\!\! {d\omega\over \pi}\,
\frac{1-e^{-i\omega t}}{\omega^2}\,\, D(\omega)\, .
\label{TimeCorrDC}$$ Neglecting fluctuations of the Ehrenfest time ($\lambda_2\to 0$ in Eq. (\[Gamma\])) and performing frequency and angle integrations in Eqs. (\[TimeCorrDC\]), (\[DDclKW\]), one obtains Eq. (\[result\]) for the momenta dispersion. Notice that in this approximation the evolution is purely classical at $t\leq 4t_E$. To account for the quantum corrections at $t\lesssim
4t_E$ one needs to keep the $\lambda_2$ term in Eq. (\[Gamma\]). The straightforward integration leads to Eq. (\[smalltime\]).
Our results, Eqs. (\[result\]), (\[smalltime\]), are expected to be quantitatively accurate if separations between the relevant time scales: $1<t_E<t_L$ are large enough. This is the case when the two dimensionless constants satisfy inequalities: ${\mathchar'26\mkern-9muk}<1<K$. (In the experiments, we are aware of [@Raizen95; @Ammann], ${\mathchar'26\mkern-9muk}\gtrsim 2$, and thus $t_E\lesssim 1$.) Another restriction has to do with the dephasing time, $\tau_\phi$ [@Cohen91]. The later may originate from non–perfect periodicity of the kicks (noise) [@SMOR00], spontaneous emission [@Ammann], as well as from many–body collisions between the atoms [@footnote1]. Whatever the nature of the dephasing time, one needs to ensure $t_E<\tau_\phi$ to observe the weak localization.
We would like to thank A. Altland, J. Liu and C. Zhang for helpful discussions. Useful comments from S. Fishman are gratefully acknowledged. C. T. and A. L. are supported by NSF Grant No. 0120702; A. K. is A. P. Sloan fellow.
For a review see e.g. S. Chu, Science [**253**]{}, 861 (1991).
F. L. Moore [*et. al.*]{}, Phys. Rev. Lett. [**75**]{}, 4598 (1995); C. F. Bharucha [*et. al.*]{}, Phys. Rev. E [**60**]{},3881 (1999).
H. Ammann [*et. al.*]{}, Phys. Rev. Lett. [**80**]{}, 4111 (1998).
M. G. Raizen, Adv. At., Mol., Opt. Phys. [**41**]{}, 43 (1999).
C. Zhang, J. Liu, M. Raizen, and Q. Niu, Phys. Rev. Lett. [**92**]{}, 054101 (2004).
The two–particle scattering mean free path is $l_\phi\approx 1/(na^{2})$, where $n$ is the atomic concentration and $a$ – the $s$-wave scattering length. The corresponding dimensionless scattering time is $\tau_\phi =
l_\phi/(Tv)$, where $v$ is a typical atomic velocity, that may be estimated as $v\approx {\hbar k_L\over {\mathchar'26\mkern-9muk}m}|l|\approx {\hbar
k_L\over {\mathchar'26\mkern-9muk}m} \sqrt{D_{cl}\tau_\phi}\,$. This leads to the self–consistent estimate of the dimensionless dephasing time: $\tau_\phi \approx (l_\phi k_L/K)^{2/3}$.
The difference with the conventional QKR, is in the boundary condition on the wave-function: $\Psi(\theta+2\pi)=\Psi(\theta)e^{i\theta l_0}$, where $l_0\in
[0,1]$ is a fractional part of an atom’s momentum in units $2\hbar
k_L$. It may be eliminated by the gauge transformation that introduces an Aharonov–Bohm flux into the kinetic energy $(\hat
l-l_0)^2/2$. Such flux does not affect dynamical localization, though averaging over $l_0$ helps to eliminate resonances [@Izrailev90].
B. V. Chirikov, Phys. Rep. [**52**]{}, 263 (1979); A. L. Lichtenberg and M. A. Lieberman, [*Regular and Chaotic Dynamics*]{} (Springer-Verlag, Berlin 1991).
A. B. Rechester and R. B. White, Phys. Rev. Lett. [**44**]{}, 1586 (1980); A. B. Rechester, M. N. Rosenbluth, and R. B. White, Phys. Rev. A [**23**]{}, 2664 (1981).
M. Khodas and S. Fishman, Phys. Rev. Lett. [**84**]{}, 2837 (2000); Erratum, [*ibid.*]{} [**84**]{}, 5918 (2000); M. Khodas, S. Fishman, and O. Agam, Phys. Rev. E [**62**]{}, 4769 (2000).
G. Casati, B. V. Chirikov, J. Ford, and F. M. Izrailev, in [*Stochastic Behavior of Classical and Quantum Hamiltonian Systems*]{}, Lecture Notes in Physics [**93**]{}, edited by G. Casati and J. Ford, 334 (Springer, New York 1979).
B. V. Chirikov, F. M. Izrailev, and D. L. Shepelyansky, Sov. Sci. Rev., Sect. C [**2**]{}, 209 (1981).
S. Fishman, D. R. Grempel, and R. E. Prange, Phys. Rev. Lett. [**49**]{}, 509 (1982).
A. Altland and M. R. Zirnbauer, Phys. Rev. Lett. [**77**]{}, 4536 (1996); [*ibid.*]{} [**80**]{}, 641 (1998); G. Casati, F. M. Israilev, and V. V. Sokolov, [*ibid.*]{} [**80**]{}, 640 (1998).
A. Altland, Phys. Rev. Lett. [**71**]{}, 69 (1993).
D. M. Basko, M. A. Skvortsov, and V. E. Kravtsov, Phys. Rev. Lett. [**90**]{}, 096801 (2003); V. E. Kravtsov, cond-mat/0312316.
D. L. Shepelyansky, Physica D [**28**]{}, 103 (1987).
A. J. Daley and A. S. Parkins, Phys. Rev. E [**66**]{}, 056210 (2002); G. Duffy [*et. al.*]{}, cond-mat/0401346.
A. I. Larkin and Yu. N. Ovchinnikov, Zh. Eksp. Teor. Fiz. [**55**]{}, 2262 (1968) \[Sov. Phys. JETP [**28**]{}, 1200 (1969)\].
G. P. Berman and G. M. Zaslavsky, Physica A [**91**]{}, 450 (1978).
G. M. Zaslavsky, Phys. Rep. [**80**]{}, 157 (1981).
F. M. Izrailev, Phys. Rep. [**196**]{}, 299 (1990).
I. L Aleiner and A. I. Larkin, Phys. Rev. B [**54**]{}, 14423 (1996).
C. Tian and A. I. Larkin, cond-mat/0310429.
M. R. Zirnbauer, in [*Supersymmetry and Trace Formulae, Chaos and Disorder*]{}, edited by I. V. Lerner, J. P. Keating and D. E. Khmelnitskii (Kluwer Academic/Plenum, New York 1999).
S. Hikami, Phys. Rev. B [**24**]{}, 2671 (1981).
In contrast to Ref. [@AL96], the Hikami box for the QKR may be found explicitly without introducing a regularization.
D. Cohen, Phys. Rev. A [**44**]{}, 2292 (1991).
D. A. Steck [*et. al.*]{}, Phys. Rev. E [**62**]{}, 3461 (2000); V. Milner [*et. al.*]{}, Phys. Rev. E [**61**]{}, 7223 (2000).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We have used deep [*GALEX*]{} observations at the infall region of the Coma cluster to measure the faintest UV luminosity functions (LFs) presented for a rich galaxy cluster thus far. The Coma UV LFs are measured to M$_{\mathrm{UV}}$ $=-10.5$ in the [*GALEX*]{} FUV and NUV bands, or 3.5 mag fainter than previous studies, and reach the dwarf early-type galaxy population in Coma for the first time. The Schechter faint-end slopes ($\alpha\approx-1.39$ in both [*GALEX*]{} bands) are shallower than reported in previous Coma UV LF studies owing to a flatter LF at faint magnitudes. A Gaussian-plus-Schechter model provides a slightly better parametrization of the UV LFs resulting in a faint-end slope of $\alpha$ $\approx$ $-1.15$ in both [*GALEX*]{} bands. The two-component model gives faint-end slopes shallower than $\alpha=-1$ (a turnover) for the LFs constructed separately for passive and star forming galaxies. The UV LFs for star forming galaxies show a turnover at M$_{\mathrm{UV}}$ $\approx$ $-14$ owing to a deficit of dwarf star forming galaxies in Coma with stellar masses below M$_{*}=10^{8}~\msun$. A similar turnover is identified in recent UV LFs measured for the Virgo cluster suggesting this may be a common feature of local galaxy clusters, whereas the field UV LFs continue to rise at faint magnitudes. We did not identify an excess of passive galaxies as would be expected if the missing dwarf star forming galaxies were quenched inside the cluster. In fact, the LFs for both dwarf passive and star forming galaxies show the same turnover at faint magnitudes. We discuss the possible origin of the missing dwarf star forming galaxies in Coma and their expected properties based on comparisons to local field galaxies.'
author:
- 'D. M. Hammer$^{1,2}$, A. E. Hornschemeier$^{2}$, S. Salim$^{3}$, R. Smith$^{4}$, L. Jenkins$^{2}$, B. Mobasher$^{5}$, N. Miller$^{6}$, and H. Ferguson$^{7}$'
bibliography:
- 'ms.bib'
title: Deep UV Luminosity Functions at the Infall Region of the Coma Cluster
---
Introduction
============
The [*nature*]{} vs. [*nurture*]{} debate is a fundamental theme of galaxy evolution studies regarding the relative importance of internal and external processes in the star formation histories of galaxies. Ultraviolet (UV) observations of galaxies are critical to this debate as they provide a direct measure of star formation by probing short-lived ($\simlt$1 Gyr) massive stars [e.g., @Kennicutt1998]. For example, the rest-frame UV luminosity function (LF) has allowed measurements of the cosmic star formation rate at high redshifts [1$\simlt$$z$$\simlt$9; e.g., @Madau1996; @Bouwens2010], and space-based UV facilities such as [*GALEX*]{} have confirmed the rapid decline of star formation from $z\sim1$ to the present day [e.g., @Arnouts2005; @Schiminovich2005]. Environment may play an important role in this decline as the galaxies that dominate the star formation density at $z\sim1$ tend to be located in dense regions [e.g., @Cooper2006; @Bell2007] and their star formation rates have declined at a relatively faster rate to the present day [@Li2011]. Moreover, models suggest that a large fraction of these galaxies now reside in local groups and clusters [@Crain2009; @Cen2011]. UV LF studies of local galaxy clusters may therefore provide important constraints on the processes that regulate star formation in dense environments, giving us a local proxy for studying these effects in the distant Universe. There is considerable debate, however, regarding the identity of the dominant processes that operate in cluster environments [see the review by @Boselli2006a]. For normal late-type galaxies, the removal of gas and the subsequent quenching of star formation within infalling cluster members is typically attributed to interactions with the hot intracluster medium . Tidal processes are invoked to explain the transformation to early-type morphology such as interactions with the cluster potential [e.g., @Merritt1983; @Byrd1990], and/or encounters with other galaxies .
Dwarf galaxies are potentially key to investigating this issue, as their shallow potential wells make them more susceptible to cluster processes. Moreover, models suggest that while giant galaxies are typically accreted into clusters as members of massive groups, and hence to have been “pre-processed" in the group environment, dwarf galaxies are more likely to have been accreted via low-mass halos [e.g., @McGee2009]. Hence the impact of the cluster should be more cleanly distinguished for dwarf galaxies.
We lack a clear understanding of the evolution of dwarf galaxies and the origin of the abundance of dwarf early-type (dE) galaxies found in local clusters [e.g., @Ferguson1994; @Jenkins2007; @Weinmann2011]. This is an important endeavor as dE galaxies dominate the faint-end slopes of cluster LFs which constrain $\Lambda$CDM models of galaxy formation and evolution [e.g., @Benson2003LFb; @Guo2011]. dE galaxies may result from recent quenching of infalling dwarf galaxies owing to processes that are active from the periphery of the cluster to the core, and possibly with galaxy groups prior to accretion [e.g., @Treu2003; @Pogg2004ICM; @Moran2007; @LiskerIII]. On the other hand, some dE galaxies may not be linked to recent quenching of the infall population, e.g., dE galaxies may survive multiple orbits inside the cluster [e.g., @Tully2002; @Pimbblet2011] or may form from the tidal debris of massive cluster members [@Duc2007].
Coma is one of the nearest examples of a rich galaxy cluster and its large dwarf galaxy population is ideal for studying the effects of external processes on star formation. Previous studies in Coma have identified the signatures of dynamical processes associated with both the ICM [e.g., @Caldwell1993; @Bravo2000; @Castander2001; @Fino2004LF; @Pogg2004ICM; @Gavazzi2006; @Miller2009a; @Smith2010; @Yagi2010] and tidal interactions [@Thompson1981; @Graham2003; @Marinova2010; @Madrid2010; @Chiboucas2010; @Peng2011]. Most of these studies were based on small galaxy samples or case studies of unique disrupted populations such as post-starburst galaxies, stripped disks, barred galaxies, ultracompact dwarf galaxies (UCDs), and intracluster globular clusters.
In this paper, we present deep UV LFs for the Coma cluster to study the environmental impact on the star formation properties of the aggregate dwarf galaxy population. Our field is located between the Coma center and the projected virial radius so that we may observe galaxies affected by the environment but before severe tidal disruption at the core. Moreover, the accretion histories of dwarf star forming and dE galaxies are better matched at the infall region as compared to the center [e.g., @Cortese2008b; @Smith2011], which allows a more straightforward comparison of their LFs in order to constrain models of dE galaxy formation. UV LFs present advantages over optical studies for this purpose as UV data allows a more reliable separation of star forming and passive galaxies, especially for dwarf galaxies [@Haines2008].
Previous Coma LFs [a complete compilation is provided in @Jenkins2007] have been measured in bands that trace star formation such as UV [@Donas1991; @Andreon1999; @Cortese2003a; @Cortese2003b; @Cortese2008b], H$\alpha$ [@Iglesias2002], optically-selected blue galaxies [@Mobasher2003LF], 24$\mu$m [@Bai2006], and radio wavelengths [@Miller2009b]. Interestingly, most studies have found little or no difference of the faint-end slope for star forming galaxies with respect to the field environment, other local clusters, or with clustercentric distance. We expect a turnover for dwarf star forming galaxies as they should be more susceptible to dynamical processes. Here, we present the UV LFs of dwarf star forming galaxies to lower masses than these other studies in an attempt to identify such a turnover, and we include the passive dwarf galaxy population in Coma for the first time in order to constrain models of dE galaxy formation. We make use of deep 26 ks [*GALEX*]{} FUV and NUV observations of the Coma infall region with SDSS photometric coverage and moderately deep redshift coverage ($r$$\sim$21.2) to extend the faint end of the Coma UV LF to 3.5 magnitudes deeper than previous efforts. We cover depths similar to the faintest [*GALEX*]{} LFs presented thus far for the (much closer) Virgo cluster [@Boselli2011]. We assume Coma is located at a distance of 100 Mpc [e.g., @Carter2008] which corresponds to a distance modulus of DM $=35.00$ mag ($z=0.023$) and angular scale of 1.667 Mpc deg$^{-1}$ for H$_{0}=71$ km s$^{-1}$ Mpc$^{-1}$, $\Omega_{\lambda}=0.73$, and $\Omega_{m}=0.27$.
Data
====
GALEX and SDSS Catalog
----------------------
In @Hammer2010a [hereafter Paper I], we describe our [*GALEX*]{} observations in detail, and the construction of a source catalog for $\sim$9700 galaxies with [*GALEX*]{} and SDSS photometry. Here we provide only a brief summary. The field is located 09 (1.5 Mpc) south-west of the Coma center and inside the cluster virial radius [r$_{vir}$=2.9 Mpc; @Lokas2003]. The spatial location of our [*GALEX*]{} field is shown in Figure \[coma\_field\]. Photometry was taken from the [*GALEX*]{} pipeline catalog for bright (m$_{\mathrm{UV}}<21$) or extended galaxies, where extended galaxies are defined as having a 90% flux radius larger than 10$\arcsec$ in the SDSS $r$-band. [*GALEX*]{} pipeline apertures were inspected for accuracy and custom photometry was performed in cases of shredded or blended objects. A PSF deblending technique [@Guillaume2006] was used to measure faint and compact sources which lessens systematic effects from object blends, source confusion, and the Eddington Bias, resulting in significantly improved detection efficiency and photometric accuracy as compared to the pipeline catalog. This hybrid approach was necessary given the nature of our very deep images that consist of many resolved cluster members superposed on a dense field of unresolved background galaxies. The catalog has a limiting depth of NUV=24.5 and FUV=25.0 mag in the central 05 [*GALEX*]{} field-of-view (FOV). The [*GALEX*]{} images extend over a larger area but we restrict faint sources to the inner region where the PSF is well behaved as required by the deblending method. We choose a magnitude limit of m$_{\mathrm{UV}}=24.5$ for both [*GALEX*]{} bands which corresponds to the 95% completeness limit of the [*GALEX*]{} images. The coverage area is larger for galaxies brighter than m$_{\mathrm{UV}}=21$ mag (the inner 06 [*GALEX*]{} FOV) as the pipeline aperture photometry is not affected by the shape of the PSF. The larger coverage area increases the number of bright cluster members in our sample which allows for better constraints on the bright-end turnover of the UV LFs. UV magnitudes are corrected for foreground Galactic extinction and are reported in the AB system. Typical photometric uncertainties span 0.1$-$0.3 mag (1$\sigma$ rms scatter) for bright and faint galaxies, respectively, in both [*GALEX*]{} bands.
We rely on the SDSS data to identify galaxies in our catalog as [*GALEX*]{} lacks the spatial resolution to perform this task. The SDSS counterparts to [*GALEX*]{} sources extend to optical magnitudes of $r$=24. For this study, however, we have imposed an optical cutoff at $r$=21.2 mag for the following reasons: (1) spectroscopic redshift coverage in our [*GALEX*]{} field is incomplete at fainter magnitudes, (2) SDSS photometric completeness drops rapidly at fainter magnitudes (see Appendix), and (3) SDSS star/galaxy classification is less reliable at fainter magnitudes (see Appendix in Paper I). As a result of this optical cutoff, our sample does not include potential Coma member galaxies that are fainter than $r$=21.2 mag, but are brighter than our UV magnitude limits at m$_{\mathrm{UV}}$=24.5 (i.e., optically-faint galaxies with very blue UV-optical colors). Galaxies with colors redder than m$_{\mathrm{UV}}$$-r$$=$3.3, which includes the entire cluster red sequence, are not affected and have full coverage to our chosen UV magnitude limits.
The photometric completeness of our catalog is addressed in the Appendix with particular attention given to the SDSS detection efficiency; the [*GALEX*]{} detection efficiency is better than 95% across the full UV magnitude range (Paper I). The SDSS detection efficiency was measured directly by comparing DR6 detections to optical source catalogs from the much deeper and higher-resolution HST-ACS images at the center of Coma [@Hammer2010b]. This comparison was performed separately for both background galaxies and Coma members. The SDSS detection efficiency for Coma members is near 100% to $r=20.5$, and drops to $\sim$70% at our adopted magnitude limit of $r=21.2$. The detection rates are slightly higher for background galaxies at a given apparent magnitude as they tend to have higher surface brightness than the dwarf galaxy population in Coma. The SDSS detection efficiency has also been converted to UV magnitudes for the galaxies in our sample. The SDSS detection rates are given in tabular form in the Appendix.
Our sample consists of 3125 galaxies after removing objects fainter than $r$=21.2 mag. We describe the spectroscopic redshift coverage in our field in Section 2.2 and remove obvious background galaxies from our galaxy sample in Section 2.3. A summary of the final catalog is given in Section 2.4.
Spectroscopic Redshift Coverage
-------------------------------
Several optical spectroscopic redshift campaigns have been conducted in the region of Coma studied here. These campaigns include surveys that span the whole [*GALEX*]{} FOV such as SDSS DR7 [$r$$<$17.77; @Adelman2008] and the spectroscopic follow-up survey of the GMP [@Godwin1983] optical photometric catalog [B$_{j}$$<$20; @Colless1996]. Smaller studies provide additional redshift measurements as compiled in NED [e.g., @Kent1982; @Caldwell1993; @vanHaarlem1993; @Biviano1995; @Colless01; @Rines2003]. A deep spectroscopic survey was performed with the William Hershel Telescope that covers $\sim$66% of the inner 05 [*GALEX*]{} FOV and a smaller fraction of the outer field as seen in Figure \[coma\_field\] [1$<$$B$-$R$$<$2 and $R$$<$20; @Mobasher2001spectra]. There is a more recent spectroscopic campaign with MMT-Hectospec that provides deeper redshift coverage ($\sim1$ hour exposures) to $r$$\sim$21.5 mag within our field and the cluster center. The MMT redshift targets ($\sim$6300 total) were selected from the SDSS DR6 photometric catalog with an emphasis on objects fainter than the SDSS spectroscopic limit at $r=17.77$. The targets were selected with minimal surface brightness and optical color bias, especially for the region of the cluster studied here (a subset of redshift targets at the center of the cluster were limited to colors bluer than $g-r=1.2$). As shown in Figure \[coma\_field\], the MMT survey covers the majority of the inner 05 [*GALEX*]{} FOV with partial coverage of the western outer field. A brief description of the MMT main redshift survey in Coma is given in several studies [@Smith2008; @Smith2009a; @Chiboucas2010].
A composite redshift catalog was created from these spectroscopic surveys that consists of 2483 redshifts in the field studied here. The majority of redshifts are taken from the SDSS and MMT surveys for galaxies brighter/fainter than $r=17.77$, respectively. We matched 1707/2483 redshifts to the 3125 galaxies in our photometric catalog described in Section 2.1. The unmatched redshifts (776/2483) are either stars or galaxies that are fainter than our chosen UV or $r$-band magnitude limits x(most unmatched redshifts are located in the outer $0\fdg5-0\fdg6$ [*GALEX*]{} FOV where we impose bright magnitude limits). We also use the partial HST-ACS coverage of our field [$\simlt$2%; @Carter2008] to determine membership for 8 additional galaxies based on morphology [see @Chiboucas2010]. We identified 5 background galaxies and 3 Coma members among the galaxies in our sample that lack redshift coverage. We consider these 8 galaxies as part of our spectroscopic sample for the remainder of the paper.
Excluding Obvious Background Galaxies
-------------------------------------
We must be aware of potential selection effects in the redshift surveys that would bias measurements of the LF. Some redshift surveys, for example, have specifically targeted cluster member galaxies via color selection. In the opposite sense, spectroscopic surveys that select targets based only on apparent magnitude may have poorer coverage of faint cluster member galaxies, i.e., dwarf galaxies tend to have lower surface brightness (thus lower S/N spectra) as compared to background galaxies for a given apparent magnitude. We can mitigate these potential selection effects by excluding galaxies at much higher redshifts that tend to have both redder optical colors and higher surface brightness than galaxies in the Coma cluster. In Figure \[uvopt\], a UV-optical color-color diagram shows the location of all galaxies in our catalog including both confirmed Coma member galaxies and background galaxies. We identify the region of the diagram where we expect Coma member galaxies using a Monte Carlo suite of multiple-burst SEDs ($\approx$100,000 models) that span a large range in metallicity, star formation history, and dust attenuation as described in [@Salim2007]; the contours in Figure \[uvopt\] show the regions of the diagram covered by 95% and 99% of the models. The galaxies located above the models are objects at higher redshifts ($z$$\simgt$0.2) that are shifted into this region owing to large $K$-corrections. We have removed the 1828 galaxies that lie above the thick line shown in Figure \[uvopt\], which was chosen conservatively to include all potential Coma members and to roughly follow the models. We retain the 1297 galaxies located below the cutoff.
Catalog Summary
---------------
The final catalog used in this study consists of 1297 objects that have [*GALEX*]{} and SDSS detections brighter than NUV=24.5 and $r$=21.2 mag, respectively. A subset (1193/1297 or 92%) have FUV detections brighter than our chosen limit at FUV$=$24.5 mag. Galaxies were identified via the SDSS star/galaxy classification. We have restricted our catalog to the region of $g$-$r$ vs. NUV-$r$ color space where we expect galaxies at the distance of Coma. Spectroscopic redshifts are available for 839 galaxies (65%) including 248 members of the Coma cluster [4000$<$v$<$10 000 km s$^{-1}$; @Colless1996]. The spatial locations of our galaxies are shown in Figure \[coma\_field\]. An analysis of the photometric completeness for our catalog is provided in the Appendix.
UV and Optical Properties of Coma Members
=========================================
In this section, we use UV-optical color diagnostics to classify the Coma galaxy population according to star formation activity. This separation is necessary for reliable measurements of the UV LF (Section 4), and allows us to assess the LFs for different galaxy populations (star forming and passive). We classify dwarf galaxies in Coma by optical magnitude ($r\geq16.5$ or M$_{r}\geq-18.5$) which corresponds to a Johnson B–band magnitude of M$_{\mathrm{B}}\sim-17.0$ [most studies adopt a cutoff between –18 $<$ M$_{\mathrm{B}}<-16$; @Ferguson1994].
FUV$-r$ Color vs. Specific Star Formation Rate
----------------------------------------------
The FUV band allows for a clean separation of star forming and passive galaxies as it is sensitive to recent star formation (a timescale of roughly 0.1 Gyr as compared to 1 Gyr for the NUV band). As such, we use FUV$-r$ color as a proxy for the specific star formation rate (SSFR) with the $r$-band roughly tracing the stellar mass. We note that NUV$-r$ color is a slightly more reliable indicator of SSFR for star forming galaxies as it is less affected by dust, but its sensitivity to metallicity does not allow for a clean separation of metal-poor passive galaxies and star forming galaxies using a simple color cut.
In Figure \[ssfr\], we compare the FUV$-r$ colors of Coma member galaxies to SSFRs taken from the MPA-JHU value-added catalog which are measured via SDSS DR7 spectroscopic emission lines [@Brinchmann2004]. The MPA-JHU values are corrected for light outside the SDSS 3$\arcsec$ fiber by fitting the SDSS photometry to stochastic SED models. For non-SF (passive) galaxies, the MPA-JHU SSFRs are estimated from the 4000 Å break [@Brinchmann2004], and the aperture corrections from SED fitting now account for previous overestimates of their SSFRs as identified by [@Salim2007]. The FUV$-r$ colors are not corrected for internal dust attenuation whereas the SSFRs are corrected for internal dust and, in some cases, for AGN emission. The diagram identifies the SSFRs that correspond to star forming and passive galaxies [taken from @Salim2007]. We are able to select star forming galaxies (Log SSFR (yr$^{-1}$) $>$ $-10.5$) using a simple color cut below FUV$-r$ $=5.5$ with few outliers. A separation based on the observed colors is necessary as we do not have reliable dust indicators for all galaxies in our sample.
For passive galaxies, the FUV$-r$ colors result from a mixture of FUV sources in each galaxy, such as possible AGN emission and/or stars from residual star formation, post-AGB stars, and low-mass helium-burning stars on the horizontal branch (HB; e.g. extreme HB stars or blue HB stars).
UV$-r$ vs. $r$ CMDs
-------------------
The FUV$-r$ vs. $r$ CMD for our catalog is presented in Figure \[fuvr\]. The diagram separates the galaxies by redshift classification: those galaxies that lack redshifts, spectroscopically confirmed background galaxies, and spectroscopically confirmed Coma members (star forming and passive). Background galaxies are more prevalent at blue colors and are the majority population at $r$$\simgt$18, whereas cluster members are the majority at brighter optical magnitudes where the redshift coverage is nearly complete. The redshift coverage spans the full color space. The cluster red sequence is the horizontal band of passive galaxies at the top of the diagram. Several Coma members are fainter than the FUV magnitude limit at FUV=24.5 but are brighter than NUV=24.5 and therefore included in our catalog. We have measured their lower FUV$-r$ color limits[^1] which allowed us to classify additional passive galaxies.
The NUV$-r$ vs. $r$ CMD is shown in Figure \[nuvr\] with the same symbols as used for the FUV CMD. The redshift coverage spans the entire NUV CMD color space. Passive Coma members trace bluer colors at faint optical magnitudes as opposed the flat red sequence seen in the FUV CMD. We performed a linear $\chi$$^{2}$ fit to the red sequence (0.36 mag scatter) as shown in Figure \[nuvr\]. The Coma member galaxies that lack FUV colors for SF/passive classification are indicated on the diagram. The majority are passive galaxies as they are located inside the 2$\sigma$ fitted red sequence (the 3 unclassified cluster members at NUV$-r$$<$3.2 are considered SF galaxies).
The tilt of the red sequence in the NUV CMD is likely a metallicity-mass effect owing to NUV emission that is dominated by main-sequence turnoff stars [MSTO; e.g., @Dorman2003], i.e., for a given stellar age, metal-poor (low-mass and optically-faint) passive galaxies have a more massive MSTO with brighter NUV emission. The average stellar ages of passive galaxies tend to be younger for lower mass galaxies so part of the tilt also reflects an age-mass relation [e.g., @Caldwell2003; @Michielsen2008; @Smith2009b and many others]. Emission from MSTO stars does not extend appreciably into the FUV band resulting in its relatively flat red sequence. Optically-bright galaxies on the red sequence ($r$$<$14) have slightly bluer colors as compared to the fitted relation. These may be UV-upturn galaxies (also visible in the FUV CMD) that are brighter in UV owing to EHB stars [e.g., @Salim2010; @Carter2011].
The average stellar masses of passive galaxies in Coma are given along the top axis of Figure \[nuvr\]. Stellar masses were fit with [KCORRECT]{} [@Blanton2007a] using 7-band photometry from [*GALEX*]{}/SDSS plus [*Spitzer*]{} IRAC 3.6 $\micron$ data when available [@Jenkins2007]. The logarithm of stellar mass is linear with $r$-band magnitude which we fit as Log M$_{*}$ $= 16.364 - 0.435$$\times$$r$ (0.06 dex scatter) for passive galaxies. The relation for SF galaxies has roughly the same slope but is shifted to lower mass by 0.25 dex with larger scatter (0.2 dex), e.g., for a given stellar mass, SF galaxies are $\sim$0.5 mag brighter in the $r$-band as compared to passive galaxies. The majority of confirmed Coma member galaxies have stellar masses larger than M$_{*}=10^{7}$ $\msun$.
Missing Dwarf Star Forming Galaxies in Coma?
--------------------------------------------
Interestingly, both CMDs show a deficit of star forming galaxies in Coma at faint optical magnitudes. This is especially evident in the region of color space labeled “gap" where there is a lack of confirmed SF galaxies. This region is tilted in the CMDs along a path that traces stellar mass in the range M$_{*}$$\sim$10$^{7.5-8.0}$ $\msun$. Although selection effects of the spectroscopic redshift coverage may create an artificial void, this seems implausible as several confirmed background galaxies are located in this region, in addition to background galaxies not shown in Figure \[nuvr\] that were excluded from this study (Section 2.3). It does not result from a magnitude or surface brightness bias in the redshift coverage of Coma members. For example, several confirmed Coma members are located at fainter optical magnitudes and the passive Coma members with stellar masses in the range M$_{*}$$\sim$10$^{7.5-8.0}$ $\msun$ would have a similar surface brightness to SF galaxies in the gap.
We have performed a [*post hoc*]{} analysis to estimate the probability that the gap results from statistical chance alone. For simplicity, we define the gap in the $r$-band interval $18.9<r<19.7$ where there is a lack of confirmed Coma members (this is a conservative definition as the gap is tilted and thus covers a slightly larger area of the CMD). The null hypothesis is that the true fraction of confirmed Coma members inside the gap among all galaxies (referred to here as the “cluster fraction”) is the average of the cluster fractions on either side. Restricting our analysis to the region of Figure \[nuvr\] with colors bluer than NUV$-r=3.5$, the expected cluster fraction inside the gap is 5% (7.5% and 2.5% for the 0.8 mag bins on either side). There are 133 total galaxies inside the gap so we expect 6-7 (133$\times0.05$) confirmed Coma members. The probability of zero confirmed Coma members inside our gap is therefore 0.1% under binomial statistics.
A more comprehensive Monte Carlo analysis shows that the probability that a similar gap appears at [*any*]{} r-band magnitude in Figure 5 is 5-10%. This worst-case scenario implies that we do not expect the gap at a particular location in the CMD. There is independent qualitative evidence, however, that a gap should appear near its current location based on the LFs presented in later sections. For example, the passive galaxy LF in Coma shows a turnover at the same magnitude where the gap extends into the red sequence, and there is suggestive evidence for a turnover in previous optical LFs in Coma at the gap [@Mobasher2003LF]. Deep Virgo UV LFs also show a turnover at magnitudes associated with the gap for both star forming and passive galaxies [@Boselli2011]. We therefore expect the random chance probability for our gap to be less than the worst-case scenario.
A physical explanation for the gap may be more plausible as suggested by its alignment with the expected evolution of galaxies in a particular stellar mass range. For example, the gap is aligned with the path that a dwarf star forming galaxy would follow after rapid quenching. We modeled this track using the GALEV evolutionary synthesis code [@Kotulla2009] for a galaxy with constant star formation (12 Gyr) followed by a 100 Myr quenching event. The model properties just prior to quenching were selected to match NGC 6822 [M$_{*}=10^{7.66}$ $\msun$, M$_{\mathrm{gas}}=10^{8.24}$ $\msun$; @Hunter2004; @Hunter2010], a galaxy in the Local Group that has UV-optical colors that place it inside the gap. The track is shown in Figure \[nuvr\] and spans 4 Gyr from the onset of quenching (the diagram shows the position of the model after 0, 0.25, 0.5, 1, and 2 Gyr).
Note that it would not be surprising if future redshift surveys identified [*some*]{} Coma members in this region. In fact, of the 16 confirmed Coma members at NUV $>21$ in the outer 05$-$06 [*GALEX*]{} FOV (thus these galaxies did not satisfy the criteria to be included in this study), we identified one SF galaxy and it is located inside the gap (albeit just bluer than the SF/passive dividing line). We therefore consider the gap as representative of an overall decline in the number of dwarf SF galaxies in Coma which is not necessarily exclusive to this region of color space. The gap is referenced throughout this study primarily as a tool for interpreting features of the UV LFs. In the Discussion section, we comment further on the possible physical origin of the gap and the SF galaxies located at fainter magnitudes.
UV$-r$ vs. UV CMDs
------------------
The UV$-r$ vs. UV CMDs for both [*GALEX*]{} bands are shown in Figure \[uvcmd\]. These CMDs help to illustrate the procedure invoked in this study for measuring the UV LFs and we will refer back to these diagrams in Section 4. Here, we make a few brief comments/predictions for the Coma UV LFs. The gap extends across a large magnitude range in both UV bands which will affect the LFs for star forming galaxies at m$_{\mathrm{UV}}$$\simgt$21, i.e., the UV LFs for SF galaxies should have a shallower slope starting at m$_{\mathrm{UV}}\approx21$ (M$_{\mathrm{UV}}=-14$ at the distance of Coma). We note that there is a high concentration of SF galaxies in Coma near the boundary that separates passive/SF galaxies, e.g., most galaxies at m$_{\mathrm{UV}}$$\simgt$21 are located in the upper half of the SF color space (NUV$-r>2.5$ and FUV$-r>3.25$). We therefore expect a sharp decline in the UV LF where the gap meets this boundary at FUV$\approx24$ (M$_{\mathrm{FUV}}=-11$) and NUV$\approx23$ (M$_{\mathrm{NUV}}=-12$).
Our catalog is progressively more incomplete for blue galaxies at magnitudes fainter than m$_{\mathrm{UV}}=22$ owing to the optical magnitude limit imposed at $r=21.2$ mag. For measurements of the Coma UV LF, we consider our coverage of SF galaxies in Coma to be complete to slightly fainter UV magnitudes (m$_{\mathrm{UV}}=23$). This fainter limit excludes only a small wedge of color space not covered by our catalog ($22<m_{\mathrm{UV}}<23$ and colors bluer than m$_{\mathrm{UV}}-r<1.8$). We do not expect many (if any) Coma members inside this region as very few cluster members have similarly blue colors, and this region corresponds to starburst galaxies[^2] which are quite rare in the region of Coma studied here [@Mahajan2010]. Passive galaxies are the majority population in Coma at m$_{\mathrm{UV}}>23$ so the imposed optical limit at $r=21.2$ has little impact on the LFs for the total cluster population.
Luminosity Function
===================
Measurements of the cluster LF, defined as the number density of galaxies for a given range of luminosity, require a reliable method of separating cluster members and foreground/background galaxies. Spectroscopic redshifts provide the most direct assessment of cluster membership but are often limited to a subset of the total population. Past UV LF studies of local clusters have determined membership for the remaining galaxies using spectroscopic completeness corrections [e.g., @Cortese2003b; @Cortese2005; @Cortese2008b; @Boselli2011], statistical subtraction of field galaxies based on a control field [e.g., @Andreon1999; @Cortese2003a], or selecting cluster member galaxies via morphology [@Cortese2008b]. Morphology selection cannot be used for this study as SDSS lacks the spatial resolution to identify dwarf galaxies at the distance of Coma. We rely on the spectroscopic completeness method for this study. We have $\sim$100% spectroscopic redshift coverage for galaxies brighter than m$_{\mathrm{UV}}=20$ mag. The completeness of the redshift coverage is shown in the top panels of Figure \[spec\_coverage\] separated into blue and red galaxies. Following the spectroscopic completeness method, the number of cluster members at fainter magnitudes is estimated statistically by assuming that the fraction of confirmed Coma members among galaxies with redshifts is representative of the entire photometric catalog [@DePropris2003; @Mobasher2003LF]. Specifically, we scale the photometric catalog by the membership fraction defined as: $$f(m)= \frac{N_{c}(m)}{N_{s}(m)}_{,}
\label{mf}$$ where N$_{c}(m)$ is the number of confirmed Coma member galaxies, and N$_{s}(m)$ is the number of galaxies with spectroscopic redshifts. The membership fractions are shown in the bottom panels of Figure \[spec\_coverage\].
The luminosity function is the number of galaxies in the [*GALEX*]{}/SDSS photometric catalog (N$_{p}$) scaled by the membership fraction, UV coverage area (A), and the magnitude bin size ($\Delta$$m$=0.5 mag): $$\Phi(m)= f(m)~\frac{N_{p}(m)}{A ~ \Delta m}
\label{lf_eqn}$$ $$\frac{\delta\Phi(m)}{ \Phi(m)} = \left[\frac{1}{N_{p}(m)} + \frac{1}{N_{c}(m)} - \frac{1}{N_{s}(m)}\right]^{1/2}_{,}
\label{lferr}$$ where the error formula assumes N$_{p}$ is a Poisson variable and N$_{c}$ is a binomial variable (i.e., the number of successes in N$_{s}$ trials with probability $f$).
We must also account for the systematic bias related to measuring the cluster LF in a different band than used for redshift target selection (the optical $r$-band). Specifically, the redshift coverage at faint UV magnitudes is better for optically-bright red galaxies (primarily cluster members) which would overestimate the membership fraction for blue galaxies that are primarily background galaxies. We note that this color selection bias would be enhanced by including galaxies fainter than the spectroscopic coverage, i.e., this would simply add more blue galaxies without redshifts. We account for the color bias by measuring the LF separately for star forming and passive galaxies using the dividing lines shown in Figure \[uvcmd\] to separate galaxies by UV$-r$ color. The total LF is a summation of the color-dependent LFs:
$$\Phi(m)=\sum_{m-r}^{sf,p}\frac{f(m | m-r)~N_{p}(m | m-r)}{A~\Delta m}.
\label{lfcoloreqn}$$
This formula reduces to eqn. (2) when the membership fraction is independent of UV-optical color. A significant difference for the LFs measured using eqn. (2) and eqn. (4), as was identified in this study, indicates bias owing to color selection effects and the color-dependent equation should be adopted. In Figure \[lf\], we show the UV LFs for Coma in both [*GALEX*]{} bands, including the separate measurements for passive and star forming galaxies. The LFs have been corrected for the Eddington Bias with errors conservatively taken as 50% of the correction [@Eddington1913], and also for imperfect SDSS and [*GALEX*]{} detection efficiency as described in the Appendix. The corrections amount to less than 7% across the full magnitude range. Separate symbols are used for SF galaxies at magnitudes fainter than M$_{\mathrm{UV}}\approx-12$ where our catalog has incomplete coverage (see Section 3.4). The LF values are given in Table 1. The total LF in each band shows a slight dip at bright magnitudes (M$_{\mathrm{FUV}}\sim-15.7$ and M$_{\mathrm{NUV}}\sim-16.7$). The galaxies brighter than this feature consist almost entirely of SF galaxies. A similar Gaussian-like distribution of bright spirals in Coma has been observed in previous UV LFs [e.g., @Andreon1999; @Cortese2008b], and in the Coma LF at radio wavelengths [@Miller2009b]. The SF LFs increase again at fainter magnitudes before another turnover at M$_{\mathrm{UV}}$$\simgt$$-14$. This faint turnover and the gap start at roughly the same magnitude as predicted in Section 3.4. The magnitude range associated with the gap is labeled on the diagrams. There is an abrupt drop in the SF LFs at fainter magnitudes (M$_{\mathrm{FUV}}\approx-11$ and M$_{\mathrm{NUV}}\approx-12$) which was also predicted in Section 3.4; these magnitudes correspond to the location in the UV CMDs (Figure \[uvcmd\]) where the gap meets the boundary that separates SF/passive galaxies, where there is otherwise a more prevalent number of SF galaxies in Coma. Interestingly, passive galaxies in the NUV band flatten at magnitudes where the gap extends into the red sequence (M$_{\mathrm{NUV}}\simgt-11.5$). Our interpretation of this feature is given in the Discussion section.
We demonstrate the importance of choosing the correct method for measuring the LF in Figure \[otherlf\]. This diagram compares Coma FUV LFs that were measured using the color-dependent spectroscopic completeness method adopted for this study (eqn. \[lfcoloreqn\]), and also using both control field subtraction and the color-independent spectroscopic completeness method (eqn. \[lf\_eqn\]). For the control field LF, we subtracted the [*GALEX*]{} field counts of [@Xu2005] from the galaxy counts in our FOV (Paper I). This method clearly overestimates the number density and faint-end slope of the LF probably due to large-scale structure behind Coma. The color-independent completeness method also overestimates the UV LF as it does not account for systematic errors from selecting redshift targets in a different band. For this method, we also included galaxies to $r$=22.2 (i.e., 1 mag fainter than the redshift coverage) to test the impact on the LF. We found that the overestimate across the last two magnitude bins is almost entirely due to including galaxies fainter than the redshift coverage. LF studies must be careful to avoid the systematic effects associated with these other methods.
LF Fitting
----------
We have modeled the Coma UV LFs as both a single Schechter function and a two-component distribution consisting of a Gaussian at bright magnitudes and a Schechter function at faint magnitudes. In Figure \[uvlf\_fit\], we show the fits to the total, star forming, and passive LFs in the FUV and NUV bands. Fits were performed using the IDL package [mpfit]{} which relies on the Levenberg-Marquardt technique to solve the least-squares problem [@Markwardt2009]. The fitted parameters for each model are listed in Table 2. The details of the fitting procedures are given below.
### Single Component Schechter Functions
We modeled the LFs as a single Schechter function [@Schechter1976] in the following parametric form: $$\Phi(M)=\phi^{*} X^{\alpha+1} e^{-X},
\label{lumfunc}$$ where $X=10^{-0.4(M-M^{*})}$, $\phi^{*}$ is the normalization, $M^{*}$ describes the bright-end turnover, and $\alpha$ is the faint-end slope.
For the total UV LFs, our data do not adequately constrain the bright-end turnover owing to the limited spatial coverage of the cluster. We therefore fixed M$^{*}$ to values reported in a wide-field study of Coma [M$^{*}_{\mathrm{NUV}}$ $=-18.5$ and M$^{*}_{\mathrm{FUV}}$ $=-18.2$; @Cortese2008b] and solved for the faint-end slope. This same procedure was applied to the star forming LFs. For the passive galaxy population, the bright-end turnover is well constrained which allowed us to solve simultaneously for the Schechter parameters.
### Two Component Gaussian + Schechter Functions
LFs are known to be a superposition of Gaussian functions for normal galaxy types (E-S0-Sa-Sb-Sc-Sd) and Schechter functions for dwarf galaxies [dEs and Irr-BCDs; @Binggeli1988; @DeLapparent2003]. The relative abundances of these galaxy types vary with environment resulting in the different shapes of field and cluster LFs. We have therefore fit the UV LFs with a two-component model that consists of a Gaussian function at bright magnitudes and a Schechter function at faint magnitudes. For the SF and passive LFs, the Schechter component is aligned with the dwarf galaxy population in order to match the expected distribution of these galaxy types. This provides a more physical interpretation of the fitted parameters. The bright Gaussian component is relatively less constrained owing to small number statistics but such galaxies are not the focus of this study. The total UV LFs include a mix of normal and dwarf galaxies at faint magnitudes so the fits do not model individual galaxy types. The two-component model does, however, provide a better parametrization of the shape of the total UV LF, e.g., the bright spiral population is clearly distinct from the other galaxies as noted in previous UV LF studies [e.g., @Cortese2008b], and is now well matched by the two-component model as seen in Figure \[uvlf\_fit\]. A more detailed description of the fitting procedure is given below.
For the total and SF LFs, we confined the center of the Gaussian to magnitudes brighter than M$_{\mathrm{UV}}=-16$, but otherwise fit the parameters without constraints. A skew term was added to the Gaussian function for the FUV LF to match the observed distribution. The last magnitude bin of the NUV SF LF drops abruptly so we performed fits both with and without this data point. The faint-end slope is $\alpha=-0.48$ and –0.72 for the full magnitude range and truncated range, respectively. The preferred version is a matter of function as the fit across the full magnitude range is a better description of our data to the completeness limit, whereas the truncated version is more consistent with the lower limits given at fainter magnitudes (the truncated fit is shown in the diagram). Both fits are given in Table 2.
The NUV LF for passive galaxies shows an inflection near M$_{\mathrm{NUV}}$ $=-14$. This location corresponds to the magnitude where dwarf galaxies become the dominant population. We therefore fit the normal passive galaxy population with a Gaussian (M$_{r}<-18.5$; shown as open diamonds in Figure \[uvlf\_fit\]), and then fit the remaining dwarf galaxies with a Schechter function. In contrast to the total and SF LFs, the shape of the passive NUV LF is matched equally well using the single or two-component model. We were unable to perform a two-component fit to the passive FUV LF as dwarf galaxies are the majority in only the last few magnitude intervals.
LF Discussion
-------------
The most notable difference between the single and two-component models are the much shallower faint-end slopes predicted by the Gauss-plus-Schechter functions for all Coma UV LFs. The two-component model removes the sensitivity of the Schechter power law to bright spiral galaxies which tend to force steeper faint-end slopes in order to match the whole LF. This model gives faint-end slopes of $\alpha$ $\approx$ –1.15 in both [*GALEX*]{} bands for the total UV LFs, as opposed to $\alpha$ $=-1.39$ for the single Schechter model. The two-component fits for the separate SF and passive galaxy populations both give faint-end slopes shallower than $\alpha$ $=-1$ (with $\sim$83% confidence for the star forming population).
A similar faint-end slope was reported for dwarf SF galaxies in an H$\alpha$ LF study at the Coma center [$\alpha$$=-0.60$; @Iglesias2002]. This turnover of the H$\alpha$ LF supported the notion that dwarf galaxies were tidally destroyed near the center of Coma resulting in a radial gradient of the faint-end slope [e.g., @Thompson1993]. That we have measured a similar faint-end slope at the infall region suggests that a radial gradient for dwarf SF galaxies is either smaller than expected or does not exist. A deep [*GALEX*]{} observation was recently performed at the center of Coma (where we have similar redshift coverage) which will allow us to test the radial dependence of the faint-end slope for dwarf galaxies that span a large range of cluster-centric distance. We may further constrain this notion by re-measuring the Coma LFs in bands that were previously associated with a radial gradient of the faint-end slope [e.g., $U$-band, 3.6$\micron$, 24$\micron$; @Beijersbergen2002; @Jenkins2007; @Bai2006], but using the more reliable measurement techniques presented here and the improved redshift coverage that is now available.
We are not aware of a previous Coma LF study that reported a faint-end slope shallower than $\alpha=-1$ for dwarf passive galaxies. However, this turnover is obvious in several previous optical LFs in Coma [see the $R$-band LF compilation in Figure 8 of @Milne2007]. It was not reported as most studies extended the Coma LF to much fainter magnitudes via the (unreliable) method of control field subtraction, which resulted in a sharp upturn of the LF after the turnover. [@Trentham1998] showed that the Coma LF is relatively flat after the turnover when considering low surface brightness (LSB) galaxies alone (a more reliable tracer of cluster membership; see their Figure 4d). This is more consistent with the Coma NUV LFs presented in this study.
In the next section, we compare our results to previous UV LF studies. We note that comparing parametrized fits are difficult owing to their dependence on magnitude, and hence are reliable only when considering the same magnitude limit (or similar galaxy types for LFs in different bands). Moreover, although the two-component models provide a better description of the Coma UV LFs, most previous studies have used a single Schechter model. For these reasons, we also rely on comparisons using the actual LF data points.
### Comparisons to Other UV LFs
In Figure \[uvlf\_other\], we compare our Coma UV LFs to previous [*GALEX*]{} LFs measured for local galaxy clusters. The other LFs are the wide-field study of Coma [r$\sim$0.1-1.5 R$_{\mathrm{vir}}$; @Cortese2008b], and studies at the centers of the relatively young Abell 1367 cluster [r$\sim$0.0–0.7 R$_{\mathrm{vir}}$; $\bar{\mathrm{r}}=0.3$ R$_{\mathrm{vir}}$; @Cortese2005] and Virgo cluster [$\bar{\mathrm{r}}=0.3$ R$_{\mathrm{vir}}$; @Boselli2011]. The UV LFs are similar at bright magnitudes despite both the different richness of the clusters and radial coverage of our study (r$\sim$0.2–0.9 R$_{\mathrm{vir}}$; $\bar{\mathrm{r}}=0.55$ R$_{\mathrm{vir}}$). For example, the faint-end slopes of our Coma LFs are nearly identical to the wide-field Coma study [$\alpha_{\mathrm{NUV}}=-1.77$ and $\alpha_{\mathrm{FUV}}=-1.61$; @Cortese2008b] and the A1367 study [$\alpha_{\mathrm{NUV}}=-1.64$ and $\alpha_{\mathrm{FUV}}=-1.56$; @Cortese2005] when we adopt the same magnitude limits at M$_{\mathrm{UV}} = -14$ and –13.5, respectively. We refit our Coma UV LFs at these limits and recovered faint-end slopes of $\alpha \simeq -1.70$ and –1.60 in both [*GALEX*]{} bands, respectively. The Coma LFs flatten at fainter magnitudes resulting in the shallower slopes that we report across our full magnitude range. The Virgo LFs reach similar magnitude limits as our study but are relatively flat, probably owing to fewer dwarf passive galaxies as reported in a recent optical LF study of local galaxy clusters [@Weinmann2011].
It is easier to assess the LFs at faint magnitudes by separating the star forming and passive galaxy populations. In Figure \[nuvlfpts\] (left panel), we compare the NUV LFs for star forming galaxies in Coma to measurements performed for Virgo and the local field [$z$$<$0.1; @Treyer2005]. We derived another NUV LF for the local field environment based on the SDSS $r$-band field LFs [$z$$\simlt$0.05; @Blanton2005]. The optical LFs were converted to the NUV band using the NUV$-r$ color LFs in @Wyder2007. The shaded region in Figure \[nuvlfpts\] represents the limits of the derived field UV LF with the bottom/top edges corresponding to the raw and completeness corrected $r$-band LFs, respectively . The field LFs are similar despite the different selection criteria and the local volumes that were sampled.
Interestingly, the LFs of Coma and Virgo have similar shapes and show the same turnover at M$_{\mathrm{NUV}}$$=-14$. We did not normalize the Virgo LF in Figure \[nuvlfpts\] thus both clusters have nearly the same number density of SF galaxies. [@Cortese2008b] suggested that the SF LFs of local clusters should be identical as they are tracing the same field population at the moment of accretion prior to rapid quenching, i.e., we detect SF galaxies in a shell surrounding the center of each cluster. This implies that although the Virgo study covered the center of the cluster, the SF galaxies for both clusters are detected in projection and are not located inside the core region. Assuming this is correct, we can make the following points. The Virgo LFs were separated by morphology, as opposed to SSFR for Coma, which suggests that the timescales for morphological transition and quenching are roughly similar. We also note that there is mild agreement between the cluster LFs at faint magnitudes (M$_{\mathrm{NUV}}$$>$–12) despite incomplete coverage of Coma SF galaxies. The incomplete coverage applies primarily to dwarf SF galaxies with relatively high SSFRs which suggests that such galaxies are rare in the region of Coma studied here, i.e., the lower limits given for Coma SF galaxies at M$_{\mathrm{NUV}}$$>-12$ are probably near the true value. The field LFs trace both clusters at magnitudes brighter than M$_{\mathrm{NUV}}=-14$ but increase across magnitudes associated with the cluster turnover. [*This is direct evidence that dwarf SF galaxies are more susceptible to environmental processes than normal galaxies.*]{} The divergence of the cluster and field LFs appears to contradict the results of [@Cortese2008b] who suggested they should be identical (albeit this relation was established at magnitudes brighter than the turnover). Two possible explanations for this discrepancy are: (1) the average field LF is not representative of the infalling population onto local galaxy clusters at faint magnitudes (i.e., a turnover exists prior to reaching the infall region studied here), or (2) the average field LF describes the infall population but the quenching rate exceeds the accretion rate for galaxies associated with the turnover (as opposed to higher-mass galaxies that may be in equilibrium). If the second scenario is correct and the number of infalling galaxies is conserved (i.e., blue galaxies are quenched but not destroyed), then we expect the LF for passive galaxies to follow the field LF and increase across magnitudes associated with the turnover for SF galaxies.
The right panel of Figure \[nuvlfpts\] shows the same diagram but now includes dwarf passive galaxies for both Coma and Virgo. For an easier comparison, we have shifted the passive LFs brighter by 2.5 mag which is the average color difference for dwarf passive/SF galaxies at the turnover (this shift roughly aligns the LFs by stellar mass). The shaded region is the expected distribution of dwarf passive galaxies in the cluster for a quenched infall population that follows the field UV LF (assuming blue galaxies are quenched but not destroyed). This model consists of the blue field LF shown in the left panel plus field galaxies that are passive before entering the cluster (shifted by 2.5 mag); the model LF was then normalized to the Coma UV LF for passive galaxies at magnitudes brighter than the turnover. The diagram shows that the passive LFs for both clusters do not trace the quenched field model at faint magnitudes. We conclude that the infall population onto local clusters is not well described by the average field LF inside the virial radius.
The more likely scenario is that the infall population shows a turnover prior to reaching the virial radius. Hence, the passive LF in Coma should roughly follow the same distribution as dwarf SF galaxies. From Figure \[nuvlfpts\], the LFs for both populations in Coma have similar shapes and both show the same turnover (after shifting the passive LF by 2.5 mag). The red fraction – the fraction of Coma member galaxies that are passive – varies between only 80-84% for dwarf galaxies measured across 3 magnitudes in NUV ($-16<$ M$_{\mathrm{NUV}}<-13$ in Figure \[nuvlfpts\]). We note that the Virgo UV LFs for both galaxy types also follow the same distribution before and after the turnover.
Discussion
==========
Perhaps the most interesting result of this study is the deficit of dwarf SF galaxies in Coma at stellar masses below M$_{*}$$\sim$10$^{8.0}$ $\msun$ as compared to the field environment. This deficit was identified via UV LFs which show a turnover for dwarf SF galaxies in Coma at M$_{\mathrm{UV}}\approx-14$ as compared to the rising field UV LFs (Figure \[nuvlfpts\]). There is suggestive evidence for a turnover in previous LF studies of Coma. For example, optical LFs for color-selected blue galaxies in Coma show a turnover at faint magnitudes [M$_{\mathrm{R}}>-16$; see Figure 9 in @Mobasher2003LF]. Although the optical turnover consists of only one data point before reaching the magnitude limit of the [@Mobasher2003LF] study, it was identified for galaxies at both the center of Coma and the infall region studied here. [@Iglesias2002] also reported a turnover for a deep H$\alpha$ LF study at the center of Coma. The deep UV LFs measured for the Virgo cluster [@Boselli2011] are remarkably similar to the Coma LFs presented here and show the same turnover at M$_{\mathrm{UV}}\approx-14$. A deficit of dwarf star forming galaxies may be a common feature of local galaxy clusters.
Interestingly, the shapes of the separate UV LFs for dwarf passive and dwarf SF galaxies in Coma are similar and both populations show the same turnover after correcting for color differences, e.g., the red fraction of dwarf galaxies is roughly constant (80–84%) across 3 magnitudes in NUV covering the magnitude range both before and after the turnover. The UV LFs for both dwarf populations in the Virgo cluster also share the same turnover [@Boselli2011]. This suggests that the dwarf subtypes share a common origin, and that recent quenching of infalling dwarf SF galaxies is sufficient to build the dE population at the infall region of Coma. Studies have offered alternative formation mechanisms for dE galaxies that are not connected to the current infall population. For example, some dE galaxies may have: joined the red sequence at much earlier epochs and are now leaving the core region on a trajectory away from the cluster center [“backsplash galaxies”; @Pimbblet2011], formed from material that was ejected from massive galaxies via tidal interactions [“tidal dwarf galaxies”; e.g., @Duc2007], or were members of the proto-cluster that were quenched by photoionization at the epoch of reionization . These alternative processes are not consistent with the matched LFs for dwarf SF and dwarf passive galaxies. We conclude that these other processes are not important at the cluster-centric distances studied here, but could potentially be relevant in the dense Coma core.
[**What are the expected properties of dwarf SF galaxies that are missing in Coma?**]{} The turnover of the Coma UV LFs is associated with a void, or gap, of dwarf SF galaxies in the UV-optical CMDs (Figures \[fuvr\] and \[nuvr\]). The gap is aligned roughly perpendicular to the red sequence in a small magnitude range at –16.5$\simlt$ M$_{r}\simlt-15.0$ which corresponds to stellar masses of M$_{*}$$\sim$10$^{7.5-8.0}$ $\msun$. The gap itself accounts for only a fraction of the total missing dwarf SF galaxies in Coma. We estimate that $\sim$50 dwarf SF galaxies are required to account for the difference between the Coma and field LFs at $-14<$ M$_{\mathrm{NUV}}<-12$, or $\sim$70% of all SF galaxies in Coma brighter than M$_{\mathrm{NUV}}=-12$. However, the turnover of the UV LF for SF galaxies begins at magnitudes associated with the gap, which suggests these two features are related. This allows us to assess the identity of the missing dwarf SF population via comparisons to field galaxy studies.
In Figure \[GAP\], we compare the UV-optical CMD of Coma member galaxies to a local sample of HI-selected dwarf SF galaxies [Sm-BCD-Im/Irr; @Hunter2010]. The Sm field galaxies tend to be located at brighter magnitudes than the gap, whereas Im/Irr galaxies (hereafter dIrr galaxies) are the majority of galaxies inside or fainter than the gap with a smaller fraction of BCDs. The dIrr galaxy population is therefore the likely candidate for missing dwarf SF galaxies in Coma. We are unable to characterize the morphology of most Coma members inside or fainter than the gap, however, as these galaxies are only marginally resolved by SDSS. At least one galaxy is a probable BCD based on its strong H$\alpha$ emission (EW \[H$\alpha$\] $\sim500$ Å; labeled on Figure \[GAP\]) and high surface brightness relative to other Coma members in this region. Interestingly, a large fraction of BCDs in the Virgo cluster are located in the same region of color space as our BCD candidate, and a few Virgo galaxies classified as “dE galaxies with blue centers" would be located inside our gap [@LiskerIII; @Kim2010]. This may indicate that the gap represents a deficit of low surface brightness SF galaxies in Coma, and those SF galaxies inside or fainter than this region tend to be more compact galaxies.
[**What processes are responsible for the lack of dwarf SF galaxies in Coma?**]{} Dynamical processes inside the cluster provide an obvious method for removing dwarf SF galaxies. ICM processes (not tidal) are the dominant mechanism at the off-center region of Coma studied here [e.g., ram-pressure stripping, viscous stripping, thermal evaporation; @Boselli2006a]. ICM processes quench star formation by removing the gas supply to galaxies resulting in a transition to the red sequence. The thermal evaporation of gas via the hot ICM [@Cowie1977] may be more efficient in Coma relative to other local clusters owing to the high temperature of its ICM [e.g., @Boselli2006a]. This process is tempered by galactic magnetic fields thus dIrr galaxies are especially vulnerable owing to their low field strengths [a factor of $\sim$3 lower than spirals; @Chyzy2011]. Preventing the infall of external gas may also effectively suppress star formation in dwarf galaxies, possibly via thermal evaporation of the extended HI gas reservoirs that envelope dIrr and BCD galaxies [e.g., @Begum2008]. The relative lack of BCD galaxies in Coma as compared to Virgo is consistent with this scenario. Regardless of the quenching process, gas ablation of infalling galaxies is consistent with the similar shapes of the LFs for dwarf passive and dwarf SF galaxies in Coma. It does not explain, however, the relative deficit of dwarf SF galaxies in Coma as compared to the field.
Alternatively, the infalling galaxy population onto Coma may have a deficit of dwarf galaxies prior to accretion, possibly owing to “pre-processing" of galaxies that flow along filaments feeding the cluster. Massive clusters such as Coma tend to be associated with larger numbers of intercluster filaments [e.g., @Pimbblet2004]. Filaments may in fact connect the Coma cluster, embedded in the Great Wall along with the A1367 cluster [@Geller1989], to a large structure of galaxies at $z\sim0.05$ and the SDSS Great Wall at $z\sim0.08$ [@Gott2005] and beyond [@Adami2009]. The infall region of Coma studied here also lies in projection with an intercluster filament that connects Coma to the A1367 cluster, and may be specifically affected by a subcluster merging process associated with the NGC 4839 group along this filament. Galaxies located in intercluster filaments show increased SF activity at the periphery of clusters ($\sim$1.5-2 R$_{\mathrm{vir}}$) possibly due to galaxy harassment with infalling substructure [@Porter2008]. The periphery of the Coma cluster, for example, hosts a large fraction of post-starburst (k+A) galaxies as compared to both the surrounding field and the intercluster filament that connects Coma and A1367 [@Mahajan2010; @Gavazzi2010]. After this initial spike in star formation, harassment can rapidly remove up to 50-90% of the stellar mass in dwarf galaxies [@Moore1999]. The remnants of dIrr galaxies associated with this process (if they survive complete disruption) would lie on the red sequence at fainter UV magnitudes than detected in our study. These remnants may be the amorphous very LSB galaxies (VLSBs) that have been detected in both Virgo and Coma at optical wavelengths [@Sabatini2005; @Hammer2010b]. Preliminary optical LFs from the HST-ACS Coma Treasury survey (Trentham et al. 2011, in preparation) show an increasing number of VLSB galaxies at magnitudes just fainter than our study. ACS coverage was limited primarily to the central region of Coma, however, so we are unable to establish whether VLSBs originated at large cluster-centric distances or via tidal processes near the core. Further ACS observations at the infall region of Coma would be useful for this purpose. We note that compact galaxies (existing either prior to the tidal interaction or afterwards owing to a strong nuclear starburst) are more likely to survive this process and reach the cluster virial radius.
The deficit of dwarf galaxies in Coma may also result from the pre-processing of galaxies in groups located in intercluster filaments (e.g., via tidal interactions). Recent studies of compact group galaxies (CGGs) have identified a deficit of SF galaxies with UV and IR properties that are otherwise common to a wide range of environments [albeit with more emphasis on normal galaxies; @Johnson2007; @Tzanavaris2010; @Walker2010]. Interestingly, [@Walker2010] found that the properties of Coma member galaxies at the infall region studied here were the best match to the CGG population. We note, however, that pre-processing in groups is likely a secondary effect as slightly less than 50% of all dwarf galaxies accreted onto local galaxy clusters originate from groups [@McGee2009]. Although studies have started to examine the UV properties of dwarf galaxies in groups and also filaments [e.g., @Haines2011; @Mahajan2011], deeper surveys are needed to constrain the SF population at the turnover identified here.
The most likely scenario in our estimation is that the deficit of dwarf galaxies in Coma results from tidal interactions with infalling substructure outside the cluster virial radius. Dwarf SF galaxies that survive this process are quenched via ICM processes inside the cluster. Deep [*GALEX*]{} observations recently performed at the center of Coma (P.I. R. Smith) will place further constraints on the evolution of dwarf galaxies in Coma. The core region also has deep HST-ACS coverage [@Carter2008] allowing for detailed structural information on cluster members.
Summary
-------
We report on a deep [*GALEX*]{} imaging survey at the infall region of the Coma cluster, which for the first time has placed constraints on the Coma FUV and NUV LFs in the magnitude range $-14.0$$\simlt$M$_{\mathrm{UV}}$$\simlt$$-10.5$. This study reports the following results:
The Coma FUV and NUV LFs for SF galaxies show a turnover at faint magnitudes (M$_{\mathrm{UV}}$$\simgt$$-14$) that is associated with a deficit of dwarf SF galaxies with stellar masses below M$_{*}\approx10^{8}$ $\msun$. The same turnover is identified in UV LFs for the Virgo cluster [@Boselli2011], whereas the field UV LFs [e.g., @Treyer2005] increase to fainter magnitudes. This is direct evidence that dwarf SF galaxies are more efficiently processed in dense environments as compared to higher-mass galaxies.
The shapes of the Coma NUV LFs for dwarf passive and dwarf SF galaxies are well matched after correcting for color differences, with both LFs showing the same turnover (dwarf passive galaxies are more numerous by a factor of $\sim$4-5). This suggests that the recent quenching of SF galaxies via ICM processes is sufficient to build the dwarf passive population at the infall region of Coma. The similar LFs also imply that the infall population has a deficit of dwarf SF galaxies prior to reaching the cluster virial radius, possibly due to severe tidal disruption of dwarf galaxies along intercluster filaments.
There is a lack of confirmed Coma member galaxies, or gap, in a small region of the UV-optical CMDs associated with dwarf SF galaxies. The gap is found over a narrow range of optical magnitudes at $-16.5\simlt$ M$_{r}\simlt-15.0$ corresponding to stellar masses of M$_{*}$$\simeq$10$^{7.5-8.0}$ $\msun$. A comparison to dwarf SF galaxies observed in the field [@Hunter2010] indicates that the gap and the overall deficit of dwarf SF galaxies in Coma may reflect a lack of low surface brightness (dIrr) galaxies in Coma.
The Coma UV LFs were fit using both a single Schechter function and a two-component Gaussian-plus-Schechter model. The two-component model provides a better parametrization of the UV LFs, especially for SF galaxies, and results in shallower faint-end slopes. Fitting the total population gives faint-end slopes of $\alpha$ $\simeq$ $-1.15$ $\pm$ $0.12$ in both [*GALEX*]{} bands as compared to $\alpha$ $=-1.39$ $\pm$ 0.06 for the single Schechter model. The two-component fits to the separate passive and star forming UV LFs were performed by aligning the Schechter component to the dwarf galaxy population [the expected distribution for dIrr and dE galaxies; e.g., @DeLapparent2003] resulting in faint-end slopes shallower than $\alpha$ $=-1$.
We have demonstrated that LF studies that estimate background contamination using a control field may significantly overestimate the number density and faint-end slope owing to large-scale structure behind Coma. The spectroscopic completeness method may also overestimate the number density and faint-end slope (for any LF measured in a different band than used to select redshift targets) by: (1) not measuring the LF separately for blue and red galaxies, or (2) including galaxies fainter than the spectroscopic redshift limit.
We thank the anonymous referee for their detailed comments and suggestions that have improved this paper. We thank R. Marzke and members of the Coma Hectospec team for their work on the redshift catalog, and M. Colless for providing an updated version of the GMP catalog redshifts. We are grateful for the LF data shared by A. Boselli, M. Treyer, and M. Blanton, and assistance with the GALEV software by R. Kotulla. We thank T. Heckman for commenting on an early draft and A. Basu-Zych, P. Tzanavaris, and B. Lehmer for helpful science discussion. This research was partially supported by the [*GALEX*]{} Cycle 2 grant 05-GALEX05-0046 (P.I. Hornschemeier). [*GALEX*]{} is a NASA Small Explorer, developed in cooperation with the Centre National d’Etudes Spatiales of France and the Korean Ministry of Science and Technology. Funding for the creation and distribution of the SDSS Archive has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, NASA, the NSF, DoE, Monbukagakusho, Max Planck Society, and the Higher Education Funding Council for England. The SDSS Web Site is http://www.sdss.org/. This study made use of the NASA Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory (Cal Tech), under contract with NASA.
[*GALEX*]{} and SDSS Photometric Completeness
=============================================
[*GALEX*]{} completeness was addressed in Paper I as simulations were performed to measure the probability that a UV source was detected for a given optical counterpart. We were able to recover the UV flux for 95% of the simulated sources brighter than NUV=24.5 and FUV=24.5 mag.
The SDSS DR6 catalog is 95% complete for point sources to $r$=22.2 mag [@Adelman2008], but this limit may be $\sim$1 mag brighter for extended galaxies owing to their lower surface brightness[^3]. We have measured the SDSS $r$-band completeness, or detection efficiency, both for Coma members and background galaxies by matching DR6 to much deeper HST-ACS source catalogs at the center of the Coma cluster [the ACS survey was performed in the F475W and F814W passbands; @Carter2008; @Hammer2010b]. Coma members in the ACS catalogs were identified from the spectroscopic redshift coverage (similar to the redshift coverage described in Section 2.2) and also using morphology selection [see @Chiboucas2010]; the majority of dwarf galaxies in Coma are low surface brightness dE and dSF galaxies, with a small fraction of compact ellipticals. In cases where SDSS did not record a detection, the $r$-band magnitude was derived from the typical SDSS/ACS colors of similar galaxies. In Table 3, we give the SDSS $r$-band detection efficiencies for background and Coma member galaxies fainter than $r$=18 (we assume 100% completeness at brighter magnitudes). We confirm that the SDSS 95% completeness limit for background galaxies is $\sim$1 mag brighter than for point sources, and drops rapidly to only a few percent for galaxies fainter than $r$=21.5 mag. The detection efficiency for Coma member galaxies is similar to the background galaxies but with a brighter and more gradual decline that starts at $r$=20.5 mag. The brighter decline results from Coma dwarf galaxies that, for a given apparent magnitude, have lower surface brightness than the average SDSS galaxy located at higher redshifts ($z$$\sim$0.1-0.2).
We have also estimated the SDSS detection efficiency for galaxies as a function of [*GALEX*]{} FUV and NUV magnitudes. This was measured in intervals of UV magnitude by summing the $r$-band detection efficiencies for each galaxy and dividing by the total number of galaxies: DE$_{UV}$=$\sum_{i=1}^{n_{gal}}$DE$_{r}$(i)/n$_{gal}$. We have further separated the Coma member and background samples into blue and red galaxies based on their UV-optical colors. These values are provided in Table 4.
![\[ssfr\] FUV$-r$ color vs. specific star formation rate (SSFR) for Coma member galaxies. SSFRs are taken from the MPA-JHU SDSS value-added catalog [e.g., @Brinchmann2004] which is limited to galaxies brighter than $r$$\approx$17.8 mag. The vertical dotted line indicates the typical separation in SSFR for galaxies classified as star forming or passive [e.g., @Salim2007]. The diagram shows that FUV$-r$ color allows for a similar separation of SF and passive galaxies in Coma with a dividing line at FUV-r=5.5 mag (horizontal dashed line).](f3.eps)
![\[fuvr\] FUV-optical CMD for all galaxies in our sample. Orange dots are galaxies that lack spectroscopic redshifts, black dots are confirmed background galaxies, and confirmed Coma members are shown as large filled symbols (red squares=passive; blue circles=star forming). We have spectroscopic redshift coverage across the full magnitude-color space. We use the horizontal dashed line at FUV$-r=5.5$ to separate star forming and passive galaxies in the Coma cluster. The top x-axis gives the corresponding absolute $r$-band magnitude for galaxies at the distance of Coma. Triangles show the lower color limits for Coma member galaxies that are detected in the NUV band but have FUV magnitudes fainter than our chosen limit at FUV$=24.5$ (gray dashed line); the lower color limits are used to classify those Coma members above FUV$-r=5.5$ as passive. The black dashed lines define a “gap" in the CMD that is void of confirmed Coma member galaxies (see Section 3.3).](f4.eps)
![\[nuvr\] NUV-optical CMD for all galaxies in our sample. Orange dots are galaxies that lack spectroscopic redshifts, black dots are confirmed background galaxies, and confirmed Coma members are shown as large symbols (red squares=passive; blue circles=star forming; black open circles=no FUV color information). We have spectroscopic redshift coverage across the full magnitude-color space. A linear fit to passive Coma member galaxies (red solid line) and the 2$\sigma$ scatter (red dotted lines) are shown. The majority of Coma members without FUV colors are passive galaxies as they are located inside the 2$\sigma$ scatter. The top x-axis gives the average stellar masses for passive Coma member galaxies as measured in this study (Log M$_{*}$ $= 16.364 - 0.435$$\times$$r$; Section 3.2); SF galaxies are $\sim$0.5 mag brighter in the $r$-band for a given stellar mass. The black dashed lines enclose a “gap" in the CMD that is void of confirmed Coma member galaxies. Inside the gap, the thick green line is the path that a dwarf SF galaxy in Coma would follow after a rapid 100 Myr quenching event (see Section 3.3); the line spans 4 Gyr from the onset of quenching and triangles show the model after 0, 0.25, 0.5, 1, and 2 Gyr. The gray long-dashed line is our chosen magnitude limit at NUV=24.5.](f5.eps)
![\[uvcmd\] UV-optical CMDs plotted against [*GALEX*]{} FUV (left panel) and NUV (right panel) magnitude. Symbols are identical to Figures \[fuvr\] and \[nuvr\]. The top x-axis gives the corresponding absolute magnitudes for galaxies at the distance of Coma. These plots are useful for illustrating the color distribution of galaxies in UV intervals and the method invoked for measuring the UV LFs in Section 4. The gray long-dashed lines show the SDSS $r$-band magnitude limit ($r$=21.2) adopted for this study. The black horizontal long-dashed lines are used to separate passive/SF galaxies for the UV LF measurements, and were selected to closely match the FUV$-r$ color classification of Coma members as shown on the diagrams. The vertical black dotted lines indicate the completeness limit adopted for SF galaxies; at fainter magnitudes, our galaxy sample is progressively more incomplete owing to the $r$-band magnitude limit at $r=21.2$. The black short-dashed lines enclose the color space affected by the gap which results in fewer SF Coma members at magnitudes fainter than M$_{\mathrm{UV}}\simgt-14$. ](f6a.eps "fig:")\
![\[uvcmd\] UV-optical CMDs plotted against [*GALEX*]{} FUV (left panel) and NUV (right panel) magnitude. Symbols are identical to Figures \[fuvr\] and \[nuvr\]. The top x-axis gives the corresponding absolute magnitudes for galaxies at the distance of Coma. These plots are useful for illustrating the color distribution of galaxies in UV intervals and the method invoked for measuring the UV LFs in Section 4. The gray long-dashed lines show the SDSS $r$-band magnitude limit ($r$=21.2) adopted for this study. The black horizontal long-dashed lines are used to separate passive/SF galaxies for the UV LF measurements, and were selected to closely match the FUV$-r$ color classification of Coma members as shown on the diagrams. The vertical black dotted lines indicate the completeness limit adopted for SF galaxies; at fainter magnitudes, our galaxy sample is progressively more incomplete owing to the $r$-band magnitude limit at $r=21.2$. The black short-dashed lines enclose the color space affected by the gap which results in fewer SF Coma members at magnitudes fainter than M$_{\mathrm{UV}}\simgt-14$. ](f6b.eps "fig:")
![\[spec\_coverage\] Properties of the spectroscopic redshift coverage at FUV (left panels) and NUV (right panels) wavelengths. The spectroscopic completeness - the fraction of all galaxies with spectroscopic redshifts - is shown in the top panels ($N_{s}/N_{p}$). The membership fraction - the fraction of galaxies with redshifts that are cluster members - is shown in the bottom panels ($N_{c}/N_{s}$). Measurements are performed separately for blue/red galaxies using the color separation defined in Figure \[uvcmd\]. The uncertainties are binomial errors from the formalism in [@Gehrels1986]; error bars are shown only at magnitudes where we use membership fractions to measure the LF (i.e., where the spectroscopic completeness is less than 100%). The top panels indicate that red galaxies have higher spectroscopic completeness as compared to blue galaxies, and bottom panels show that a significantly higher fraction of red galaxies are Coma members. Given the color dependence of the spectroscopic redshift coverage, we construct the Coma UV LF as a summation of the separate blue/red LFs in order to avoid systematic color selection effects.](f7.eps)
[^1]: FUV lower limits were measured on the background-subtracted [*GALEX*]{} pipeline images inside a circular aperture covering the FWHM PSF. The flux was scaled by a factor of 2 (the galaxies are nearly point sources) and corrected for Galactic extinction. Magnitudes were taken at the lower 2$\sigma$ flux value (95% confidence interval) or set to FUV=24.5 if photometry failed.
[^2]: We consider starburst galaxies to have specific star formation rates (SSFR) larger than Log \[SSFR\] $=-9$ yr$^{-1}$. Coma members in this region of color space would be classified as starburst galaxies based on estimates of the SSFR using the stellar mass relation described in §3.2 and the [@Kennicutt1998] UV SFR conversion assuming negligible dust extinction.
[^3]: www.sdss.org/dr6/products/general/completeness.html
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'A. Chen'
- 'M. Kübel'
- 'B. Bergues'
- 'M. F. Kling'
- 'A. Emmanouilidou'
title: 'Non-sequential double ionization with near-single cycle laser pulses'
---
Introduction {#introduction .unnumbered}
============
Non-sequential double ionization (NSDI) in intense near-infrared laser fields is a fundamental process with electron-electron correlation playing a key role [@Corkum; @NSDI1; @NSDI2]. Considerable information regarding NSDI has been obtained from kinematically complete experiments, i.e., the momenta of the escaping electrons and ions are measured in coincidence [@kinematic1]. Most of these experiments employ multi-cycle laser pulses allowing for multiple recollisions to occur before both electrons ionize. Multiple recollisions complicate the electron dynamics and render the comparison with theory difficult. Recently, however, kinematically complete experiments succeeded in confining NSDI to a single laser cycle by using carrier-envelope phase (CEP)-controlled few- and near-single-cycle pulses [@Camus; @Kling1]. These experiments with near-single-cycle pulses allow for an easier comparison between theory and experiment.
To interpret the double ionization spectra of driven Ar measured using near-single-cycle laser pulses, a simple one-dimensional (1D) classical model was put forth [@Kling1; @Kling2; @Bergues]. This model relies on the assumption that the dominant pathways of double ionization are, for small and intermediate intensities, delayed non-sequential ionization and, for higher intensities, sequential ionization. This model neglects the contribution of another major pathway of double ionization, namely, direct ionization as well as the Coulomb potential. This 1D model did not achieve a quantitative agreement with the complete set of available experimental data over the whole intensity range. Delayed ionization—also referred to as recollision-induced excitation with subsequent field ionization, RESI [@RESI1; @RESI2], and direct ionization are two main pathways of NSDI. An interesting finding of these near-single cycle experiments was that the correlated momenta components of the two escaping electrons along the direction of the laser field have a cross-shaped pattern for an intensity around 10$^{14}$ W/cm$^2$ [@Kling1; @Kling2; @Bergues]. A cross-shaped correlated momenta pattern due to the delayed double ionization mechanism was previously identified in the context of strongly-driven He at an intensity 9$\times$10$^{14}$ W/cm$^{2}$ and a wavelength 400 nm [@Emmanouilidou1]. In the context of strongly-driven Ar, the above described 1D model attributed the cross-shaped pattern of the correlated momenta to the delayed pathway of double ionization [@Kling2; @Bergues]. Furthermore, a quantum mechanical calculation, which only considers the delayed pathway of double ionization and neglects the Coulomb potential, identified the key role that the symmetry of the excited state plays in the final shape of the correlated momenta [@Faria].
In this work, using a three-dimensional (3D) semiclassical model, NSDI of Ar is studied when Ar is driven by 750 nm near-single-cycle laser pulses at intensities ranging from 0.85$\times$10$^{14}$ W/cm$^2$ to 5$\times$10$^{14}$ W/cm$^2$. In this 3D model the only approximation is in the initial state. There is no approximation during the time propagation. That is, all Coulomb forces and the interaction of each electron with the laser field are fully accounted for. No assumptions are made regarding the prevailing mechanism of double ionization and there is no free parameter. This is not the case for the 1D model [@Bergues]. Moreover, the Coulomb singularity is fully accounted for using regularized coordinates [@KS]. This is an advantage over models which soften the Coulomb potential [@Huang]. Previous successes of this 3D model include identifying the mechanism responsible for the fingerlike structure [@Agapi1], which was predicted theoretically [@Taylor1] and was observed experimentally for He driven by 800 nm laser fields [@vshape1; @vshape2]. Moreover, this model was used to investigate direct versus delayed pathways of NSDI for He driven by a 400 nm laser field while achieving excellent agreement with fully ab-initio quantum mechanical calculations [@Agapi2]. Using this model, in this work, several observables are computed for different intensities of strongly-driven Ar. These observables are the sum of the two electron momentum components along the direction of the polarization of the laser field and the double differential probability of the two electron momentum components along the polarization direction of the laser field. Furthermore, the amplitude and the phase of the asymmetry parameter that determines the difference of the ions escaping with positive versus negative momentum along the polarization direction of the laser field are computed as a function of the carrier envelope phase (CEP) and the intensity. The computed results using the 3D semiclassical model are found to be in better agreement with the experimental results over the whole intensity range [@Kling1; @Kling2; @Bergues] compared to the results obtained with the 1D model in ref. [@Bergues]. Motivated by the good agreement between theory and experiment, the strength of the 3D semiclassical model in fully accounting for the electron dynamics is utilized to identify the prevailing pathway of double ionization as a function of intensity. In addition, for a small intensity around 10$^{14}$ W/cm$^{2}$, the dependence of the double ionization pathways on CEP is computed using the 3D semiclassical model. Finally, the transition from strong to soft recollisions is identified as the main reason for the experimentally observed escape of the two electrons with opposite velocities at higher intensities [@Agapi4].
Method
======
For the current studies, a 3D semiclassical model is employed that is formulated in the framework of the dipole approximation [@Agapi1]. The time propagation is determined by the three-body Hamiltonian of the two electrons with the nucleus kept fixed. All Coulomb forces are accounted for: the interaction of each electron with the nucleus and the laser field and the electron-electron interaction are all included in the time propagation. The laser field is given by
$$\begin{aligned}
\vec{E}(t)=E_0e^{\left(-2ln2\left(\frac{t}{\tau}\right)^2\right)}cos(\omega t+\phi)\hat{z},\end{aligned}$$
where $\tau=4fs$ is the full-width-half-maximum pulse duration, $\omega$=0.061 a.u (750 nm) the frequency, $E_{0}$ the strength and $\phi$ the CEP of the laser field. In this work linearly polarized laser fields are considered. The initial state in the 3D model entails one electron tunneling through the field-lowered Coulomb potential with the Ammosov-Delone-Krainov (ADK) formula [@A1; @A2]. To obtain the tunnel ionization rate for Ar, in the ADK formula the first ionization energy of Ar, i.e. $I_{p_{1}}=0.579$ a.u. and the effective charge $Z=1$ are used. The momentum along the direction of the electric field is zero while the transverse one is given by a Gaussian distribution [@A1; @A2]. The remaining electron is initially described by a microcanonical distribution [@Abrimes]. The microcanonical distribution is obtained using the second ionization energy of Ar, i.e. $I_{p_{2}}=1.02$ a.u. and an effective charge equal to $Z=2$ a.u. During the time propagation each electron is interacting with the nucleus with charge $Z=2$. In what follows, the tunneling and bound electron are denoted as electrons 1 and 2, respectively.
The intensities considered for the results presented range from 0.85$\times$10$^{14}$ W/cm$^2$ to 5$\times$10$^{14}$ W/cm$^2$. For the smallest intensity, 12 CEPs are considered ranging from $\phi=15^{\circ}$ to $\phi=345^{\circ}$ in steps of 30$^{\circ}$. For all other intensities, 24 CEPs are considered ranging from $\phi=0^{\circ}$ to $\phi=360^{\circ}$ in steps of 15$^{\circ}$. For the results presented regarding total double ionization the average has been taken over all CEPs for each intensity. It is noted that the computations required are challenging, since, it is time-consuming to obtain enough double ionization events that render the statistical error very small for each of the 12 or 24 CEPs for each intensity. Therefore, computations were performed for six intensities in the range from 0.85$\times$10$^{14}$ W/cm$^2$ to 5$\times$10$^{14}$ W/cm$^2$. Using the results obtained at these six intensities an average over the focal volume is performed [@laser_volume_effect] to directly compare with experiment. It is, however, noted that computations at a larger number of intensities are needed to account more accurately for the focal volume effect. For the results presented it is stated explicitly when focal volume averaging is included and when it is not.
Results
=======
Double ionization and pathways
------------------------------
In [Fig. \[ratioDISI\]]{}, the ratio of double to single ionization events is computed as a function of the laser intensity and compared to the experimental results [@Bergues]. It is found that the computed ratio of double to single ionization events reproduces well the overall pattern of the observed ratio. The computed ratio is found to be at most a factor of two smaller than the observed ratio and by a factor of 3.5 when the focal volume effect is accounted for. This difference possibly suggests that the effective charge of $Z=2$ used to model the attractive Coulomb potential in the 3D semiclassical model during time propagation overestimates the Coulomb attraction.
\[ht!\] ![Ratio of double to single ionization as a function of intensity. Experimental results are denoted by dark blue squares and light blue crosses and computed results are presented by a solid line with black circles when the focal volume effect is not accounted for and by a dashed-line with triangles when the focal volume effect is accounted for. The difference in the two experimental sets results from slightly different averaging over the focal volume [@Bergues].[]{data-label="ratioDISI"}](Fig1_Exp_DI_SI_Volume_interp "fig:"){width="40.00000%"}
Once the doubly ionizing events are obtained using the 3D model, an analysis of the trajectories is performed in time in order to identify the contribution of the direct and the delayed pathway of NSDI as a function of the laser intensity. The main two double ionization energy transfer pathways are identified by using the time difference between the recollision time $\mathrm{t_{rec}}$ and the ionization time of each electron $\mathrm{t_{ion}^{i}}$, with $\mathrm{i=1,2}$, for each doubly ionizing classical trajectory. The recollision time is defined as the time of minimum approach of the two electrons and is identified by the maximum in the electron pair potential energy. The ionization time for each electron is defined as the time when the sum of the electrons’ kinetic energies (using the canonical momentum) and the potential energy due to the electron’s interaction with the nucleus becomes positive and remains positive thereafter. The canonical momentum of an electron is given by $\mathrm{{\bf p}-{\bf A}}$, with ${\bf A}$ the vector potential. The ionization time of the tunneling electron is, thus, not necessarily the time this electron tunnels at the start of the time propagation. This energy is referred to as compensated energy and was introduced in ref. [@Leopold]. A doubly ionized trajectory is labeled as delayed or direct depending on the time differences $\mathrm{t_{ion}^{1}-t_{rec}}$ and $\mathrm{t_{ion}^{2}-t_{rec}}$. Specifically, if the conditions $$\begin{aligned}
|t_{ion}^{1}-t_{rec}|<t_{diff}, \hspace{2pt} t_{ion}^{2}<t_{ion}^{1} \hspace{5pt} or \hspace{5pt} |t_{ion}^{2}-t_{rec}|<t_{diff}, \hspace{2pt}t_{ion}^{1}<t_{ion}^{2}\end{aligned}$$ are satisfied then the trajectory is labeled as direct. If the conditions $$\begin{aligned}
t_{ion}^{1}-t_{rec}>t_{diff}, \hspace{2pt} t_{ion}^{2}-t_{rec}<t_{diff} \hspace{5pt} or \hspace{5pt} t_{ion}^{2}-t_{rec}>t_{diff}, \hspace{2pt} t_{ion}^{1}-t_{rec}<t_{diff} \end{aligned}$$ are satisfied then the trajectory is labeled as delayed. If the conditions $$\begin{aligned}
t_{ion}^{1}-t_{rec}>t_{diff} \hspace{5pt} and \hspace{5pt} t_{ion}^{2}-t_{rec}>t_{diff}\end{aligned}$$ are satisfied then the trajectory is labeled as double delayed. The percentage of delayed and direct trajectories depends on the choice of the time difference $\mathrm{t_{diff}}$. This is shown in [Fig. \[probpathways\]]{} where the probability of direct and delayed double ionization events is plotted for $\mathrm{t_{diff}}$ equal to 1/10 T, 1/20 T and 1/40 T. The probability of the direct (delayed) ionization pathway is obtained by dividing the number of doubly ionizing trajectories labeled as direct (delayed) with the total number of doubly ionizing trajectories. It is found that the percentage contribution of the direct and the delayed double ionization pathways as a function of intensity display general trends that do not significantly depend on the choice of $\mathrm{t_{diff}}$. Both the direct and the delayed pathways of double ionization significantly contribute at all intensities. Thus the direct pathway can not be neglected as was done in previous models. The direct pathway contributes the most for intermediate intensities. In [Fig. \[probpathways\]]{}, at a high intensity above 4$\times$10$^{14}$ W/cm$^2$, it is shown that the percentage contribution of the direct pathway of double ionization starts decreasing. At this high intensity a transition from strong to soft recollisions takes place, as discussed in the following. $\mathrm{t_{diff}}$ can not be chosen neither very large, such as 1/4 T, or very small such as 1/40 T. Choices in between are reasonable and lead to similar trends of the two prevailing pathways of double ionization. $\mathrm{t_{diff}=1/10}$ T is chosen for the results presented in this work. It is found that double delayed trajectories contribute no more than 15% for the smallest intensity even when the time difference is chosen small and equal to 1/40 T. If instead of the compensated energy the energy of each electron is used to identify the ionization time different results are obtained. Namely, one finds that at an intensity of 0.85$\times$10$^{14}$ W/cm$^2$ almost all trajectories are identified as double delayed. This was the conclusion in ref. [@Camus]. Using the actual energy to identify the ionization time at an intensity of 3$\times$10$^{14}$ W/cm$^2$ results in the direct pathway of double ionization still only contributing 20%. However, this is not a reasonable result. At 3$\times$10$^{14}$ W/cm$^2$ 3.17 U$_{p}$ is equal to 50 eV which is much higher than the second ionization energy of Ar. Moreover, the recollsion at this intensity is strong, which is discussed in [**2.4**]{}, and so the direct pathway of double ionization should contribute significantly. Thus, the compensated energy is employed to identify the ionization time in this work which leads to both the direct and delayed pathway being the main pathways of double ionization in agreement with ref. [@Huang] for the smallest intensity.
![Probability of direct (blue circles) and delayed (red triangles) pathways of double ionization as a function of the laser intensity for a delay time of 1/10 T (solid lines), 1/20 T (dashed lines) and 1/40T (dotted lines). The focal volume effect is not accounted for.[]{data-label="probpathways"}](Fig2_Diect_Delay_1o10_20_40){width="40.00000%"}
Distribution of the sum of the momenta
--------------------------------------
In [Fig. \[summomenta1\]]{}, the sum of the two electron momentum components along the polarization direction of the laser field are presented for intensities from 0.85$\times$10$^{14}$ W/cm$^{2}$ to 5$\times$10$^{14}$W/cm$^2$. In [Fig. \[summomenta1\]]{}, the contribution to the sum of the momenta of the direct and the delayed pathways of double ionization is also shown; the focal volume effect is not accounted for. As expected, it is found that the delayed pathway’s contribution is a distribution concentrated around zero while the direct pathway’s contribution is a doubly-peaked distribution. The direct pathway’s distribution of the sum of the momenta is the broadest one. Therefore, including only the delayed pathway of double ionization would result in a narrower distribution of the sum of the momenta than the observed one. Indeed, the 1D model described in ref.[@Bergues] which accounts only for the delayed pathway of double ionization results in a narrower distribution of the sum of the momenta than the observed one.
\[ht!\] ![Distribution of the sum of the two electron momentum components parallel to the polarization of the laser field (black solid lines) for laser field intensities from 0.85$\times$10$^{14}$ W/cm$^{2}$ to 5$\times$10$^{14}$W/cm$^2$. For each intensity, the distribution of the sum of the momenta of the delayed pathway (red dotted lines) and of the direct pathway (blue dashed lines) are also plotted. The focal volume effect is not accounted for.[]{data-label="summomenta1"}](Fig4_mom_1D_panels "fig:"){width="50.00000%"}
In [Fig. \[summomenta2\]]{}, the experimental results for the sum of the momenta in ref.[@Bergues] are compared with one set of computed results that account for the focal volume effect (black dashed lines) and one that does not (black solid lines). It is found that the computed results are in good agreement with the observed ones. Specifically, it is found that, for each intensity, the computed sum of the momenta extends over a range that is very similar to the experimental one. For instance, for an intensity of 0.85$\times$10$^{14}$ W/cm$^{2}$, the computed sum of the momenta extends over a range from roughly -2 a.u. to 2 a.u., while, for an intensity of 5$\times$10$^{14}$ W/cm$^{2}$, it extends from -4 a.u. to 4 a.u.; for both intensities these ranges are in agreement with the experimental results [@Bergues]. It is noted that a difference of the computed distributions of the sum of the momenta with the experimental ones is that the computed ones have smaller values around zero. This is more so the case for the computed results that account for the focal volume effect. This difference suggests that the current 3D model underestimates the contribution of the delayed pathway of double ionization.
\[ht!\] ![Distribution of the sum of the two electron momentum components parallel to the polarization of the laser field for laser field intensities from 0.85$\times$10$^{14}$ W/cm$^{2}$ to 5$\times$10$^{14}$W/cm$^2$. The computed results with the focal volume effect not accounted for are denoted by black solid lines and when it is accounted for by black dashed lines; the blue crosses denote the experimental results. []{data-label="summomenta2"}](Fig5_mom_1D_panels_Volume "fig:"){width="50.00000%"}
Asymmetry parameter
-------------------
The asymmetry parameter $$A(I,\phi)=\frac{N_{+}(I,\phi)-N_{-}(I,\phi)}{N_{+}(I,\phi)+N_{-}(I,\phi)}
\label{eqasym1}$$ is computed as a function of the intensity I and the CEP ($\phi$). $N_{+}(I,\phi)$ and $N_{-}(I,\phi)$ denote the percentage of double ionization events with ions escaping with positive and negative momentum, respectively, along the direction of the polarization of the laser field. Since in the 3D semiclassical model the nucleus is fixed, $N_{+}(I,\phi)$ and $N_{-}(I,\phi)$ correspond to the percentage of double ionization events where the sum of the two electrons momentum components along the direction of the laser field polarization are negative and positive, respectively. For each intensity, $A(I,\phi)$ is fitted with the sinusoidal function $$A(I,\phi)=A_{0}(I)\sin{(\phi+\phi_{0}(I))}.$$ The resulting asymmetry amplitude $A_{0}(I)$ and offset phase $\phi_{0}(I)$ are plotted in [Fig. \[A0as\]]{} (a) and (b), respectively, and compared with two sets of experimentally obtained asymmetry parameters [@Bergues]. The comparison shows that the 3D semiclassical model reproduces well the decreasing pattern of $A_{0}$ and the increasing pattern of $\phi_{0}$ with increasing intensity. However, the computed values for these asymmetry parameters are higher than the ones obtained from the experimental results. In addition, in [Fig. \[A0as\]]{} (a) and (b) the asymmetry parameters are plotted for each of the main two pathways of NSDI. It is shown that for both pathways the asymmetry parameter $\phi_{0}(I)$ has a similar pattern. The asymmetry parameter $A_{0}(I)$ for the delayed pathway is generally smaller which suggests that for the delayed pathway the two electron momentum components are more spread out in all four quadrants than for the direct ionization pathway. This is indeed shown in the correlated momenta presented in [Fig. \[correlated\]]{} in the next section.
\[ht!\] ![Asymmetry parameters $A_{0}$ (a) and $\phi_{0}$ (b) as a function of intensity. The computed results for all double ionization events when the focal volume effect is not accounted for are denoted by black solid lines with circles and when it is accounted for by black dashed lines with triangles. The delayed pathway of double ionization is denoted by red solid lines with triangles and the direct pathway by blue solid lines with circles. For the direct and delayed pathways of double ionization the focal volume effect is not accounted for. Experimental results are denoted by light blue crosses and dark blue squares. []{data-label="A0as"}](Fig6_A0_phi0_Volume "fig:"){width="80.00000%"}
Correlated momenta as a function of intensity: transition from strong to soft recollisions
------------------------------------------------------------------------------------------
\[ht!\] ![First row: measured correlated momenta. Second row: computed correlated momenta for all double ionization events with the focal volume effect accounted for. Third row: correlated momenta for the direct pathway of double ionization with the focal volume effect not accounted for. Fourth row: correlated momenta for the delayed pathway of double ionization with the volume effect not accounted for. Fifth row: correlated momenta for the double delayed pathway of double ionization with the volume effect not accounted for.[]{data-label="correlated"}](Fig7_mom_2D_panels "fig:"){width="90.00000%"}
\[ht!\] ![Correlated momenta at an intensity 0.85$\times$10$^{14}$ W/cm$^{2}$ for the delayed pathway of double ionization for the case when the electron that ionizes second is electron 2 (a) and electron 1 (b).[]{data-label="correlated3"}](mom_Delay1_Delay2_1o10_E049 "fig:"){width="45.00000%"}
For intensities ranging from 0.85$\times$10$^{14}$ W/cm$^{2}$ to 5$\times$10$^{14}$W/cm$^2$, the correlated momenta for the direct and delayed double ionization pathways are plotted in [Fig. \[correlated\]]{}. For the smaller intensity of 0.85$\times$10$^{14}$ W/cm$^{2}$ the computed correlated momenta resemble but do not quite have the cross-shape pattern of the measured results, see [Fig. \[correlated\]]{}. There are fewer double ionization events with both momenta being close to zero than in the experimentally obtained cross-shaped pattern [@Kling1; @Bergues]. These close-to-zero momenta also reduce the asymmetry parameter. Thus, the reduced number of such events produced in our calculation is in accord with computing an asymmetry parameter $A_{0}(I)$ that is larger than the experimental one, see [Fig. \[A0as\]]{}. The cross-shaped pattern is better reproduced by the delayed double ionization pathway, see [Fig. \[correlated\]]{}. In addition, from an analysis of the classical trajectories it is found that for 63% of the delayed pathway of double ionization the initially bound electron ionizes after the tunneling electron following recollision while for 37% it is the other way around. The electron that ionizes second does so generally with a smaller momentum. It is found that the correlated momenta when the tunneling electron ionizes second resembles more a cross-shaped pattern, see [Fig. \[correlated3\]]{}. Moreover, as in the observed correlated momenta in ref.[@Bergues], the computed correlated momenta transition to the well-known pattern for intermediate intensities [@Weber]. This pattern involves both electrons escaping in the same direction either parallel or antiparallel to the laser field, thus, giving rise to a much higher probability density in the first and third quadrants of the correlated momenta, rather than the second and fourth ones. This is the pattern of the correlated momenta for the direct pathway of double ionization at intensities 2-4$\times$10$^{14}$ W/cm$^{2}$. At these intensities the direct pathway of double ionization is the prevailing one.
\[ht!\] ![Distribution of the tunneling time of electron 1 (blue line) and of the recollision time (red line) for intensities 10$^{14}$ W/cm$^{2}$, 3$\times$10$^{14}$ W/cm$^{2}$ and 5$\times$10$^{14}$ W/cm$^{2}$ and for two different CEPs, $\phi=15^{\circ}$ and $\phi=105^{\circ}$, for each intensity. Similar results hold for all other CEPs. Grey line denotes the laser field. The light blue arrows indicate the mapping of a tunneling-time peak to a recollision-time peak.[]{data-label="colltimes"}](Fig8_collision_t0 "fig:"){width="80.00000%"}
A less known pattern is that observed experimentally and retrieved computationally with the 3D semiclassical model for intensities above 4$\times$10$^{14}$ W/cm$^{2}$, see [Fig. \[correlated\]]{}. For these higher intensities, it is found that the two electrons escape mostly with opposite momenta for a significant number of double ionization events. To identify the reason for this shift, in [Fig. \[colltimes\]]{}, the time electron 1 tunnel-ionizes and the recollision time are plotted for three different intensities, namely, 10$^{14}$ W/cm$^{2}$, 3$\times$10$^{14}$ W/cm$^{2}$ and 5$\times$10$^{14}$ W/cm$^{2}$ and for two different CEP’s, namely, $\phi=15^{\circ}$ and $\phi=105^{\circ}$ for each intensity. The tunneling time of electron 1 is found to be close to the times corresponding to the extrema of the laser field for all three intensities. However, the distribution of the recollision time is found to shift from times corresponding roughly to zeros of the laser field for an intensity of 10$^{14}$ W/cm$^{2}$ to times corresponding to the extrema of the laser field for an intensity of 5$\times$10$^{14}$ W/cm$^{2}$. The transfer of energy from electron 1 to electron 2 is much smaller for the soft recollisions. For these higher intensities, where soft recollisions prevail, the momentum of electron 1 is mostly determined from the vector potential at the tunneling time. The momentum of electron 2 is determined by the vector potential shortly after recollision takes place which is roughly half a laser cycle after electron 1 tunnel-ionizes. As a result the two electrons escape mostly with opposite momenta. This mechanism of soft recollisions for higher intensities was first identified in a theoretical study of strongly-driven N$_{2}$ with fixed nuclei [@Agapi4].
The contribution to the correlated momenta of the direct and delayed pathways of double ionization is shown in [Fig. \[correlated\]]{}. For the direct pathway, for intensities from 0.85$\times$10$^{14}$ W/cm$^{2}$ to 4$\times$10$^{14}$W/cm$^2$ the probability density is significantly higher in the first and third quadrants of the correlated momenta than in the second and fourth ones. This is a known pattern. For strong recollisions, the two electron momentum components along the direction of the laser field are both determined from the vector potential at times just larger than the recollision time. Thus, both electrons escape with similar momenta in the direction along the polarization of the laser field. For an intensity of 5$\times$10$^{14}$ W/cm$^{2}$ soft recollisions prevail and, as discussed above, the two electrons escape mostly with opposite momenta along the direction of the polarization of the laser field. For the delayed pathway, this opposite momenta pattern, demonstrated with much higher probability density in the second and fourth quadrants of the correlated momenta, sets in at lower intensities, 3$\times$10$^{14}$ W/cm$^{2}$, see [Fig. \[correlated\]]{}. This is consistent with a smaller transfer of energy taking place from electron 1 to electron 2 at the recollision time in the delayed pathway compared to the energy transfer in the direct pathway. Regarding small intensities, in [Fig. \[correlated\]]{} it is shown that the computed correlated momenta due to the delayed pathway have a pattern similar to but not quite cross-shaped.
Correlated momenta and double ionization pathways as a function of CEP
----------------------------------------------------------------------
In what follows, the dependence of the correlated momenta on the CEP is investigated at an intensity of 0.85$\times$10$^{14}$ W/cm$^{2}$. In [Fig. \[CEPcor\]]{}, the correlated momenta are plotted for $\phi$ ranging from $15^{\circ}$ to $345^{\circ}$ with a step of $30^{\circ}$. Due to the symmetry of the Hamiltonian, when $\phi \rightarrow \phi+180^{\circ}$ then ${\bf p} \rightarrow {\bf -p}$. With this symmetry in mind, the experimental results from ref.[@Kling3] are plotted in [Fig. \[CEPcor\]]{} for $\phi$ ranging from $15^{\circ}$ to $165^{\circ}$ and the computed results are plotted for $\phi$ ranging from $195^{\circ}$ to $345^{\circ}$. A very good agreement is found between the computed and the experimental results. Specifically, the computed correlated momenta correctly reproduce the overall observed pattern for each individual CEP. A difference between the computed and the experimental results is that the former results are more concentrated on the first and third quadrants suggesting that the computations overestimate the contribution of the direct pathway. To better understand the change of the correlated momenta pattern as a function of the CEP plotted in [Fig. \[CEPcor\]]{}, in [Fig. \[CEPprob\]]{} the percentage of the direct and delayed pathways of double ionization events are plotted as a function of the CEP. At an intensity of 0.85$\times$10$^{14}$ W/cm$^{2}$ the delayed pathway has the largest contribution for CEPs $\phi=45^{\circ}$ and $\phi=225^{\circ}$, while it has the smallest for CEPs $\phi=165^{\circ}$ and $\phi=345^{\circ}$. In [Fig. \[CEPcor\]]{}, a comparison of the correlated momenta between $\phi=45^{\circ}$ and $\phi=165^{\circ}$ shows that there is higher probability density for both electrons to ionize with the same large momentum for $\phi=165^{\circ}$ than for $\phi=45^{\circ}$. This is indeed consistent with the direct ionization pathway having a larger contribution for $\phi=165^{\circ}$ than for $\phi=45^{\circ}$ as shown in [Fig. \[CEPprob\]]{}. Similar conclusions can be drawn by comparing the correlated momenta for $\phi=345^{\circ}$ with the one for $\phi=225^{\circ}$. From [Fig. \[CEPprob\]]{}, it is found that the contribution of each of the two main pathways of double ionization vary roughly by 20% as a function of the CEP for the smallest intensity of 0.85$\times$10$^{14}$ W/cm$^{2}$.
\[ht!\] ![Correlated momenta at an intensity of 0.85$\times$10$^{14}$ W/cm$^{2}$ for $\phi$ ranging from $15^{\circ}$ to $345^{\circ}$ with a step of $30^{\circ}$. For $\phi$ ranging from $15^{\circ}$ to $165^{\circ}$ the experimental results from ref.[@Kling3] are displayed. For $\phi$ ranging from $195^{\circ}$ to $345^{\circ}$ the computed results are displayed. The average over all $\phi$s using the computed results is the correlated momentum plot in the middle. The dark circle in each plot denotes the origin of the axes.[]{data-label="CEPcor"}](fig_exp_theory_3 "fig:"){width="70.00000%"}
\[ht!\] ![% of the direct and the delayed pathways of double ionization events as a function of the CEP for an intensity of 0.85$\times$10$^{14}$ W/cm$^{2}$. []{data-label="CEPprob"}](Fig10_prob_1o10 "fig:"){width="40.00000%"}
Conclusions
===========
Using a 3D semiclassical model we investigated the dependence of double ionization observables on the intensity and on the carrier envelope phase of a near-single-cycle near-infrared laser field employed to drive Ar. The good agreement of the computed results with recent experiments employing near-single-cycle laser pulses [@Kling1; @Kling3; @Bergues], adds to previous successes of this 3D model in identifying features of non-sequential double ionization of two-electron atoms when driven by many-cycle laser pulses [@Agapi1; @Agapi2; @Emmanouilidou1]. A difference between the computed and the experimental results was found to be a lower value of the distribution of the sum of the two electron momentum components along the direction of the polarization of the laser field and a lower value around zero of the correlated momenta. This seems to suggest that the current 3D model overestimates the Coulomb attraction of each electron from the nucleus. Future studies can improve on the 3D model for many electron atoms such as Ar by using more accurate effective potentials for the time propagation. Moreover, it was demonstrated that the main pathways of double ionization, that is, the direct and the delayed pathways, both significantly contribute at all intensities currently under consideration. Furthermore, the prevalence of the direct versus the delayed pathway was investigated as a function of the CEP for an intensity of 0.85$\times$10$^{14}$ W/cm$^{2}$ and it was shown that the results obtained are consistent with features of the observed correlated momenta [@Kling3]. Finally, a previously-predicted in the context of a strongly-driven fixed-nuclei N$_{2}$ unexpected anti-correlation momentum pattern at higher intensities [@Agapi4], is observed experimentally in the context of strongly-driven Ar [@Bergues] and also reproduced in the current work for strongly-driven Ar by a near-single-cycle laser field. It is shown that this anti-correlation pattern is due to soft recollisions with recollision times close to the extrema of the laser field.
[10]{} Corkum, P. B. Plasma perspective on strong field multiphoton ionization. [*Phys. Rev. Lett.*]{} [**71**]{}, 1994 (1993). Taylor, K. T., Parker, J. S., Dundas, D. & Meharg, K. J. Theory of laser-driven double-ionization of atoms at Ti:sapphire laser wavelengths. [*J. Mod. Opt*]{}. [**54**]{}, 1959 (2007). Becker, A., D[" o]{}rner, R. & Moshammer, R. Multiple fragmentation of atoms in femtosecond laser pulses. [*J. Phys. B: At. Mol. Opt. Phys.*]{} [**38**]{}, S753 (2006). Ullrich, J. [*et al*]{}. Recoil-ion and electron momentum spectroscopy: reaction-microscopes. [*Rep. Prog. Phys.*]{} [**66**]{}, 1463 (2003). N. Camus, [*et al*]{}. Attosecond correlated dynamics of two electrons passing through a transition state. [*Phys. Rev. Lett*]{}. [**108**]{}, 073003 (2012). Bergues, B. [*et al*]{}. Attosecond tracing of correlated electron-emission in non-sequential double ionization. [*Nat. Commun.*]{} [**3**]{}, 813 (2012). Johnson, N. G. [*et al*]{}. Single-shot carrier-envelope-phase-tagged ion-momentum imaging of nonsequential double ionization of argon in intense 4-fs laser fields. [*Phys. Rev. A*]{} [**83**]{}, 013412 (2011). K[" u]{}bel, M. [*et al*]{}. Complete characterization of single-cycle double ionization of argon from the nonsequential to the sequential ionization regime. [*Phys. Rev. A*]{} [**93**]{}, 053422 (2016). Kopold, R., Becker, W., Rottke, H. & Sandner, W. Routes to nonsequential double ionization. [*Phys. Rev. Lett.*]{} [**85**]{}, 3781 (2000). Feuerstein, B. [*et al*]{}. Separation of recollision mechanisms in nonsequential strong field double ionization of Ar: the role of excitation tunneling. [*Phys. Rev. Lett.*]{} [**87**]{}, 043003 (2001). Emmanouilidou, A. Prevalence of different double ionization pathways and traces of three-body interactions in strongly driven helium. [*Phys. Rev. A* ]{} [**83**]{}, 023403 (2011). Maxwell, A. S. & Figueira de Morisson Faria, C. Controlling below-threshold nonsequential double ionization via quantum interference. [*Phys. Rev. Lett.*]{} [**116**]{}, 143001 (2016). Kustaanheimo, P. & Stiefel, E. Perturbation theory of Kepler motion based on spinor regularization. [*J. Reine Angew. Math.*]{} [**218**]{}, 204 (1965). Huang, C., Zhou, Y., Zhang, Q., & Lu, P. Contribution of recollision ionization to the cross-shaped structure in nonsequential double ionization. [*Optics Express*]{} [**21**]{}, 11390 (2013). Emmanouilidou, A. Recoil collisions as a portal to field-assisted ionization at near-uv frequencies in the strong-field double ionization of helium. [*Phys. Rev. A*]{} [**78**]{}, 023411 (2008). Parker, J. S. [*et al*]{}. High-energy cutoff in the spectrum of strong-field nonsequential double ionization. [*Phys. Rev. Lett.*]{} [**96**]{}, 133001 (2006). Staudte, A. [*et al*]{}. Binary and recoil collisions in strong field double ionization of helium. [*Phys. Rev. Lett.*]{} [**99**]{}, 263002 (2007). Rudenko, A. [*et al*]{}. Correlated two-electron momentum spectra for strong-field nonsequential double ionization of He at 800 nm. [*Phys. Rev. Lett.*]{} [**99**]{}, 263003 (2007). Emmanouilidou, A., Parker, J. S., Moore, L. R. & Taylor, K. Direct versus delayed pathways in strong-field non-sequential double ionization. [*New J. Phys.*]{} [**13**]{}, 043001 (2011). Emmanouilidou, A. & Staudte, A. Intensity dependence of strong-field double-ionization mechanisms: From field-assisted recollision ionization to recollision-assisted field ionization. [*Phys. Rev. A*]{} [**80**]{}, 053415 (2009).
Landau, L. D. & Lifshitz, E. M. [*Quantum Mechanics*]{} (Pergamon Press, New York, 1977).
Delone, N. B. & Krainov, V. P. Energy and angular electron spectra for the tunnel ionization of atoms by strong low-frequency radiation. [*J. Opt. Soc. Am. B*]{} [**8**]{}, 1207 (1991). Abrines, R. & Percival, I. C. Classical theory of charge transfer and ionization of hydrogen atoms by protons. [*Proc. Phys. Soc.*]{} [**88**]{}, 861 (1966).
Wang, P., Sayler, A. M., Carnes, K. D., Esry, B. D. & Ben-Itzhak, I. Disentangling the volume effect through intensity-difference spectra: application to laser-induced dissociation of H$_2^+$. [*Opt. Lett.*]{} [**30**]{}, 664 (2005).
Leopold, J. G., & Percival, I. C. Ionisation of highly excited atoms by electric fields. III. Microwave ionisation and excitation. [*J. Phys. B*]{} [**12**]{}, 709 (1979).
Weber, Th. [*et al*]{}. Correlated electron emission in multiphoton double ionization. [*Nature*]{} [**405**]{}, 658 (2000). Bergues, B., K[" u]{}bel, M., Kling, N. G., Burger, C. & Kling, M. F. Single-cycle non-sequential double ionization. [*IEEE J. Sel. Top. Quantum Electron.*]{}, [**21**]{}, 8701009, (2015).
Acknowledgements {#acknowledgements .unnumbered}
================
A.E. acknowledges the EPSRC grant no. J0171831 and the use of the computational resources of Legion at UCL. M.K., M.F.K., and B.B. acknowledge support from the Max Planck Society and the DFG cluster of excellence “Munich Centre for Advanced Photonics (MAP)".
Author contributions statement {#author-contributions-statement .unnumbered}
==============================
A.C. performed the analysis of the computations involved and contributed to the ideas involved. A.E. provided the codes used for the computations and the analysis and was responsible for the main ideas involved in the theoretical analysis. M.K., M.F.K., and B.B provided the experimental results and contributed to the discussion in the paper concerning the comparison of the computations with the experimental results. All authors reviewed the manuscript.
Additional information {#additional-information .unnumbered}
======================
**Competing financial interests:** The authors declare no competing financial interests.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We numerically simulate the low-energy properties of interacting electrons in a random potential using the Hartree-Fock based exact diagonalization method. In particular, we investigate how the transport properties are influenced by the combined effects of disorder and correlations in the presence of the electron spin. To this end we calculate the participation number of many-particle states in Fock space, the return probability of single-particle excitations, and the Kubo-Greenwood conductance. It turns out that in the strongly localized regime interactions increase the conductance whereas for weak disorder interactions decrease the conductance. In contrast, single-particle excitations in general experience a localizing influence of the interactions.'
address: 'Institut für Physik, Technische Universität Chemnitz, D-09107 Chemnitz, Germany'
author:
- Frank Epperlein
- Svetlana Kilina
- Michael Schreiber
- Sergey Uldanov
- Thomas Vojta
title: 'Fock space localization, return probability, and conductance of disordered interacting electrons'
---
,
,
,
, and
electronic transport, disorder, localization, metal-insulator transitions
Introduction
============
The interplay between Anderson localization and electron-electron interactions belongs to the most important unsolved problems of today’s condensed matter physics. The standard theoretical approach to this problem is based on perturbation theory in both disorder and interactions. In the metallic regime, the interactions lead to quantum corrections to the conductivity in addition to the weak-localization corrections [@AA79]. A scaling theory [@Finkelstein83; @BK94] was developed on the basis of these perturbative results. It predicts that a phase transition between an insulator and a [*normal*]{} metal, i.e., a disordered Fermi liquid, is possible only for dimensions $d>2$. In $d=2$ the situation is inconclusive. Under renormalization the disorder flows to zero but the interaction strength diverges, i.e., the perturbation theory becomes uncontrolled.
The attention to these questions resurged in the mid nineties, mainly due to two surprising developments. First, Shepelyansky suggested [@TIP] that in the localized regime two interacting particles can form a pair whose localization length is much larger than that of a single particle. Second, measurements [@2DMIT] on high-quality Si-MOSFETs revealed indications of a true metal-insulator transition (MIT) in two dimensions. Since these experiments are carried out at low electron density where the Coulomb interaction is particularly strong compared to the Fermi energy, interaction effects are a likely reason for this phenomenon. A complete understanding has, however, not yet been obtained. In a number of theories new quantum states of matter have been postulated, either non-Fermi liquid metals or unusual superconductors [@runaway; @newquantumstates; @SIT]. However, other explanations based on more conventional physics like temperature-dependent disorder [@AM] or impurity screening [@screening], imply that the seeming MIT is a transient phenomenon, and the true ground state is always an insulator.
We have recently developed [@Dointer; @HFD] an efficient method to investigate disordered interacting electrons numerically, the Hartree-Fock (HF) based exact diagonalization (HFD). It is related to the quantum-chemical configuration interaction approach. We have used this method to study the influence of interactions on the conductance of spinless fermions in one, two, and three dimensions. We found a delocalizing tendency of the interactions for strong disorder but a localizing one for weak disorder. In this work we extend our HFD investigations to electrons with spin. In particular, we calculate the participation number of the many-particle eigenstates in Fock space, the localization properties of single-particle excitations, and the Kubo-Greenwood conductance.
Model and method {#sec:II}
================
The generic model for [*spinless*]{} interacting disordered electrons is the quantum Coulomb glass [@qcg; @epperhf]. In this paper we use a generalization of the quantum Coulomb glass to electrons with spin. It is defined on a regular square lattice with $\mathcal{N}=L^2$ sites occupied by $N=N_\uparrow +N_\downarrow=2K \mathcal{N}$ electrons ($0\!<\!K\!<\!1$). To ensure charge neutrality each lattice site carries a compensating positive charge of $2Ke$. The Hamiltonian is given by $$\begin{aligned}
H &=& -t \sum_{\langle ij\rangle, \sigma} (c_{i\sigma}^\dagger c_{j\sigma}
+ h.c.)+
\sum_{i,\sigma} \varphi_i n_{i\sigma} \\
&&+~\frac{1}{2}\sum_{i\not=j,
\sigma,\sigma'}U_{ij}~(n_{i\sigma}-K)(n_{j\sigma'}-K)\nonumber\\
&&+~ U_H \sum_{i} n_{i\uparrow} n_{i\downarrow}\nonumber
\label{eq:Hamiltonian}\end{aligned}$$ where $c_{i\sigma}^\dagger$, $c_{i\sigma}$, and $n_{i\sigma}$ are the creation, annihilation and occupation number operators at site $i$ and spin $\sigma$. $\langle ij \rangle$ denotes all pairs of nearest-neighbor sites. $t$ is the strength of the hopping term, i.e., the kinetic energy. We parametrize the interaction $U_{ij} = e^2/r_{ij}$ by its value $U$ between nearest-neighbor sites and include a Hubbard interaction $U_H$. The random potential values $\varphi_i$ are chosen independently from a box distribution of width $2
W_0$ and zero mean. The boundary conditions are periodic and the Coulomb interaction is treated in the minimum image convention (which implies a cut-off at a distance of $L/2$).
A numerically exact solution of a quantum many-particle system requires the diagonalization of a matrix whose dimension increases exponentially with system size. This severely limits the possible sizes. To overcome this problem we have developed the HFD method. The basic idea is to work in a truncated Hilbert space consisting of the HF ground state and the low-lying excited Slater states. For each disorder configuration three steps are performed: (i) find the HF solution of the problem, (ii) determine the $B$ Slater states with the lowest energies, and (iii) calculate and diagonalize the Hamilton matrix in the subspace spanned by these states. The number $B$ of basis states determines the quality of the approximation, reasonable values have to be found empirically.
Results {#sec:III}
=======
To characterize, to what extent the interactions introduce non-trivial correlations (beyond the HF level) into the system we calculate the participation number of the many-particle ground state with respect to the HF Slater determinants. The results for a system of $4 \times 4$ sites at half filling are shown in Fig. \[Fig:fockpn\].
For $U=0$ the Fock space participation number is 1 because the HF approximation is (trivially) exact for non-interacting electrons. Increasing the interaction mixes the single-particle states, and the inverse Fock space participation number is reduced. Nontrivial interaction effects, if any, become manifest mostly in this region. For even larger interactions the eigenstates become again more localized in Fock space since the HF approximation becomes exact again for $U \to \infty$.
Analogous information can be extracted from the single-particle survival probability $$Z(\omega) = \frac 1 N \sum_{ij} \lim_{\delta \to 0} \frac \delta \pi \, G_{ij}^R(\omega + i \delta)
\, G_{ji}^A(\omega - i \delta),$$ which averages the sum of the transmission probabilities of a single-particle excitation to all sites. Here $G_{ij}^{R,A}(\omega)$ are the retarded and advanced single-particle Greens functions. The survival probability at the Fermi energy, shown in Fig. \[Fig:quasi\], is related to the square of the quasiparticle weight.
We now turn to the single-particle localization properties which we characterize by the return probability of the single-particle excitations, $$R_p(\omega) = \frac 1 N \sum_{i} \lim_{\delta \to 0} \frac \delta \pi \, G_{ii}^R(\omega + i \delta)
\, G_{ii}^A(\omega - i \delta).$$ However, in this quantity localization and decay of the quasiparticles are entangled. To extract the localization properties we normalize $R_p$ by the survival probability $Z$. The data in Fig. \[Fig:ret\_norm\] show
that, in general, interactions tend to localize the [*single-particle*]{} excitations. This is in agreement with earlier HF results [@epperhf] which showed that the Coulomb gap in the [*single-particle*]{} density of states is responsible for the trend towards localization. In Fig. \[Fig:ret\_norm\] a tiny delocalization seems to occur at low interaction strength and small kinetic energy. However, the changes are within the statistical errors of our study.
In a real transport experiment the most accessible observable is the conductance. Theoretically, it can be obtained from linear-response theory. It is essentially determined by the current-current correlation function of the ground state. The real (dissipative) part of the conductance (in units of $e^2/h$) at frequency $\omega$ is given by the Kubo-Greenwood formula [@K57], $${\rm Re} ~ G^{xx}(\omega) = \frac {2 \pi^2} {\omega} \sum_{\nu} |\langle 0 | j^x|\nu \rangle |^2
\delta(\omega+E_0-E_{\nu}).
\label{eq:kubo}$$ $j^x$ is the $x$ component of the current operator and $|\nu\rangle$ is an eigenstate of the Hamiltonian. Eq. (\[eq:kubo\]) describes an isolated system while in a real d.c. transport experiment the sample is connected to contacts and leads. This results in a finite life time $\tau$ of the eigenstates leading to an inhomogeneous broadening $\gamma = \tau^{-1}$ of the $\delta$ functions in (\[eq:kubo\]) [@datta]. To suppress the discreteness of the spectrum of a finite system, $\gamma$ should be at least of the order of the single-particle level spacing. Here, this requires comparatively large $\gamma \ge 0.05$. We tested different $\gamma$ and found that the conductance [*values*]{} depend on $\gamma$ but the qualitative results do not.
The typical d.c. conductance is shown in Fig. \[Fig:conductance\]. Since the logarithm of the conductance rather than the conductance itself is a self-averaging quantity in a disordered system we calculate the typical conductance by averaging the logarithms of the conductance values of 1000 disorder configurations.
We find that, as in the spinless case, there are two qualitatively different regimes. For large kinetic energy $t$ the interactions always reduce the d.c. conductance, while for small kinetic energy, i.e. in the localized regime, small and moderate interactions significantly enhance the d.c. conductance. For larger interaction strength the conductance drops, indicating the crossover to a Wigner crystal or Wigner glass.
A comparison with analogous results for spinless fermions [@Dointer] shows that the delocalizing influence of the interactions is significantly stronger for electrons with spin.
This work was supported in part by the DFG under Grant No. SFB393/C2 and by the NSF under Grant No. DMR-98-70597.
[99]{} See, e.g., B. L. Altshuler and A. G. Aronov, in A. L. Efros and M. Pollak (Eds.) [*Electron-electron interactions in disordered systems*]{}, North-Holland, Amsterdam (1985). A. M. Finkelstein, Zh. Eksp. Teor. Fiz. [**84**]{}, 168 (1983) \[Sov. Phys. JETP [**57**]{}, 97 (1983)\]. D. Belitz and T. R. Kirkpatrick, Rev. Mod. Phys. [**66**]{}, 261 (1994). D. L. Shepelyansky, Phys. Rev. Lett. [**73**]{}, 2607 (1994). S. V. Kravchenko et al., Phys. Rev. Lett. [**77**]{}, 4938 (1996). C. Castellani, C. DiCastro and P. A. Lee, Phys. Rev. B [**57**]{}, R9381 (1998). Q. Si and C. M. Varma, Phys. Rev. Lett. [**81**]{}, 4951 (1998); S. Chakravarty et al., Phil. Mag. B [**79**]{}, 859 (1999); G. Benenti, X. Waintal, and J.-L. Pichard, Phys. Rev. Lett. [**83**]{}, 1826 (1999). F. C. Zhang and T. M. Rice, cond-mat/9708050; D. Belitz and T. R. Kirkpatrick, Phys. Rev. B [**58**]{}, 8214 (1998); P. Phillips et al., Nature [**395**]{}, 253 (1998). B. L. Altshuler and D. L. Maslov, Phys. Rev. Lett. [**82**]{} 145 (1999). A. Gold and V. T. Dolgopolov, Phys. Rev. B [**33**]{}, 1076 (1986); S. Das Sarma and E. H. Hwang, Phys. Rev. Lett. [**83**]{}, 164 (1999). T. Vojta, F. Epperlein, and M. Schreiber, Phys. Rev. Lett. [**81**]{}, 4212 (1998). T. Vojta, F. Epperlein, and M. Schreiber, Computer Phys. Commun. [**121–122**]{}, 489 (1999). A. L. Efros and F. G. Pikus, Solid State Commun. [**96**]{}, 183 (1995); J. Talamantes, M. Pollak, and L. Elam, Europhys. Lett. [**35**]{}, 511 (1996). F. Epperlein, M. Schreiber, and T. Vojta, Phys. Rev. B [**56**]{}, 5890 (1997). K. Kubo, J. Phys. Soc. Jpn. [**12**]{}, 576 (1957); D. A. Greenwood, Proc. Phys. Soc. [**71**]{}, 585 (1958). See, e.g., S. Datta, [*Electronic transport in mesoscopic systems*]{}, Cambridge University Press, Cambridge (1997).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The continuous-variable quantum key distribution with entanglement in the middle, a semi-device-independent protocol, places the source in the untrusted third party between Alice and Bob, and thus has the advantage of high levels of security with the purpose of eliminating the assumptions about the source device. However, previous works considered the collective-attack analysis, which inevitably assumes that the states of the source has an identical and independently distributed (i.i.d) structure, and limits the application of the protocol. To solve this problem, we modify the original protocol by exploiting an energy test to monitor the potential high energy attacks an adversary may use. Our analysis removes the assumptions of the light source and the modified protocol can therefore be called source-device-independent protocol. Moreover, we analyze the security of the continuous-variable source-device-independent quantum key distribution protocol with a homodyne-homodyne structure against general coherent attacks by adapting a state-independent entropic uncertainty relation. The simulation results indicate that, in the universal composable security framework, the protocol can still achieve high key rates against coherent attacks under the condition of achievable block lengths.'
author:
- 'Yichen Zhang$^{1}$'
- 'Ziyang Chen$^{2}$'
- Christian Weedbrook$^3$
- 'Song Yu$^{1}$'
- 'Hong Guo$^{2}$'
title: |
Continuous-variable source-device-independent quantum key distribution\
against general attacks
---
[^1]
Introduction
============
Quantum key distribution (QKD) [@RevModPhys.74.145; @RevModPhys.81.1301], as one of the most practical quantum cryptography technology, allows two users (traditionally called Alice and Bob) to establish a set of secret keys exploiting both quantum mechanics and classical post-processing methods. This can provide information-theoretic security even against existing potential eavesdroppers.
Continuous-variable (CV) QKD protocol [@RevModPhys.84.621; @Entropy.17.6072], of which the characteristic is that the information is encoded on the quadratures of the light field and measured with coherent measurement methods, e.g., homodyne [@Phys.Rev.Lett.88.057902.2002] and heterodyne detection [@Phys.Rev.Lett.93.170504.2004], has developed rapidly. There are two main reasons resulting in CV-QKD attracting so much attention in recent years: it can be easily implemented with standard telecom components and compatible with wavelength division multiplexing [@New.J.Phys.12.103042.2010; @New.J.Phys.17.043027.2015], and it can achieve high key rate in metropolitan distance [@field_test], which has advantages of short-range implementation.
There are plenty of CV-QKD protocols proposed to deal with different scenarios. In the case of fully trusted-device protocols, it is always assumed that both Alice and Bob are honest, and Eve can only control the quantum channels rather than the devices at the two parties. A large number of distinctive trusted-device protocols, including discrete modulation CV-QKD protocols [@Phys.Rev.Lett.102.180504.2009; @Phys.Rev.A.83.042312.2011; @arXiv.1805.04249], two-way protocols [@Nat.Phys.4.726.2008; @Int.J.Quantum.Inf.10.1250059.2012; @J.Phys.B.47.035501.2014; @Phys.Rev.A.92.062323.2015; @Sci.Rep.6.22225.2016; @J.Phys.B.At.Mol.Opt.Phys.50.035501.2017] and so forth, have been put forward to enrich the protocol design. However, because of the imperfection of the practical source and detection devices, a QKD system may be attacked by an potential eavesdropper, and it compromise the security of a protocol [@New.J.Phys.20.103016.2018]. To eliminate all the loopholes of devices, fully device-independent protocols are proposed, which do not make any assumptions for all experimental devices and allows Eve to control them all . Nevertheless, those protocols face experimental challenges because they have to perform a detection-loophole-free Bell test [@Phys.Rev.Lett.120.040406.2018].
As a compromise, semi-device-independent (semi-DI) protocols are proposed, such as measurement-device-independent (MDI) [@Phys.RevA.89.052301; @Phys.RevA.90.052325; @Nat.Photon.9.397.2015] and one-sided device-independent (1sDI) [@Nat.Commun.6.8795.2015; @Optica.3.634.2016] QKD protocols, to give a trade off between the security of some devices and the performance of a protocol, which regard that part of the protocol is honest and the other part is untrusted. Remarkably, both CV-MDI [@LS.JW4A.33; @Phys.Rev.A.97.052327.2018; @Phys.Rev.A.98.012314.2018] and CV-1sDI protocols [@Nat.Commun.6.8795.2015; @Phys.Rev.Lett.109.100502.2012; @Phys.Rev.A.90.042325.2014] have been analyzed against general coherent attacks, which improves the security analysis of the protocols.
CV-QKD with entanglement in the middle [@Phys.Rev.A.87.022308.2013] is the protocol of which the source is placed with the untrusted third party in the middle and controlled by the malicious eavesdropper. Alice and Bob then measure one of the modes they received separately, with either homodyne or heterodyne detection. The goal of the protocol is that we do not need to give assumptions on the source, which is sometimes ill-characterised and unsafe in communication. Nevertheless, the security analysis of the CV-QKD with entanglement in the middle protocol is only confined to the collective attack cases, which inevitably assumes that the states of the source has identical and independently distributed (i.i.d) structure, i.e., ${\rho _{{A^n}{B^n}}} = \sigma _{AB}^{ \otimes n}$, leading to the protocol not being able to reach the original idea of source-device-independent (SDI).
Inspired by the security analysis technique used in the 1sDI protocol by F. Furrer *et al.* [@Phys.Rev.Lett.109.100502.2012; @Phys.Rev.A.90.042325.2014], we adapt one type of state-independent entropic uncertainty relation with CVs to analyse the security of the CV-QKD with entanglement in the middle protocol under coherent attacks and only consider the case that both Alice and Bob perform homodyne detections. We modify the original protocol by exploiting an energy test at the reconciliation side (Bob’s side for reverse reconciliation as an example) to monitor the potential high energy attacks an adversary may use. By properly quantifying the correlation between Alice’s and Bob’s data, which could be used for estimating Eve’s knowledge of the raw key, we obtain the secret key rate of a finite number of exchanged signals supposing that the strategy Eve exploiting is a coherent attack. Our analysis removes the assumptions of the light source and assumes that the sampling process performed in Alice’s and Bob’s sides are i.i.d, which is needed for exploiting the entropic uncertainty relation. Therefore, The modified protocol can be called CV-SDI QKD protocol. Finally, simulation shows that even when the coherent attack is considered, CV-QKD with entanglement in the middle can still reach a non-zero key rate over short distance, without giving any constrains of the source.
The paper is organized as follows. In Sec. \[Composable\_security\_definition\], we give a short description on the definition of composable security in QKD. In Sec. \[The\_protocol\_and\_the\_security\_against\_collective\_attacks\], we give an introduction of the original CV-QKD protocol with entanglement in the middle under collective attacks. In Sec. \[The\_security\_against\_general\_coherent\_attacks\], we exploit a version of state-independent entropic uncertainty relation to analyse the security of the modified CV-SDI QKD protocol against coherent attacks. Then, in Sec. \[Simulation\_and\_discussion\], we present the simulation results of the secret key rates in both the collective and coherent attack cases. Finally, we summarize the paper in Sec. \[Conclusion\].
Composable security definition {#Composable_security_definition}
==============================
A detailed discussion of the composability in quantum cryptography can be seen in Ref. [@New.J.Phys.11.085006.2009]. For better understanding of the security criteria in QKD, we first need to know what is perfectly secure for a QKD protocol. A perfectly secure QKD protocol should satisfy three criteria: correctness, secrecy and robustness [@New.J.Phys.11.085006.2009].
We denote $S_{A}$ is the key in Alice’s side and $S_{B}$ is the key in Bob’s side. Correctness requires that every key outputted from Alice and Bob’s side is identical, namely $S_{A} = S_{B} = S$. Secrecy means that the output key $S$ is uniformly distributed and independent of Eve’s auxiliary systems, and the classical-quantum (cq) state of $S$ and $E$ can be described as ${\rho _{sE}}{\rm{ = }}\sum\limits_{s \in S} {{p_s}\left| s \right\rangle \left\langle s \right| \otimes \rho _E^s}$, where $p_s$ is the probability that the key satisfies $S=s$ and ${\left\{ {\left| s \right\rangle } \right\}_{s \in S}}$ is a family of orthonormal vectors. Here the uniform distribution ensures that the eavesdropper is all equal probability for guessing any result, and there is no optimal guessing strategies. Being independent of Eve means Eve cannot increase the guessing probability with the help of its auxiliary systems. Robustness means that the protocol can usually generate a key, i.e., $S \ne \bot$. Hence if a protocol both satisfies correctness, secrecy and robustness, then the protocol achieves perfect security, and we can describe the perfectly secure classical-quantum-state of $S$ and $E$ as $$\
\rho _{SE}^{perfect} = \left( {1 - {p_ \bot }} \right)\sum\limits_{s \in S} {\frac{1}{{\left| S \right|}}\left| s \right\rangle \left\langle s \right| \otimes {{\rho '}_E}} + {p_ \bot }\left| \bot \right\rangle \left\langle \bot \right| \otimes {\rho ''_E},$$ where ${p_ \bot } \in \left[ {0,1} \right]$ is the probability that the protocol cannot generate a key and ${{{\rho '}_E}}$ and ${\rho ''_E}$ are the density operators of Eve.
However, a practical QKD protocol cannot be designed perfectly, so the perfect security definition is hard to fulfill. If a protocol is “similar" to a perfectly secure protocol, we can define the $\epsilon$-security by indistinguishability to describe a protocol of which security is approximately the same as the ideal protocol. The requirement of $\epsilon$-security is that a protocol should be $\epsilon_c$-correct, $\epsilon_s$-secrecy and $\epsilon_r$-robust [@New.J.Phys.11.085006.2009].
$\epsilon_c$-correctness requires that the output keys $S_A$ and $S_B$ from Alice and Bob’s side should be the same at most times expect for a very small probability $\epsilon_c$, which could be described by $$\
\Pr \left[ {{S_A} \ne {S_B}} \right] \le {\epsilon_c}.$$ The requirement of $\epsilon_r$-robustness is that the probability of a protocol which cannot generate keys is not larger than a very small value $\epsilon_r$, which means $$\
\Pr \left[ {{S_A} = \bot } \right] \le {\epsilon_r},$$ and a protocol is $\epsilon_s$-secret if and only if the distance between the classical-quantum-state ${\rho _{SE}}$ and $\rho _{SE}^{perfect}$ is smaller than $\epsilon_s$, which could be represented by $$\
\frac{1}{2}{\left\| {{\rho _{SE}} - \rho _{SE}^{perfect}} \right\|_1} \le \epsilon_s,$$ where ${{\rho _{SE}}}$ is the joint state of a output key $S$ and the auxiliary system held by Eve.
If a protocol satisfies both $\epsilon_c$-correct, $\epsilon_s$-secrecy and $\epsilon_r$-robust, then we can treat the protocol $\epsilon$-secure, where $\epsilon = \epsilon_c+\epsilon_s+\epsilon_r$.
The original CV-QKD protocol with entanglement in the middle against collective attacks {#The_protocol_and_the_security_against_collective_attacks}
=======================================================================================
We begin by describing the CV-QKD protocol with entanglement in the middle, which was originally proposed in Ref. [@Phys.Rev.A.87.022308.2013]. A two-mode squeezed vacuum state EPR, with an unknown variance $V$, is prepared by the untrusted third party, see Fig. \[entangling\_in\_the\_middle\]. The EPR source can be created either by a untrusted communication party Charlie or by the potential adversary Eve. The two modes of an EPR source, e.g., EPR$_1$ and EPR$_2$, are sent to Alice and Bob separately though quantum channels. As the general assumption in QKD is that both of the two quantum channels could be totally controlled by potential eavesdropper Eve; leading to the introduction of loss and noise to the states after transmission. Assuming the quadratures of the two modes of the EPR source are ${\hat X_{EP{R_{\rm{1}}}}}$ and ${\hat X_{EP{R_{\rm{2}}}}}$ with the covariance matrix (CM)
![(Color online) Schematic of the entanglement-in-the-middle CV-QKD protocol [@Phys.Rev.A.87.022308.2013]. EPR: untrusted two-mode squeezed state with variance $V$. Hom: homodyne detection. QM: quantum memory. Only the homodye detections are discussed here and Eve’s attacks are considered as two correlated modes attacks without loss of generality.[]{data-label="entangling_in_the_middle"}](entangling_in_the_middle.eps "fig:"){width="\linewidth"}\
$$\
{\gamma _{EPR}{\rm{ = }}\left( {\begin{array}{*{20}{c}}
{V\textbf{I}}&{\sqrt {{V^2} - 1} \textbf{Z}}\\
{\sqrt {{V^2} - 1} \textbf{Z}}&{V\textbf{I}}
\end{array}} \right)},$$
where $\textbf{I} = \rm{diag[1,1]}$ and $\textbf{Z} = \rm{diag[1,-1]}$, and the transmissivities of two channels are ${\tau _A}$ and ${\tau _B}$ respectively, then we have the quadratures after transmissions, given by $$\begin{aligned}
\
{\hat X_{\rm{A}}}{\rm{ = }}\sqrt {{\tau _A}} {\hat X_{EP{R_1}}}{\rm{ + }}\sqrt {{\rm{1 - }}{\tau _A}} {\hat X_{{E_1}}}, \notag\\
{\hat X_{\rm{B}}}{\rm{ = }}\sqrt {{\tau _B}} {\hat X_{EP{R_2}}}{\rm{ + }}\sqrt {{\rm{1 - }}{\tau _B}} {\hat X_{{E_2}}},\end{aligned}$$where $E_{1}$ and $E_{2}$ are the ancillary systems which Eve inject into the links to attack the protocol. The two-correlated-modes eavesdropping strategy is considered here, which is the general two-modes attack strategy, where the CM ${\gamma _{{E_1}{E_2}}}$ of the two correlated modes is $$\
{\gamma _{{E_1}{E_2}}} = \left( {\begin{array}{*{20}{c}}
{{\omega _A}\rm{\textbf{I}}}&\rm{\textbf{G}}\\
\rm{\textbf{G}}&{{\omega _B}\rm{\textbf{I}}}
\end{array}} \right),$$ where ${\omega _A}$ and ${\omega _B}$ are the variance of modes $\rm{{E_1}}$ and $\rm{{E_2}}$, and the correlation term $\rm{\textbf{G}}= \rm{diag}\left[ {g,g'} \right]$ with the correlation parameters $g$ and $g'$ satisfying the constraints shown in Ref. [@New.J.Phys.15.113046.2013]. The attack is optimal by setting modes $\rm{{E_1}}$ and $\rm{{E_2}}$ as coherent given in Refs. [@Nat.Photon.9.397.2015; @Phys.Rev.A.91.022320.2015; @arXiv:1509.04144].
Originally, Alice and Bob perform quadrature measurements via homodyne or heterodyne detections, and in this paper, we only consider the scenario that both Alice and Bob employ homodyne detections to get one measurement result, i.e., quadrature $x$ or $p$. After finishing the state preparation and measurement phases, both Alice and Bob announce which quadrature they choose through an authenticated pubic channel to sift their keys. They hold the data for which the selected quadratures are the same and discard the rest. Finally, the two communication parties proceed with classical data post-processing, namely parameter estimation, error correction and privacy amplification to distill their keys.
In the case of collective attacks setting, the state ${\rho _{{A^N}{B^N}{E^N}}}$ after all runs can be considered as a tensor product state, namely ${\rho _{{A^N}{B^N}{E^N}}} = \rho _{ABE}^{ \otimes N}$, where $N$ is the total number of quantum signals exchanged during the protocol. In this paper, we only focus on the asymptotic case under collective attacks to show the ideal performance of the protocol, where the total number of quantum states $N$ tends to infinite. The asymptotic secret key rate $K_{collective}^{asym}$ (for reverse reconciliation) is given by the Devetak-Winter formula [@Proc.Roy.Soc.A.461.207.2005], which reads $$\
K_{collective}^{asym} = \max \left\{ {\beta I\left( {A:B} \right) - \chi \left( {B:E} \right),0} \right\},
\label{D_W_relation}$$ where $\beta$ is the reconciliation efficiency, $I\left( {A:B} \right)$ is the classical mutual information between Alice’s and Bob’s data, and $\chi \left( {B:E} \right)$ is the Holevo information between Bob’s data and the eavesdropper [@Probl.Inf.Transm.9.177.1973]. This is given by $\chi \left( {B:E} \right) = S\left( E \right) - S\left( {E|B} \right)$, where $S\left( E \right)$ is the von Neumann entropy of Eve and $S\left( {E|B} \right)$ is the conditional von Neumann entropy of Eve given Bob’s information.
$\chi \left( {B:E} \right)$ can be bounded with the help of the Gaussian state extramelity theorem [@Phys.Rev.Lett.97.190503.2006; @Phys.Rev.Lett.96.080502.2006] in the case of collective attacks, hence we assume that the state ${\rho _{AB}}$ is Gaussian to minimize the final secret key rates, which can be calculated from the CM. A detailed derivation of the CM and the key rate can be seen in Appendix \[CMSKRUCA\].
The modified CV-SDI QKD protocol against general coherent attacks {#The_security_against_general_coherent_attacks}
=================================================================
In the case of general coherent attacks, the assumption that ${\rho _{{A^N}{B^N}{E^N}}}$ has tensor product structure is invalid, so we cannot apply Eq. (\[D\_W\_relation\]) directly to bound the security key rate after finite runs of the protocol. There are in general two main security-proof techniques developed in CV-QKD to handle coherent attack issues. One method is the de Finetti theorem [@Phys.Rev.Lett.114.070501.2015; @Phys.Rev.Lett.118.200501.2017], which have the ability to reduce the security from coherent attacks to collective attacks, and it was successfully employed to analyse the protocol which has some symmetric properties [@Phys.Rev.A.97.052327.2018]. The alternative is the entropic uncertainty relation [@Phys.Rev.Lett.109.100502.2012; @Phys.Rev.A.90.042325.2014], which requires that the protocol needs to randomly measure between two quadratures and perform the sifting process [@Nat.Commun.6.8795.2015; @Phys.Rev.A.98.012314.2018]. We exploit the latter tool in this paper to obtain the security of the entanglement-in-the-middle protocol with homodyne-homodyne structure against coherent attacks.
We analyse the protocol under general coherent attacks with untrusted source in the middle by adapting the approach described in Ref. [@Phys.Rev.A.90.042325.2014]. Thanks to the composable security framework, we have the ability to study the protocol considering some imperfect situation, such as the practical detection model, the energy test and finite-size effect, which allows us to modify the protocol in coherent-attack case.
The practical detection model and the measurement phase
-------------------------------------------------------
We model the practical detector as an ideal homodyne detector followed by an analog-to-digital converter (ADC) with finite sampling range, and therefore the measurement process can be divided into two steps.
In Step 1, Alice and Bob exploit ideal homodyne detectors to measure the input signal with infinite ranges and resolutions. The measurement quadratures are ideal continuous variables with infinite dimensions, hence the measurement results are also continuous. Assuming that the sifting process is done, we denote the outputs of ideal homodyne detectors as ${Q_A}$ and ${Q_B}$ in two sides. In general CV-QKD scenario, the statistical distribution of each outcome should follow a Gaussian distribution.
In order to obtain a tight bound using the entropic uncertainty relation, we need to rescale one of two results, ${Q_A}$ or ${Q_B}$, and ensure that Alice’s and Bob’s measurement outcomes have high correlations after transmission through untrusted channels. We use the transformations below (using Alice as an example) to scale the quadrature measurements: $$\
{Q_A} \to {\tilde Q_A} = {t_q}{Q_A},
\label{rescale}$$ where $t_q$ denotes the rescaling factor related to the channel losses of Alice and Bob, which can be determined by matching the variances of Alice’s and Bob’s measurement results. Supposing that $m$ signals are randomly chosen to do the parameter estimation, the average value of quadrature measurement results both in Alice’s and Bob’s sides can be estimated by $$\
\hat E\left( {{Q_A}} \right) = \frac{1}{m}\sum\limits_{i = 1}^m {Q_A^i} ,\quad \hat E\left( {{Q_B}} \right) = \frac{1}{m}\sum\limits_{i = 1}^m {Q_B^i} ,$$ where ${Q_A} = \left\{ {Q_A^i} \right\}_{i = 1}^m$ and ${Q_B} = \left\{ {Q_B^i} \right\}_{i = 1}^m$, and it is easy to estimate the parameter $t_q$ by [@Phys.Rev.A.98.012314.2018] $$\
{{\hat t_q}} = \sqrt {\frac{{\sum\limits_{i = 1}^m {{{\left( {Q_B^i - \hat E\left( {{Q_B}} \right)} \right)}^2}} }}{{\sum\limits_{i = 1}^m {{{\left( {Q_A^i - \hat E\left( {{Q_A}} \right)} \right)}^2}} }}}.$$
In the symmetric case, where the channel losses and noises of Alice and Bob are approximately the same, we can simplify the analysis by assuming that ${t_q} \approx 1$.
In Step 2, the ADCs with finite range and finite precision followed by homodyne detectors are exploited to discretize continuous measuring intervals into discrete intervals, and the continuous variables ${\tilde Q_A}$ and ${Q_B}$ are also discretized. The measurement results are grouped into intervals: $$\
\left( { - \infty , - \alpha } \right], \left( { - \alpha , - \alpha + \delta } \right], ....,\left( { \alpha - \delta , \alpha } \right], \left( {\alpha ,\infty } \right),$$ where $\alpha $ is the maximum discretization range of the ADCs, which takes the finite range of detectors into consideration, and $\delta$ denotes the resolution of the measurement, which shows how much detail the detector can detect. The corresponding outcome alphabet is denoted by $\chi = \left\{ {1,2,...,{{2\alpha } \mathord{\left/{\vphantom {{2\alpha } \delta }} \right.\kern-\nulldelimiterspace} \delta }} \right\}$, where we assume ${{2\alpha } \mathord{\left/{\vphantom {{2\alpha } \delta }} \right.\kern-\nulldelimiterspace} \delta } \in \mathbb{N}$ and every measurement outcome corresponds to one of the intervals. After this step, the continuous outcomes are replaced by the discrete results, which are denoted by $$\
{\tilde Q_A}\mathop \to \limits^{discrete} {X_A}, \quad {Q_B}\mathop \to \limits^{discrete} {X_B}.$$
This detection model can effectively illustrate the practical detector with finite range and resolution, without considering the efficiency of the detector, which could be modeled by a beam splitter with transmissivity $T_{d} $ [@Phys.Rev.Lett.102.130501.2009]. However, the “discretization” process may cause security issues when compared with the ideal detection case since the detection results are missing information about the quadratures. One issue is that any measurement outcomes inside one of the equal-length intervals $\left( { - \alpha , - \alpha + \delta } \right], ....,\left( { \alpha - \delta , \alpha } \right]$ will map to the same value and it may cause a reduction in the information about the state within each sampling interval due to the finite sampling bits. This effect can be suppressed by increasing the number of sampling bits. The other problem is caused by two intervals with infinite length, namely $\left( { - \infty , - \alpha } \right]$ and $\left[ {\alpha ,\infty } \right)$, and users cannot know the full information about the state outside the detection range. In other word, users cannot distinguish whether the energy of the measured pulse is low or high, which may leave some loopholes for eavesdropping. This problem can be solved by the energy test solution.
The energy test
---------------
![(Color online) Schematic of the energy test at Bob’s side. Bob uses a beam splitter with transmissivity $T$ to split the incoming signal into two parts. The transmission mode $\rm{{X'_B}}$ is used for generating Bob’s data and the reflection mode $a'$ is exploited to perform the energy test. $a$ and $b$ are two vacuum modes induced by beam splitters. Modes ${t_1}$ and ${t_2}$ are the output modes of the balanced beam splitter used for checking whether $\left| {{q_{{t_1}}}} \right|$ and $\left| {{p_{{t_2}}}} \right|$ are below a certain threshold.[]{data-label="energy_test"}](energy_test.eps "fig:"){width="\linewidth"}\
For fear of the large energy attack that Eve may exploit during the communication process, the protocol should be modified by adding the energy test step to ensure that the energy of measured states is below a certain threshold. We adapt the energy test method proposed in Ref. [@Phys.Rev.A.90.042325.2014] to study entanglement-in-the-middle protocol to remove the assumption of the source in the security analysis, which should be considered in trusted source scenario [@Phys.Rev.Lett.109.100502.2012], hence this protocol also can be called *source-device-independent* protocol.
Assuming that the protocol is performed with reverse reconciliation, the energy test is exploited in Bob’s side before Bob perform the measurement step, which is described in Fig. \[energy\_test\]. Bob use a beam splitter with almost perfect transmissivity $T$ to split incoming mode $\rm{X_{B}}$ into two parts, and $a$ is the vacuum mode introduced by the other port of the beam splitter. Mode $\rm{{X'_B}}$ is the transmitted mode of the output used for generating Bob’s raw data using a homodyne detector, and $a'$ is the reflected mode for the energy test. The reflected mode $a'$ is measured by a heterodyne detector, which consists of a balanced beam splitter and two homodyne detectors. Modes ${t_1}$ and ${t_2}$ are the output modes of the balanced beam splitter used for checking whether the amplitude of one output $\left| {{q_{{t_1}}}} \right|$ and the phase of the other output $\left| {{p_{{t_2}}}} \right|$ are below a certain threshold. If for every measured signal, both the amplitude $\left| {{q_{{t_1}}}} \right|$ and the phase $\left| {{p_{{t_2}}}} \right|$ are not larger than the threshold ${M_{th}}$, we say that the energy test passes; and the protocol aborts otherwise. The probability that Bob measures with homodyne detection larger than the detection range $\alpha$ can be bounded by the function $\Gamma \left( {\alpha,T,{M_{th}} } \right)$, which reads [@Phys.Rev.A.90.042325.2014] $$\
\Gamma \!\left( {\alpha,T,{M_{th}} } \right): =\! \frac{{\sqrt {1 \!+\! \lambda } \! +\! \sqrt {1\! +\! {\lambda \!^{ - 1}}} }}{2}\exp \!\left( { - \frac{{{{\left( {\mu \alpha \! -\! {M_{th}}} \right)}^2}}}{{T\left( {1\! +\! \lambda } \right)/2}}} \right),$$ where $\mu = \sqrt {\frac{{1 - T}}{{2T}}}$ and $\lambda = {\left( {\frac{{2T - 1}}{T}} \right)^2}$. The smoothness of the energy test ${\tilde \epsilon }$ further can be bounded by $$\
\tilde \epsilon = \sqrt {\frac{{2n\Gamma \left( {\alpha,T,M_{th} } \right)}}{{{p_{pass}}}}}.
\label{energy_test_parameter}$$
Finite-size effect and the key rate
-----------------------------------
In the coherent-attack scenario, due to the leftover hash lemma, the ${\epsilon _c}$-correct and ${\epsilon _s}$-secret key of length $\ell _{\sec }$ can be extracted [@arXiv.0512258.2006], which can be expressed by $$\
{\ell _{\sec }} \le H_{\min }^\epsilon {\left( {{X_B^{key}}|E} \right)_\rho } - {\ell _{EC}} - {\log _2}\frac{1}{{\epsilon _1^2{\epsilon _c}}} + 2,
\label{EUR_key_rata}$$ where ${\ell _{EC}}$ denotes the leaked information in error correction step, and it can be estimated before the error correction begin during the parameter estimation phase, $H_{\min }^\epsilon \left( {{X_B^{key}}|E} \right)$ is the smooth conditional min-entropy of data $X_B^{key}$ with smoothing parameter $\epsilon$, conditioned on the information Eve may have, which quantifies Eve’s uncertainty about the Alice’s measurement outcomes. $\epsilon$ satisfies $\epsilon \le {{\left( {{\epsilon _s} - {\epsilon _{\rm{1}}}} \right)} \mathord{\left/{\vphantom {{\left( {{\epsilon _s} - {\epsilon _{\rm{1}}}} \right)} {2{p_{pass}} - 2\tilde \epsilon }}} \right. \kern-\nulldelimiterspace} {2{p_{pass}} - 2\tilde \epsilon }}$, where ${\tilde \epsilon}$ is the security parameter related to the energy test given in Eq. (\[energy\_test\_parameter\]) and we choose ${\epsilon _1} = {{\epsilon _s} \mathord{\left/
{\vphantom {e {\rm{2}}}} \right.\kern-\nulldelimiterspace} {\rm{2}}}$ for simplification [@New.J.Phys.15.053049.2013]. Equation (\[EUR\_key\_rata\]) is a CV type key formula considering the quantum side information $E$ in infinite-dimensional Hilbert space [@Phys.Rev.Lett.109.100502.2012].
The parameter ${\ell _{EC}}$ can be easily obtained by publishing some of Bob’s data (in the reverse reconciliation case), which is $$\
{\ell _{EC}} = H\left( {{X_B}} \right) - \beta I\left( {{X_B}:{X_A}} \right),$$ where $H\left( {{X_B}} \right)$ denotes the discrete Shannon entropy of the data in Bob side, which can be described by $$H\left( {{X_B}} \right) = - \sum\limits_{i = 1}^n {p\left( {{x_i}} \right)\delta {\log _2} p\left( {{x_i}} \right) - {\log _2} \delta },$$ and $I\left( {{X_B}:{X_A}} \right)$ is the mutual information between Alice and Bob.
Our target is to bound the smooth min-entropy $H_{\min }^\epsilon \left( {{X_B^{key}}|E} \right)$ in the presence of quantum adversaries. The entropic uncertainty relations were originally introduced in discrete variable QKD to bound the min-entropy and to show the protocols’ security [@Nat.Phys.6.659.2010; @Phys.Rev.Lett.106.110506.2011]. They were thereafter extended to infinite dimensions by Furrer et al. [@Commun.Math.Phys.306.165.2011; @J.Math.Phys.55.122205.2014]. Therefore we exploit one type of uncertainty relation formula shown in Ref. [@Phys.Rev.A.90.042325.2014] to bound the min-entropy in the entanglement-in-the-middle protocol, and the feature of the entropic uncertainty relation together with the energy test, resulting in the protocols being source-device-independent.
Entropic uncertainty relation gives a bound of guessing the uncertainty that the eavesdropper may have, when both communication parties randomly measure in two bases. The relationship between smooth min- and max- entropies is given by $$\
H_{\min }^\epsilon {\left( {X_B^{key}|E} \right)_\omega } \ge n\log \frac{1}{{c\left( \delta \right)}} - H_{\max }^{\epsilon}{\left( {{X_B^{key}}|{A^n}} \right)_\omega },
\label{inequality_uncertainty}$$ where $c\left( \delta \right)$ quantifies the overlap of the two measurements and is independent of the measured states, which considers the detectors’ discretization process and has the form: $$\
c\left( \delta \right) = \frac{1}{{2\pi }}{\delta ^2} \cdot S_0^{\left( 1 \right)}{\left( {1,\frac{{{\delta ^2}}}{4}} \right)^2},$$ where $S_0^{\left( 1 \right)}$ is the ${0^{th}}$ radial prolate spheroidal wave function of the first kind [@J.Math.Phys.51.072105.2010], which can be well approximated by $c\left( \delta \right) \approx {{{\delta ^2}} \mathord{\left/ {\vphantom {{{\delta ^2}} {\left( {2\pi } \right)}}} \right.\kern-\nulldelimiterspace} {\left( {2\pi } \right)}}$ if the interval length $\delta $ is not large. $H_{\max }^{\epsilon}{\left( {{X_B^{key}}|{A^n}} \right)_\omega }$ is the smooth max-entropy between Bob’s data and Alice’s system with smoothing parameter $\epsilon$. In Eq. (\[inequality\_uncertainty\]), we assume that the random sampling of detections are i.i.d. The goal of estimating the smooth min-entropy $H_{\min }^\epsilon \left( {{X_B^{key}}|E} \right)$ is to give an upper bound of the smooth max-entropy $H_{\max }^{\epsilon}{\left( {{X_B^{key}}|{A^n}} \right)_\omega }$.
To estimate the upper bound of $H_{\max }^{\epsilon}{\left( {{X_B^{key}}|{A^n}} \right)_\omega }$, first due to the data processing inequality [@IEEE.Trans.Inf.Theory.56.4674.2010], we can obtain that
$$\
H_{\max }^{\epsilon}{\left( {{X_B^{key}}|{A^n}} \right)_\omega } \le H_{\max }^{\epsilon}{\left( {{X_B^{key}}|{X_A^{key}}} \right)_\omega },$$
and we need to bound the correlation between data $X_B^{key}$ and $X_A^{key}$. For that we exploit the average distance, $$\
d\left( {X,Y} \right) = \frac{1}{n}\sum\limits_{i = 1}^n {\left| {{{ {X} }_i} - {{ Y }_i}} \right|},$$ to give the bound of the smooth max-entropy. It has been shown in Ref. [@Phys.Rev.Lett.109.100502.2012] that if $\Pr \left[ {d\left( {X_B^{key},X_A^{key}} \right) \ge d} \right] \le {\epsilon ^2}$ hold, we can always give a bound by $$\
H_{\max }^\epsilon \left( {{{X_B^{key}}|{X_A^{key}}}} \right) \le n{\log _2}\gamma \left( d \right).$$ where $\gamma $ is a function arising from a large deviation consideration, which reads $$\
\gamma \left( t \right) = \left( {t + \sqrt {{t^2} + 1} } \right){\left[ {{t \mathord{\left/
{\vphantom {t {\left( {\sqrt {{t^2} + 1} - 1} \right)}}} \right.
\kern-\nulldelimiterspace} {\left( {\sqrt {{t^2} + 1} - 1} \right)}}} \right]^t}.$$ However, we have only data $X_A^{pe}$ and $X_B^{pe}$ with $m$ length to perform the parameter estimation rather than data $X_A^{key}$ and $X_B^{key}$, thus parameter $d$ needs to be bounded by exploiting the data only consumed in parameter estimation step. Two functions need to be defined first, one is the average second moment of the difference between two sequences, which reads $$\
{d_2}\left( {X,Y} \right) = \frac{1}{N}\sum\limits_{k = 1}^N {{{\left| {{X^k} - {Y^k}} \right|}^2}},
\label{average_second_moment}$$ and the other is the average second moment for the discretized data measurements, which is denoted by $$\
{m_2}\left( X \right) = \frac{1}{N}\sum\limits_{k = 1}^N {{{\left| {{X^k} - {\alpha \mathord{\left/
{\vphantom {\alpha \delta }} \right.
\kern-\nulldelimiterspace} \delta }} \right|}^2}}.
\label{average_second_moment_discretized_data}$$ Then we check whether the average distance ${d^{PE}} = d\left( {X_A^{pe},X_B^{pe}} \right)$ is not larger than a certain threshold $d_0$. They continue the protocol if ${d^{PE}} \le {d_0}$ and abort the protocol otherwise. In the case of the protocol proceeding, Alice and Bob calculate the average second moments of their data respectively, which denote $V_{{X_A}}^{PE} = {m_2}\left( {X_A^{pe}} \right)$ and $V_{{X_B}}^{PE} = {m_2}\left( {X_B^{pe}} \right)$ according to Eq. (\[average\_second\_moment\_discretized\_data\]), and they also compute the average second moments between their data by $V_d^{PE} = {d_2}\left( {X_A^{pe},X_B^{pe}} \right)$ according to Eq. (\[average\_second\_moment\]).
With the help of Serfling’s large deviation bound [@Ann.Stat.2.39.1974], we can finally bound the max-entropy by $$\
H_{\max }^\epsilon \left( {{{X_B^{key}}|{X_A^{key}}}} \right) \le n{\log _2}\gamma \left( d_{0} + \mu \right),$$ where $\mu$ describes the statistical fluctuation deviating from $d\left( {X_B^{key},X_A^{key}} \right)$, which denotes $$\
\mu = \sqrt {2{{\log }_2}{\xi ^{ - 1}}} \frac{{N{\sigma _*}}}{{m\sqrt n }} + \frac{{4\left( {{\alpha \mathord{\left/
{\vphantom {\alpha \delta }} \right.
\kern-\nulldelimiterspace} \delta }} \right){{\log }_2}{\xi ^{ - 1}}}}{3}\frac{N}{{nm}},$$ with $$\begin{aligned}
\sigma _*^2 =& \frac{m}{N}\left( {V_d^{PE} - \frac{m}{N}{{\left( {{d^{PE}}} \right)}^2}} \right) + \frac{m}{N}\left( {V_{{X_A}}^{PE} + V_{{X_B}}^{PE} + 2\frac{\nu }{{{\delta ^2}}}} \right) \notag\\
&+ 2\frac{m}{N}\sqrt {\left( {V_{{X_A}}^{PE} + \frac{\nu }{{{\delta ^2}}}} \right)\left( {V_{{X_B}}^{PE} + \frac{\nu }{{{\delta ^2}}}} \right)},\end{aligned}$$ and $$\begin{aligned}
\xi =& {\left( {{\epsilon _s} - {\epsilon _1} - 2\sqrt {2n\Gamma \left( {\alpha ,T,{M_{th}}} \right)} } \right)^2} \notag\\
& - 2\exp \left( { - 2{{\left( {{\nu \mathord{\left/{\vphantom {\nu \alpha }} \right.\kern-\nulldelimiterspace} \alpha }} \right)}^2}\frac{{n{m^2}}}{{N\left( {m + 1} \right)}}} \right).\end{aligned}$$ $\nu$ is the smallest real number making $\xi$ positive. If there exist $\nu$ such that $\xi$ is positive and ${\epsilon _1} - 2\sqrt {2\Gamma \left( {\alpha ,T,{M_{th}}} \right)} < {\epsilon _s}$ is satisfied, the final secret key rate under coherent attacks can be written as $$\
{K_{coherent}} = {{{\ell _{Low}}} \mathord{\left/{\vphantom {{{\ell _{Low}}} N}} \right.\kern-\nulldelimiterspace} N},$$ where ${\ell _{Low}}$ is the lower bound of the secure key length, which reads $$\
{\ell _{Low}} =n\left[ {\log \frac{1}{{c\left( \delta \right)}} - \log \gamma \left( {{d_0} + \mu } \right)} \right] - {\ell _{EC}} - \log \frac{1}{{\varepsilon _1^2{\varepsilon _c}}}+2.
\label{keyrate}$$ Otherwise, we denote that the key rate ${K_{coherent}} = 0$. The detailed proof of this section can be seen in Ref. [@Phys.Rev.A.90.042325.2014].
![(Color online) Secret key rates of the CV-SDI QKD protocol. The protocol is under symmetric configuration with ${\tau _A} = {\tau _B} = \sqrt {{T}}$ where $T$ is the total transmissivity of the channel. We consider the protocol with perfect reconciliation efficiency $\beta = 1$ and ideal modulation variance $V = {10^5}$. We also set the excess noise as $\xi = 0.001$ in each channel and the overall security parameter is smaller than ${10^{ - 20}}$. The gray dot line is the PLOB bound [@Nat.Commun.8.15043.2017] and the black solid line is the key rate under collective attacks. The red solid line is the key rate under coherent attacks with infinite exchanged signals. The four dashed lines, from top to bottom, are the secret key rates under coherent attacks, with the block lengths from ${10^{10}}$ to ${10^{7}}$.[]{data-label="simulation_key_rates"}](simulation_key_rates.eps "fig:"){width="\linewidth"}\
Simulation and discussion {#Simulation_and_discussion}
=========================
Using the results in the previous section, we can plot the secret key rate as a function of the total transmission distance focusing on the symmetric configuration where we set ${\tau _A} = {\tau _B} = \sqrt {{T}}$ and $T$ is the transmissivity of the channel. The simulations are under two-mode optimal attacks to show the performance of the protocol and both collective and coherent attack scenarios are discussed shown in Fig. \[simulation\_key\_rates\].
In order to facilitate the analysis of the performance of the protocol, we simulate the key rate with some ideal parameters. For instance, we assume that the protocol has an ideal modulation variance $V = {10^5}$ (which could replace an infinite modulation variance) and perfect reconciliation efficiency $\beta = 1$. Also, to get the lower bound of the protocol, we set $g = \min \left\{ {\sqrt {\left( {{\omega _A} - 1} \right)\left( {{\omega _B} + 1} \right)} ,\sqrt {\left( {{\omega _A} + 1} \right)\left( {{\omega _B} - 1} \right)} } \right\}$ and ${\omega _A} = {\omega _B} = 1 + {{T\xi } \mathord{\left/
{\vphantom {{T\xi } {\left( {1 - T} \right)}}} \right. \kern-\nulldelimiterspace} {\left( {1 - T} \right)}}$ with excess noise $\xi$ in one channel for two-mode optimal attacks. In the coherent attack cases, we set the interval parameter $\alpha$ to $52$ [@Phys.Rev.Lett.109.100502.2012] and the overall security parameter is smaller than ${10^{ - 20}}$. Meanwhile, the parameter ${M_{th}}$ is set to $12$ to ensure that the energy test fails with probability smaller than ${10^{ - 20}}$.
![(Color online) Secret key rates as functions of block size of the CV-SDI QKD protocol. The black solid line shows the performance with the distance of 5km. The red dot-dashed line and the blue dashed line are the key rates of the protocol with distances of 10km and 14km, respectively. The other parameters are as in Fig. \[simulation\_key\_rates\][]{data-label="N_change"}](N_change.eps "fig:"){width="\linewidth"}\
In Fig. \[simulation\_key\_rates\], the gray dot line shows the Pirandola-Laurenza-Ottaviani-Banchi (PLOB) bound [@Nat.Commun.8.15043.2017], which gives an upper bound of the secret key capacity of the lossy channel. The black solid line is the asymptotic key rate under collective attacks, and the longest transmission distance is over 18km, which is a little shorter than that of the two-mode individual attacks case (where the correlation parameter $g=0$) [@Phys.Rev.A.87.022308.2013]. The other five curves, from top to bottom, describe the key rates under coherent attacks. The red solid curve is obtained for $N \to \infty$, and the other dashed lines describe the rate for $N = {10^{10}}$ to $N = {10^{7}}$ with finite exchanged signals. In Fig. \[N\_change\], we also plot the secret key rate under coherent attacks as a function of block size for different distances. The distances are 5km, 10km and 14km, respectively. We point out that when the block size reduces, the secret key rate decreases, and it is not achievable if the block size is below ${10^7}$.
We notice that there is a gap between the performance of CV-QKD protocol with entanglement in the middle under collective attacks and that under asymptotic coherent attacks cases. The reason is that the bound given by the entropic uncertainty relation is not very tight especially in the high losses regime, which has been shown in Ref. [@Phys.Rev.A.90.042325.2014].
Conclusion {#Conclusion}
==========
In conclusion, we have analyzed the security of continuous-variable source-device-independent quantum key distribution protocol against general coherent attacks, where the source of the protocol is untrusted and may be controlled by the malicious adversary. By exploiting the state-independent entropic uncertainty relation together with the energy test, our analysis has no assumptions on the source, making the protocol source-device-independent even under coherent attacks. The simulation results indicate that, in the universal composable security framework, the protocol is still secure, achieving high key rates against coherent attacks under the condition of achievable block lengths ($N$ from ${10^7}$ to ${10^{10}}$).
This work was supported in part by the Key Program of National Natural Science Foundation of China under Grants 61531003, the National Science Fund for Distinguished Young Scholars of China under Grant No. 61225003, the National Natural Science Foundation under Grants 61427813, China Postdoctoral Science Foundation under Grant 2018M630116, and the Fund of State Key Laboratory of Information Photonics and Optical Communications.
Covariance matrix and the secret key rate under collective attacks {#CMSKRUCA}
==================================================================
The final bipartite quantum state ${\rho _{AB}}$ of Alice and Bob has the CM with the form $$\
{\gamma _{AB}} = \left( {\begin{array}{*{20}{c}}
{a\rm{\textbf{I}}}&{c\rm{\textbf{Z}}}\\
{c\rm{\textbf{Z}}}&{b\rm{\textbf{I}}}
\end{array}} \right),$$ where $$\begin{aligned}
&a = {\tau _A}V + \left( {1 - {\tau _A}} \right){\omega _A}, \notag\\
&b = {\tau _B}V + \left( {1 - {\tau _B}} \right){\omega _B}, \notag\\
&c = \sqrt {{\tau _A}{\tau _B}} \sqrt {{V^2} - 1} - g\sqrt {1 - {\tau _A}} \sqrt {1 - {\tau _B}},\end{aligned}$$ and we let $$\
g = \min \left\{ {\sqrt {\left( {{\omega _A} - 1} \right)\left( {{\omega _B} + 1} \right)} ,\sqrt {\left( {{\omega _A} + 1} \right)\left( {{\omega _B} - 1} \right)} } \right\}$$ by setting modes ${E_1}$ and ${E_2}$ are coherent. Then the secret key rate $K_{collective}^{asym}$ can be calculated by Eq. (\[D\_W\_relation\]) if we restrict our discussion in reverse reconciliation cases. The mutual information between Alice’s and Bob’s data can be described as $$\
I\left( {A:B} \right) = \frac{1}{2}{\log _2}\left( {\frac{a}{{a - {{{c^2}} \mathord{\left/
{\vphantom {{{c^2}} b}} \right.
\kern-\nulldelimiterspace} b}}}} \right).$$ To obtain the von Neumann entropy $S\left( E \right)$ and $S\left( {E|B} \right)$, we always assume that Eve can purify the whole system in order to maximize her information, thus we have $S\left( E \right) = S\left( {AB} \right)$ and $S\left( {E|B} \right) = S\left( {A|B} \right)$. $S_{AB}$ is a function of the symplectic eigenvalues ${\lambda _{1,2}}$ of ${\gamma _{AB}}$, which reads $$\
S\left( {AB} \right) = G\left[ {{{\left( {{\lambda _1} - 1} \right)} \mathord{\left/
{\vphantom {{\left( {{\lambda _1} - 1} \right)} 2}} \right.
\kern-\nulldelimiterspace} 2}} \right] + G\left[ {{{\left( {{\lambda _2} - 1} \right)} \mathord{\left/
{\vphantom {{\left( {{\lambda _2} - 1} \right)} 2}} \right.
\kern-\nulldelimiterspace} 2}} \right],$$ where $$\
G\left( x \right) = \left( {x + 1} \right){\log _2}\left( {x + 1} \right) - x\log x,$$ and $$\
\lambda _{1,2}^2 = \frac{1}{2}\left[ {\Delta \pm \sqrt {{\Delta ^2} - 4{D^2}} } \right],$$ where we use the notations that $\Delta = {a^2} + {b^2} - 2{c^2}$ and $D = ab - {c^2}$. After Bob performs homodyne detection, Alice’s CM conditioned on Bob’s measurement results will transform to $$\
\gamma _A^{{x_b}} = {\gamma _A} - \Sigma _C^T{\left( {X{\gamma _B}X} \right)^{ - 1}}{\Sigma _C},$$ where ${\gamma _A} = a\rm{\textbf{I}}$, ${\gamma _B} = b\rm{\textbf{I}}$, ${\Sigma _C} = c\rm{\textbf{Z}}$ and $X = \left[ {1,0;0,0} \right]$. $S\left( {A|B} \right) = G\left[ {{{\left( {{\lambda _3} - 1} \right)} \mathord{\left/
{\vphantom {{\left( {{\lambda _3} - 1} \right)} 2}} \right. \kern-\nulldelimiterspace} 2}} \right]$ is a function of the symplectic eigenvalue ${{\lambda _3}}$ of the covariance matrix $\gamma _A^{{x_b}}$ with ${\lambda _3} = \sqrt {a\left( {a - {{{c^2}} \mathord{\left/
{\vphantom {{{c^2}} b}} \right.\kern-\nulldelimiterspace} b}} \right)}$. Therefore, the secret key rate under collective attacks when the reverse reconciliation is performed is $$\
K_{collective}^{asym} = \beta I\left( {A:B} \right) - \left[ {S\left( {AB} \right) - S\left( {A|B} \right)} \right].$$
[99]{} N. Gisin, G. Ribordy, W. Tittel, and H. Zbinden, Rev. Mod. Phys. **74**, 145 (2002).
V. Scarani, H. Bechmann-Pasquinucci, N. J. Cerf, M. Dušek, N. Lütkenhaus, and M. Peev, Rev. Mod. Phys. **81**, 1301 (2009).
C. Weedbrook, S. Pirandola, R. García-Patrón, N. J. Cerf, T. C. Ralph, J. H. Shapiro, and S. Lloyd, Rev. Mod. Phys. **84**, 621 (2012).
E. Diamanti and A. Leverrier, Entropy **17**, 6072 (2015).
F. Grosshans, and P. Grangier, Phys. Rev. Lett. **88**, 057902 (2002).
C. Weedbrook, A. M. Lance, W. P. Bowen, T. Symul, T. C. Ralph, P. K. Lam, Phys. Rev. Lett. **93**, 170504 (2004).
B. Qi, W. Zhu, L. Qian, and H. K. Lo, New J. Phys. **12**, 103042 (2010).
R. Kumar, H. Qin and R. Alleáume, New J. Phys. **17**, 043027 (2015).
Y. Zhang, Z. Li, Z. Chen, C. Weedbrook, Y. Zhao, X. Wang, C. Xu, X. Zhang, Z. Wang, M. Li, X. Zhang, Z. Zheng, B. Chu, X. Gao, N. Meng, W. Cai, Z. Wang, G. Wang, S. Yu, and H. Guo, arXiv: 1709.04618 (2017).
A. Leverrier and P. Grangier, Phys. Rev. Lett. **102**, 180504 (2009).
A. Leverrier and P. Grangier, Phys. Rev. A **83**, 042312 (2011).
Z. Li, Y. Zhang, and H. Guo, arXiv:1805.04249.
S. Pirandola, S. Mancini, S. Lloyd, and S. L. Braunstein, Nat. Phys. **4**, 726 (2008).
M. Sun, X. Peng, Y. Shen, and H. Guo, Int. J. Quantum Inf. **10**, 1250059 (2012).
Y. C. Zhang, Z. Li, C. Weedbrook, S. Yu, W. Gu, M. Sun, X. Peng, and H. Guo, J. Phys. B **47**, 035501 (2014).
C. Ottaviani, S.Mancini, and S. Pirandola, Phys. Rev. A **92**, 062323 (2015).
C. Ottaviani and S. Pirandola, Sci. Rep. **6**, 22225 (2016).
Y. Zhang, Z. Li, Y. Zhao, S. Yu and H. Guo, J. Phys. B: At. Mol. Opt. Phys. **50**, 035501 (2017).
A. Huang, S. Barz, E. Andersson and V. Makarov, New J. Phys. **20**, 103016 (2018).
O. Thearle, J. Janousek, S. Armstrong, S. Hosseini, M. Schünemann (Mraz), S. Assad, T. Symul, M. R. James, E. Huntington, T. C. Ralph, and P. K. Lam, Phys. Rev. Lett. **120**, 040406 (2018).
Z. Li, Y.-C. Zhang, F. Xu, X. Peng, and H. Guo, Phys. Rev. A **89**, 052301 (2014).
Y.-C. Zhang, Z. Li, S. Yu, W. Gu, X. Peng, and H. Guo, Phys. Rev. A **90**, 052325 (2014).
S. Pirandola, C. Ottaviani, G. Spedalieri, C. Weedbrook, S. L. Braunstein, S. Lloyd, T. Gehring, C. S. Jacobsen, and U. L. Andersen, Nat. Photon. **9**, 397 (2015).
T. Gehring, V. Handchen, J. Duhme, F. Furrer, T. Franz, C. Pacher, R. F. Werner, and R. Schnabel, Nat. Commun. **6**, 8795 (2015).
N. Walk, S. Hosseini, J. Geng, O. Thearle, J. Y. Haw, S. Armstrong, S. M. Assad, J. Janousek, T. C. Ralph, T. Symul, H. M. Wiseman, and P. K. Lam, Optica 3, 634 (2016).
Y. Zhang, Z. Li, S. Yu, and H. Guo, Optical Society of America, Laser Science **JW4A.33** (2017).
C. Lupo, C. Ottaviani, P. Papanastasiou, and S. Pirandola, Phys. Rev. A **97**, 052327 (2018).
Z. Chen, Y. Zhang, G. Wang, Z. Li and H. Guo, Phys. Rev. A **98**, 012314 (2018).
F. Furrer, T. Franz, M. Berta, A. Leverrier, V. B. Scholz, M. Tomamichel, and R. F. Werner, Phys. Rev. Lett. **109**, 100502 (2012).
F. Furrer, Phys. Rev. A **90**, 042325 (2014).
C. Weedbrook, Phys. Rev. A **87**, 022308 (2013).
J. Müller-Quade and R. Renner, New J. Phys. **11**, 085006 (2009).
S. Pirandola, New J. Phys. **15**, 113046 (2013).
C. Ottaviani, G. Spedalieri, S. L. Braunstein, and S. Pirandola, Phys. Rev. A **91**, 022320 (2015).
C. Ottaviani, G. Spedalieri, S. L. Braunstein, and S. Pirandola, arXiv:1509.04144 (2015).
I. Devetak, and A. Winter, Proc. Roy. Soc. A **461**, 207 (2005).
A. S. Holevo, Probl. Inf. Transm. **9**, 177 (1973).
R. García-Patrón and N. J. Cerf, Phys. Rev. Lett. **97**, 190503 (2006).
M. M. Wolf, G. Giedke, and J. I. Cirac, Phys. Rev. Lett. **96**, 080502 (2006).
A. Leverrier, Phys. Rev. Lett. **114**, 070501 (2015).
A. Leverrier, Phys. Rev. Lett. **118**, 200501 (2017).
R. García-Patrón and N. J. Cerf, Phys. Rev. Lett. **102**, 130501 (2009).
R. Renner Ph.D. thesis, Swiss Federal Institute of Technology (ETH) Zurich, 2006; arXiv:quant-ph/0512258.
T. Eberle, V. Händchen, J. Duhme, T. Franz, F. Furrer, R. Schnabel and R. F. Werner, New J. Phys. **15**, 053049 (2013).
M. Berta, M. Christandl, R. Colbeck, J. M. Renes, and R. Renner. Nat. Phys. **6**, 659 (2010).
M. Tomamichel, R. Renner, Phys. Rev. Lett. **106**, 110506 (2011).
F. Furrer, J. [Å]{}berg, R. Renner, Commun. Math. Phys. **306**, 165 (2011).
F. Furrer, M. Berta, M. Tomamichel, V. B. Scholz and M. Christandl, J. Math. Phys. **55**, 122205 (2014).
J. Kiukas and R. F. Werner, J. Math. Phys. **51**, 072105 (2010).
M. Tomamichel, R. Colbeck, and R. Renner, IEEE Trans. Inf. Theory **56**, 4674 (2010).
R. J. Serfling, Ann. Stat. **2**, 39 (1974).
S. Pirandola, R. Laurenza, C. Ottaviani, and L. Banchi, Nat. Commun. **8**, 15043 (2017).
[^1]: [email protected]
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
An exact parity replicates the Standard Model giving a Mirror Standard Model, SM $\leftrightarrow$ SM$'$. This “Higgs Parity" and the mirror electroweak symmetry are spontaneously broken by the mirror Higgs, ${ \left\langle {H'} \right\rangle } = v' \gg { \left\langle {H} \right\rangle }$, yielding the Standard Model Higgs as a Pseudo-Nambu-Goldstone Boson of an approximate $SU(4)$ symmetry, with a quartic coupling $\lambda_{SM}(v') \sim 10^{-3}$. Mirror electromagnetism is unbroken and dark matter is composed of $e'$ and $\bar{e}'$. Direct detection may be possible via the kinetic mixing portal, and in unified theories this rate is correlated with the proton decay rate. With a high reheat temperature after inflation, the $e'$ dark matter abundance is determined by freeze-out followed by dilution from decays of mirror neutrinos, $\nu' \rightarrow \ell H$. Remarkably, this requires $v' \sim (10^8 - 10^{10})$ GeV, consistent with the Higgs mass, and a Standard Model neutrino mass of $(10^{-2} - 10^{-1})$ eV, consistent with observed neutrino masses. The mirror QCD sector exhibits a first order phase transition producing gravitational waves that may be detected by future observations. Mirror glueballs decay to mirror photons giving dark radiation with $\Delta N_{\rm eff} \sim 0.03 - 0.4$. With a low reheat temperature after inflation, the $e'$ dark matter abundance is determined by freeze-in from the SM sector by either the Higgs or kinetic mixing portal.
[*Dedicated to the memory of Ann Nelson.*]{}
author:
- David Dunsky
- 'Lawrence J. Hall'
- Keisuke Harigaya
bibliography:
- 'mirrorDM.bib'
title: |
Dark Matter, Dark Radiation and Gravitational Waves\
from\
Mirror Higgs Parity
---
c ł u v Ł ¶ §
ø
Ø
Introduction
============
At high energy colliders, precision measurements of the electroweak symmetry breaking sector of the Standard Model (SM) have been pursued for decades, but so far there has been no discovery of any physics that would lead to a natural explanation of the weak scale. If the SM Effective Field Theory is valid well above the weak scale, at what mass scale will it finally break down? A possible answer has been provided by the LHC: perhaps new physics enters at the scale where the SM Higgs quartic coupling passes through zero. For example, this new physics could be the breaking of PQ symmetry [@Redi:2012ad] or of supersymmetry [@Hall:2013eko; @Ibe:2013rpa; @Hall:2014vga; @Fox:2014moa].
Another possibility for this new physics is the breaking of a discrete symmetry, “Higgs Parity", that interchanges the SM Higgs, $H$, a doublet under the weak $SU(2)$, with a partner Higgs, $H'$, a doublet under some $SU(2)'$ [@Hall:2018let]. There are many implementations of this idea. One elegant possibility is that $SU(2)'$ is identified as the $SU(2)_R$ under which the right-handed quarks and leptons transform as doublets. In this case Higgs Parity may include spacetime parity and lead to a solution of the strong CP problem [@Hall:2018let]: parity forces $\theta$ to vanish and the quark Yukawa matrices to be Hermitian [@Beg:1978mt; @Mohapatra:1978fy; @Babu:1988mw; @Babu:1989rb]. Furthermore, since the breaking of $SU(3) \times SU(2)_L \times SU(2)_R \times U(1)_{B-L}$ occurs at the scale where the SM Higgs quartic vanishes, a remarkably successful unification of couplings results [@Hall:2018let; @Hall:2019qwx]. However, the theory needs extending to incorporate dark matter (DM).
In another class of theories, Higgs parity transforms SM quarks and leptons, $(q,u,d,l,e)$, into mirror quarks and leptons, $(q', u', d', l', e')$. We have recently explored such a theory where the electroweak group is doubled, but QCD is not, so both ordinary and mirror quarks are colored [@Dunsky:2019api]. This theory solves the strong CP problem, with mirror quark contributions to $\bar{\theta}$ cancelling contributions from the ordinary quarks [@Barr:1991qx]. Although there is no immediate path to gauge coupling unification, the theory does have the interesting possibility of $e'$ dark matter that is within reach of direct detection. However, hadrons containing the $u'$ quark are also stable, and since the bounds on such heavy hadron dark matter are very strong, the $e'$ production mechanism must be non-thermal rather than thermal.
In this paper we study a complete mirror sector where Higgs Parity doubles the entire Standard Model: ${\rm SM} \leftrightarrow {\rm SM}'$. In this theory $e'$ and $u'$ are again stable and DM candidates; but since now $u'$ does not couple to QCD, it is much less constrained by direct detection, allowing successful DM production via Freeze-Out with dilution or via Freeze-In. Long ago, a mirror copy of the SM with an unbroken parity was introduced as a way to restore space-time inversion symmetry [@Lee:1956qn; @Kobzarev:1966qya; @Pavsic:1974rq; @Foot:1991bp].
This Mirror Higgs Parity theory is highly constrained: the parameters in the SM$'$ Lagrangian are the same as in the SM Lagrangian, so that the only new parameters are the ones describing portal interactions: one for kinetic mixing, one for the Higgs portal and several for the neutrino portal. Although the doubling of QCD implies that Higgs Parity can no longer solve the strong CP problem, there is now a gravity wave (GW) signal from the QCD$'$ transition. In the case of Freeze-Out DM, once the neutrino portal parameters are chosen to give the observed DM abundance, the GW signal can be computed entirely in terms of measured SM parameters. This paper is devoted to the DM, dark radiation (DR) and GW signals and their relation.
In section \[sec:Z2\] we review how Higgs Parity predicts the vanishing Higgs quartic coupling at a high energy scale. Section \[sec:MSM\] introduces the mirror copy of the SM with Higgs Parity and the mass spectrum of the mirror sector. Direct detection of DM and, in unified theories, its relation to the proton decay rate is discussed in section \[sec:DD\]. The constraint from long-lived mirror glueballs is investigated in section \[sec:BBNandDR\]. In section \[sec:cosmoAbundances\], we compute the relic abundance of $e'/u'$ dark matter and dark radiation. The spectrum of the GWs from the mirror QCD phase transition is estimated in section \[sec:GW\]. The final section is devoted to conclusions and discussions.
Vanishing Higgs Quartic from a $Z_2$ Symmetry {#sec:Z2}
=============================================
In this section we review the framework of [@Hall:2018let] that yields the near vanishing of the SM Higgs quartic coupling at a high energy scale. Consider a $Z_2$ symmetry that exchanges the $SU(2)$ weak gauge interaction with a new $SU(2)'$ gauge interaction, and the Higgs field $H(2,1)$ with its partner $H'(1,2)$, where the brackets show the $(SU(2), SU(2)')$ representation. We call the $Z_2$ symmetry as Higgs Parity. The scalar potential for $H$ and $H'$ is $$\begin{aligned}
\label{eq:potential}
V(H,H') = - m^2 (H^\dagger H + H'^\dagger H') + \frac{\lambda}{2} (H^\dagger H + H'^\dagger H')^2 + \lambda' H^\dagger H H'^\dagger H' .\end{aligned}$$ We assume that the mass scale $m$ is much larger than the electroweak scale. With $m^2$ positive, the Higgs parity is spontaneously broken and $H'$ acquires a large vacuum expectation value of ${ \left\langle {H'} \right\rangle } = v'$, with $v'^2 = m^2/\lambda$. After integrating out $H'$ at tree-level, the Low Energy potential in the effective theory for $H$ is $$\begin{aligned}
\label{eq:potentialLE}
V_{LE}(H) = \lambda' \; v'^2 \; H^\dagger H - \lambda' \left(1 + \frac{\lambda'}{2 \lambda} \right) (H^\dagger H)^2 .\end{aligned}$$ To obtain the hierarchy ${ \left\langle {H} \right\rangle } = v \ll v'$, it is necessary to tune $\lambda'$ to a very small value $\lambda' \sim - v^2/v'^2$; the quartic coupling of the Higgs $H$, $\lambda_{\rm SM}$, is then extremely small.
The vanishing quartic can be understood by an accidental $SU(4)$ symmetry under which $(H, H')$ is in a fundamental representation. For $|\lambda'| \ll 1$, necessary for $v \ll v'$, the potential in Eq. (\[eq:potential\]) becomes $SU(4)$ symmetric. After $H'$ obtains a vacuum expectation value, the SM Higgs is understood as a Nambu-Goldstone boson with a vanishing potential. Note that in this limit of extremely small $\lambda'$, the vacuum alignment in the SU(4) space is determined by the Coleman-Weinberg potential. The top contribution beats the gauge contribution so that the true vacuum is the asymmetric one, where the entire condensate lies in $H'$ (or in $H$, which is physically equivalent). (The $SU(4)$ symmetry implies that the Higgs boson contribution to the Coleman-Weinberg potential does not affect the vacuum orientation.)
Below the scale $v'$, quantum corrections from SM particles renormalize the quartic coupling, and it becomes positive. From the perspective of running from low to high energies, the scale at which the SM Higgs quartic coupling vanishes is identified with $v'$. The threshold correction to $\lambda_{\rm SM}(v')$ is calculated in the next section.
Although the scale $v'$ is much smaller than the Planck scale and the typical unification scale, the theory is no more fine-tuned than the SM because of Higgs Parity. The required fine-tuning of the theory is $$\begin{aligned}
\frac{m^2}{\Lambda^2} \times \frac{v^2}{m^2} = \frac{v^2}{\Lambda^2},\end{aligned}$$ where the first factor in the left hand side is the fine-tuning to obtain the scale $m$ much smaller than the cut off scale $\Lambda$, and the second one is the fine-tuning in $\lambda'$ to obtain the electroweak scale from $m$. The total tuning is the same as in the SM, $v^2 / \Lambda^2$, and may be explained by environment requirements [@Agrawal:1997gf; @Hall:2014dfa].
The Mirror Standard Model {#sec:MSM}
=========================
The phenomenology of the theory crucially depends on the action of Higgs Parity on the SM gauge group. Refs. [@Hall:2018let; @Hall:2019qwx] considers the case where the $SU(3)_c \times U(1)_Y$ gauge group is not replicated. The theory solves the strong CP problem and can be embedded into $SO(10)$ unification. Ref. [@Dunsky:2019api] replicates the $U(1)_Y$ gauge group. The theory solves the strong CP problem and has an interesting dark matter candidate. In this paper we study a theory where the SM gauge group is entirely replicated by a $Z_2$ symmetry which maps $$\begin{aligned}
\label{eq:Paction}
SU(3) \times SU(2) \times U(1) \; \; \leftrightarrow \; \; & SU(3)' \times SU(2)' \times U(1)' \nonumber \\
q, \bar{u}, \bar{d}, \ell, \bar{e} \; \; \leftrightarrow \; \; & q', \bar{u}', \bar{d}', \ell', \bar{e} \nonumber \\
H \; \; \leftrightarrow \; \; & H'.\end{aligned}$$ where matter is described by 2-component, left-handed, Weyl fields. [^1]
The Lagrangian
--------------
The most general gauge and Higgs Parity invariant Lagrangian up to dimension $5$ is $$\begin{split}
\label{eq:L}
{\cal L} \; =& \; {\cal L}_{SM} (q, \bar{u},\bar{d},l,\bar{e},H) + {\cal L}_{SM'}(q', \bar{u}',\bar{d}',l',\bar{e}',H') + \lambda'' (H^\dagger H)(H'^\dagger H') + \frac{\epsilon}{2} \; B^{\mu \nu} B'_{\mu \nu} \\
& + \, (\ell \, \eta \, \ell) \frac{H^2}{M_M} + (\ell' \, \eta \, \ell') \frac{H^{'2}}{M_M} + (\ell \, \xi \, \ell') \frac{H H'}{M_D}
+ {\rm h.c.}
\end{split}$$ where ${\cal L}_{SM}$ is the SM Lagrangian up to dimension $4$ and ${\cal L}_{SM'}$ its $Z_2$ mirror. The next two terms of link the SM and mirror sectors: $\lambda''=\lambda + \lambda'$ describes mixing between the ordinary and mirror Higgs doublets and $\epsilon$ kinetic mixing between ordinary and mirror hypercharge. The dimension 5 operators in the second line of $\eqref{eq:L}$ describe the neutrino sector. $M_{M,D}$ are large mass scales and $\eta$ and $\xi$ are $3 \times 3$ dimensionless flavor matrices.
The Mirror Spectrum
-------------------
The charged mirror fermions acquire a mass $m_{f'} = y_{f'}v'$ from the vacuum expectation value of the mirror Higgs, $v'$. The $Z_2$ symmetry sets $y_{f'} = y_f$ at the scale $\mu = v'$, so that mirror fermion masses are larger than their SM counterparts by a factor of approximately $v'/v$, as shown in Fig. \[fig:mirrorSpectrum\].
Mirror electrons and up quarks are the lightest fermions charged under $U(1)'_{EM}$ and $SU(3)'$, respectively, and thus stable and viable DM candidates. We explore $e'$ and $u'$ DM in Sec. \[sec:cosmoAbundances\].
![Mass spectrum of key mirror particles. The purple band shows the range of mirror neutrino masses for SM neutrino masses betwen $0.01 - 0.10 ~\rm{eV}$.[]{data-label="fig:mirrorSpectrum"}](Figures/mirrorSpectrum.pdf){width="70.00000%"}
Unlike mirror quarks, mirror glueballs, $S'$, acquire mass chiefly from $SU(3)'$ nonperturbative effects, with mass [@Borsanyi:2012ve; @Chen:2005mg] $$\begin{aligned}
m_S' \simeq 6.8 \Lambda_{\rm{QCD}}' \gg \Lambda_{\rm{QCD}}.\end{aligned}$$ The mirror QCD confinement scale, $\Lambda_{\rm{QCD}}'$, is not a free parameter, but is determined by running $\alpha_S({m_Z}) \simeq .1181$ up to the $Z_2$ restoration scale $v'$, equating $\alpha_S(v') = \alpha_{S'}(v')$, and then running $\alpha_{S'}$ down to lower scales until it diverges at the scale $\Lambda_{\rm{QCD}}'$. In the $\overline{\rm MS}$ scheme the dynamical scale is given by $$\begin{aligned}
\Lambda_{\rm{QCD}}' \simeq 190~\GeV \left(\frac{v'}{10^{10} ~\GeV}\right)^{4/11}.\end{aligned}$$ Mirror glueballs are unstable and dominantly decay to $\gamma' \gamma'$, and if heavy enough, subdominantly to $H H^\dagger$. The latter are visible decays which may occur during BBN if $S'$ is long-lived. We investigate such constraints in Sec. \[sec:BBNandDR\].
Standard and mirror neutrinos obtain mass from the dimension 5 operators on the second line of . We will be interested in small mixing between $\nu'$ and $\nu$ with $M_D \gg M_M$ so that $m_{\nu'}/m_\nu \simeq (v'/v)^2$, giving $$\begin{aligned}
\label{eq:mnu}
m_{\nu'} \simeq 10^5 \; \GeV \left( \frac{m_\nu}{0.03 \; \mbox{eV}} \right) \left( \frac{v'}{10^{10}~{\rm GeV}} \right)^2\end{aligned}$$ as shown in Fig. \[fig:mirrorSpectrum\] for two values of $m_\nu$. Mirror neutrinos are unstable and decay to $\ell H$ or if heavy enough, beta decay to $e', u', d'$. Long-lived $\nu'$ may come to dominate the energy density of the universe and release significant entropy into the SM thermal bath upon decaying. We investigate the effect of such entropy dilution on freeze-out $e'$ and $u'$ DM in Sec. \[sec:fo\].
Prediction for $v'$
-------------------
Between the electroweak scale and the scale $v'$, the running of the Higgs quartic coupling $\lambda_{\rm SM}$ is exactly the same as in the SM. We follow the computation in [@Buttazzo:2013uya] and show the running in the left panel of Fig. \[fig:vpPrediction\] for a range of top quark mass $m_t = (173.0\pm 0.4)$ GeV, QCD coupling constant at the $Z$ boson mass $\alpha_S(m_Z) = (0.1181 \pm 0.0011)$, and Higgs mass $m_h = (125.18 \pm 0.16)$ GeV.
The value of the SM quartic coupling at the scale $v'$ is not exactly zero because of the threshold correction [@Dunsky:2019api], $$\begin{aligned}
\lambda_{\rm SM}(v') \simeq - \frac{3}{8\pi^2} y_t^4 \, {\rm ln} \frac{e}{y_t} + \frac{3}{128\pi^2} (g^2 + {g'}^2)^2 \, {\rm ln} \frac{e}{\sqrt{(g^2 + {g'}^2) / 2 }} + \frac{3}{64\pi^2} g^4 \, {\rm ln} \frac{e}{ g/ \sqrt{2}},\end{aligned}$$ where the $\overline{\rm MS}$ scheme is assumed. The prediction for the scale $v'$ is shown in the right panel of Fig. \[fig:vpPrediction\]. For each top quark mass and QCD coupling constant, the range of the prediction corresponds to the 1-sigma uncertainty in the measured Higgs mass, $m_h = (125.18 \pm 0.16)$ GeV. Within the uncertainties, $v'$ as small as few $10^8$ GeV is possible. Future measurements can pin down the scale $v'$ with an accuracy of few tens percent [@Dunsky:2019api].
![(Left) Running of the SM quartic coupling. (Right) Predictions for the scale $v'$ as a function of $m_t$.[]{data-label="fig:vpPrediction"}](Figures/lambdavpPlot.pdf){width=".95\textwidth"}
![(Left) Running of the SM quartic coupling. (Right) Predictions for the scale $v'$ as a function of $m_t$.[]{data-label="fig:vpPrediction"}](Figures/vpvsmtPlot.pdf){width=".95\textwidth"}
Kinetic Mixing {#sec:kineticMixing}
--------------
Even though quantum corrections to the kinetic mixing are small, [^2] no symmetry forbids a tree-level $\epsilon$ from being order unity in the effective Lagrangian . However, as shown in Fig. \[fig:epsilonConstraintPlot\], mirror electron DM with $\epsilon \gtrsim 10^{-8}$ is strongly constrained by nuclear and electron recoil experiments, ionization signals, and cosmology ([@Dunsky:2018mqs] and references therein.)
![Constraints on kinetic mixing if DM is composed of mirror electrons.[]{data-label="fig:epsilonConstraintPlot"}](Figures/mX_q_constraint.pdf){width="90.00000%"}
A natural explanation for such a small $\epsilon$ is that $SU(3) \times SU(2) \times U(1) \times SU(3)' \times SU(2)' \times U(1)'$ unifies into a larger gauge group with no abelian factors. Consequently, $\epsilon$ must vanish above the unification scale $v_{G}$ by gauge invariance.
For example, consider a theory where the SM gauge group and the mirror gauge group separately unify to $G \times G'$ at scale $v_{G}$, shown qualitatively in Fig. \[fig:EFT\]. Above $v_{G}$ the operators that induce kinetic mixing between the standard and mirror sectors are: $$\begin{aligned}
\frac{1}{2}\frac{c_6}{M_{\rm{Pl}}^2}(\Sigma F)(\Sigma' F') + \frac{1}{2}\frac{c_8}{M_{\rm{Pl}}^4} (\Sigma^2 F)(\Sigma'^2 F') + \mathcal{O}(1/M_{\rm{Pl}}^6)
\label{eq:GUTkineticmixing}\end{aligned}$$ where $F,F'$ are the gauge field strengths and $\Sigma,\Sigma'$ the Higgs fields. The first term is absent if $\Sigma$ is not an adjoint representation of $G$ or charged under some symmetry. When $\Sigma$ and $\Sigma'$ acquire a vacuum expectation value $v_{G}$, [^3] the higher dimensional operators in induce a kinetic mixing $\epsilon$ $$\begin{aligned}
\epsilon \simeq 3.5 \times 10^{-5} \, c_6 \left(\frac{v_{G}}{10^{16} ~\GeV}\right)^2 + 6.0 \times 10^{-10} \, c_8 \left(\frac{v_{G}}{10^{16} ~\GeV}\right)^4 + \mathcal{O}{(v_G^6/M_{\rm{Pl}}^6)}.
\label{eq:eps1}\end{aligned}$$
![Qualitative picture of the effective field theory at scales $v, v'$, and $v_{G}$. The gauge groups $G$ and $G'$ do not contain any abelian factors so that kinetic mixing can only be radiatively generated at the scale $v_G$ and below, or be induced by higher dimensional operators at $v_G$.[]{data-label="fig:EFT"}](Figures/EFT_vert.pdf){width="80.00000%"}
It is possible to freeze-in $e'$ as DM via the induced kinetic mixing of . As shown in Fig. \[fig:EFT\], the correct DM abundance can be produced for a kinetic mixing parameter $\epsilon \simeq 4 \times 10^{-11}$, essentially independent of DM mass. If the dim-6 coefficient $c_6$ is non-zero, the correct $e'$ DM abundance can be produced for the unification scale $v_{G} \simeq 1 \times 10^{13} \,{c_6}^{-1/2} ~ \GeV$. If $c_6$ vanishes, and the dim-8 coefficient $c_8$ is non-zero, the correct $e'$ DM abundance can be produced for $v_G \simeq 5 \times 10^{15} \,{c_8}^{-1/4} ~ \GeV$.
Direct detection and the Correlation with Proton Decay {#sec:DD}
======================================================
Direct Detection by Nuclear Recoils
-----------------------------------
Kinetic mixing induced from higher dimensional operators allows $e'$ dark matter to scatter electromagnetically with a nucleus. The Rutherford cross section for scattering between $e'$ and a nucleus of mass $m_N$ and atomic number $Z$, with relative velocity $v_{\rm rel}$ is given by $$\begin{aligned}
\frac{d \sigma} {dq} = \frac{8\pi \alpha^2 Z^2 \epsilon^2}{v_{\rm rel}^2 q^3} |F(q)|^2,\end{aligned}$$ where $q$ is the momentum transfer and $F(q)$ is the nuclear form factor. The number of expected events in a direct detection experiment with an energy threshold $E_{\rm th}$, a total target mass $M_{\rm tar}$, an exposure time $T$, and atomic weight $A$ is $$\begin{aligned}
N_{\rm event} = 1.6 \times \left( \frac{\epsilon}{ 10^{-8}} \right)^2 \frac{10^7 \, {\rm GeV}}{m_{e'}} \left( \frac{Z}{54} \right)^2 \left( \frac{131}{A} \right)^2 \frac{10{\rm \, keV}}{E_{\rm th}} \frac{f(E_{\rm th})}{0.3} \frac{M_{\rm tar} t}{{\rm ton}\times {\rm year}},
\label{eq:Nevent}\end{aligned}$$ where we assume a local DM density of $0.3$ GeV/cm$^3$, as well as a velocity distribution of $$\begin{aligned}
d v f(v) = dv \frac{4}{\sqrt{\pi}} \frac{v^2}{v_0^3} \, {\rm exp}(- v^2/v_0^2),~~v_0= 220 \, {\rm km}/{\rm s}.\end{aligned}$$ Here $f(E_{\rm th})$ takes into account the suppression of the scattering by the form factor, $$\begin{aligned}
f(E_{\rm th}) = \left[ \int_{q_{\rm th}}^{q_{\rm max}} d q |F(q)|^2 q^{-3} \right] / \left[ \int_{q_{\rm th}}^{q_{\rm max}} d q q^{-3} \right], \nonumber \\
q_{\rm th} = \sqrt{2 m_N E_{\rm th}},~ q_{\rm max}= 2 m_N v_{\rm rel}.\end{aligned}$$ Assuming the Helm form factor [@Helm:1956zz; @Lewin:1995rx], we find $f(E_{\rm th})\simeq 0.3$.
XENON1T searches for a recoil between DM and Xenon with a threshold energy around 10 keV [@Aprile:2018dbl]. The bound obtained there can be interpreted as an upper bound of 16 on the expected number of the events. Currently, the strongest bound on $\epsilon$ for $m_{e'} > 10^2 ~\GeV$ comes from XENON1T [@Dunsky:2018mqs], requiring $$\begin{aligned}
\epsilon < 1 \times 10^{-10} \, \left( \frac{m_{e'}}{10^2 ~\GeV} \right)^{1/2}\end{aligned}$$ as shown in Fig. \[fig:epsilonConstraintPlot\]. If $\epsilon$ is close to this bound, future experiments may detect $e'$ dark matter.
Correlation Between Proton Decay and Direct Detection
-----------------------------------------------------
Let us consider a case where the SM gauge group is embedded into a unified gauge group with heavy gauge bosons mediating proton decay. The proton decay rate is $$\begin{aligned}
\Gamma^{-1}(p \rightarrow \pi^0 e^+) \simeq 3 \times 10^{35}{\rm years} \; \left( \frac{v_G}{10^{16}~{\rm GeV}} \right)^4 \left(\frac{0.103 ~\GeV^2}{W_0}\right)^2,\end{aligned}$$ where $|W_0| = 0.103 \pm 0.041~\GeV^2$ encodes the relevant hadronic matrix element extracted from a lattice computation [@Aoki:2013yxa]. We also assume that below the heavy gauge boson mass scale the gauge group contains a $U(1)$ factor which eventually joins the $U(1)_Y$ gauge group. (This case excludes, for example, the Pati-Salam gauge group breaking at an intermediate scale.) The kinetic mixing is given by Eq. and we assume $c_6=0$. The direct detection rate $N_{\rm event}/M_{\rm tar}t$ of (\[eq:Nevent\]) and the proton decay rate are correlated with each other, $$\begin{aligned}
\Gamma^{-1}(p \rightarrow \pi^0 e^+) \simeq 3 \times 10^{35} {\rm years} \; \left( \frac{N_{\rm event}}{M_{\rm tar} t} \frac{{\rm ton} \times {\rm year}}{10} \right)^{1/2} \frac{1}{c_8} \left( \frac{v'}{2 \times 10^{9}~{\rm GeV}} \right)^{1/2},\end{aligned}$$ as shown in Fig. \[fig:ddProtonDecay\]. The blue region shows that if XENON1T were to detect a nuclear recoil signal, the proton lifetime would generally be longer than Hyper-Kamiokande could detect, for $c_8 = 1$. The orange region shows the analgous signal region for LZ. For $v' \leq 10^9 ~\GeV$, Hyper-Kamiokande and LZ both can detect correlating proton decay and nuclear recoil signals, respectively. If $c_8 > 1$, the kinetic mixing parameter is stronger for fixed $v_G$ so that nuclear recoil experiments and proton detect experiments may find correlating signals for $v' \gtrsim 10^9 ~\GeV$. For example, the dashed blue and orange contours of Fig. \[fig:ddProtonDecay\] show the reach of XENON1T and LZ, respectively, for $c_8 = 10$.
![Correlation between the proton decay rate and the DM-nuclear scattering rate as a function of $v'$. The rates are related as they both depend on the unification scale $v_G$ via higher-dimensional operators.[]{data-label="fig:ddProtonDecay"}](Figures/ddPlot.pdf){width="70.00000%"}
High and Low Reheat Scenarios; BBN and Dark Radiation {#sec:BBNandDR}
=====================================================
Since all the parameters of the SM have been determined, the only free parameters that affect the cosmology of the Mirror Higgs Parity theory are the reheat temperature after inflation and the portal parameters that connect the SM and mirror sectors. A key question is whether the two sectors were brought into thermal equilibrium after inflation.
At sufficiently high temperatures, the SM and mirror sectors are kept in thermal equilibrium by the Higgs portal; the sectors then decouple at a temperature $$\begin{aligned}
\label{eq:Tdec}
\frac{ T_{\rm dec} }{ v' } \simeq 10^{-3} \left(\frac{v'}{10^{9} ~\GeV}\right)^{1/3}.\end{aligned}$$ Our two cosmological scenarios correspond to whether the reheat temperature after inflation, $T_{\rm{RH}}$ is above or below $T_{\rm dec}$, and lead to very different mechanisms for the abundance of $e'$ and $u'$ dark matter. For $T_{\rm{RH}} > T_{\rm dec}$, the $u'$ and $e'$ abundances are given by freeze-out as the temperature drops below their masses, followed by dilution from $\nu'$ decay; for $T_{\rm{RH}} < T_{\rm dec}$ we assume that only the SM sector is reheated, so that DM arises from freeze-in. These two schemes for DM production are discussed in the next section.
In both high and low reheating cosmologies, long-lived mirror glueballs are produced whose decay products may yield substantial dark radiation or alter the relic abundances of light elements. In this section we study the general constraints on the maximum production of mirror glueballs. These results will be used in the next section to place limits on the high $T_{\rm{RH}}$ scheme and identify regions of parameter space that give signals of dark radiation and perturbed light element abundances. The mirror QCD confinement transition occurs when the mirror thermal bath cools to a temperature $T'_c = 1.26 \, \Lambda'_{\rm{QCD}}$ [@Borsanyi:2012ve]. At this point, the mirror bath contains only $\gamma'$ and $g'$ so that the ratio of entropies of the two sectors at $T'_c$ is about $r = (16/106.75)(T'_c/T_c)^3$. If the reheat temperature after inflation is greater than $T_{\rm{dec}}$, the two sectors were initially in thermal equilibrium and $r = (8/9)(g'_*(T_{\rm{dec}})/106.75)$. On the other hand, if the reheat temperature after inflation is below $T_{\rm{dec}}$, the two sectors were never in thermal equilibrium and ratio of temperatures $T'/T$ is generally much less than one.
![Mirror glueball decay to $\gamma' \gamma'$ (left) and $H, H^\dagger$ (right).[]{data-label="fig:glueballDecay"}](Figures/glueballDecayGammaGammaCropped.pdf){width="100.00000%"}
![Mirror glueball decay to $\gamma' \gamma'$ (left) and $H, H^\dagger$ (right).[]{data-label="fig:glueballDecay"}](Figures/glueballDecayHiggsHiggsCropped.pdf){width="100.00000%"}
After the mirror QCD transition, $g'$ confine to form mirror glueballs $S'$, whose energy density normalized by the entropy is given by $$\begin{aligned}
\frac{\rho_{S'}}{s} =\frac{3}{4} \, A \, r \, T'_c.\end{aligned}$$ The factor $A$ takes into account the non-trivial dynamics before and after the phase transition and is estimated in Appendix \[sec:rhos\]. $A=1$ corresponds to the limit where massless ideal gas of mirror gluons suddenly becomes pressure-less mirror glueballs at $T_c'$ and the mirror glueball number density is conserved afterward.
Mirror glueballs are typically long-lived. The lifetime of the mirror glueball is dominantly set by its decay rate to mirror photons, described by the dimension-8 operator $F'F' G'G'$, generated by a loop of mirror quarks of mass $m_{q'}$ and charge $Q'$ as shown in the left panel of Fig. \[fig:glueballDecay\]. After confinement this becomes a dimension-5 operator connecting $S'$ to $\gamma' \gamma'$ [@Juknevich:2009ji] $$\begin{aligned}
\Delta \mathcal{L}_{S' \to \gamma'\gamma'} \, &= \, \frac{Q'^2}{240\pi} \, \frac{ \alpha}{m_{q'}^4} \; \mathbf{F_{0^{++}}^{S'}} \; \; F'_{\mu \nu}F'^{\mu \nu} S'\end{aligned}$$ with matrix element $\mathbf{F_{0^{++}}^{S'}} = \langle 0 | 1/2 g_s^2 G^a_{\mu \nu} G_a^{\mu\nu} | 0^{++} \rangle \simeq 2.7 m_{S'}^3$ [@Meyer:2008tr]. Since the amplitude is dominated by the smallest $m_{q'}$, we take $q' = u'$ giving $Q' = 2/3$, so that the mirror glueball decay rate to mirror photons is $$\begin{aligned}
\Gamma_{S' \to \gamma' \gamma'} \, \simeq \, \frac{1}{16\pi}\left(\frac{2.7 \alpha}{270 \pi } \right)^2 \frac{m_{S'}^9}{m_{u'}^8}.\end{aligned}$$ The mirror glueball can also decay to the SM sector via the Higgs portal as shown by the right panel of Fig. \[fig:glueballDecay\]. The decay rate to $H H^\dagger$ is given by $$\begin{aligned}
\Gamma_{S' \to H H^\dagger} \, \simeq \, \frac{1}{8\pi}\left(\frac{2.7}{16 \pi^2}\right)^2 \frac{m_{S'}^5}{v'^4}.\end{aligned}$$ If its lifetime, $\Gamma_{S'}^{-1} \simeq (\Gamma_{S' \to \gamma' \gamma'} + \Gamma_{S' \to H H^\dagger})^{-1}$, exceeds about $1~\rm{s}$, $S'$ decays during BBN. If this occurs, $S'$ may inject substantial energy density, $\rho_{vis}$, into the SM hadronic sector altering the neutron to proton ratio before nucleosynthesis or disassociating light elements immediately after, leading to the constraint [@Kawasaki:2004qu] $$\begin{aligned}
\frac{\rho_{vis}}{s} = \frac{\Gamma_{S' \to H H^\dagger}}{\Gamma_{S'}} \, \frac{3}{4} \, A \,\frac{r}{D} \, T'_c \; \lesssim \; 10^{-14} ~\GeV.\end{aligned}$$ Here, $D$ is a generic dilution factor which may arise if there exists a particle which comes to dominate the energy density of the universe and decays before BBN, thereby injecting entropy into the SM thermal bath.
In the cosmology with $T_{\rm{RH}} > T_{\rm{dec}}$, mirror neutrinos are a natural candidate to provide such dilution since they are abundantly produced, decouple from the mirror bath while relativistic, and are long-lived. In this scenario, $D = T_{\rm{MD}, \nu'}/T_{\rm{RH},\nu'}$, where $T_{\rm{MD}, \nu'}$ is the temperature of the SM bath when $\nu'$ induced matter-domination begins and $T_{\rm{RH},\nu'}$ when it ends. If $T_{\rm{RH}} < T_{\rm{dec}}$, there is no particle in the mirror standard model to provide such dilution and $D = 1$. We show the BBN constraints as a function of $v'$ in Fig. \[fig:glueballConstraints\] in orange using the precise energy yield constraints calculated in [@Kawasaki:2004qu]. When $T_{\rm{RH}} > T_{\rm{dec}}$, $r$ is known so $D$ is constrained, as shown in the left panel of Fig. \[fig:glueballConstraints\]. When $T_{\rm{RH}} < T_{\rm{dec}}$, $D$ is known so $r$ is constrained, as shown in the right panel of Fig. \[fig:glueballConstraints\].
In addition, the energy deposited by $S'$ into mirror photons is constrained, even if $S'$ does not decay during BBN. The mirror photons behave as dark radiation, whose energy density is conventionally expressed as an excess in the effective number of neutrino $\Delta N_{\rm{eff}}$. For the high $T_{\rm{RH}}$ cosmology, with $\nu'$ decay leading to a dilution factor $D$, $\Delta N_{\rm{eff}}$ depends on whether $S'$ decays before, during, or after the $\nu'$ matter-dominated era $$\begin{split}
\Delta N_{\rm{eff}} &\simeq \frac{\Gamma_{S' \to \gamma' \gamma'}}{\Gamma_{S'}} \frac{4}{7}\left(\frac{43}{4}\right)^{4/3}\frac{r}{D} \frac{T'_c}{\sqrt{\Gamma_{S'}M_{\rm{Pl}}}} A
\\& \times
\begin{dcases}
\left(\frac{\pi^2}{10}\right)^{1/4}\frac{g_*(T_{\Gamma_{S'}})^{1/4}}{g_{*S}(T_{\Gamma_{S'}})^{1/3}} \frac{1}{D^{1/3}} & S' \text{ decays before MD}\\
\left(\frac{\pi^2}{10}\right)^{1/3} \left(\frac{T_{\rm{RH},\nu'}}{\sqrt{\Gamma_{S'}M_{\rm{Pl}}}}\right)^{1/3} & S' \text{ decays during MD} \\
\left(\frac{\pi^2}{10}\right)^{1/4}\frac{g_*(T_{\Gamma_{S'}})^{1/4}}{g_{*S}(T_{\Gamma_{S'}})^{1/3}} & S' \text{ decays after MD.}
\end{dcases}
\label{eq:Ycrit}
\end{split}$$ For the low $T_{\rm{RH}}$ cosmology few $\nu'$ are produced, so they do not give a matter dominated era and $D=1$. Contours of the dark radiation abundance produced from $S' \to \gamma' \gamma'$ are shown in Fig. \[fig:glueballConstraints\].
![$\Delta N_{\rm{eff}}$ contours (purple) and BBN constraints (orange) from $S' \to \gamma' \gamma', HH^\dag$. In the left (right) panel the two sectors were (were not) initially thermally coupled so that DM is thermally produced via freeze-out and dilution (freeze-in). The temperature ratio of the two sectors, $T'/T$, is evaluated at the mirror confinement temperature. For clarity, we take $A = 1$.[]{data-label="fig:glueballConstraints"}](Figures/FO_DNeff.pdf){width=".95\textwidth"}
![$\Delta N_{\rm{eff}}$ contours (purple) and BBN constraints (orange) from $S' \to \gamma' \gamma', HH^\dag$. In the left (right) panel the two sectors were (were not) initially thermally coupled so that DM is thermally produced via freeze-out and dilution (freeze-in). The temperature ratio of the two sectors, $T'/T$, is evaluated at the mirror confinement temperature. For clarity, we take $A = 1$.[]{data-label="fig:glueballConstraints"}](Figures/FI_DNeff.pdf){width=".95\textwidth"}
Cosmological Abundance of Mirror Dark Matter {#sec:cosmoAbundances}
============================================
Freeze-Out and Dilution from $\nu'$ Decay {#sec:fo}
-----------------------------------------
![Temperatures of the mirror bath around which each mirror fermion freezes-out (solid) and decays (dashed). Mirror temperatures of sector decoupling, $\nu'$ decoupling, as well as the mirror QCD phase transition, are shown as dotted lines. []{data-label="fig:tempFOplot"}](Figures/tempFOplot.pdf){width="80.00000%"}
In this section, we take the reheat temperature of the universe larger than the temperature at which the two sectors decouple, $T_{\rm{RH}} > T_{\rm dec}$. In this case, the relic abundances of mirror $e'$ and $u'$ dark matter are set by freeze-out followed by dilution from the late decays of $\nu'$.[^4] As the temperature of the universe drops, unstable mirror particles decay, while stable $e'$ and $u'$ annihilate and freeze-out. Although heavier mirror charged leptons and quarks are unstable, their decay widths are much smaller than their masses because of the large mirror electroweak scale. Fig. \[fig:tempFOplot\] shows the temperatures around which each particle freezes-out (solid lines) and decays (dashed lines). Here we ignore the effects caused by late decays of mirror neutrinos, and include them momentarily. For $v'$ in the range of $(10^8-10^{11})$ GeV, the $e'$ and $u'$ abundances are determined by the following processes in chronological order:
1. $b'$ freezes-out.
2. $c'$, $\mu'$ and $s'$ freeze-out. During these annihilations, $b'$ and $c'$ decay producing $c'$, $\mu'$ and $s'$. The annihilations also produce $e'$, $u'$ and $d'$, but they thermalize quickly.
3. $d'$, $u'$ and $e'$ freeze-out. During these annihilations, $s'$ and $\mu'$ partially decay producing $e'$, $u'$ and $d'$.
4. QCD’ phase transition occurs. Mirror hadrons composed of $s'$, $u'$ and $d'$ quickly annihilate. Mirror hadrons composed of $s'$ and $d'$ decay into $u'u'u'$.
We note that $\tau'$ is short-lived and does not affect the above processes. A set of Boltzmann equations describing the freeze-out dynamics is shown in Appendix \[sec:boltzmann\].
We elaborate on the fourth process. After the mirror QCD phase transition, mirror quarks are tied with each other by strings and form bound states. For $v' < 10^{10} ~\GeV$, the Coulomb binding energy of mirror hadrons containing a $u'$ or $d'$ is comparable to $T'_c$ [@Harigaya:2016nlg], and so an $O(1)$ fraction of these mirror quarks form loosely bound states with large radii $\sim \Lambda_{\rm QCD}'$. With such a large cross-section, these mirror hadrons scatter among themselves efficiently, rearranging their quark constituent until they contain a $q' \bar{q}'$ pair, and subsequently annihilate into $\gamma'$ [@Kang:2006yd; @Harigaya:2016nlg]. For $v' > 10^{10} ~\GeV$ the Coulomb binding energy of mirror hadrons is larger than $T'_c$, and so most of the mirror quarks initially form tightly bound states with a smaller radius $\sim (m_{q'} \alpha_S')^{-1}$ [@DeLuca:2018mzn]. Nevertheless, these tightly bound states still have a relatively large radius and scatter and annihilate relatively efficiently. The mirror baryon containing only mirror strange quarks, $s's's'$, generally forms a tightly bound state for all $v'$. Still, $s'$ annihilates efficiently so that its beta decay contributions to $e'$ are small. The thermal abundances of $e'$ and $u'$ are shown in Fig. \[fig:abundances\]. The solid lines conservatively assume that the annihilation cross-section of mirror hadrons is $\pi / (m_{q'} \alpha_S')^2$. The abundance of $e'$ does not change even if the cross-section is as large as $ \Lambda_{\rm QCD}'^{-2}$. For comparison, the dashed line assumes mirror hadrons completely cease annihilating after confinement. Even though the annihilation cross-section of $e'$ does not change in either case, the relic abundance of $e'$ drops when annihilations of mirror hadrons continue after the QCD’ phase transition since any beta decays from $s'$ or $d'$ that produce $e'$ below $T'_c$ are effectively absent (see Fig. \[fig:tempFOplot\]). To the left of the vertical dotted line, the QCD’ phase transition occurs before $u'$ freezes-out, which is why its abundance dramatically increases if hadronic annihilations are assumed to cease below $T'_c$.
![The cosmological abundance of mirror electrons and up quarks from freeze-out and from decays of heavier charged mirror fermions. Dilution from mirror neutrino decays is not included.[]{data-label="fig:abundances"}](Figures/foTotalPlot.pdf){width="80.00000%"}
We see from the solid lines of Fig. \[fig:abundances\] that $e'$ is the dominant component of DM. On the other hand, efficient annihilations after the QCD$'$ phase transition make $u'$ a small component of DM, which exists today in the form of mirror hadrons like $u'u'u'$. For all $v' > 6 \times 10^7 ~\GeV$, the thermal abundance of $e'$ is too large to be DM. This is problematic as such a low $v'$ requires $m_t$ and $\alpha_S(m_Z)$ to lie beyond their current $3 \sigma$ experimental values.
Nevertheless, in the above discussion, we have ignored mirror neutrinos which are cosmologically stable if $m_{\nu'} < m_{e'} + m_{u'} + m_{d'}$ and $M_D$ of (\[eq:L\]) is sufficiently large. The former prevents decays to the mirror sector, due to mirror fermion number and mirror electromagnetic charge conservation, and the latter suppresses decays to the SM sector. However, as $M_D$ is reduced, mirror neutrinos can decay well after they becoming non-relativistic to SM particles, thereby diluting $e'$ and $u'$. Consequently, the $v'$ required to produce $e'$ DM shifts to higher scales.
Shortly after the two sectors decouple at $T_{\rm{dec}}$, $\nu'$ decouple from the mirror thermal bath as the mirror weak interaction rate drops below the Hubble expansion rate, as shown in Fig. \[fig:tempFOplot\]. Since $T_{\rm{dec}} = T'_{\rm{dec}} \gg m_{\nu'}$, $\nu'$ decouple while relativistic with an initial yield $Y_{\nu'} \simeq n_{\nu'}(T_{\rm{dec}})/s(T_{\rm{dec}}) = 0.0123$. With this initial abundance, if $\nu'$ are sufficiently long-lived they dominate the energy density of the universe prior to decaying.
### One generation of long-lived $\nu'$
For our first example, we assume that two flavors of $\nu'$ decay rapidly and study $e'$ dilution from decays of the single long-lived flavor. The long-lived $\nu'$ decays to $\ell H$ via the neutrino portal operator of (\[eq:L\])[^5] with a decay rate $$\begin{aligned}
\Gamma_{\nu' \to l h} = \frac{m_{\nu'}}{8 \pi}\frac{v'^2}{M_D^2}.
\label{eq:nulhdecay}\end{aligned}$$ The mass of the mirror neutrino is given by Eq. (\[eq:mnu\]), and for sufficiently large $v'$, the mirror neutrino is massive enough that it can beta decay into $e', u'$ and $\bar{d}'$, with a decay rate $$\begin{aligned}
\Gamma_{\nu' \to e'u'\bar{d}'} = \frac{3}{8}\frac{1}{192 \pi^3}\frac{m_{\nu'}^5}{v'^4}.\end{aligned}$$ When $\nu'$ dominantly decay into the SM sector, the decay products heat up the SM thermal bath, thereby diluting the frozen-out abundance of $e'$ and $u'$ relative to $n_\gamma$ by a factor $$\begin{aligned}
D = \frac{T_{\rm{MD},\nu'}}{T_{\rm{RH},\nu'}} \simeq \frac{m_{\nu'}Y_{\nu'}}{1.2 (\Gamma_{\nu'}M_{\rm{Pl}})^{1/2}}\left(\frac{\pi^2}{10g_{*RH}}\right)^{1/4}.
\label{eq:dil}\end{aligned}$$ ($\Gamma_{\nu'})^{-1} = (\Gamma_{\nu' \to l h} + \Gamma_{\nu' \to e'u'\bar{d}'})^{-1}$ is the lifetime of the mirror neutrino. The numerical factor of $1.2$ is taken from [@Harigaya:2018ooc]. We solve the Boltzmann equation for the abundance of mirror fermions in Appendix \[sec:boltzmann\], including freeze-out, the change of the expansion rate during the mirror neutrino matter-dominated era, and dilution from $\nu'$ decays. An approximation for the resulting $e'$ yield from freeze-out and dilution is $$\begin{aligned}
\frac{\rho_{e',\rm{FO}}}{s} \approx 35 \, \frac{m_{e'}^2}{\pi \alpha^2}\frac{1}{M_{\rm{Pl}}} \frac{g_*^{1/2}}{g_{*S}} \frac{1}{D} \approx 5 \times 10^{-6} \frac{v'^2 v}{\sqrt{M_{\rm{Pl}} m_\nu}}\frac{1}{M_D}
\label{eq:FOest}\end{aligned}$$ where $D$ is the dilution factor provided by mirror neutrino decays .
![Constraints on $(v', m_{\nu})$ when $e'$ dark matter arises from freeze-out and dilution from one long-lived species of $\nu'$. Here $m_\nu$ is the mass of the neutrino that is the Higgs Parity partner of the long-lived $\nu'$. Purple contours show $\Delta N_{\rm eff}$ resulting from decays of $S'$ to $\gamma'$. Vertical gray contours show $v'$ when $m_t$ and $\alpha_S(m_Z)$ deviate from their central values by $0$ to $3\sigma$.[]{data-label="fig:vp-mnu"}](Figures/mnuConstraintPlotep_1gen.pdf){width="70.00000%"}
For a given $(v', m_{\nu}')$, the parameter $M_D$ is determined to yield the correct $e'$ DM abundance. Furthermore, the resulting values of $M_D$ are large enough that $m_{\nu}'$ can be mapped to $m_\nu$ by the scaling $$\begin{aligned}
m_\nu = m_{\nu'} \frac{v^2}{v'^2}.\end{aligned}$$ Further constraints on this scenario are shown in the $(v', m_{\nu})$ plane in Fig. \[fig:vp-mnu\].
In the allowed white region, we find $M_D$ must lie within the range $(10^{18} - 10^{23}) ~\GeV$. In the red-shaded region, the $e'$ abundance is smaller than the dark matter abundance without dilution. For too small a neutrino mass, the required $T_{\rm{RH},\nu'} \approx \sqrt{\Gamma_{\nu'}M_{\rm{Pl}}}$ to reproduce the dark matter abundance is below the MeV scale and affects BBN as well as the effective number of neutrinos [@Kawasaki:1999na; @Kawasaki:2000en]. We adopt the bound $T_{\rm{RH},\nu'} > 4$ MeV [@deSalas:2015glj], excluding the pink-shaded region. In the blue-shaded region, the mirror beta decay $\nu'\rightarrow e' u'\bar{d}'$ is kinematically allowed, creating too much $e'$ and $u'$ abundance. In the orange-shaded region the sum of the SM neutrino masses are above 0.3 eV, which is disfavored by the observations of the Cosmic Microwave Background (CMB) [@Aghanim:2018eyx]. The gray-shaded region is excluded at the $3\sigma$ level from measurement of $\alpha_S$ and the Higgs and top masses. If the long-lived species is the lightest $\nu'$ then beta decay to $\nu' e'\bar{e}'$ cannot occur. However, if the long-lived $\nu'$ is one of the heavier states, then the lightly green-shaded region of Fig. \[fig:vp-mnu\] is also excluded since the long-lived $\nu'$ creates $e'$. The corresponding SM neutrino mass should be above $\Delta m^2_{31(23)}$, excluding the lightly yellow-shaded region. The allowed region is not large: $m_t$ should be above its present central value and, remarkably, the neutrino mass must be within a factor of 10 of its present upper bound of $0.1$ eV.
In the resulting allowed region of parameter space for $e'$ dark matter, the purple contours show our prediction for $\Delta N_{\rm{eff}}$ from decays of mirror glueballs, produced at the QCD$'$ confining transition, to mirror photons. Throughout the entire region $\Delta N_{\rm{eff}}$ is in the range $0.03 \mathchar`- 0.4$, allowed by Planck [@Aghanim:2018eyx] and within range of the sensitivities of CMB Stage IV experiments [@Abazajian:2016yjj].
### Universal coupling strength of neutrino portal
As a second illustration of $e'$ freeze-out and dilution from $\nu'$ decays, we take the strength of the neutrino portal coupling to be independent of generation. Thus, in a neutrino mass basis, we take $\nu' \rightarrow lH$ decays to be given by (\[eq:nulhdecay\]) for all three generations of $\nu'$. To avoid overproducing $e'$, all three $\nu'$ must be light enough that beta decay is forbidden. Thus the total decay rate of each mirror neutrino is given by and is proportional to $m_{\nu'}$. Consequently, the dilution is dominated by the heaviest mirror neutrino. For a normal hierarchy ($m_{\nu_1} \ll m_{\nu_2} < m_{\nu_3}$) of SM neutrinos, the mirror neutrino responsible for dilution is $\nu'_3$; for an inverted hierarchy ($m_{\nu_3} \ll m_{\nu_1}<m_{\nu_2}$), $\nu'_{2,1}$ give comparable dilutions; and for a quasi-degenerate spectrum $\nu'_{3,2,1}$ all give comparable dilutions.
The bounds from BBN, too much dark matter from $\nu'\rightarrow e' u'\bar{d}'$ decay, and too little dark matter from freeze-out are approximately as in Fig. \[fig:vp-mnu\], with the vertical axis interpreted as the heaviest neutrino, which is constrained by oscillation data to be at or above 0.05 eV. Thus the larger values of $v'$ and $\Delta N_{\rm{eff}}$ are excluded in this case. The upper bound on the heaviest neutrino from the cosmological limit on the sum of the neutrino masses is 0.1 eV.
![Constraints on $(v', m_{\nu})$ when $e'$ dark matter arises from freeze-out and dilution from $\nu'$ with universal neutrino portal couplings. Here $m_\nu$ is the mass of the [*lightest*]{} neutrino. Purple contours show $\Delta N_{\rm eff}$ resulting from decays of $S'$ to $\gamma'$. Vertical gray contours show $v'$ when $m_t$ and $\alpha_S(m_Z)$ deviate from their central values by $0$ to $3\sigma$. In the allowed white region, $\Delta N_{\rm{eff}}$ is always greater than $0.03$, which will be probed by CMB Stage IV [@Abazajian:2016yjj].[]{data-label="fig:vp-mnu3"}](Figures/mnuConstraintPlotep_3gen.pdf){width="70.00000%"}
In addition to these bounds, there is a constraint from the decay $\nu_3' \rightarrow \nu_{1,2}' e'\bar{e}'$ for a normal hierarchy or $\nu_{2,1}' \rightarrow \nu_{3}' e'\bar{e}'$ for an inverted hierarchy. In either case, too much $e'$ is produced. Regardless of whether the SM neutrinos obey a normal or inverted hierarchy, this constraint can be translated to a bound on the *lightest* SM neutrino: $$\begin{aligned}
m_{\nu,{\rm lightest}} > \frac{\Delta m_{31}^2}{4 m_e} \frac{v'}{v} - m_e \frac{v}{v'} .
\label{eq:nu'cascade}\end{aligned}$$ $\Delta m_{31}^2 \equiv |m_3^2 - m_1^2| \simeq (0.05 ~{\rm eV})^2$ is the atmospheric neutrino mass difference squared and $m_e$ is the electron mass. We have made the good approximation that $\Delta m_{31}^2$ is also the mass squared difference between the lightest and heaviest SM neutrino in an inverted hierarchy. This bound is shown in the yellow hatched region of Fig. \[fig:vp-mnu\].
The constraints on this scheme for $e'$ dark matter are shown in Fig. \[fig:vp-mnu3\], where the vertical axis is the [*lightest*]{} SM neutrino mass. The bound of appears in green. If $v'$ turns out to be larger than $4\times 10^9$ GeV, the lightest neutrino mass is predicted to be in a narrow range. The lightest mirror neutrino is longer-lived than the heaviest mirror neutrino for a universal $M_D$, but decays before the onset of the BBN for $m_\nu > 10^{-3}$ eV.
The sum of the masses of the three neutrinos can be constrained through its imprint on the structure of the universe. Future measurements of the CMB, BAO, and 21 cm emission are expected to determine the sum of the masses with an uncertainty of $10$ meV [@Font-Ribera:2013rwa; @Allison:2015qca; @Archidiacono:2016lnv]. One can check the consistency of the the measurements and the bounds we have obtained.
During the matter dominated era by $\nu'$, cosmic perturbations of massive components can grow. Since $e'$ tightly couples to mirror photons, the perturbation of $e'$ does not grow by itself. The perturbation of mirror glueballs grows, decays into mirror photons, which scatter with $e'$ and grow the perturbation of $e'$, like the growth of a weakly interacting massive particle during a matter dominated era [@Choi:2015yma]. We will discuss the implication of the growth to the future searches for ultra compact mini halo elsewhere.
Freeze-In from Higgs Portal and Kinetic Mixing {#sec:fi}
----------------------------------------------
In this section, we consider the relic abundances of mirror $e'$ when the reheat temperature of the universe is below $T_{\rm dec}$ and only the SM sector is reheated. Since the SM and mirror sectors are weakly coupled below $T_{\rm{dec}}$, mirror DM is produced via freeze-in through the Higgs portal, as shown in Fig. \[fig:fpHiggsPortal\].
![Freeze-in production of mirror fermions (left) and mirror gauge bosons (right) through the Higgs portal.[]{data-label="fig:fpHiggsPortal"}](Figures/HHee_Cropped.pdf){width=".675\textwidth"}
![Freeze-in production of mirror fermions (left) and mirror gauge bosons (right) through the Higgs portal.[]{data-label="fig:fpHiggsPortal"}](Figures/HHFF_Cropped.pdf){width="100.00000%"}
Although the mirror fermion and gauge boson production rates are UV-dominated, the entropy production during reheating negates far-UV production so that the dominant production occurs around $T_{\rm{RH}}$. Reheat temperatures below the mirror electron mass yield insufficient $e'$ to be DM since the small $e'$ freeze-in abundance is further diluted by $(m_{e'}/T_{\rm{RH}})^6$ as production almost ceases below $T \approx m_{e'}$.[^6] Consequently, we focus on $T_{\rm{RH}} \gtrsim m_{e'}$. A set of Boltzmann equations describing the freeze-in dynamics is shown in Appendix \[sec:boltzmann\]. The thermal evolution of the mirror electrons is as follows:
At $T_{\rm{RH}}$, the mirror electrons carry a typical energy $T_{\rm{RH}}$ and a freeze-in number density[^7] $$\begin{aligned}
n(T_{\rm{RH}}) \, = \, \frac{4}{9} \frac{n_H(T_{\rm{RH}})^2}{H(T_{\rm{RH}})} \; \langle \sigma v (T_{\rm{RH}})\rangle.
\label{eq:nFI}\end{aligned}$$ $n_H$ is the SM Higgs thermal number density, $H$ is Hubble, and $\langle \sigma v\rangle$ is the freeze-in cross-section given by $$\begin{aligned}
\langle \sigma v(T_{\rm{RH}})\rangle = \frac{1}{8 \pi} \frac{y_e^2}{v'^2}.\end{aligned}$$ For all $v'$, the frozen-in abundance of $e'$ at $T_{\rm{RH}}$ exceeds that of dark matter for $T_{\rm{RH}} \gtrsim m_{e'}$. For $v' \gtrsim 4 \times 10^8 ~\GeV$, annihilations of $e'$ are ineffective during subsequent freeze-out. The freeze-in yield of $e'$ from the Higgs portal is $$\begin{aligned}
\frac{\rho_{e',\rm{FI}}}{s} &\approx 0.01 \frac{1}{(g_*)^{1/2} g_{*S}} \frac{y_e^3}{v'}T_{\rm{RH}} M_{\rm{Pl}} &\text{(Higgs Portal)}
\label{eq:hPortalFI}\end{aligned}$$ In this regime, a reheat temperature approximately equal to the mirror electron mass reproduces the correct DM abundance, as shown in Fig. \[fig:vp-trh\].
For $v' \lesssim 4 \times 10^8 ~\GeV$, annihilations of $e'$ are effective during subsequent freeze-out and the allowed $T_{\rm{RH}}$ rises, as shown in Fig. \[fig:vp-trh\]. However, as $T_{\rm{RH}}$ increases, mirror fermions heavier than $e'$ are produced at $T_{\rm{RH}}$, which transfer much of their abundance to $\gamma'$ and $e'$ as they annihilate and thermalize via $2 \to 2$ and $2 \to 3$ processes as discussed in Appendix \[sec:boltzmann\].
For $T_{\rm{RH}} \geq T_{\rm{dec}}$, the two sectors were once in thermal equilibrium and the situation reverts to traditional freeze-out discussed in Sec. \[sec:fo\]. $\Delta N_{\rm eff}$ and BBN constraints from frozen-in mirror glueball decays are not shown in Fig. \[fig:vp-trh\] as they are much weaker than the bound on overproduction of $e'$ DM.
![Constraints on the mirror electroweak scale $v'$ and the reheat temperature $T_{\rm{RH}}$ of the universe. In the blue region, $e'$ is overproduced via freeze-in from the Higgs portal. In the red region, the required $\epsilon$ to freeze-in $e'$ as DM via the kinetic mixing portal (shown by the dotted counters) is large enough to produce nuclear recoil signals in XENON1T. In the orange region, the reheat temperature is high enough that the two sectors were originally thermally coupled and the freeze-in regime reduces to the freeze-out regime (see Sec. \[sec:fo\]).[]{data-label="fig:vp-trh"}](Figures/totalvpvsTRHPlot_glue.pdf){width="60.00000%"}
Finally, as mentioned in Sec. \[sec:kineticMixing\], $e'$ DM can also be frozen-in via kinetic mixing induced from higher dimensional operators . On one hand, the freeze-in abundance of $e'$ through the Higgs portal is dominantly set by its yukawa coupling, which is fixed and whose smallness prevents sufficient $e'$ to be produced as DM for $T_{\rm{RH}} < m_{e'}$. On the other hand, the freeze-in abundance of $e'$ through kinetic mixing is set by $\epsilon$, which is a free parameter (indirectly set by the unification scale $v_{G}$), and whose value can be chosen to sufficiently produce $e'$ DM for reheat temperatures as low as $\sim m_{e'}/25$.
For $T_{\rm{RH}} < m_{e'}$, the freeze-in yield of $e'$ from kinetic mixing is $$\begin{aligned}
\frac{\rho_{e',\rm{FI}}}{s} &\approx 0.02 \, \pi \alpha^2 \epsilon^2 M_{\rm{Pl}}\left(\frac{m_{e'}}{T_{\rm{RH}}}\right)^2 \exp\left(-\frac{2m_{e'}}{T_{\rm{RH}}}\right) &\text{(Kinetic Mixing)}
\label{eq:kineticMixingFI}\end{aligned}$$ The black dotted contours in the region $T_{\rm{RH}} < m_{e'}$ of Fig. \[fig:vp-trh\] show the $\epsilon$ necessary for $e'$ to be frozen-in as DM. The shaded red region is excluded if $e'$ is the DM since the required $\epsilon$ to freeze-in $e'$ DM via kinetic mixing is large enough to already produce recoil signals at XENON1T[^8]. A similar calculation for the proposed LZ experiment, which can probe $\epsilon$ an order of magnitude smaller, produces the green contour ‘LZ’. For low $v'$, LZ has the potential to probe nearly all reheat temperatures capable of freezing-in $e'$.
Gravitational Waves from Mirror QCD phase transition {#sec:GW}
====================================================
In the range of $v'$ consistent with the observed top quark mass, mirror quark masses are much larger than the mirror QCD scale. The mirror QCD phase transition is then first order [@Yaffe:1982qf; @Svetitsky:1982gs]. The phase transition proceeds by nucleation of bubbles, which collide with each other and produce gravitational waves [@Witten:1984rs].
We consider the case where the $e'$ dark matter abundance is set by freeze-out followed by dilution from late $\nu'$ decays. The abundance of gravitational waves $\Omega_{\rm GW,col} h^2$ directly produced by the bubble collisions as a function of a frequency $f$ is given by $$\begin{aligned}
\frac{d \, \Omega_{\rm GW,col} \, h^2}{d \, {\rm ln} f} \; \simeq \; & 2 \times 10^{-8} \; \frac{ (f / f_p)^3}{ 0.3 + (f/f_p)^4} \, \left( \frac{10}{\beta/H} \right)^2 D^{-4/3} \left( \frac{\rho_{g'}/\rho_{\rm tot}}{2/3} \frac{\rho_{\rm lat}}{\rho_{g'}} \frac{\rho_{\rm kin}}{\rho_{\rm lat}} \right)^2 \frac{\rho_{\rm tot} / \rho_{\rm SM}}{3} , \label{eq:GW}\\
f_p \; \simeq \; & 2\times 10^{-5} \, {\rm Hz} \left(\frac{\beta/H}{10} \right) \, \left( \frac{T'_c}{ 100\, {\rm GeV}} \right) D^{-1/3} \; \left( \frac{g_{\rm dec}'}{60} \frac{100}{g_{\rm dec}} \right)^{1/3} \; \left( \frac{ \rho_{\rm tot} / \rho_{g'} } {3/2} \right)^{1/2} \left( \frac{b'}{0.5} \right)^{1/6}.\end{aligned}$$ $f_p$ is close to the frequency at the peak of the distribution and $T'_c \simeq 1.3 \, \Lambda'_{\rm QCD}$ is the temperature of the mirror QCD phase transition. Here we use the results of Ref. [@Huber:2008hg], assuming that the velocity of the bubble wall is the speed of light, and take into account the dilution $D$ from $\nu'$ decay. The ratio $(\beta/H)$ parametrizes the duration of the phase transition $\beta^{-1}$ in comparison with the Hubble time scale $H^{-1}$. $\rho_{\rm tot}$ is the total energy density, $\rho_{g'}$ is the energy density of the mirror gluon bath, $\rho_{\rm lat}$ is the latent heat of the phase transition, $\rho_{\rm kin}$ is the kinetic energy of the bubble wall and $\rho_{\rm SM}$ is the energy density of the SM bath, all of which are evaluated at the phase transition. $g_{\rm dec}$ and $g_{\rm dec}'$ are the degrees of freedom of the SM and the mirror sector at the decoupling of the two sectors, respectively. $b'$ parametrizes the energy density of the mirror gluon just before the phase transition, $\rho_{g'} = b' {T_{\rm QCD}'}^{4}$. The ratio $\rho_{\rm SM} / \rho_{g'}$ is estimated in Appendix \[sec:rhos\].
![Gravitational wave spectrum generated by the mirror QCD phase transition for $\beta/H = 10$ (left) and $\beta/H = 100$ (right). Future gravitational wave detectors such as LISA and BBO may detect a signal if $m_t$ and $\alpha_S(m_Z)$ lie more than $2\sigma$ away from their current central values.[]{data-label="fig:GW"}](Figures/gwSpectrumPlot.pdf){width="100.00000%"}
Gravitational waves are also produced by the turbulent motion of fluids induced by the bubbles [@Kamionkowski:1993fg]. The abundance of such gravitational waves $\Omega_{\rm GW,tub} h^2$ is $$\begin{aligned}
\frac{d \, \Omega_{\rm GW,tub} \, h^2}{d \, {\rm ln} f} \; \simeq \; & 4 \times 10^{-9} \; \frac{9 (f/f_p)^3}{( f / f_p + 0.02 H/\beta) ( f / f_p + 0.8 )^{11/3}} \left( \frac{10}{\beta/H} \right)^2 D^{-4/3} \nonumber \\
& \times \left( \frac{\rho_{g'} / \rho_{\rm tot}}{2/3} \frac{\rho_{\rm lat}}{\rho_{g'}} \frac{\rho_{\rm kin}}{\rho_{\rm lat}} \right)^{3/2} \frac{\rho_{\rm tot} / \rho_{\rm SM}}{3}
\label{eq:GW2} \\
f_p \; \simeq \; & 1\times 10^{-4} {\rm Hz} \;\; \left(\frac{\beta/H}{10} \right) \, \left( \frac{T'_c}{ 100\, {\rm GeV}} \right) D^{-1/3} \; \left( \frac{g_{\rm dec}'}{60} \frac{100}{g_{\rm dec}} \right)^{1/3} \; \left( \frac{ \rho_{\rm tot} / \rho_{g'} } {3/2} \right)^{1/2} \left( \frac{b'}{0.5} \right)^{1/6}.\end{aligned}$$ Here we use the results of Refs. [@Caprini:2009yp; @Caprini:2010xv] assuming that the bubble walls expand at the speed of light.[^9] Numerically, this contribution is smaller than the one from the bubble collision.
The prediction (\[eq:GW\], \[eq:GW2\]) for the gravity wave spectrum depends on $v'$ via $T'_c$ and especially $D$. With $v'$ determined by the top quark mass, we show in Fig. \[fig:GW\] the prediction for the spectrum of the gravitational waves for various $m_t$, taking $\beta/H$ of (10,100) in the (left, right) panel. The dashed and dotted lines show the contribution from the bubble collision and the turbulent motion respectively, and the solid lines show the sum of them. In the blue shaded region, the freeze-out followed by the dilution from $\nu'$ fails as is shown in Fig. \[fig:vp-mnu\]. The ratio $(\beta/H)$ is likely to be $O(100)$ [@Hogan:1984hx]. If the top quark mass is large enough, gravitational waves can be detected by future experiments such as LISA, DECIGO and BBO [@Moore:2014lga]. We note that prediction for the gravitational wave spectrum assumes that the phase transition occurs before the $\nu'$ matter-dominated era. This condition is satisifed in the region where future experiments may detect the gravitational wave spectrum, that is, at the $2-3 \, \sigma$ level for $m_t$ and $\alpha_S(m_Z)$.
We also note that many aspects of the phase transition in QCD-like theories, such as ($\beta/H$) and $\rho_{\rm kin} / \rho_{\rm lat}$, are not well-understood because of the non-perturbative nature. Once the phase transition is well-understood, it will become possible to check the consistency of future measurements of the top quark mass and the gravitational wave spectrum.
Conclusions and Discussions
===========================
We have introduced the Mirror Higgs Parity theory, described by (\[eq:L\]). The entire SM Lagrangian, including dimension 5 operators for neutrino masses, is replicated by Higgs parity and the only unknown parameters are those of the kinetic mixing, Higgs and neutrino portals that connect the two sectors. The spectrum of the mirror sector is a scaled up version of the SM spectrum, as shown for the light mirror particles in Fig. \[fig:mirrorSpectrum\]. The scaling depends only on the Higgs Parity breaking scale $v'$, which sets the scale at which the SM Higgs quartic vanishes and will become better determined by precision measurements of $(m_t, \alpha_S)$.
There are several interesting theories containing the Higgs Parity mechanism for the vanishing of the Higgs quartic at high energies. Mirror Higgs Parity is the simplest theory where the Higgs Parity partner of the electron, $e'$, is dark matter, with an abundance set by thermal mechanisms. Direct detection of $e'$ dark matter can occur via kinetic mixing and leads to a recoil spectrum characteristic of photon exchange. The present bound from XENON1T and the future reach of LZ on the kinetic mixing parameter $\epsilon$ are shown in Fig. \[fig:epsilonConstraintPlot\].
If the SM gauge group is unified at scale $v_G$ into a group such as SU(5), the proton decay rate scales as $\Gamma_p \propto 1/v_G^4$. Furthermore, since kinetic mixing vanishes in the unified theory, it may arise from a higher dimensional operators, such as in Eq. (\[eq:GUTkineticmixing\]), leading to $\epsilon \propto v_G^n$, where $n$ is a model-dependent, positive integer. Thus proton decay excludes small $v_G$ and direct detection excludes large $v_G$. The correlation of these two rates for $n=4$ is shown in Fig. \[fig:ddProtonDecay\]. A large fraction of the allowed parameter space of the theory will be probed by a combination of Hyper-Kamiokande and LZ.
For large values of the reheat temperature after inflation, $T_{\rm{RH}}$, the SM and mirror sectors reach thermal equilibrium via the Higgs portal interaction. The $e'$ relic abundance arises first from freeze-out and is then diluted by $\nu'$ decay to $\ell H$. Fixing the neutrino portal parameters to obtain the observed abundance, the remaining relevant parameters are $v'$, which determines $m_{e'}$, and $m_\nu$ which determine $m_{\nu'}$. The constraints on this scheme for dark matter are shown in the $(v', m_\nu)$ plane in Fig. \[fig:vp-mnu\], for the case that dilution is dominated by a single $\nu'$. Remarkably, the corresponding neutrino is required to have a mass larger than 0.01 eV, in the range of masses determined from oscillation data. Furthermore, $v'$ must be in the range of $(10^8 - 10^{10})$ GeV, overlapping the allowed range determined by requiring the Higgs quartic to vanish at $v'$. Within this allowed unshaded region of Fig. \[fig:vp-mnu\], we predict the contribution to dark radiation arising from decays of mirror glueballs to mirror photons. The resulting $\Delta N_{\rm eff}$, shown by purple contours, varies from about 0.04 to 0.4, and is highly correlated with $v'$ and therefore with $m_t$.
Since all the mirror quarks are much heavier than the mirror confining scale, the mirror QCD phase transition, which occurs at $T' \sim (40-1000)$ GeV for $v' = (10^8 - 10^{12})$ GeV, is first order and produces gravitational waves from bubble dynamics and turbulent fluid motion at the transition. The spectral energy density today, normalized to the critical energy density, is then obtained by including the $\nu'$ decay dilution factor, and is shown in Fig. \[fig:GW\]. Part of the allowed region of the theory can be probed by LISA, DECIGO and BBO, and a gravity wave signal in these experiments would be correlated with $m_t$ and $\Delta N_{\rm eff}$.
For low values of the reheat temperature after inflation, $T_{\rm{RH}}$, $e'$ DM can arise via freeze-in production. The observed DM abundance may be obtained anywhere in the unshaded region of Fig. \[fig:vp-trh\]. On the edge of the blue shaded region this occurs via the Higgs portal, which is UV dominated around $T_{\rm{RH}}$. In the rest of the unshaded region this occurs via kinetic mixing, dominated at temperatures near $m_{e'}$, for a suitable value of $\epsilon$.
Mirror Higgs Parity exchanges $SU(3)$ with $SU(3)'$ and hence does not solve the strong CP problem. One possible solution is to introduce a QCD axion [@Peccei:1977hh; @Peccei:1977ur; @Weinberg:1977ma; @Wilczek:1977pj]. If Higgs Parity transforms the QCD axion into a mirror QCD axion, the mirror QCD axion is an axion-like-particle with a mass $$\begin{aligned}
m_{a'} = 0.6 \frac{\Lambda_{\rm QCD}^{'2}}{f_a} = 0.4~{\rm keV} \left( \frac{v'}{10^9~{\rm GeV}} \right)^{8/11} \frac{10^{10}~{\rm GeV}}{f_a},
\label{eq:ma'}\end{aligned}$$ where the topological susceptibility is taken from [@Durr:2006ky]. The mass is correlated with $v'$ and hence with the top quark mass. Both axions may contribute to the dark matter density.
Alternatively, if the QCD axion is neutral under Higgs Parity it couples to QCD and mirror QCD with the same decay constant. Since Higgs Parity ensures the equality of the theta angles in the two sectors, the strong CP problem is still solved [@Rubakov:1997vp; @Berezhiani:2000gh; @Hook:2014cda; @Fukuda:2015ana]. The mass is given by Eq. (\[eq:ma’\]). An advantage of such a heavy axion is that it is easier to understand the PQ symmetry as an accidental symmetry [@Fukuda:2015ana]. In this case, it is even possible to have a small decay constant $\ll 10^9$ GeV, since the large mass prevents the production of axions in stellar objects and meson decays. We will discuss the phenomenology of axion dark matter in Mirror Higgs Parity in future works.
Acknowledgement {#acknowledgement .unnumbered}
===============
This work was supported in part by the Director, Office of Science, Office of High Energy and Nuclear Physics, of the US Department of Energy under Contracts DE-AC02-05CH11231 and DE-SC0009988 (KH), as well as by the National Science Foundation under grants PHY-1316783 and PHY-1521446. This work was performed in part at Aspen Center for Physics, which is supported by National Science Foundation grant PHY-1607611.
Boltzmann equations for the $e'$ and $u'$ abundance {#sec:boltzmann}
===================================================
In this appendix we show the Boltzmann equations governing the thermal relic abundance of $e'$ and $u'$. To simplify the expression, we omit the superscript $'$ except for the titles of sections and the mirror temperature $T'$. The number densities are that per color.
Freeze-Out
----------
For $T_{\rm{RH}} > T_{\rm{dec}}$, the relic abundances of $e$ and $u$ are set by freeze-out.
$b'$ freeze-out
---------------
During the freeze-out of $b$, the decay of $b$ is negligible and we solve the following equation, $$\begin{aligned}
\dot{n}_b + 3Hn_b = -{ \left\langle {\sigma_b v_{\rm rel}} \right\rangle } (n_b^2 - n_{b,eq}),\end{aligned}$$ ${ \left\langle {\sigma_b v} \right\rangle }$ is the thermal average of the annihilation cross section times the relative velocity of $b\bar{b}$. We include the Sommerfeld effect [@Sommerfeld], $$\begin{aligned}
\sigma_q v_{\rm rel} = \frac{2\pi \alpha_{3q,UV}^2}{27 m_q^2 } f (\frac{2\pi c_1 \alpha_{3q,IR}}{v_{\rm rel}}) + \frac{(5 + 6N_{<q} )\pi \alpha_{3q,UV}^2}{27 m_q^2 } f (\frac{2\pi c_8 \alpha_{3q,IR}}{v_{\rm rel}}),\nonumber \\
f(x) = \frac{x}{e^x -1},~c_1 = - \frac{4}{3},~c_8 = \frac{1}{6}, \nonumber \\
\alpha_{3q,UV} \equiv \alpha_3(m_q),~ \alpha_{3q,IR} = \alpha_3(m_q \alpha_3(m_q)),
\label{eq:qannihilate}\end{aligned}$$ where $N_{<q}$ is the total number of quarks and mirror quarks lighter than the mirror quark $q$ (e.g. $N_{<b}=4$). Here $\alpha_{3q,UV}$ is used for the process with a momentum exchange around the mass of $q$, namely the annihilation, while $\alpha_{3q,IR}$ is used for the process with a momentum exchange around the inverse of the Bohr radius of the $q\bar{q}$ bound state, namely the soft gluon exchange to attract $q\bar{q}$.
$c'$, $\mu'$ and $s'$ freeze-out
--------------------------------
During the freeze-out of $c$, $\mu$ and $s$, the decays of $\mu$ and $s$ are negligible. We solve the following equations, $$\begin{aligned}
\dot{n}_b + 3 H n_b = & - 8 |V_{cb}|^2 \Gamma_b n_b, \\
\dot{n}_c + 3H n_c = & -{ \left\langle {\sigma_c v} \right\rangle } (n_c^2 - n_{c,eq}) - 5 \Gamma_c n_c + 11 |V_{cb}|^2 \Gamma_b n_b, \\
\dot{n}_\mu + 3H n_\mu = & -{ \left\langle {\sigma_\mu v} \right\rangle } (n_\mu^2 - n_{\mu,eq}) + 3 |V_{cb}|^2 \Gamma_b n_b + 3 \Gamma_c n_c, \\
\dot{n}_s + 3H n_s = & -{ \left\langle {\sigma_s v} \right\rangle } (n_s^2 - n_{s,eq}) + 3 |V_{cb}|^2 \Gamma_b n_b + 5 \Gamma_c n_c,\end{aligned}$$ Here $\Gamma_f$ is defined by $$\begin{aligned}
\Gamma_f = \frac{m_f^5}{1536 \pi^3 v^4}.\end{aligned}$$ The annihilation cross section of a mirror lepton $\ell$ are $$\begin{aligned}
\sigma_\ell v_{\rm rel} = (1 + \sum_{f<\ell} q_f^2 )\frac{\pi \alpha^2}{m_\ell^2} f(-\frac{2\pi \alpha}{v_{\rm rel}}),
\label{eq:lannihilate}\end{aligned}$$ where the summation is taken for mirror fermions lighter than $\ell$ with a charge $q_f$.
$d'$, $u'$ and $e'$ freeze-out
------------------------------
During the freeze-out of $d$, $u$ and $e$, the decay of $d$ is negligible. The Boltzmann equation is given by $$\begin{aligned}
\dot{n}_\mu + 3H n_\mu = & - 4 \Gamma_\mu n_\mu,\\
\dot{n}_s + 3H n_s = & - 4 |V_{us}|^2 \Gamma_s n_s,\\
\dot{n}_d + 3H n_d = & -{ \left\langle {\sigma_d v_{\rm rel}} \right\rangle } (n_d^2 - n_{d,eq}) + \Gamma_\mu n_\mu + 3 |V_{us}|^2 \Gamma_s n_s,\\
\dot{n}_u + 3H n_u = & -{ \left\langle {\sigma_u v_{\rm rel}} \right\rangle } (n_u^2 - n_{u,eq}) + \Gamma_\mu n_\mu + 7 |V_{us}|^2 \Gamma_s n_s,\\
\dot{n}_e + 3H n_e = & -{ \left\langle {\sigma_e v_{\rm rel}} \right\rangle } (n_e^2 - n_{e,eq}) + \Gamma_\mu n_\mu + 3 |V_{us}|^2 \Gamma_s n_s.\end{aligned}$$ The freeze-out abundance of $d$ is transferred into the abundance of $u$ and $e$ by the mirror beta decay.
Freeze-In
---------
For $T_{\rm{RH}} < T_{\rm{dec}}$, the relic abundances of $e$ and $u$ are set by freeze-in. During the reheating era, the Boltzmann equations are given by $$\begin{aligned}
\dot{n}_f + 3H n_f =& { \left\langle {\sigma_{HH^\dagger \rightarrow f \bar{f}} \, v_{\rm rel}} \right\rangle }(n_H^2 - n_f ^2) + { \left\langle {\sigma_{\rm{therm}}\, v_{\rm rel}} \right\rangle }(n_{g}^2 - n_f^2)\Theta(T' - m_f) + ~ \\ &{ \left\langle {\sigma_{\rm{therm}}\, v_{\rm rel}} \right\rangle }(n_{\gamma}^2 - n_f^2)\Theta(T' - m_f) + { \left\langle {\sigma_{\rm{f}}\, v_{\rm rel}} \right\rangle }(n_{\gamma,eq}^2(m_{f}/T',\mu_\gamma) - n_f^2)\Theta(m_f-T') \nonumber, \\
\dot{n}_e + 3H n_e =& { \left\langle {\sigma_{HH^\dagger \rightarrow e \bar{e}} \, v_{\rm rel}} \right\rangle }(n_H^2 - n_e ^2) + ~ \\ &{ \left\langle {\sigma_{\rm{therm}}\, v_{\rm rel}} \right\rangle }(n_\gamma^2 - n_e^2)\Theta(T' - m_e) + { \left\langle {\sigma_{\rm{e}}\, v_{\rm rel}} \right\rangle }(n_{\gamma,eq}^2(m_{e}/T',\mu_\gamma) - n_e^2)\Theta(m_e-T')\nonumber, \\
\dot{n}_{\gamma} + 3H n_{\gamma} =& { \left\langle {\sigma_{HH^\dagger \rightarrow 2{\gamma}} \, v_{\rm rel}} \right\rangle }(n_H^2 - n_{\gamma} ^2) + { \left\langle {\sigma_{\rm{2 \rightarrow 3}}\, v_{\rm rel}} \right\rangle }(n_f^2 - n_f^2\frac{n_{\gamma}}{n_{\gamma,eq}(T',\mu = 0)}) + ~
\\ &{ \left\langle {\sigma_{\rm{therm}}\, v_{\rm rel}} \right\rangle }(n_f^2 - n_{\gamma}^2)\Theta(T' - m_f) + { \left\langle {\sigma_{\rm{therm}}\, v_{\rm rel}} \right\rangle }(n_e^2 - n_\gamma^2)\Theta(T' - m_e) \nonumber, \\
\dot{n}_{g} + 3H n_{g} =& { \left\langle {\sigma_{HH^\dagger \rightarrow 2{g}} \, v_{\rm rel}} \right\rangle }(n_H^2 - n_{g} ^2) + { \left\langle {\sigma_{\rm{therm}}\, v_{\rm rel}} \right\rangle }(n_f^2 - n_{g}^2)\Theta(T' - m_f) + ~\\ &{ \left\langle {\sigma_{\rm{2 \rightarrow 3}}\, v_{\rm rel}} \right\rangle }(n_f^2 - n_f^2\frac{n_{g}}{n_{g,eq}(T',\mu = 0)} + n_g^2 - n_g^2\frac{n_{g}}{n_{g,eq}(T',\mu = 0)}). \nonumber\end{aligned}$$ $f$ is the mirror fermion with the largest mass below $T_{\rm{RH}}$ and subscript $H$ is the SM Higgs. The production cross sections from the SM Higgs are [@ELLIS1976292; @PhysRevD.8.172] $$\begin{aligned}
{ \left\langle {\sigma_{HH^\dagger \rightarrow f \bar{f}} \, v_{\rm rel}} \right\rangle } \simeq & \frac{1}{8 \pi} \frac{y_f^2}{v'^2} \\
{ \left\langle {\sigma_{HH^\dagger \rightarrow 2{\gamma}} \, v_{\rm rel}} \right\rangle } \simeq & \frac{1}{16\pi}\left(\frac{\alpha}{4 \pi}\right)^2 \frac{T^2}{v'^4} \left(\sum_f \frac{Q_f^2}{3} \right)^2 \\
{ \left\langle {\sigma_{HH^\dagger \rightarrow 2{g}} \, v_{\rm rel}} \right\rangle } \simeq & \frac{1}{2\pi}\left(\frac{\alpha_S}{4 \pi}\right)^2 \frac{T^2}{v'^4} \left(\sum_q \frac{1}{6} \right)^2,\end{aligned}$$ where the summation on $f$ and $q$ is taken for mirror fermions and quarks with masses greater than $T$. Initially possessing a typical energy $\sim T$, the thermalization cross-section among mirror charged fermions is given by $$\begin{aligned}
{ \left\langle {\sigma_{\rm{therm}}\, v_{\rm rel}} \right\rangle } \approx \frac{4 \pi \alpha_i^2}{{T'}^2}.
\label{eq:2to2}\end{aligned}$$ while the soft, number-changing $(f \bar{f} \rightarrow f \bar{f} \gamma, \, f \bar{f} \rightarrow f \bar{f}g, \, gg\rightarrow ggg)$ bremsstrahlung cross-sections are given by $$\begin{aligned}
{ \left\langle {\sigma_{\rm{2 \rightarrow 3}}\, v_{\rm rel}} \right\rangle } \approx \frac{\alpha_i^3}{2}\left(\frac{\alpha_i n_i}{T'}\right)^{-1} \ln \left(\frac{T'^3}{\alpha_i n_i}\right),
\label{eq:2to3}\end{aligned}$$ and $$\begin{aligned}
H = \frac{5}{18} \left(\frac{\pi^2}{10} g_*\right)^{1/2} \frac{T^4}{T_{\rm{RH}}^2 M_{\rm{Pl}}}\end{aligned}$$ is the Hubble scale during the reheating matter-dominated era. Here, $\alpha_i$ equals $\alpha_{\rm EM}$ or $\alpha_S(T')$ and $n_i$ equals $n_e$ or $n_f$ depending on whether the exchange involves mirror photons or gluons.
Soft-scattering keeps the mirror bath in kinetic equilibrium (but not necessarily chemical equilibrium), establishing an effective temperature $$\begin{aligned}
T' = \frac{1}{3}\frac{\rho'_{\rm{tot}}(T)}{n'_{\rm{tot}}(T)}\end{aligned}$$ where $\rho'_{\rm{tot}}(T)$ is the total energy density of the mirror sector frozen in via the Higgs portal when the universe is at a temperature $T$, and $n'_{\rm{tot}}$ is the total number density of the mirror sector determined from the Boltzmann equations. For mirror photons, $\gamma$, and gluons, $g$, the equilbrium number densities are $$\begin{aligned}
n_{eq}\left(\frac{m}{T'}, \,\mu \right) &= g \left(\frac{mT'}{2\pi}\right)^{3/2}\exp \left(-\frac{m}{T'} + \frac{\mu}{T'}\right) = \sqrt{\frac{\pi}{8}}\left(\frac{m}{T'}\right)^{3/2} \exp \left(-\frac{m}{T'}\right) n \\
n_{eq}(T',\mu = 0) &= \frac{2 g}{\pi^2}T'^3.\end{aligned}$$ For low $v'$ and high $T_{\rm{RH}}$, thermalization of $e$ and $\gamma$ via $2 \to 3$ and $2 \to 2$ processes are effective, thereby increasing $n'_{\rm{tot}}$ and decreasing $T'$. This thermalization acts to cool the mirror bath so that mirror particles freeze-out instantly with an annihilation cross-section ${ \left\langle {\sigma_{\rm{f}}\, v_{\rm rel}} \right\rangle }$ given by if a quark, and if a lepton. Nevertheless, these frozen-out particles are then continually replenished by fresh particles from the Higgs portal. Since freeze-in production is maximized at $T_{\rm{RH}}$ and any pre-thermalized contribution is typically small, the most important contributions to the present-day abundance of $e'$ occurs at and below $T_{\rm{RH}}$, discussed below -.
For $T < T_{\rm{RH}}$, the universe is radiation dominated. The mirror bath remains in kinetic equilibrium (not necessarily chemical equilibrium), establishing an effective temperature $$\begin{aligned}
T' = \frac{1}{3}\frac{\rho'_{\rm{tot}}}{n'_{\rm{tot}}} \simeq \frac{1}{3}\frac{\rho'_{\rm{tot}}(T_{\rm{RH}})}{n_f + n_e + n_\gamma + n_g} \left(\frac{T}{T_{\rm{RH}}}\right)^{4}.
\label{eq:tpTRH}\end{aligned}$$ The Boltzmann equations for $m_{e} < T' < T_{\rm{RH}}$ determine the evolution of $n_f, n_e, n_g$, and $n_\gamma$, and are given by $$\begin{aligned}
\dot{n}_f + 3H n_f =& { \left\langle {\sigma_{HH^\dagger \rightarrow f \bar{f}} \, v_{\rm rel}} \right\rangle }(n_{H,eq}^2(m_{f}/T) - n_f ^2) + ~ \\
&{ \left\langle {\sigma_f \, v_{\rm rel}} \right\rangle }(n_{\gamma,eq}^2(m_{f}/T', \,\mu_\gamma) - n_f^2) + { \left\langle {\sigma_f \, v_{\rm rel}} \right\rangle }(n_{g,eq}^2(m_{f}/T', \,\mu_g) - n_f^2), \\
\dot{n}_e + 3H n_e =& { \left\langle {\sigma_{HH^\dagger \rightarrow e \bar{e}} \, v_{\rm rel}} \right\rangle }(n_H^2 - n_e ^2) + { \left\langle {\sigma_{\rm{therm}}\, v_{\rm rel}} \right\rangle }(n_\gamma^2 - n_e^2), \\
\dot{n}_\gamma + 3H n_\gamma =& { \left\langle {\sigma_{HH^\dagger \rightarrow 2\gamma} \, v_{\rm rel}} \right\rangle }(n_H^2 - n_\gamma ^2) + { \left\langle {\sigma_f \, v_{\rm rel}} \right\rangle }(n_f^2 - n_{\gamma,eq}^2(m_{f}/T', \,\mu_\gamma)) +~ \\ &{ \left\langle {\sigma_{\rm{2 \rightarrow 3}}\, v_{\rm rel}} \right\rangle }(n_f^2 - n_f^2\frac{n_\gamma}{n_{\gamma,eq}(T',\mu = 0)}) + { \left\langle {\sigma_{\rm{therm}}\, v_{\rm rel}} \right\rangle }(n_e^2 - n_\gamma^2)\Theta(T' - m_e). \nonumber \\
\dot{n}_g + 3H n_g =& { \left\langle {\sigma_{HH^\dagger \rightarrow 2g} \, v_{\rm rel}} \right\rangle }(n_H^2 - n_g ^2) + { \left\langle {\sigma_f \, v_{\rm rel}} \right\rangle }(n_f^2 - n_{g,eq}^2(m_{f}/T', \,\mu_g)) +~ \\ &{ \left\langle {\sigma_{\rm{2 \rightarrow 3}}\, v_{\rm rel}} \right\rangle }(n_f^2 - n_f^2\frac{n_g}{n_{g,eq}(T',\mu = 0)}+ n_g^2 - n_g^2\frac{n_{g}}{n_{g,eq}(T',\mu = 0)}))\nonumber\end{aligned}$$ Last, $e'$ freezes-out when $T'$ drops below its mass. The Boltzmann equation for $T' < m_{e}$ is $$\begin{aligned}
\dot{n}_e + 3H n_e =& { \left\langle {\sigma_{e}\, v_{\rm rel}} \right\rangle }(n_{\gamma,eq}^2(m_{e}/T', \,\mu_\gamma) - n_e ^2)
\label{eq:boltzmannTRHf}\end{aligned}$$
Energy densities of the mirror QCD bath {#sec:rhos}
=======================================
In this appendix we estimate the energy density of the mirror QCD bath. We derive the energy density at the phase transition, which is used to estimate the magnitude of gravitational waves, and the energy density of the mirror glueballs after the transition, which is used to estimate the dark radiation abundance. We assume entropy conservation around the mirror QCD phase transition. Entropy production via super-cooling will result in enhancement of the signals.
The SM and mirror sectors decouple from each other at the temperature shown in Eq. (\[eq:Tdec\]). Around this temperature, $e'$, $\mu'$, $u'$, $d'$, $s'$, $g'$, and $\gamma'$ are in the thermal bath; the effective number of degrees of freedom of the mirror sector is $g_{\rm dec}' \simeq 60$. After decoupling, the entropies of the two sectors are separately conserved. Around the mirror QCD phase transition, the mirror gluon bath is nearly pressureless. Parametrizing the energy density of the mirror gluon bath by $\rho_{g'} = b \, {T'}^4$, the ratio of the temperatures of the two sectors is $$\begin{aligned}
\frac{T_{\rm SM}}{T_{g'}} = 0.3 \left( \frac{g_{\rm dec}}{g_c} \frac{60}{g_{\rm dec}'} \frac{b}{0.5} \right)^{1/3},\end{aligned}$$ where $g_c$ is the effective number of degrees of freedom of the SM bath at the mirror QCD phase transition. The ratio of the energy densities is $$\begin{aligned}
\frac{\rho_{\rm SM}}{ \rho_{g'}} = 0.5 \left( \frac{106.75}{g_c} \frac{b}{0.5} \right)^{1/3} \left( \frac{g_{\rm dec}}{106.75} \; \frac{60}{g_{\rm dec}'} \right)^{4/3}.\end{aligned}$$ For $T' \lesssim 0.7 T'_c$, the energy and the entropy density of the mirror QCD bath is well-approximated by that of the ideal gas of the lightest mirror glueballs with a mass $m_{S'}\simeq 5.3 T_c'$ [@Borsanyi:2012ve]. Entropy conservation within this decoupled mirror bath implies its entropy density scales as $\propto a^{-3}$. $3 \to 2$ annihilations keep warm the mirror glueballs so that their temperature falls approximately as $\propto \ln a$ and energy density as $\propto a^{-3} \,(\ln{a})^{-1}$ until they decouple or decay [@Carlson:1992fn; @Hochberg:2014dra; @Forestell:2016qhc]. Here, $a$ is the scale factor of the universe. The $3 \to 2$ cross-section is given by [@Forestell:2016qhc] $$\begin{aligned}
\langle\sigma_{3 \to 2}v^2 \rangle \simeq \frac{B}{(4 \pi)^3}\left(\frac{4 \pi}{3}\right)^6\frac{1}{m_{S'}^5},\end{aligned}$$ where $B$ is an $\mathcal{O}(1)$ number whose value weakly affects $a_f$. We take $B = 1$.
As discussed in Sec. \[sec:BBNandDR\], the non-trivial dynamics around the mirror QCD phase transition are encoded in the modification factor $A$, the ratio of the actual mirror glueball energy density to that derived by a non-interacting ideal gas approximation and the glueball number conservation, $$\begin{aligned}
A = \frac{4 T_f'}{3 T_c'} = \frac{4}{3} \frac{2 m_{S'}}{T_c'} W\left(\frac{2}{(2 \pi)^3} \left(\frac{45}{32 \pi^2}\right)^2 \left( \frac{m_{S'}}{T_c'} \, \frac{a_f}{a_c}\right)^6\right)^{-1} {\mathrel{\vcenter{
\offinterlineskip\halign{\hfil$##$\cr
\propto\cr\noalign{\kern2pt}\sim\cr\noalign{\kern-2pt}}}}}\left( \ln\frac{a_f}{a_c}\right)^{-1}.
\label{eq:gBallFudge}\end{aligned}$$ Here, $W(x)$ is the product-log function, which is a solution of $W e^W = x$. $a_c$ is the would-be scale factor at $T'= T_c'$ if the mirror gluons remain an ideal gas until the phase transition, and $a_f$ is the scale factor of the universe when the $3 \to 2$ reactions among mirror glueballs freeze-out, or the mirror glueballs decay. For $v' > 10^9$ GeV, $a_f$ is determined by the former and otherwise by the latter.
![The QCD$'$ modification factor $A$ as a function of $a_f/a_c$. $A$ is defined as the energy density ratio of the actual glueball gas to that derived by a non-interacting ideal gas approximation and glueball number conservation.[]{data-label="fig:aFactor"}](Figures/aFactorPlot.pdf){width="70.00000%"}
For $0.7 T'_c \lesssim T' \lesssim T'_c$, the energy density of the mirror glueball bath deviates from that of a weakly-interacting ideal gas composed of the lightest mirror glueballs, and hence the second equality of $\eqref{eq:gBallFudge}$ is invalid. In this strongly interacting regime, $A$ is determined by taking the lattice result for $\rho_{g'}(T'_f/T'_c)$ from [@Borsanyi:2012ve] and equating it with $s_{g'} T'_f$ - an excellent approximation since the glueball gas is nearly pressureless. Here, $s_{g'} = 32 \pi^2/45 {T'_c}^3 (a_c/a_f)^3$ is the entropy density of the mirror glueball bath. $T'_f/T'_c$ is then numerically solved for as a function of $a_f/a_c$ and inserted into $\eqref{eq:gBallFudge}$ to determine $A$ as function of $a_f/a_c$ as shown for both regimes in Fig. \[fig:aFactor\].
[^1]: The $Z_2$ mapping described in is not unique. For example, the $Z_2$ symmetry can be extended to spacetime parity if space is inverted and SM fields are mapped to their Hermitian conjugated mirrors.
[^2]: Diagrams contributing to kinetic mixing via the Higgs portal only occur beyond four loops, likely inducing an $\epsilon \ll 10^{-12}$.
[^3]: Since the $Z_2$ symmetry is unbroken above $v'$, $\langle \Sigma \rangle = \langle \Sigma' \rangle = v_{G}$.
[^4]: Furthermore, the maximum temperature of the universe after inflation is taken less than the mirror electroweak scale, to avoid domain wall problems from the spontaneous breaking of Higgs Parity. Generically the maximal temperature is higher than the reheat temperature. See [@Harigaya:2013vwa; @Mukaida:2015ria] for a recent estimation of the maximal temperature.
[^5]: We take $\xi = \eta =1$.
[^6]: Some $e'$ production still occurs for $T_{\rm{RH}} < T < m_{e'}$ by scatterings involving highly energetic particles produced by inflatons [@Harigaya:2014waa; @Harigaya:2019tzu], which we find is not efficient enough to reproduce the DM abundance.
[^7]: For low $v'$ and high $T_{\rm{RH}}$, $e'$ and $\gamma'$ may thermalize during reheating, altering (\[eq:nFI\]). Thermalization cools the mirror bath so that mirror particles freeze-out instantly but are then replenished by the Higgs portal. Since freeze-in production is maximized at $T_{\rm{RH}}$, any pre-thermalized contribution is typically small. Even so, we consider this effect in Appendix \[sec:boltzmann\].
[^8]: If $e'$ is not the DM, or is produced in a non-thermal way, the red region is not applicable and the $SM \times SM'$ model is not necessarily excluded.
[^9]: Since the mirror QCD bath couples to the standard model particles very weakly, bubbles only induce turbulent motion of mirror glueballs. In particular, a turbulent magnetic field is not induced. For a phase transition generating magnetic turbulence, Ref. [@Caprini:2009yp] finds a spectrum of gravitational waves produced by turbulent magnetic fields similar to that from turbulent motion of fluids, and hence we simply use the fitting provided in Ref. [@Caprini:2010xv].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'At variance with fully inclusive quantities, which have been computed already at the two- or three-loop level, most exclusive observables are still known only at one-loop, as further progress was hampered so far by the greater computational problems encountered in the study of multi-leg amplitudes beyond one loop. We show in this paper how the use of tools already employed in inclusive calculations can be suitably extended to the computation of loop integrals appearing in the virtual corrections to exclusive observables, namely two-loop four-point functions with massless propagators and up to one off-shell leg. We find that multi-leg integrals, in addition to integration-by-parts identities, obey also identities resulting from Lorentz-invariance. The combined set of these identities can be used to reduce the large number of integrals appearing in an actual calculation to a small number of master integrals. We then write down explicitly the differential equations in the external invariants fulfilled by these master integrals, and point out that the equations can be used as an efficient method of evaluating the master integrals themselves. We outline strategies for the solution of the differential equations, and demonstrate the application of the method on several examples.'
---
TTP99–49\
December 1999
3.5cm
[**Differential Equations for Two-Loop Four-Point Functions**]{} 1.cm [T. Gehrmann]{} and [E. Remiddi]{}[^1] .7cm [*Institut für Theoretische Teilchenphysik, Universität Karlsruhe, D-76128 Karlsruhe, Germany*]{}
2.6cm
Introduction
============
Perturbative corrections to many inclusive quantities have been computed to the two- and three-loop level in past years. From the technical point of view, these inclusive calculations correspond to the computation of multi-loop two-point functions, for which many elaborate calculational tools have been developed. In contrast, corrections to exclusive observables, such as jet production rates, could up to now only be computed at the one-loop level. These calculations require the computation of multi-leg amplitudes to the required number of loops, which beyond the one-loop level turn out to be a calculational challenge obstructing further progress. Despite considerable progress made in recent times, many of the two-loop integrals relevant for the calculation of jet observables beyond next-to-leading order are still unknown. One particular class of yet unknown integrals appearing in the two-loop corrections to three jet production in electron-positron collisions, to two-plus-one jet production in electron-proton collisions and to vector boson plus jet production in proton-proton collisions are two-loop four-point functions with massless internal propagators and one external leg off-shell.
Taking two-loop four-point integrals arising in the calculation of Feynman diagrams in covariant gauges (non-covariant gauges introduce integrals of a structure beyond the treatment of this paper) as an example, we elaborate on in this paper several techniques to compute multi-leg amplitudes beyond one loop. We demonstrate how integration-by-parts identities (already known to be a very valuable tool in inclusive calculations) and identities following from Lorentz-invariance (which are non-trivial only for integrals depending on at least two independent external momenta) can be used to reduce the large number of different integrals appearing in an actual calculation to a small number of master integrals. This reduction can be carried out mechanically (by means of a small chain of computer programs), without explicit reference to the actual structure of the integrals under consideration and can also be used for the reduction of tensor integrals beyond one loop.
The master integrals themselves, however, can not be computed from these identities. We derive differential equations in the external momenta for them. Solving these differential equations, it is possible to compute the master integrals without explicitly carrying out any loop integration, so that this technique appears to be a valuable alternative to conventional approaches for the computation of multi-loop integrals.
The plan of the paper is as follows. In Section \[sec:master\] we review the derivation of the integration by parts (IBP) identities and introduce the Lorentz invariance (LI) identities. In Section \[sec:diffeq\] the differential equations for the master amplitudes are obtained. The practical application of these tools is outlined in detail in Section \[sec:oneloop\] on a self-contained rederivation of the one-loop massless box integral with one off-shell leg. Section \[sec:twoloop\] contains examples of massless two-loop four-point functions with one off-shell leg, evaluated for arbitrary space-time dimension. We show which of these functions can be reduced to simpler functions and which are genuine master integrals, and compute some of the master integrals by solving the corresponding differential equations. Finally, Section \[sec:conc\] contains conclusions and an outlook on potential future applications of the tools developed here. The higher transcendental functions appearing in our results for the one-loop and two-loop integrals are summarised in an Appendix, where we also discuss how these functions can be expanded around the physical number of space-time dimensions.
Reduction to Master Integrals {#sec:master}
=============================
Any scalar massless two-loop integral can be brought into the form $$I(p_1,\ldots,p_n) = \int \frac{\d^d k}{(2\pi)^d}\frac{\d^d l}{(2\pi)^d}
\frac{1}{D_1^{m_1}\ldots D_{t}^{m_t}} S_1^{n_1}
\ldots S_q^{n_q} \; ,
\label{eq:generic}$$ where the $D_i$ are massless scalar propagators, depending on $k$, $l$ and the external momenta $p_1,\ldots,p_n$ while $S_i$ are scalar products of a loop momentum with an external momentum or of the two loop momenta. The topology (interconnection of propagators and external momenta) of the integral is uniquely determined by specifying the set $(D_1,\ldots,D_t)$ of $t$ different propagators in the graph. The integral itself is then specified by the powers $m_i$ of all propagators and by the selection $(S_1,\ldots,S_q)$ of scalar products and their powers $(n_1,\ldots,n_q)$ . (all the $m_i$ are positive integers greater or equal to 1, while the $n_i$ are greater or equal to 0). Integrals of the same topology with the same dimension $r=\sum_i m_i$ of the denominator and same total number $s=\sum_i n_i$ of scalar products are denoted as a class of integrals $I_{t,r,s}$. The integration measure and scalar products appearing the above expression are in Minkowskian space, with the usual causal prescription for all propagators. The loop integrations are carried out for arbitrary space-time dimension $d$, which acts as a regulator for divergencies appearing due to the ultraviolet or infrared behaviour of the integrand (dimensional regularisation, [@dreg; @hv]).
Any four-point function depends on three linearly independent external momenta, $p_1$, $p_2$ and $p_3$. At the two-loop level, one can combine the two loop momenta $k$ and $l$ and these external momenta to form 9 different scalar products involving $k$ or $l$. As the propagators present in the graph are (linearly independent) combinations of scalar products, only $9-t$ different scalar products can appear explicitly in an integral with $t$ different propagators. Since a two-loop four-point function can have at most seven different propagators, as can be found by considering the insertion of a propagator into a one-loop four-point function, one has in general $t\leq 7$, while the minimum number of massless propagators in a two-loop graph is $t=3$, corresponding to a two-point function. The number of different two-loop four-point integrals for given $t$ (number of different propagators), $r$ (sum of powers of all propagators) and $s$ (sum of powers of all scalar products) can be computed from simple combinatorics: $$N(I_{t,r,s}) = { r-1 \choose r-t} {8-t+s \choose s} \; .
\label{eq:intnum}$$
The number $N(I_{t,r,s})$ of the integrals grows quickly as $r, s$ increase, but the integrals are related among themselves by various identities. One class of identities follows from the fact that the integral over the total derivative with respect to any loop momentum vanishes in dimensional regularisation $$\int \frac{\d^d k}{(2\pi)^d} \frac{\partial}{\partial k^{\mu}}
J(k,\ldots) = 0,$$ where $J$ is any combination of propagators, scalar products and loop momentum vectors. $J$ can be a vector or tensor of any rank. The resulting identities [@hv; @chet] are called integration-by-parts (IBP) identities and can for two-loop integrals be cast into the form $$\begin{aligned}
\int \frac{\d^d k}{(2\pi)^d} \frac{\d^d l}{(2\pi)^d}
\frac{\partial}{\partial k^{\mu}} v^{\mu} f(k,l,p_i) & = & 0, \nonumber \\
\int \frac{\d^d k}{(2\pi)^d} \frac{\d^d l}{(2\pi)^d}
\frac{\partial}{\partial l^{\mu}} v^{\mu} f(k,l,p_i) & = & 0, \end{aligned}$$ where the integrand $f(k,l,p_i)$ is a scalar function, containing propagators and scalar products and $v_{\mu}$ can be any external or loop momentum vector. As a consequence, one obtains for a graph with $m$ loops and $n$ independent external momenta a total number of $N_{{\rm IBP}} = m(n+m)$. For a two-loop four-point function, this results in ten IBP identities for each integrand.
In addition to the IBP identities, one can also exploit the fact that all integrals under consideration are Lorentz scalars (or, perhaps more precisely, “$d$-rotational” scalars) , which are invariant under a Lorentz (or $d$-rotational) transformation of the external momenta. In order to derive the resulting Lorentz invariance (LI) identities, we consider an infinitesimal Lorentz transformation $$p^{\mu} \to p^{\mu} + \delta p^{\mu} =
p^{\mu} + \delta \epsilon^{\mu}_{\nu} p^{\nu} \qquad
\mbox{with} \qquad \delta \epsilon^{\mu}_{\nu} = - \delta
\epsilon^{\nu}_{\mu}\;,$$ which should not change the scalar Feynman integral $$I(p_1+\delta p_1,
\ldots , p_n+\delta p_n)=
I(p_1,\ldots,p_n) \; .$$ Expanding $$I(p_1+\delta p_1,
\ldots , p_n+\delta p_n) =
I(p_1,\ldots,p_n) + \delta p_1^{\mu}
\frac{\partial}{\partial p_1^{\mu}}
I(p_1,\ldots,p_n) + \ldots + \delta p_n^{\mu}
\frac{\partial}{\partial p_n^{\mu}}
I(p_1,\ldots,p_n)\; ,$$ one arrives at $$\delta \epsilon^{\mu}_{\nu} \left( p_1^{\nu}\frac{\partial}{\partial
p_1^{\mu}} + \ldots + p_n^{\nu}\frac{\partial}{\partial
p_n^{\mu}} \right) I(p_1,\ldots,p_n) = 0\; .$$ Since $\delta \epsilon^{\mu}_{\nu}$ has six independent components, the above equation contains up to six LI identities. These are however not always linearly independent. To determine the maximum number of linearly independent identities, one uses the antisymmetry of $\delta
\epsilon^{\mu}_{\nu}$ to obtain $$\left(p_1^{\nu}\frac{\partial}{\partial
p_{1\mu}} - p_1^{\mu}\frac{\partial}{\partial
p_{1\nu}} + \ldots + p_n^{\nu}\frac{\partial}{\partial
p_{n\mu}} - p_n^{\mu}\frac{\partial}{\partial
p_{n\nu}}\right) I(p_1,\ldots,p_n) = 0 \;.
\label{eq:li}$$ This equation can be contracted with all possible antisymmetric combinations of $p_{i\mu}p_{j\nu}$ to yield LI identities for $I$. For a three-point vertex, two of the external momenta are linearly independent ($n=2$), and only one antisymmetric combination of them can be constructed, resulting in one LI identity ($N_{{\rm LI}}=1$). A four-point function depends on three external momenta ($n=3$), allowing us to construct three linearly independent antisymmetric combinations, which yield three LI identities ($N_{{\rm LI}}=3$). The full potential of the LI identities can only be exploited for integrals involving five or more external legs, which allow to construct six linearly independent antisymmetric combinations of external momenta, thus projecting out all six LI identities ($N_{{\rm LI}}=6$).
Since $I$ is a scalar, it can not depend on the momenta $p_i$ itself, but only on scalar products $s_{ij} = 2 p_i\cdot p_j$ of the external momenta. Replacing $$\frac{\partial}{\partial p_{i\mu}} = \sum_j 2\left(p_{i\mu}+p_{j\mu}\right)
\frac{\partial}{\partial s_{ij}}\;,$$ one finds that (\[eq:li\]) becomes a trivial identity, independent of $I$. However, the derivatives in (\[eq:li\]) can be interchanged with the loop integrations in $I$, such that they do not act anymore on the integral $I$, but on the integrand of $I$. After this interchange, (\[eq:li\]) becomes a non-trivial relation between different integrals.
In the case of two-loop four-point functions, one has a total of 13 equations (10 IBP + 3 LI) for each integrand corresponding to an integral of class $I_{t,r,s}$, relating integrals of the same topology with up to $s+1$ scalar products and $r+1$ denominators, plus integrals of simpler topologies ([*i.e.*]{} with a smaller number of different denominators). The 13 identities obtained starting from an integral $I_{t,r,s}$ do contain integrals of the following types:
- $I_{t,r,s}$: the integral itself.
- $I_{t-1,r,s}$: simpler topology.
- $I_{t,r+1,s}, I_{t,r+1,s+1}$ : same topology, more complicated than $I_{t,r,s}$.
- $I_{t,r-1,s}, I_{t,r-1,s-1}$: same topology, simpler than $I_{t,r,s}$.
Quite in general, single identities of the above kind can be used to obtain the reduction of $I_{t,r+1,s+1}$ or $I_{t,r+1,s}$ integrals in terms of $I_{t,r,s}$ and simpler integrals - rather than to get information on the $I_{t,r,s}$ themselves.
Integrations-by-parts identities are widely applied in multi-loop calculations of inclusive quantities (see e.g. [@krev] for a review), which are related to two-point functions. In these calculations, only a relatively small number of different topologies has to be considered, but the integrals appearing in the calculation can bear large powers of propagators and scalar products, arising for example from expansions in masses or momenta. In these calculations, it is desirable to have reduction formulae for arbitrary powers of propagators and scalar products. These can be obtained from IBP identities derived for an integral with arbitrary powers (to be treated symbolically) of scalar products and propagators; the derivation of these symbolic reduction formulae requires a lot of ingenuity, based on the direct inspection of the explicit form of the IBP identities for each considered topology, and could not be carried out mechanically.
For loop integrals with a large number of external legs, IBP identities are needed for a large number of different topologies, but in general for relatively small powers of propagators and scalar products. In this case, it would therefore be desirable to have a mechanical procedure for solving, for any given topology, IBP and LI identities for integrals with fixed powers of the propagators and scalar products.
If one considers the set of all the identities obtained starting from the integrand of all the $N(I_{t,r,s})$ integrals of class $I_{t,r,s}$, one obtains $(N_{{\rm IBP}}+ N_{{\rm LI}}) N(I_{t,r,s})$ identities which contain $N(I_{t,r+1,s+1})+N(I_{t,r+1,s})$ integrals of more complicated structure. From (\[eq:intnum\]) it can be read off that with increasing $r$ and $s$ the number of identities grows faster than the number of new unknown integrals[^2]. As a consequence, if for a given $t$-topology one considers the set of all the possible equations obtained by considering all the integrands up to certain values $r^*, s^*$ of $r, s$, for large enough $r^*, s^*$ the resulting system of equations is overconstrained and can be used for expressing the more complicated integrals, with greater values of $r, s$ in terms of simpler ones, with smaller values of $r, s$. (Let us observe that, the system being overconstrained, the equations cannot be all independent; it is not [*a priori*]{} known how many equations are in fact linearly independent and, correspondingly, how many integrals of the topology under consideration will remain after reduction).
The required values $r^*$ and $s^*$ for $r$ and $s$ can be found by counting the number of accumulated equations (equations for all integrals with $r\leq r^*$ and $s\leq s^*$) and comparing them with the number of accumulated unknown integrals, with $(r\leq r^* + 1,s \leq s^* + 1)$, but excluding $(r=t,s=s^* + 1)$. As an example, we list in Table \[tab:unknown\] the number of equations and unknowns for two-loop four-point functions with seven denominators $t=7$. It can be seen that a complete reduction requires at least one of the combinations ($r^*,s^*$): (7,2); (8,1); (9,0).
\
\
The above table illustrates that typically hundreds of equations have to be solved in order to obtain a reduction towards simpler integrals. The task is performed automatically (and independently of the topology!) by a computer program invoking repeatedly the computer algebra packages FORM [@form] and MAPLE [@maple]. For any given four-point two-loop topology, this procedure can result either in a reduction towards a small number (typically one or two) of integrals of the topology under consideration and integrals of simpler topology (less different denominators), or even in a complete reduction of all integrals of the topology under consideration towards integrals with simpler topology. Left-over integrals of the topology under consideration are called irreducible master integrals or just master integrals. If a topology turns out to contain irreducible master integrals, one is in principle free to choose which integrals are taken as master integrals, as long as the chosen integrals are not related by the IBP and LI identities. In our reduction, we choose $I_{t,t,0}$ for topologies with one master integral and $I_{t,t,0}$ together with the required number of integrals of type $I_{t,t+1,0}$ for topologies with more than one master integral.
As a final point, it is worth noting that the procedure described above can be used to reduce a tensor integral towards scalar integrals. All integrals appearing in the projection of an arbitrary tensorial integral onto a tensor basis will take the form (\[eq:generic\]), i.e. they can be classified as $I_{t,r,s}$ and reduced to master integrals.
In the following section, we will demonstrate that the established reduction of all the integrals to a few master integrals can be used also to write differential equations in the external invariants $s_{ij}$ for the master integrals themselves, and then how the differential equations can be used to compute these master integrals.
Differential Equations for Master Integrals {#sec:diffeq}
===========================================
The IBP and LI identities discussed in the previous section allow us to express integrals of the form (\[eq:generic\]) as a linear combination of a few master integrals, [*i.e.*]{} integrals which are not further reducible, but have to be computed by some different method.
For the case of massless two-loop four-point functions, several techniques have been proposed in the literature, such as for example the application of a Mellin-Barnes transformation to all propagators [@smirnov; @tausk] or the negative dimension approach [@glover]. Both techniques rely on an explicit integration over the loop momenta, with differences mainly in the representation used for the propagators. So far, these techniques were only applied to a limited number of master integrals: Smirnov [@smirnov] has recently used the Mellin-Barnes method to compute the planar double box integral for the case of all external legs on shell (massless case); the same method has been applied by Tausk [@tausk] for the computation of the non-planar on-shell double box integral; the negative dimension approach has been applied by Anastasiou, Glover and Oleari [@glover] to compute the class of two-loop box integrals which correspond to a one-loop bubble insertion in one of the propagators of the one-loop box. A general method for the computation of the master integrals appearing in two-loop four-point functions has however not yet been found. So far, it has also not even been clear (apart from the planar double box topology, where Smirnov and Veretin have recently demonstrated the reducibility of any integral of this topology to two master integrals [@smir2]) how many master integrals exist for a given topology. Solving the identities discussed in the previous section, we are now able to identify the irreducible master integrals. A list of reducible two-loop four-point topologies will be given in Section \[sec:twoloop\].
A method for the computation of master integrals avoiding the explicit integration over the loop momenta is to derive differential equations in internal propagator masses or in external momenta for the master integral, and to solve these with appropriate boundary conditions. This method has first been suggested by Kotikov [@kotikov] to relate loop integrals with internal masses to massless loop integrals. It has been elaborated in detail and generalised to differential equations in external momenta in [@remiddi]; first applications were presented in [@appl]. In the case of four-point functions with one external off-shell leg and no internal masses, one has three independent invariants, resulting in three differential equations.
The derivatives in the invariants $s_{ij}=(p_i+p_j)^2$ can be expressed by derivatives in the external momenta: $$\begin{aligned}
\sab \frac{\partial}{\partial \sab} & = & \frac{1}{2} \left( +
p_1^{\mu} \frac{\partial}{\partial p_1^{\mu}} +
p_2^{\mu} \frac{\partial}{\partial p_2^{\mu}} -
p_3^{\mu} \frac{\partial}{\partial p_3^{\mu}}\right) \nonumber \\
\sac \frac{\partial}{\partial \sac} & = & \frac{1}{2} \left( +
p_1^{\mu} \frac{\partial}{\partial p_1^{\mu}} -
p_2^{\mu} \frac{\partial}{\partial p_2^{\mu}} +
p_3^{\mu} \frac{\partial}{\partial p_3^{\mu}}\right)
\label{eq:derivatives} \\
\sbc \frac{\partial}{\partial \sbc} & = & \frac{1}{2} \left( -
p_1^{\mu} \frac{\partial}{\partial p_1^{\mu}} +
p_2^{\mu} \frac{\partial}{\partial p_2^{\mu}} +
p_3^{\mu} \frac{\partial}{\partial p_3^{\mu}}\right) \nonumber \end{aligned}$$ The combinations of derivatives and momenta appearing on the right hand side of (\[eq:derivatives\]) are obviously linearly independent from the combinations appearing in the LI identities (\[eq:li\]) (which vanish identically when acting on a function depending on the scalars $\sab,\sac,\sbc$). The three derivatives of an integral $I_{t,r,s}(\sab,\sac,\sbc,d)$ are not linearly independent, but related due to the properties of $I$ under rescaling of all external momenta: $$I_{t,r,s}(\sab,\sac,\sbc,d) = \lambda^{-\alpha(d,r,s)}
I_{t,r,s}(\lambda^2\sab,\lambda^2\sac,\lambda^2\sbc,d)\; ,$$ where $\alpha(d,r,s)$ is the mass dimension of the integral. For a $m$-loop integral in $d$ space-time dimensions with $r$ powers of denominators and $s$ scalar products, one finds $\alpha(d,r,s) = m d + 2s -2r$. The above equation yields the rescaling relation $$\left[ - \frac{\alpha}{2} +
\sab \frac{\partial}{\partial\sab} + \sac
\frac{\partial}{\partial\sac} +
\sbc \frac{\partial}{\partial\sbc} \right] I_{t,r,s}(\sab,\sac,\sbc,d)
=0 \; .
\label{eq:rescale}$$
Let us now first consider the case of a topology with only one master integral, which is chosen to be $I_{t,t,0}$, defined as $$I_{t,t,0}(\sij,s_{jk},s_{ki},d) =
\int \frac{\d^d k}{(2\pi)^d}\frac{\d^d l}{(2\pi)^d}
f_{t,t,0}(k,l,p_i) \; ,
\label{eq:tt0}$$ where $ f_{t,t,0}(k,l,p_i) $ is a suitable integrand of the form appearing in (\[eq:generic\]). It is evident that acting with the right hand sides of (\[eq:derivatives\]) on the right hand side of (\[eq:tt0\]) will, after interchange of derivative and integration, yield a a combination of integrals of the same type as appearing in the IBP and LI identities for $I_{t,t,0}$, including integrals of type $I_{t,t+1,1}$ and $I_{t,t+1,0}$. Consequently, the scalar derivatives (on left hand side of (\[eq:derivatives\])) of $I_{t,t,0}$ can be expressed by a linear combination of integrals up to $I_{t,t+1,1}$ and $I_{t,t+1,0}$. These can all be reduced to $I_{t,t,0}$ and to integrals of simpler topology by applying the IBP and LI identities. This reduction results in differential equations for $I_{t,t,0}$ of the form: $$\begin{aligned}
\sij \frac{\partial}{\partial \sij} I_{t,t,0}(\sij,s_{jk},s_{ki},d) & = &
A(\sij,s_{jk},s_{ki},d)
I_{t,t,0}(\sij,s_{jk},s_{ki},d) \nonumber \\
& & +
F(\sij,s_{jk},s_{ki},d,I_{t-1,r,s}(\sij,s_{jk},s_{ki},d))\;,
\label{eq:master1}\end{aligned}$$ where $\sij$, $s_{jk}$, $s_{ki}$ are the three invariants, $A(\sij,s_{jk},s_{ki},d)$ turns out to be a rational function of the invariants and of $d$, $F(I_{t-1,r,s})$ is a linear combination (with coefficients depending on $\sij,s_{jk},s_{ki}$ and $d$) of integrals of type $I_{t-1,r,s}$, containing only topologies simpler than $I_{t,t,0}$, but potentially with high powers of some denominators and scalar products. These $I_{t-1,r,s}$, which refer to a simpler topology and therefore can be considered as known in a bottom-up approach, play the role of an inhomogeneous term in the equation; one can then look for the proper solution of (\[eq:master1\]) in a straightforward way.
The master integral $I_{t,t,0}(\sij,s_{jk},s_{ki},d)$ can indeed be obtained by matching the general solution of (\[eq:master1\]) to an appropriate boundary condition. Quite in general, finding a boundary condition is a simpler problem than evaluating the whole integral, since it depends on a smaller number of kinematical variables. In some cases, the boundary condition can even be determined from the differential equation itself: for $\sij=0$, (\[eq:master1\]) yields, if $A(0,s_{jk},s_{ki},d)\neq 0$, $$I_{t,t,0}(0,s_{jk},s_{ki},d) = -
\left[A(0,s_{jk},s_{ki},d)\right]^{-1}
F(0,s_{jk},s_{ki},d,I_{t-1,r,s}(0,s_{jk},s_{ki},d))\; .$$ For $A$ vanishing at $\sij=0$ one can consider $I_{t,t,0}$ in the limit where one of the external momenta vanishes, corresponding to the vanishing of both invariants involving this momentum, e.g. $\sij=0$ and $s_{ki}=0$ for $p^{\mu}_i=0$. In this case, $I_{t,t,0}$ reduces to a three-point vertex function with one off-shell external leg. All these functions, which might be determined by iterating the procedure just described for the considered 4-point function, have actually already been computed at the two-loop level in [@kl] using IBP identities to reduce all possible topologies to a few master integrals, which, in this case, can be computed straightforwardly using Feynman parameters (cf. Section \[sec:twoloop\]). Starting from the boundary condition in $\sij=s_{ki}=0$, one can determine $I_{t,t,0}(0,s_{jk},s_{ki},d)$ by solving the differential equation in $s_{ki}$ – this provides the desired boundary condition in $\sij=0$.
For topologies with more than one master integral, (\[eq:master1\]) will be replaced by a system of coupled, linear, first order differential equations for all master integrals of the topology under consideration. The determination of the master integrals from these equations follows the same lines as discussed above, with the only difference that the general solution for the system of coupled differential equations is harder to obtain than for a single equation. In case the coupled equations can not be decoupled by an appropriate choice of variables, several mathematical techniques can be employed here [@kamke]: the system of $n$ coupled first order differential equations can for example be rewritten into one $n$-th order differential equation, which is then solved with standard methods. In some cases, the system can also be transformed into a form which is known to be solved by generalised hypergeometric series [@kamke; @bateman; @grad; @exton].
It is clear from the above discussion, that the determination of a master integral of a certain topology with $t$ different denominators requires the knowledge of all the integrals appearing in the inhomogeneous term. These integrals are subtopologies of the topology of the integral under consideration, and contain at most $t-1$ different denominators. The determination of master integrals has therefore to proceed bottom-up from simpler topologies with a small number of different denominators towards more complicated topologies with an increasing number of different denominators. For the case under special consideration, massless two-loop four-point functions with up to one external leg off-shell, this implies that one has to progress from the simplest master integrals with $t=3$ (off-shell two-point function) to construct all master integrals up to $t=7$.
A Pedagogical Example: the One-Loop Four-Point Function {#sec:oneloop}
=======================================================
To illustrate how the method explained above works in practice, we present in this section a detailed and self-contained derivation of the one-loop four-point function $$\boxLO = \int \frac{\d^d k}{(2\pi)^d} \frac{1}{k^2 (k-p_2)^2
(k-p_2-p_3)^2 (k-p_1-p_2-p_3)^2}\; .
\label{eq:boxLOdef}$$
The topology of this integral is given by the set of its four propagators, it is the only $t=4$ topology at one loop. In the notation introduced above, (\[eq:boxLOdef\]) corresponds to an integral $I_{4,4,0}$, the simplest integral of this topology with all propagators appearing in first power and with no scalar products. For the one-loop four-point function, one has four independent scalar products involving $k_{\mu}$ ($k^2$ and $k_{\mu}p_i^{\mu}$) and four linear independent denominators. This implies that any scalar product involving $k_{\mu}$ can be rewritten as linear combination of propagators and invariants $s_{ij} \equiv 2 p_{i \mu}
p_j^{\mu}$. Consequently, no integrals with scalar products in the numerator can appear for this topology.
The reduction of all four integrals with one squared propagator ($I_{4,5,0}$) can be carried out by considering the IBP identities for $I_{4,4,0}$: $$\int \frac{\d^d k}{(2\pi)^d} \frac{\partial}{\partial k^{\mu}}
\frac{v^{\mu}}{k^2 (k-p_2)^2
(k-p_2-p_3)^2 (k-p_1-p_2-p_3)^2} = 0 \;,
\label{eq:IBPLO}$$ where $v^{\mu}$ can be the loop momentum $k^{\mu}$ or any of the external momenta $p_i^{\mu}$, thus yielding four identities. It turns out that LI identities and IBP identities for integrals with higher powers of the propagators do not contain additional information which would allow a reduction of $I_{4,4,0}$. The integral (\[eq:boxLOdef\]) is therefore a master integral. It is the only master integral for this topology.
To proceed towards the differential equations in the invariants $\sab$, $\sac$ and $\sbc$ for this master integral, let us consider the derivatives in the external momenta: $$\begin{aligned}
p_1^{\mu} \frac{\partial}{\partial p_1^{\mu}} \boxLO & = &
\int \frac{\d^d k}{(2\pi)^d} \frac{1}{k^2 (k-p_2)^2
(k-p_2-p_3)^2 (k-p_1-p_2-p_3)^2} \nonumber \\
& &
\left( \frac{2p_1^{\mu}
(k-p_1-p_2-p_3)_{\mu}}{(k-p_1-p_2-p_3)^2}\right)\;, \\
p_2^{\mu} \frac{\partial}{\partial p_2^{\mu}} \boxLO & = &
\int \frac{\d^d k}{(2\pi)^d} \frac{1}{k^2 (k-p_2)^2
(k-p_2-p_3)^2 (k-p_1-p_2-p_3)^2}\nonumber \\
& &
\left(\frac{2p_2^{\mu}
(k-p_1-p_2-p_3)_{\mu}}{(k-p_1-p_2-p_3)^2} + \frac{2p_2^{\mu}
(k-p_2-p_3)_{\mu}}{(k-p_2-p_3)^2} + \frac{2p_2^{\mu}
(k-p_2)_{\mu}}{(k-p_2)^2}\right)\;, \\
p_3^{\mu} \frac{\partial}{\partial p_3^{\mu}} \boxLO & = &
\int \frac{\d^d k}{(2\pi)^d} \frac{1}{k^2 (k-p_2)^2
(k-p_2-p_3)^2 (k-p_1-p_2-p_3)^2} \nonumber \\
& &
\left(\frac{2p_3^{\mu}
(k-p_1-p_2-p_3)_{\mu}}{(k-p_1-p_2-p_3)^2} + \frac{2p_3^{\mu}
(k-p_2-p_3)_{\mu}}{(k-p_2-p_3)^2} \right)\; . \end{aligned}$$
The right hand sides of the above equations contain terms with
- four different propagators with one squared propagator, no scalar product
- four different propagators, no squared propagator, no scalar product
- three different propagators, one squared propagator.
The terms of type (i) can now be reduced to type (ii) and (iii) by using the integration-by-parts identities (\[eq:IBPLO\]). One obtains: $$\begin{aligned}
p_1^{\mu} \frac{\partial}{\partial p_1^{\mu}} \boxLO & = & - \boxLO
+ \boxLOpamb\; , \nonumber \\
p_2^{\mu} \frac{\partial}{\partial p_2^{\mu}} \boxLO & = & - \boxLO
+ \boxLOpdmc\; , \nonumber \\
p_3^{\mu} \frac{\partial}{\partial p_3^{\mu}} \boxLO & = & (d-6) \boxLO
- \boxLOpamb - \boxLOpdmc\; , \nonumber \end{aligned}$$ where $(\bullet)$ denotes a squared propagator and $(\times)$ stands for a pinched (cancelled) propagator.
A check on these equations is provided by the rescaling relation (\[eq:rescale\]) $$\left[ 4-\frac{d}{2} + \sab \frac{\partial}{\partial\sab} + \sac
\frac{\partial}{\partial\sac} +
\sbc \frac{\partial}{\partial\sbc} \right] \boxLO = 0\; ,$$ which is related to the above equations by $$\sab \frac{\partial}{\partial\sab} + \sac
\frac{\partial}{\partial\sac} +
\sbc \frac{\partial}{\partial\sbc} = \frac{1}{2} \sum_{i=1}^3 p_i^{\mu}
\frac{\partial}{\partial p_i^{\mu}} \; .$$ Inserting the derivatives obtained above, one finds that the rescaling relation is indeed fulfilled.
The three-propagator terms appearing in the above equations can be further reduced by using the IBP identities for the corresponding vertex; the vertex amplitudes, in turn, can all be expressed in terms of bubble integrals, and one finally finds $$\boxLOpamb = \frac{d-3}{p_2\cdot (p_1+p_3)} \left[
\frac{1}{(p_1+p_2+p_3)^2} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill} - \frac{1}{(p_1+p_3)^2}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{13}$}}
\end{picture}
}}
\hfill} \right]\; .$$ A similar identity is obtained by exchanging $p_1
\leftrightarrow p_2$.
The differential equations for the one-loop box integral then follow from (\[eq:derivatives\]). The set of differential equations reads: $$\begin{aligned}
\sab \frac{\partial}{\partial \sab}
\boxLO
& = & - \frac{d-4}{2} \boxLO \nonumber \\
& & + \frac{2(d-3)}{\sab+\sac}\left[\frac{1}{\sabc} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill}
-\frac{1}{\sbc} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{23}$}}
\end{picture}
}}
\hfill} \right] \nonumber \\
& & + \frac{2(d-3)}{\sab+\sbc}\left[\frac{1}{\sabc} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill}
-\frac{1}{\sac} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{13}$}}
\end{picture}
}}
\hfill} \right]\;, \label{eq:boxLOsab}\\
\sac \frac{\partial}{\partial \sac}
\boxLO
& = & \frac{d-6}{2} \boxLO \nonumber \\
& & - \frac{2(d-3)}{\sab+\sac}\left[\frac{1}{\sabc} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill}
-\frac{1}{\sbc} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{23}$}}
\end{picture}
}}
\hfill} \right]\;, \label{eq:boxLOsac}\\
\sbc \frac{\partial}{\partial \sbc}
\boxLO
& = & \frac{d-6}{2} \boxLO \nonumber \\
& & - \frac{2(d-3)}{\sab+\sbc}\left[\frac{1}{\sabc} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill}
-\frac{1}{\sac} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{13}$}}
\end{picture}
}}
\hfill} \right]\;, \label{eq:boxLOsbc}\end{aligned}$$ where $\sabc = \sab+\sac + \sbc$. The one-loop bubble diagrams in the inhomogenous term yield: $${
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p$}}
\end{picture}
}}
\hfill} =
\left[\frac{(4\pi)^{\frac{4-d}{2}}}{16\pi^2}\frac{ \Gamma (3-d/2)
\Gamma^2 (d/2-1)}{ \Gamma (d-3)} \right] \; \frac{-2i}{(d-4)(d-3)}
\left( -p^2\right)^{\frac{d-4}{2}} \equiv A_{2,LO}
\left( -p^2\right)^{\frac{d-4}{2}} \;.$$
The boundary conditions in $\sij=0$ can be readily read off from the above: $$\begin{aligned}
\boxLO(\sab=0) & = & \frac{4(d-3)}{(d-4)}\frac{1}{\sac\sbc}\Bigg[
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill} \nonumber \\
& & \hspace{1.3cm}
- {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{13}$}}
\end{picture}
}}
\hfill} -
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{23}$}}
\end{picture}
}}
\hfill} \Bigg]\;, \label{eq:boundLOsab} \\
\boxLO(\sac=0) & = & \frac{4(d-3)}{(d-6)}\frac{1}{\sab}\Bigg[\frac{1}{\sabc}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill} - \frac{1}{\sbc} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{23}$}}
\end{picture}
}}
\hfill} \Bigg]\;, \\
\boxLO(\sbc=0) & = & \frac{4(d-3)}{(d-6)}\frac{1}{\sab}\Bigg[\frac{1}{\sabc}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill} - \frac{1}{\sac} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{13}$}}
\end{picture}
}}
\hfill} \Bigg] \;.\end{aligned}$$ The result for the one-loop box integral can in principle be obtained by integrating any of the differential equations (\[eq:boxLOsab\])–(\[eq:boxLOsbc\]). In practice, it turns out to be more appropriate to introduce a new set of variables, namely $\sac$, $\sbc$ and $\sabc=\sab+\sac+\sbc$, corresponding to the arguments appearing in the two-point functions in the inhomogeneous terms. This transformation yields a differential equation in $\sabc$, which will be used for integration. Note that this transformation also modifies the differential equations in $\sac$ and $\sbc$. The differential equation in $\sabc$ reads: $$\begin{aligned}
\lefteqn{\frac{\partial}{\partial \sabc}
\boxLO
+ \frac{d-4}{2(\sabc-\sac-\sbc)} \boxLO = } \nonumber \\
& & + \frac{2(d-3)}{(\sabc-\sbc)(\sabc-\sac-\sbc)}
\left[\frac{1}{\sabc} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill}
-\frac{1}{\sbc} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{23}$}}
\end{picture}
}}
\hfill} \right] \nonumber \\
& & + \frac{2(d-3)}{(\sabc-\sac)(\sabc-\sac-\sbc)}
\left[\frac{1}{\sabc} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill}
-\frac{1}{\sac} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{13}$}}
\end{picture}
}}
\hfill} \right] \label{eq:boxLOq}\; .
\label{eq:boxLOmaster}\end{aligned}$$ The boundary condition in $\sabc=0$ can not be trivially determined from this equation, reflecting the fact that the massless one-loop box integral with all external legs on shell is not reducible to simpler subtopologies by IBP identities. The boundary condition in $\sabc=-\sac-\sbc$ can however be determined from (\[eq:boundLOsab\]).
Equation (\[eq:boxLOmaster\]) is a linear, inhomogeneous first order differential equation of the form $$\frac{\partial y(x)}{\partial x} + f(x) y(x) = g(x),$$ which can be solved by introducing an integrating factor (see for instance [@bronstein] or any standard book on differential equations) $$M(x) = e^{\int f(x) \d x},$$ such that $y(x)=1/M(x)$ solves the homogenous differential equation ($g(x)=0$). This yields the general solution of the inhomogenous equation as $$y(x) = \frac{1}{M(x)} \left( \int g(x) M(x) \d x + C\right),$$ where the integration constant $C$ can be adjusted to match the boundary conditions.
For (\[eq:boxLOmaster\]), we have at once the integrating factor $$M(\sabc) = (\sac + \sbc - \sabc )^{\frac{d-4}{2}}\; .
\label{eq:intfac}$$ This factor is not unambiguous, since $$M'(\sabc) = (\sabc -\sac - \sbc)^{\frac{d-4}{2}}\; .
\label{eq:intfacal}$$ would also be a valid integrating factor. We select (\[eq:intfac\]) by requiring a real integrating factor in the region: $-\sabc \ge -\sac - \sbc \ge 0$. The final result for the box integral does not depend on the selection of the integrating factor; using (\[eq:intfacal\]), one must however be more careful in applying analytic continuation formulae and in multiplying non-integer powers of the invariants.
With the integrating factor (\[eq:intfac\]), the one-loop box integral reads: $$\begin{aligned}
\boxLO (\sabc,\sac,\sbc) &=&
2(d-3) A_{2,LO} \left( \sac + \sbc - \sabc
\right)^{2-\frac{d}{2}}
\int^{\sabc}\hspace{-0.25cm} \d \sabcp \left( \sac + \sbc - \sabcp
\right)^{\frac{d}{2}-3}\nonumber \\
& & \hspace{-0.9cm}
\Bigg[ \frac{\left(-\sac\right)^{\frac{d}{2}-3}}{ \sac - \sabcp }
+\frac{\left(-\sbc\right)^{\frac{d}{2}-3}}{\sbc - \sabcp}
- \frac{2\sabcp - \sac - \sbc}{(\sac - \sabcp)(\sbc - \sabcp) }
\left(-\sabcp\right)^{\frac{d}{2}-3} \Bigg].
\label{eq:boxLOintform}\end{aligned}$$ From a computational point of view, the evaluation of the box amplitude of (\[eq:boxLOdef\]) has been reduced to the one dimensional integration corresponding to the solution of (\[eq:boxLOmaster\]).
The lower boundary of the integral is independent of $\sabc$ and can be adjusted arbitrarily. The first two terms in (\[eq:boxLOintform\]) can be easily integrated by shifting the integration variable to $\sabcp - \sac - \sbc$, which is then integrated between $0$ and $\sabc - \sac - \sbc$. To integrate the last term, one introduces a new variable $\sabcp (\sabcp - \sac - \sbc)$, which is integrated between $0$ and $\sabc (\sabc - \sac - \sbc)$. The resulting integrals yield can be identified as integral representation of the hypergeometric function $\,_2F_1$. With this choice of variables and boundaries, no constant term is required to match the boundary conditions. The result for the one-loop box integral then reads: $$\begin{aligned}
\lefteqn{\boxLO = - \frac{4(d-3)}{d-4} A_{2,LO}
\frac{1}{\sac\sbc}} \nonumber \\
& & \hspace{0.46cm}
\Bigg[ \left(\frac{\sac \sbc}{\sac - \sabc}\right)^{\frac{d}{2}-2}
\,_2F_1\left(d/2-2, d/2-2; d/2-1; \frac{\sabc - \sac - \sbc}{\sabc -\sac}
\right)\nonumber \\
& & \hspace{0.4cm}
+ \left(\frac{\sac \sbc}{\sbc - \sabc}\right)^{\frac{d}{2}-2}
\,_2F_1\left(d/2-2, d/2-2; d/2-1; \frac{\sabc - \sac - \sbc}{\sabc -\sbc}
\right) \nonumber \\
& & \hspace{0.4cm} - \left( \frac{-\sabc \sac \sbc}{(\sac - \sabc)
(\sbc - \sabc)}
\right)^{\frac{d}{2}-2} \,_2F_1\left(d/2-2, d/2-2; d/2-1;
\frac{\sabc (\sabc - \sac - \sbc)}{(\sabc -\sac)(\sabc - \sbc)}\right) \Bigg]
\;.
\nonumber \\\end{aligned}$$ The invariants appearing in this expression can be safely continued from the region $-\sabc \ge -\sac - \sbc \ge 0$ to the physical region $\sabc \ge \sac + \sbc \ge 0$. The arguments of the hypergeometric functions are ratios of invariants, they are not changed by the analytic continuation. The non-integer powers of invariants appearing as coefficients acquire imaginary parts, their signs are uniquely determined by the convention $-p^2 = -p^2 - i0$, thus fixing the imaginary part of the whole expression. The above equation reproduces the well-known result from the literature, e.g. [@ert]. It should be kept in mind that in applying dimensional regularisation, no distinction between infrared and ultraviolet poles, which both show up as $1/(d-4)$ in the above equation and in the coefficient $A_{2,LO}$, is made. Needless to say, those singularities are an intrinsic feature of the dimensional regularisation, and by non means an artifact of the differential equation approach. In particular, the most singular part of the above one loop integral, which is ultraviolet finite, as well as of the two loop integrals discussed in the following section, arises from soft configurations, and can in principle be re-derived by applying a strong ordering procedure [@catani] to the integrands.
When the box integral is expressed in the above form, where no expansion around $d=4$ has yet been performed, analytic continuations, e.g. to the on-shell case $\sabc=0$ or to collinear and soft limits $s_{ij}=0$, can be made with ease.
Results on Two-Loop Four-Point Functions {#sec:twoloop}
========================================
In the following, we shall outline how the techniques derived in Sections \[sec:master\] and \[sec:diffeq\] can be applied to the computation of two-loop integrals appearing in amplitudes for the decay of one massive into three massless particles: two-loop four-point functions with one off-shell leg. The main purpose of this section is to illustrate applications of the tools developed above to non-trivial problems; the list of integrals given here is far from complete. We provide a comprehensive list of master integrals only up to $t=5$, for $t=6$ and $t=7$ only reducible integrals are quoted.
A prefactor common to all massless scalar integrals is $$S_d = \left[(4\pi)^{\frac{4-d}{2}}\frac{ \Gamma (3-d/2)
\Gamma^2 (d/2-1)}{ \Gamma (d-3)} \right] \; ,$$ which is also appearing in all counterterms in the $\overline{{\rm
MS}}$–scheme.
In the following, the notation of external momenta is as follows: $p_i$ denotes an on-shell momentum, $p_{ij(k)}$ denotes an off-shell momentum, being the sum of two (three) on-shell momenta $p_i$, $p_j$(, $p_k$) with $s_{ij(k)}=(p_{ij(k)})^2$. $p$ is an arbitrary momentum.
$t=3$
-----
For two-loop integrals with $t=3$, only one topology exists: the two-loop vacuum bubble. The corresponding integral fulfils a homogeneous differential equation, which can not be used to infer any boundary condition. The integral can however be computed using Feynman parameters: $$\begin{aligned}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p$}}
\end{picture}
}}
\hfill} &=& \left(\frac{S_d}{16 \pi^2}\right)^2
\frac{\Gamma(5-d)\Gamma^2(d-3)}{\Gamma^2(3-d/2)\Gamma(d/2-1)\Gamma(3d/2-3)}
\frac{1}{(d-3)(d-4)}
\left( -p^2 \right)^{d-3} \nonumber \\
&\equiv & A_3 \left( -p^2 \right)^{d-3}\;.\end{aligned}$$
$t=4$
-----
Several different two-loop topologies exist for $t=4$. Two types of two-point functions are encountered. The first can be reduced to $${
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1.5,1){\oval(1,1)[bl]}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p$}}
\end{picture}
}}
\hfill}= \frac{3d-8}{d-4} \frac{1}{p^2} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p$}}
\end{picture}
}}
\hfill}$$ using IBP identities. The second is the product of two one-loop bubble integrals and yields $${
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(2.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(2,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p$}}
\end{picture}
}}
\hfill} = \left( A_{2,LO} \right)^2 (-p^2)^{d-4}\; ,$$ which generalises trivially to a three-point function $${
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.5)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(2.75,0.5){\vector(1,0){0.1}}
\put(1.5,1.0){\vector(0,1){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(0,1){0.6}}
\put(2.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(2,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\put(2.75,0.7){\makebox(0,0)[b]{$p_{12}$}}
\put(1.5,1.25){\makebox(0,0)[b]{$p_3$}}
\end{picture}
}}
\hfill}
= \left( A_{2,LO} \right)^2
(-\sabc)^{\frac{d}{2}-2}(-\sab)^{\frac{d}{2}-2}\; .
\label{eq:doublebubblex}$$
Only one of the master integrals at $t=4$ fulfils a homogeneous differential equation: a two-loop vertex integral with one off-shell leg. This integral can also be computed using Feynman parameters: $$\begin{aligned}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{12}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_2$}}
\end{picture}
}}
\hfill}
&=& \left(\frac{S_d}{16 \pi^2}\right)^2 \frac{ \Gamma^3(d-3)
\Gamma(5-d)}{\Gamma(3-d/2)\Gamma^2(d/2-1)
\Gamma(3d/2-4)}\frac{-2}{(d-3)(d-4)^2}
\left(-s_{12}\right)^{d-4} \nonumber\\
&\equiv & A_4 \left(-s_{12}\right)^{d-4}\; .\end{aligned}$$ The other vertex integral topology with one off-shell leg can be reduced using IBP identities: $${
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,1.2){\oval(1,1)[br]}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{12}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_2$}}
\end{picture}
}}
\hfill} = -\frac{3d-8}{d-4} \frac{1}{s_{12}}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{12}$}}
\end{picture}
}}
\hfill} \; .$$
Among the three vertex integrals with two off-shell legs, only one can be reduced using IBP and LI identities: $${
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,1.2){\oval(1,1)[br]}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{12}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} = \frac{3d-8}{d-4} \frac{1}{s_{123}
- s_{12}} \left[ {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{12}$}}
\end{picture}
}}
\hfill} - {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill} \right] \;.$$ The two remaining ones are master integrals. Written out in terms of propagators, they read: $$\begin{aligned}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{12}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} & =&
\int \frac{\d^d k}{(2\pi)^d}\frac{\d^d l}{(2\pi)^d} \frac{1}{
k^2 l^2 (k-p_{123})^2 (k-l-p_{12})^2} \; ,
\label{eq:firstMI} \\
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,1.2){\oval(1,1)[br]}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_3$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{12}$}}
\end{picture}
}}
\hfill} & =&
\int \frac{\d^d k}{(2\pi)^d}\frac{\d^d l}{(2\pi)^d} \frac{1}{
k^2 l^2 (l-p_{12})^2 (k-l-p_{3})^2} \; . \end{aligned}$$ Both fulfil inhomogeneous differential equations. For a vertex $p_{123} \to p_{12} + p_3$, the appropriate variables for the differential equations are $s_{123}$ and $s_{12}$. To illustrate the structure of the differential equations, we quote them for (\[eq:firstMI\]): $$\begin{aligned}
s_{123} \frac{\partial}{\partial s_{123}}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{12}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} & = & \frac{d-4}{2}\; \frac{2\sabc -
\sab}{\sabc - \sab} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{12}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber \\
& &-
\frac{3d-8}{2}\; \frac{1}{\sabc - \sab} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{12}$}}
\end{picture}
}}
\hfill}\;, \nonumber \\
\sab \frac{\partial}{\partial \sab} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{12}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill}
& = & - \frac{d-4}{2}\; \frac{\sab}{\sabc - \sab}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{12}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber \\
&& +
\frac{3d-8}{2}\; \frac{1}{\sabc - \sab} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{12}$}}
\end{picture}
}}
\hfill}\;.\end{aligned}$$ The boundary conditions for $\sabc=0$ or $\sab=0$ are obtained directly from the vertex integrals with one off-shell leg quoted above. Using these, one finds $$\!{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{12}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill}\! = A_4 \left( \sab -
\sabc\right)^{\frac{d}{2}-2} \left(-\sabc\right)^{\frac{d}{2}-2}
- \frac{3d-8}{2(d-3)} A_3 \frac{\left(-\sab\right)^{d-3}}{-\sabc}
\,_2F_1\left(\frac{d}{2}-1, 1; d-2; \frac{\sab}{\sabc}\right).$$ The second master integral can be obtained from this by analytic continuation of the hypergeometric function: $${
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,1.2){\oval(1,1)[br]}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_3$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{12}$}}
\end{picture}
}}
\hfill} = -\frac{3d-8}{d-4} A_3
\left( -\sab \right)^{\frac{d}{2}-2} \left(-\sabc\right)^{\frac{d}{2}-2}
\,_2F_1\left(\frac{d}{2}-1, 2-\frac{d}{2}; 3-\frac{d}{2};
\frac{\sabc-\sab}{\sabc}\right)\; .$$
Vertex integrals with three off-shell legs can not appear as subtopologies in two-loop four-point functions with one off-shell leg.
$t=5$
-----
The two-loop two-point function with $t=5$ is a well known example [@chet; @krev] for the application of IBP identities: $${
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(1,0){0.5}}
\put(1,1){\line(0,-1){1}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p$}}
\end{picture}
}}
\hfill} =
\frac{2(3d-8)(3d-10)}{(d-4)^2} \frac{1}{\left(p^2\right)^2}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p$}}
\end{picture}
}}
\hfill}- \frac{2(d-3)}{d-4} \frac{1}{p^2} {
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(2.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(2,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p$}}
\end{picture}
}}
\hfill}\; .$$
The four different $t=5$ three-point functions with one off-shell leg can also be reduced by using IBP and LI identities: $$\begin{aligned}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.9,0.2){\vector(1,0){0.1}}
\put(1.9,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,0.7){\line(2,1){1}}
\put(0.5,0.7){\line(2,-1){1}}
\put(1.5,1.2){\line(0,-1){1}}
\put(1.5,1.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(1,0){0.5}}
\put(0.5,0.7){\line(1,0){1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{12}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_2$}}
\end{picture}
}}
\hfill} &= & \frac{(d-3)(3d-8)(3d-10)}{(d-4)^3}
\;\frac{1}{(\sab)^2} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{12}$}}
\end{picture}
}}
\hfill}\;,\\
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.9,0.2){\vector(1,0){0.1}}
\put(1.9,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,0.7){\line(2,1){1}}
\put(0.5,0.7){\line(2,-1){1}}
\put(1.5,1.2){\line(0,-1){1}}
\put(1.5,1.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(-1,2){0.4}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{12}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_2$}}
\end{picture}
}}
\hfill} &= & \frac{(d-3)(3d-10)}{(d-4)^2}
\; \frac{1}{\sab} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{12}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_2$}}
\end{picture}
}}
\hfill} \nonumber \\
&& -\frac{(d-3)(3d-8)(3d-10)}{(d-4)^3}
\;\frac{1}{(\sab)^2} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{12}$}}
\end{picture}
}}
\hfill}\;, \\
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.9,0.2){\vector(1,0){0.1}}
\put(1.9,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,0.7){\line(2,-1){1}}
\put(0.5,0.7){\line(1,1){0.5}}
\put(1.5,1.2){\line(0,-1){1}}
\put(1.5,1.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(1,0){0.5}}
\put(1.25,1.2){\circle{0.5}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{12}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_2$}}
\end{picture}
}}
\hfill} &= & -\frac{(3d-8)(3d-10)}{(d-4)^2}
\;\frac{1}{(\sab)^2} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{12}$}}
\end{picture}
}}
\hfill}\;,\\
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.2){\vector(1,0){0.1}}
\put(1.9,0.2){\vector(1,0){0.1}}
\put(1.9,1.2){\vector(1,0){0.1}}
\put(0,0.2){\line(1,0){0.5}}
\put(0.5,0.2){\line(1,1){1}}
\put(0.5,0.2){\line(1,0){1}}
\put(1.5,1.2){\line(0,-1){0.5}}
\put(1.5,1.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(1,0){0.5}}
\put(1.5,0.45){\circle{0.5}}
\put(0.25,0.4){\makebox(0,0)[b]{$p_{12}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_2$}}
\end{picture}
}}
\hfill} &= & -\frac{3d-10}{2(d-4)}
\; \frac{1}{\sab} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{12}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_2$}}
\end{picture}
}}
\hfill}\;.\end{aligned}$$
By applying IBP and LI identities, it is likewise possible to reduce all but one $t=5$ three-point function with two off-shell legs: $$\begin{aligned}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.9,0.2){\vector(1,0){0.1}}
\put(1.9,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,0.7){\line(2,1){1}}
\put(0.5,0.7){\line(2,-1){1}}
\put(1.5,1.2){\line(0,-1){1}}
\put(1.5,1.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(1,0){0.5}}
\put(0.5,0.7){\line(1,0){1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{12}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} &= & - \frac{(d-3)(3d-10)}{(d-4)^2}
\; \frac{1}{\sabc - \sab} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,1.2){\oval(1,1)[br]}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_3$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{12}$}}
\end{picture}
}}
\hfill} \nonumber\\
&& + \frac{(d-3)(3d-8)(3d-10)}{(d-4)^3}
\;\frac{1}{\sab(\sabc-\sab)} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{12}$}}
\end{picture}
}}
\hfill}\;, \\
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.9,0.2){\vector(1,0){0.1}}
\put(1.9,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,0.7){\line(2,1){1}}
\put(0.5,0.7){\line(2,-1){1}}
\put(1.5,1.2){\line(0,-1){1}}
\put(1.5,1.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(-1,2){0.4}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_3$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{12}$}}
\end{picture}
}}
\hfill} &= &
\frac{(d-3)(3d-10)}{(d-4)^2}
\; \frac{1}{\sabc - \sab} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{12}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber\\
&& - \frac{(d-3)(3d-8)(3d-10)}{(d-4)^3}
\;\frac{1}{\sabc(\sabc-\sab)} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill}\;, \\
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.9,0.2){\vector(1,0){0.1}}
\put(1.9,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,0.7){\line(2,-1){1}}
\put(0.5,0.7){\line(1,1){0.5}}
\put(1.5,1.2){\line(0,-1){1}}
\put(1.5,1.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(1,0){0.5}}
\put(1.25,1.2){\circle{0.5}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{12}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} &= & \frac{(3d-8)(3d-10)}{(d-4)^2}
\; \frac{1}{\sab(\sabc-\sab)} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{12}$}}
\end{picture}
}}
\hfill}
\nonumber \\
&& - \frac{(3d-8)(3d-10)}{(d-4)^2}\;\frac{1}{\sabc(\sabc-\sab)}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill}\;, \\
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.9,0.2){\vector(1,0){0.1}}
\put(1.9,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,0.7){\line(2,-1){1}}
\put(0.5,0.7){\line(1,1){0.5}}
\put(1.5,1.2){\line(0,-1){1}}
\put(1.5,1.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(1,0){0.5}}
\put(1.25,1.2){\circle{0.5}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_3$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{12}$}}
\end{picture}
}}
\hfill} &= & \frac{3d-10}{2(d-4)}
\; \frac{1}{\sabc-\sab} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,1.2){\oval(1,1)[br]}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_3$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{12}$}}
\end{picture}
}}
\hfill} \nonumber\\
&& - \frac{(3d-8)(3d-10)}{2(d-4)^2}\;\frac{1}{\sabc(\sabc-\sab)}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill}\;, \\
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.2){\vector(1,0){0.1}}
\put(1.9,0.2){\vector(1,0){0.1}}
\put(1.9,1.2){\vector(1,0){0.1}}
\put(0,0.2){\line(1,0){0.5}}
\put(0.5,0.2){\line(1,1){1}}
\put(0.5,0.2){\line(1,0){1}}
\put(1.5,1.2){\line(0,-1){0.5}}
\put(1.5,1.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(1,0){0.5}}
\put(1.5,0.45){\circle{0.5}}
\put(0.25,0.4){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{12}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} &= & - \frac{3d-10}{2(d-4)}
\; \frac{1}{\sabc-\sab} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{12}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill}\nonumber\\
&&+ \frac{(3d-8)(3d-10)}{2(d-4)^2}\;\frac{1}{\sab(\sabc-\sab)}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{12}$}}
\end{picture}
}}
\hfill}\; .\end{aligned}$$
The remaining three-point function is a master integral, which can be found by solving the corresponding differential equations: $$\begin{aligned}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.9,0.2){\vector(1,0){0.1}}
\put(1.9,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,0.7){\line(2,1){1}}
\put(0.5,0.7){\line(2,-1){1}}
\put(1.5,1.2){\line(0,-1){1}}
\put(1.5,1.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(-1,2){0.4}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{12}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} &= &\frac{2(d-3)}{d-4}\,
\left(A_{2,LO}\right)^2\;
\left(-\sab\right)^{d-4} \frac{1}{-\sabc}
\,_2F_1\left(
1-\frac{d}{2},d-3; d-2; \frac{\sabc-\sab}{\sabc} \right) \nonumber \\
& & - \frac{(3d-8)(3d-10)}{(d-4)^2} \, A_3 \nonumber \\
& &
\Bigg[
\frac{1}{\sab}\, \left(-\sabc\right)^{d-4}
\,_3F_2\left( \frac{d}{2}-1,1,d-3; 3-\frac{d}{2},d-2;
\frac{\sab-\sabc}{\sab} \right) \nonumber \\
&& + \frac{1}{\sabc}\, \left(-\sab\right)^{d-4}
\,_3F_2\left( \frac{d}{2}-1,1,d-3; 3-\frac{d}{2},d-2;
\frac{\sabc-\sab}{\sabc} \right)\Bigg] \; .\end{aligned}$$
For $t=5$, one finds four different topologies for two-loop four-point functions. These are all master integrals obeying inhomogeneous differential equations in the external invariants. Solving these equations and matching the boundary conditions, all master integrals can be determined. $$\begin{aligned}
{
\mbox{\parbox{4cm}{\hspace{0.25cm}
\begin{picture}(3.5,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.8,0.2){\vector(1,0){0.1}}
\put(2.8,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2.5}}
\put(0.5,1.2){\line(1,0){2.5}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.7){\circle{1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(3.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(3.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} & = & \frac{3d-10}{d-4}
\, A_4 \, \frac{1}{\sbc}\left(-\sac\right)^{d-4}
\left(\frac{\sab+\sac}{\sbc}\right)^{3-d}\nonumber \\
&& \hspace{1.4cm}
\,_2F_1\left(d-3,\frac{d}{2}-2;\frac{d}{2}-1;\frac{\sab}{\sab+\sac}\right)
\nonumber\\
&& -\frac{3d-10}{2(d-4)}\, A_4 \, \frac{1}{\sab}
\left( -\sab-\sbc\right)^{d-4}
\left(\frac{\sab+\sac}{\sab}\right)^{\frac{d}{2}-3}\nonumber \\
&& \hspace{1.4cm}
\,_2F_1\left(3-\frac{d}{2},4-d;5-d;\frac{\sac\sbc}{(\sab+\sac)(\sab+\sbc)}
\right)\nonumber \\
&&\hspace{-1cm}
-\frac{(3d-8)(3d-10)}{4(d-3)(d-4)} \,A_3\,
\frac{1}{\sab\sabc}
\left(-\sac\right)^{d-3}
\left(\frac{\sab+\sbc}{\sabc}\right)^{\frac{d}{2}-2}
\left(\frac{\sab+\sac}{\sab}\right)^{\frac{d}{2}-3} \nonumber\\
&& \hspace{-1cm}
\,_2F_1\left(\frac{d}{2}-1,1;d-2;\frac{\sac}{\sabc}\right)
\,_2F_1\left(3-\frac{d}{2},4-d;5-d;\frac{\sac\sbc}{(\sab+\sac)(\sab+\sbc)
}\right)\nonumber \\
&&\hspace{-1cm}-\frac{(3d-8)(3d-10)}{2(d-4)^2}\, A_3 \, \frac{1}{-\sab-\sbc}
\left(-\sac\right)^{d-4}
\left(\frac{\sab+\sbc}{\sab+\sac}\right)^{\frac{d}{2}-1}
\left(\frac{\sabc}{\sab}\right)^{\frac{d}{2}-2}\nonumber \\
&& F_1\left(4-d,2-\frac{d}{2},2-\frac{d}{2},5-d,\frac{\sbc}{\sab+\sbc},
\frac{\sbc}{\sabc}\right) \;, \label{eq:firstFIVE}
\\
{
\mbox{\parbox{4cm}{\hspace{0.25cm}
\begin{picture}(3.5,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.8,0.2){\vector(1,0){0.1}}
\put(2.8,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2.5}}
\put(0.5,1.2){\line(1,0){2.5}}
\put(2.5,0.2){\line(0,1){1}}
\put(1.5,0.7){\circle{1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(3.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(3.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} & = &
\frac{(3d-10)(3d-8)}{(d-4)(d-6)}\, A_3\, \frac{1}{\sab+\sbc} \bigg[
\left(-\sac\right)^{d-4}
\,_2F_1\left(1,1;4-\frac{d}{2};\frac{\sbc}{\sab+\sbc} \right) \nonumber \\
&& \hspace{0.5cm} - \left(-\sabc\right)^{d-4}
F_1\left(1,1,4-d,4-\frac{d}{2};\frac{\sbc}{\sab+\sbc},
\frac{\sbc}{\sabc}\right) \bigg] \nonumber\\
&& - \frac{2(3d-8)(3d-10)}{(d-4)(d-6)}\, A_3\, \left(-\sabc\right)^{d-5}
\,_2F_1\left(1,2-\frac{d}{2};\frac{d}{2}-1;\frac{\sab}{\sab+\sac}\right)
\nonumber \\
&& \hspace{-0.7cm} \left[ \,_2F_1\left(1,5-d;4-\frac{d}{2}; \frac{\sbc}{\sabc}\right) +
\frac{d-6}{d-4}
\,_2F_1\left(1,5-d;3-\frac{d}{2}; \frac{\sab+\sac}{\sabc}\right) \right]
\!\! , \label{eq:secondFIVE}
\\
{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(2,0.2){\line(-1,1){1}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} & = & \frac{(3d-8)(3d-10)}{(d-4)^2}\;
A_3 \;\frac{1}{\sac\sbc} \;\Bigg[ \nonumber \\
&& \hspace{-2cm} -\left(\frac{\sac\sbc}{-(\sab+\sbc)}\right)^{d-3}
\,_2F_1\left( d-3,d-3;d-2;\frac{\sab}{\sab+\sbc}\right) \nonumber \\
&& \hspace{-2cm}-\left(\frac{\sac\sbc}{-(\sab+\sac)}\right)^{d-3}
\,_2F_1\left( d-3,d-3;d-2;\frac{\sab}{\sab+\sac}\right)
\label{eq:thirdFIVE}\\
&&\hspace{-2cm}
+\left(\frac{-\sabc\sac\sbc}{(\sab+\sbc)(\sab+\sac)}\right)^{d-3}
\,_2F_1\left(
d-3,d-3;d-2;\frac{\sab\sabc}{(\sab+\sac)(\sab+\sbc)}\right)\Bigg] \;,
\nonumber\\
{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(1,0.2){\line(1,1){1}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} & = & - A_4 \frac{(d-3)(3d-10)}{(d-4)^2}
\left( -\sabc \right)^{d-4} \left(-\sac-\sbc\right)^{3-d} \nonumber \\
&& \Bigg[ \left(-\sbc\right)^{d-4} F_1\left(4-d,d-3,2-\frac{d}{2},5-d,
\frac{\sac}{\sac+\sbc},\frac{\sac}{\sabc}\right)\nonumber \\
&& +\left(-\sac\right)^{d-4} F_1\left(4-d,d-3,2-\frac{d}{2},5-d,
\frac{\sbc}{\sac+\sbc},\frac{\sbc}{\sabc}\right)\Bigg]\nonumber \\
&& -A_3\frac{(3d-8)(3d-10)}{2(d-4)^2} \frac{1}{-\sabc} \nonumber \\
&& \Bigg[ \left(-\sac\right)^{d-4} S_1\left(\frac{d}{2}-1,1,1,d-2,5-d,
\frac{\sbc}{\sabc},\frac{\sac}{\sabc}\right)\nonumber \\
&&
+\left(-\sbc\right)^{d-4} S_1\left(\frac{d}{2}-1,1,1,d-2,5-d,
\frac{\sac}{\sabc},\frac{\sbc}{\sabc}\right)\Bigg]\;. \label{eq:fourthFIVE}\end{aligned}$$ The integrals (\[eq:firstFIVE\]) and (\[eq:secondFIVE\]) are one-loop bubble insertions into the one-loop box and have already been computed for arbitrary powers of the propagators in [@glover]. The integrals (\[eq:thirdFIVE\]) and (\[eq:fourthFIVE\]) were, to our knowledge, not known up to now. In the reduction of integrals of the topology (\[eq:fourthFIVE\]), one finds two master integrals, whose differential equations decouple in the variable $\Delta\equiv\sac-\sbc$. The second master integral for this topology can be found by rearranging one of the differential equations: $$\begin{aligned}
\lefteqn{{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(1,0.2){\line(1,1){1}}
\put(1.5,0.7){\circle*{0.2}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} =}\nonumber \\
&& - \frac{4\sab\sabc}{\sac\sbc}
\frac{\partial}{\partial\sab} {
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(1,0.2){\line(1,1){1}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill}\nonumber \\
&& + (d-4)\frac{3\sab+\sabc}{\sac\sbc} {
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(1,0.2){\line(1,1){1}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill}
\nonumber \\
&& -\frac{(d-3)(3d-10)}{d-4}\frac{1}{\sac\sbc} \left(
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{13}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_2$}}
\end{picture}
}}
\hfill} + {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{23}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_1$}}
\end{picture}
}}
\hfill}\right)
\nonumber \\
&& +\frac{2(d-3)(3d-8)(3d-10)}{(d-4)^2}\frac{1}{\sac\sbc}\left(
\frac{1}{\sac} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{13}$}}
\end{picture}
}}
\hfill} + \frac{1}{\sbc}{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{23}$}}
\end{picture}
}}
\hfill}
\right)\; .\end{aligned}$$
Finally, products of one-loop vertex with one-loop bubble integrals also yield topologies with $t=5$. These can all be reduced to (\[eq:doublebubblex\]) and are not quoted explicitly.
The complete list of integrals at $t=5$ which were derived in this section can now be used to compute all integrals at $t=6$ and $t=7$ which can be reduced using IBP and LI identities. The results of this reduction are summarised in the following.
$t=6$
-----
Two-loop integrals with $t=6$ arising in calculations in covariant gauges must be three- or four-point functions. Since we are concerned with subgraphs that can appear in the reduction of four-point functions with one off-shell leg, we need to consider three-point functions with up to two off-shell legs. For general three-point functions at $t=6$, one finds three distinct topologies: two planar and one crossed arrangement of the loop momenta. The crossed graphs correspond to master integrals, while the planar graphs are reducible, as first pointed out in [@kl], where the three-point integral with one off-shell leg was computed. We reproduce these results: $$\begin{aligned}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.9,0.2){\vector(1,0){0.1}}
\put(1.9,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,0.7){\line(1,1){0.5}}
\put(0.5,0.7){\line(1,-1){0.5}}
\put(1.5,1.2){\line(0,-1){1}}
\put(1.0,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{12}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_2$}}
\end{picture}
}}
\hfill} &= & \frac{3(d-3)(3d-10)}{(d-4)^2}
\frac{1}{\sab^2} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{12}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_2$}}
\end{picture}
}}
\hfill} \nonumber \\
&& + \frac{4(d-3)^2}{(d-4)^2} \frac{1}{\sab^2} {
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(2.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(2,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{12}$}}
\end{picture}
}}
\hfill}
\nonumber \\
&&-\frac{6(d-3)(3d-8)(3d-10)}{(d-4)^3} \frac{1}{\sab^3} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{12}$}}
\end{picture}
}}
\hfill}
\; ,\\
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.9,0.2){\vector(1,0){0.1}}
\put(1.9,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,0.7){\line(2,1){1}}
\put(0.5,0.7){\line(2,-1){1}}
\put(1.5,1.2){\line(0,-1){1}}
\put(1.5,1.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(1,0){0.5}}
\put(1.5,0.7){\line(-2,-1){0.5}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{12}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_2$}}
\end{picture}
}}
\hfill} &= & -\frac{3(d-3)(3d-10)}{2(d-4)(d-5)}
\frac{1}{\sab^2} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{12}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_2$}}
\end{picture}
}}
\hfill} \nonumber \\
&&+ \frac{3(d-3)(3d-8)(3d-10)}{(d-4)^2(d-5)}
\frac{1}{\sab^3} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{12}$}}
\end{picture}
}}
\hfill}\; .\end{aligned}$$ The results for two off-shell legs read: $$\begin{aligned}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.9,0.2){\vector(1,0){0.1}}
\put(1.9,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,0.7){\line(1,1){0.5}}
\put(0.5,0.7){\line(1,-1){0.5}}
\put(1.5,1.2){\line(0,-1){1}}
\put(1.0,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{23}$}}
\end{picture}
}}
\hfill} &= & \frac{1}{\sabc}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.9,0.2){\vector(1,0){0.1}}
\put(1.9,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,0.7){\line(2,1){1}}
\put(0.5,0.7){\line(2,-1){1}}
\put(1.5,1.2){\line(0,-1){1}}
\put(1.5,1.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(-1,2){0.4}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{23}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_1$}}
\end{picture}
}}
\hfill} + \frac{(d-3)(3d-10)}{(d-4)^2}
\frac{1}{(\sab+\sac)\sabc} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,1.2){\oval(1,1)[br]}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{23}$}}
\end{picture}
}}
\hfill} \nonumber \\
&&+\frac{2(d-3)(3d-10)}{(d-4)^2} \frac{1}{(\sab+\sac)\sabc}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{23}$}}
\end{picture}
}}
\hfill}\nonumber \\
&& -\frac{4(d-3)^2}{(d-4)^2} \frac{1}{(\sab+\sac)\sabc}
{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.5)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(2.75,0.5){\vector(1,0){0.1}}
\put(1.5,1.0){\vector(0,1){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(1.5,0.5){\line(0,1){0.6}}
\put(2.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(2,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\put(2.75,0.7){\makebox(0,0)[b]{$p_{23}$}}
\put(1.5,1.25){\makebox(0,0)[b]{$p_1$}}
\end{picture}
}}
\hfill}\nonumber \\
&& +\frac{4(d-3)^2}{(d-4)^2} \frac{1}{(\sab+\sac)\sabc}\; ,
{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){0.5}}
\put(2.5,0.5){\line(1,0){0.5}}
\put(1,0.5){\circle{1}}
\put(2,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill}\nonumber \\
&& +\frac{(d-3)(3d-8)(3d-10)}{(d-4)^3}
\frac{1}{\sbc(\sab+\sac)\sabc} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{23}$}}
\end{picture}
}}
\hfill} \nonumber \\
&& -\frac{4(d-3)(3d-8)(3d-10)}{(d-4)^3}
\frac{1}{(\sab+\sac)\sabc^2}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill}\; ,\\
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.9,0.2){\vector(1,0){0.1}}
\put(1.9,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,0.7){\line(2,1){1}}
\put(0.5,0.7){\line(2,-1){1}}
\put(1.5,1.2){\line(0,-1){1}}
\put(1.5,1.2){\line(1,0){0.5}}
\put(1.5,0.2){\line(1,0){0.5}}
\put(1.5,0.7){\line(-2,-1){0.5}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{23}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_1$}}
\end{picture}
}}
\hfill} &= &-\frac{3(d-3)(3d-10)}{2(d-4)(d-5)}
\frac{1}{(\sab+\sac)^2} \left({
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{23}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_1$}}
\end{picture}
}}
\hfill}
+\!\!{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,1.2){\oval(1,1)[br]}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{23}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_1$}}
\end{picture}
}}
\hfill}\right) \nonumber \\
&& +\frac{3(d-3)(3d-8)(3d-10)}{2(d-4)^2(d-5)}\frac{1}{\sbc(\sab+\sac)^2}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{23}$}}
\end{picture}
}}
\hfill} \nonumber \\
&& +\frac{3(d-3)(3d-8)(3d-10)}{2(d-4)^2(d-5)}\frac{1}{\sabc(\sab+\sac)^2}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill} \; .\end{aligned}$$
For $t=6$, both one-loop bubble insertions on propagators of the one-loop box can be reduced: $$\begin{aligned}
{
\mbox{\parbox{4cm}{\hspace{0.25cm}
\begin{picture}(3.5,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.8,0.2){\vector(1,0){0.1}}
\put(2.8,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2.5}}
\put(0.5,1.2){\line(1,0){2.5}}
\put(1,0.2){\line(0,1){1}}
\put(2,1.2){\line(0,-1){0.5}}
\put(2,0.45){\circle{0.5}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(3.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(3.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} & = & -\frac{3(\sab+\sac)}{\sac\sbc}
{
\mbox{\parbox{4cm}{\hspace{0.25cm}
\begin{picture}(3.5,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.8,0.2){\vector(1,0){0.1}}
\put(2.8,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2.5}}
\put(0.5,1.2){\line(1,0){2.5}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.7){\circle{1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(3.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(3.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber \\
&& + \frac{3(3d-10)}{2(d-4)} \frac{1}{\sac\sbc}\left(
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{23}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill}- {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{13}$}}
\end{picture}
}}
\hfill}
\right)\nonumber \\
&& -\frac{3(3d-8)(3d-10)}{2(d-4)^2} \frac{1}{\sac^2\sbc}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{13}$}}
\end{picture}
}}
\hfill}\; ,\\
{
\mbox{\parbox{4cm}{\hspace{0.25cm}
\begin{picture}(3.5,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.8,0.2){\vector(1,0){0.1}}
\put(2.8,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2.5}}
\put(0.5,1.2){\line(1,0){2.5}}
\put(2.5,0.2){\line(0,1){1}}
\put(1.5,1.2){\line(0,-1){0.5}}
\put(1.5,0.45){\circle{0.5}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(3.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(3.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} & = &
\frac{3}{\sac} {
\mbox{\parbox{4cm}{\hspace{0.25cm}
\begin{picture}(3.5,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.8,0.2){\vector(1,0){0.1}}
\put(2.8,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2.5}}
\put(0.5,1.2){\line(1,0){2.5}}
\put(2.5,0.2){\line(0,1){1}}
\put(1.5,0.7){\circle{1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(3.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(3.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill}
+ \frac{3(3d-10)}{2(d-4)} \frac{1}{\sac\sbc}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,1.2){\oval(1,1)[br]}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{23}$}}
\end{picture}
}}
\hfill} \nonumber \\
&& +\frac{3(3d-8)(3d-10)}{2(d-4)^2} \frac{1}{\sac\sbc\sabc}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill}\;.\end{aligned}$$
Among the remaining four planar diagrams at $t=6$, three are reducible to simpler subtopologies: $$\begin{aligned}
{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(1.5,0.2){\line(-1,2){0.5}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} & = & -3 \frac{\sab}{\sac\sbc}
{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(2,0.2){\line(-1,1){1}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber \\
&& -\frac{6(d-3)}{d-4}\frac{1}{\sac}
{
\mbox{\parbox{4cm}{\hspace{0.25cm}
\begin{picture}(3.5,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.8,0.2){\vector(1,0){0.1}}
\put(2.8,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2.5}}
\put(0.5,1.2){\line(1,0){2.5}}
\put(2.5,0.2){\line(0,1){1}}
\put(1.5,0.7){\circle{1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(3.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(3.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber \\
&& -\frac{3(d-3)(3d-10)}{(d-4)^2} \frac{1}{\sac\sbc}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,1.2){\oval(1,1)[br]}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{23}$}}
\end{picture}
}}
\hfill} \nonumber \\
&& -\frac{3(d-3)(3d-8)(3d-10)}{(d-4)^3} \nonumber \\
&& \hspace{1cm} \left(\frac{1}{\sac^2\sbc}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{13}$}}
\end{picture}
}}
\hfill}+ \frac{1}{\sac\sbc^2} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{23}$}}
\end{picture}
}}
\hfill}\right)\; , \\
{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(2,0.2){\line(-1,2){0.5}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} & = & -3 \frac{\sab}{\sac\sbc}
{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(2,0.2){\line(-1,1){1}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber \\
&& +\frac{6(d-3)}{d-4} \frac{\sab+\sac}{\sac\sbc}
{
\mbox{\parbox{4cm}{\hspace{0.25cm}
\begin{picture}(3.5,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.8,0.2){\vector(1,0){0.1}}
\put(2.8,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2.5}}
\put(0.5,1.2){\line(1,0){2.5}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.7){\circle{1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(3.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(3.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber \\
&& -\frac{3(d-3)(3d-10)}{(d-4)^2} \frac{1}{\sac\sbc}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{23}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber \\
&& +\frac{3(d-3)(3d-10)}{(d-4)^2} \frac{1}{\sac\sbc}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{13}$}}
\end{picture}
}}
\hfill} \nonumber \\
&& -\frac{3(d-3)(3d-8)(3d-10)}{(d-4)^3} \nonumber \\
&& \hspace{1cm} \left(\frac{1}{\sac\sbc^2}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{23}$}}
\end{picture}
}}
\hfill} - \frac{1}{\sac\sbc\sabc} {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill}\right)\; , \\
{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(1.5,0.2){\line(1,2){0.5}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} & = & -\frac{1}{d-4}
{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(1,0.2){\line(1,1){1}}
\put(1.5,0.7){\circle*{0.2}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill}\nonumber\\
&& + \frac{6(d-3)}{d-4} \frac{\sab+\sac}{\sac\sbc}
{
\mbox{\parbox{4cm}{\hspace{0.25cm}
\begin{picture}(3.5,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.8,0.2){\vector(1,0){0.1}}
\put(2.8,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2.5}}
\put(0.5,1.2){\line(1,0){2.5}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.7){\circle{1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(3.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(3.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber \\
&& - \frac{3(d-3)(3d-10)}{(d-4)^2} \frac{1}{\sac\sbc}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{23}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber \\
&& +\frac{3(d-3)(3d-10)}{(d-4)^2} \frac{1}{\sac\sbc}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{13}$}}
\end{picture}
}}
\hfill} \nonumber \\
&& +\frac{3(d-3)(3d-8)(3d-10)}{(d-4)^3} \frac{1}{\sac^2\sbc}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{13}$}}
\end{picture}
}}
\hfill}\; .\end{aligned}$$
One of the two remaining non-planar diagrams is also reducible, the other non-planar topology contains two master integrals. The reducible integral reads: $$\begin{aligned}
{
\mbox{\parbox{4.5cm}{\hspace{0.25cm}
\begin{picture}(4,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(1.5,0.2){\line(1,2){0.2}}
\put(2,1.2){\line(-1,-2){0.2}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(1,2){0.5}}
\put(1.5,1.2){\line(1,-2){0.5}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} & = &
\hspace{0.28cm}
\frac{3(d-4)}{2d-9}\, \frac{\sab}{\sac\sbc}{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(2,0.2){\line(-1,1){1}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill}
\nonumber \\
&&+\frac{3(d-4)}{2d-9}\, \frac{\sbc}{\sab\sac}{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(2,0.2){\line(-1,1){1}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_3$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_1$}}
\end{picture}
}}
\hfill}
\nonumber \\
&&+\frac{3(d-4)}{2d-9}\, \frac{\sac}{\sab\sbc}{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(2,0.2){\line(-1,1){1}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_3$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_2$}}
\end{picture}
}}
\hfill}
\nonumber \\
&&+ \frac{3(d-3)(3d-8)(3d-10)}{(d-4)^2(2d-9)}\,
\frac{\sac+\sbc}{\sab^2\sac\sbc}{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{12}$}}
\end{picture}
}}
\hfill} \nonumber \\
&&+ \frac{3(d-3)(3d-8)(3d-10)}{(d-4)^2(2d-9)}\,
\frac{\sab+\sbc}{\sab\sac^2\sbc}{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{13}$}}
\end{picture}
}}
\hfill} \nonumber \\
&&+ \frac{3(d-3)(3d-8)(3d-10)}{(d-4)^2(2d-9)}\,
\frac{\sab+\sac}{\sab\sac\sbc^2}{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{23}$}}
\end{picture}
}}
\hfill} \nonumber \\
&&- \frac{3(d-3)(3d-8)(3d-10)}{(d-4)^2(2d-9)}\,
\frac{1}{\sab\sac\sbc}{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill}\; .\end{aligned}$$
$t=7$
-----
At $t=7$ different one finds six different topologies. Three of them are triangle insertions to the one-loop box. These three integrals are all reducible, two of them contain only master integrals up to $t=5$ in their reduction: $$\begin{aligned}
\lefteqn{{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(1.5,1.2){\line(1,-1){0.5}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} =}\nonumber \\
&& -\frac{6(d-3)(3d-14)}{(d-4)(d-6)}\, \frac{1}{\sbc^2}\,
{
\mbox{\parbox{4cm}{\hspace{0.25cm}
\begin{picture}(3.5,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.8,0.2){\vector(1,0){0.1}}
\put(2.8,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2.5}}
\put(0.5,1.2){\line(1,0){2.5}}
\put(2.5,0.2){\line(0,1){1}}
\put(1.5,0.7){\circle{1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_2$}}
\put(3.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(3.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill}\nonumber \\
&& + \frac{3(3d-14)}{(d-6)} \, \frac{\sab^2}{\sac^2\sbc^2}\,
{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(2,0.2){\line(-1,1){1}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber \\
&& -\frac{6(d-3)(3d-14)}{(d-4)(d-6)}\, \frac{(\sab+\sac)^2}{\sac^2\sbc^2}\,
{
\mbox{\parbox{4cm}{\hspace{0.25cm}
\begin{picture}(3.5,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.8,0.2){\vector(1,0){0.1}}
\put(2.8,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2.5}}
\put(0.5,1.2){\line(1,0){2.5}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.7){\circle{1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(3.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(3.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber \\
&& - \frac{3(d-3)(3d-10)(3d-14)}{2(d-4)^2(d-5)(d-6)} \, \frac{(2d-10)\sab +
(3d-14) \sbc}{\sac\sbc^2(\sab+\sbc)} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(0.5,1.2){\oval(1,1)[br]}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{13}$}}
\end{picture}
}}
\hfill}
\nonumber \\
&& + \frac{3(d-3)(3d-10)(3d-14)}{2(d-4)^2(d-5)(d-6)} \, \frac{(2d-10)\sab +
(3d-14) \sac}{\sac^2\sbc^2} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{23}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber\\
&& - \frac{3(d-3)(3d-10)(3d-14)}{2(d-4)^2(d-5)(d-6)} \, \frac{(2d-10)
\sab\sabc +(3d-14) \sac\sbc }{\sac^2\sbc^2(\sab+\sbc)}
{
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_{13}$}}
\end{picture}
}}
\hfill} \nonumber \\
&& - \frac{3(d-3)(3d-8)(3d-10)(3d-14)}{2(d-4)^3(d-5)(d-6)} \,
\frac{(2d-10)\sab + (d-6)\sbc}{\sac^2\sbc^2(\sab+\sbc)}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{13}$}}
\end{picture}
}}
\hfill}\nonumber\\
&& + \frac{3(d-3)(3d-8)(3d-10)(3d-14)}{(d-4)^3(d-5)(d-6)}\,
\frac{(d-5)\sab - (d-4)\sac}{ \sac^2\sbc^3} \, {
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{23}$}}
\end{picture}
}}
\hfill}\nonumber\\
&& + \frac{3(d-3)(3d-8)(3d-10)(3d-14)}{2(d-4)^3(d-5)(d-6)}\,
\frac{(2d-10)\sab\sabc + (d-6)
\sac\sbc}{\sac^2\sbc^2\sabc(\sab+\sbc)}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{123}$}}
\end{picture}
}}
\hfill}\; ,\\
\lefteqn{{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(1.5,0.2){\line(1,1){0.5}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} =}\nonumber \\
&& \frac{2(2d-9)}{(d-4)(d-6)}\,\frac{2\sab+\sac+\sbc}{\sac\sbc}
{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(1,0.2){\line(1,1){1}}
\put(1.5,0.7){\circle*{0.2}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill}\nonumber \\
&& -\frac{3(d-4)}{d-6}\, \frac{(\sac+\sbc)^2}{\sac^2\sbc^2}
{
\mbox{\parbox{3.5cm}{\hspace{0.25cm}
\begin{picture}(3,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.3,0.2){\vector(1,0){0.1}}
\put(2.3,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2}}
\put(1,0.2){\line(1,1){1}}
\put(0.5,1.2){\line(1,0){2}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.2){\line(0,1){1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(2.55,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.55,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill}\nonumber \\
&& -\frac{6(d-3)(3d-14)}{(d-4)(d-6)}\, \frac{(\sab+\sbc)^2}{\sac^2\sbc^2}
{
\mbox{\parbox{4cm}{\hspace{0.25cm}
\begin{picture}(3.5,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.8,0.2){\vector(1,0){0.1}}
\put(2.8,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2.5}}
\put(0.5,1.2){\line(1,0){2.5}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.7){\circle{1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_2$}}
\put(3.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(3.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber\\
&& -\frac{6(d-3)(3d-14)}{(d-4)(d-6)}\, \frac{(\sab+\sac)^2}{\sac^2\sbc^2}
{
\mbox{\parbox{4cm}{\hspace{0.25cm}
\begin{picture}(3.5,1.4)
\thicklines
\put(0.7,0.2){\vector(-1,0){0.1}}
\put(2.8,0.2){\vector(1,0){0.1}}
\put(2.8,1.2){\vector(1,0){0.1}}
\put(0.8,1.2){\vector(1,0){0.1}}
\put(0.5,0.2){\line(1,0){2.5}}
\put(0.5,1.2){\line(1,0){2.5}}
\put(1,0.2){\line(0,1){1}}
\put(2,0.7){\circle{1}}
\put(0.45,1.2){\makebox(0,0)[r]{$p_{123}$}}
\put(0.45,0.2){\makebox(0,0)[r]{$p_1$}}
\put(3.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(3.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill} \nonumber\\
&& +\frac{3(d-3)(3d-10)(3d-14)}{2(d-4)^2(d-5)(d-6)}\,
\frac{(2d-10)\sab+(3d-14)\sbc}{\sac^2\sbc^2} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{13}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_1$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill}
\nonumber \\
&& +\frac{3(d-3)(3d-10)(3d-14)}{2(d-4)^2(d-5)(d-6)}\,
\frac{(2d-10)\sab+(3d-14)\sac}{\sac^2\sbc^2} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{23}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_2$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_3$}}
\end{picture}
}}
\hfill}
\nonumber \\
&& -\frac{3(d-3)(3d-10)}{(d-4)^2(d-6)}\, \frac{(3d-14)\sab + (4d-18)\sac
-(d-4)\sbc }{\sac^2\sbc^2} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{13}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_2$}}
\end{picture}
}}
\hfill}
\nonumber \\
&& -\frac{3(d-3)(3d-10)}{(d-4)^2(d-6)}\, \frac{(3d-14)\sab -(d-4)\sac
+ (4d-18)\sbc }{\sac^2\sbc^2} {
\mbox{\parbox{3cm}{\hspace{0.25cm}
\begin{picture}(2.5,1.4)
\thicklines
\put(0.3,0.7){\vector(1,0){0.1}}
\put(1.7,0.2){\vector(1,0){0.1}}
\put(1.7,1.2){\vector(1,0){0.1}}
\put(0,0.7){\line(1,0){0.5}}
\put(1,1.2){\line(0,-1){1}}
\put(1,1.2){\line(1,0){1}}
\put(1,0.2){\line(1,0){1}}
\put(1,0.7){\circle{1}}
\put(0.25,0.9){\makebox(0,0)[b]{$p_{123}$}}
\put(2.05,1.2){\makebox(0,0)[l]{$p_{23}$}}
\put(2.05,0.2){\makebox(0,0)[l]{$p_1$}}
\end{picture}
}}
\hfill}
\nonumber \\
&& -\frac{3(d-3)(3d-8)(3d-10)}{(d-4)^3(d-5)(d-6)}
\frac{(d-5)(3d-14)(\sab+\sac)+(d-4)^2\sbc}{\sac^3\sbc^2}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{13}$}}
\end{picture}
}}
\hfill}\nonumber \\
&& -\frac{3(d-3)(3d-8)(3d-10)}{(d-4)^3(d-5)(d-6)}
\frac{(d-5)(3d-14)(\sab+\sbc)+(d-4)^2\sac}{\sac^2\sbc^3}
{
\mbox{\parbox{2.5cm}{\hspace{0.25cm}
\begin{picture}(2,1)
\thicklines
\put(0.3,0.5){\vector(1,0){0.1}}
\put(0,0.5){\line(1,0){2}}
\put(1,0.5){\circle{1}}
\put(0.25,0.7){\makebox(0,0)[b]{$p_{23}$}}
\end{picture}
}}
\hfill}\;. \end{aligned}$$ The remaining three topologies are the double box and two different momentum arrangements of the crossed box. These topologies contain each two master integrals.
Conclusions and Outlook {#sec:conc}
=======================
Progress in the computation of exclusive observables, such as for example jet production rates, beyond the next-to-leading order has up to now been hampered mainly by difficulties in the calculation of virtual two-loop integrals with more than two external legs. In contrast to this, many inclusive observables (which correspond from the calculational point of view to two-point functions) are known to next-to-next-to-leading order and even beyond. These higher order calculations relied on a variety of elaborate technical tools for the computation of the virtual integrals. In this paper, we outline how techniques known from multi-loop calculations of two-point integrals can be modified and extended towards the computation of integrals with a larger number of external legs.
We demonstrate that the large number of different two-loop integrals appearing in an actual calculation can be reduced to a small number of scalar master integrals by using the well-known integration-by-parts identities [@hv; @chet] together with identities following from Lorentz-invariance which are unique to multi-leg integrals. As a by-product of this reduction, one is also able to reduce two-loop integrals with tensorial structure to scalar integrals. In contrast to two-point integrals, where only a few topologically different graphs can appear with potentially large powers of propagators and scalar products, one finds that the reduction of three- and four-point integrals gives rise to a large number of topologically different graphs, which appear however only with small powers of propagators and scalar products. The reduction of two-point functions usually proceeds via solving manually the integration-by-parts identities for arbitrary powers of propagators and denominators in a given graph topology; this procedure seems to be not practicable for multi-leg integrals. To accomplish the reduction of these, we developed an algebraic program which automatically derives and solves the integration-by-parts and Lorentz-invariance identities for a given graph up to some pre-selected fixed number of powers in denominators and scalar products independent of the topology.
To compute the scalar master integrals, we derive differential equations in the external momenta [@remiddi] for them; the boundary conditions of these differential equations correspond to simpler integrals, where for example one of the external momenta vanishes. These differential equations can be solved (for arbitrary space-time dimensions) by employing standard mathematical methods. We observe that the differential equations for the master integrals we considered up to now are solved by generalised hypergeometric functions. We illustrate the application of this method in detail on the example of the one-loop four-point function with one off-shell leg.
Using the differential equation method, we provide a complete list of all master integrals with up to $t=5$ different denominators that can appear in the reduction of two-loop four-point functions with one off-shell leg. We also list all reducible integrals with $t=6$ and $t=7$ different propagators. The computation of the master integrals with $t=6$ and $t=7$ is still an outstanding task.
The differential equations for these outstanding master integrals are of similar structure as the differential equations for master integrals with a smaller number of different propagators. The main problem towards a complete computation of these integrals is at present the integration of the inhomogeneous term, containing itself already hypergeometric functions arising from the subtopologies.
It is worthwhile to point out similarities and differences between the differential equation method employed in this paper and other methods employed for similar calculations in the literature. Both the negative dimension approach of [@glover] and the Mellin-Barnes transformation method employed in [@smirnov; @tausk] rely on choosing a particular assignment of momentum vectors to the internal loop propagators. After this assignment, a representation of the propagators in terms of a multiple sum (negative dimension approach) or an integral transformation (Mellin-Barnes method) is employed, such that the integral over the loop momentum can be carried out explicitly. The final result for the integral is then retrieved by resummation of a multiple sum or by an inverse integral transformation. Both methods, when employed for arbitrary space-time dimension, give rise to generalised hypergeometric functions, which can be represented as multiple sums as well as inverse Mellin-Barnes integrals [@bateman]. In the differential equation method, one assigns momentum vectors to the loop propagators only for the sake of deriving the differential equations and the IBP and LI identities. After applying these identities to simplify the differential equations, one obtains a relation between the derivative (with respect to an external momentum) of a master integral, the master integral itself and other master integrals with simpler topology, independent of the parametrisation chosen for the internal propagators. These differential equations can then be solved analytically by integration; the resulting integrals correspond to the integral representations of generalised hypergeometric functions [@bateman]. Using the differential equation method, one can therefore circumvent the explicit loop momentum integration needed in the other methods and one arrives at a representation of the hypergeometric functions, which is presumably more transparent than a multiple sum or an inverse integral transformation. In the integral representation, it is in particular straightforward to identify linear combinations of different hypergeometric functions, which are difficult to disentangle in the other representations. At present, it should however not be claimed that any of the methods is superior, since none of them could yet be employed to compute all outstanding two-loop four-point master integrals.
As a final point, we note that the methods derived in this paper contain a high level of redundancy, which allows for a number of non-trivial checks on the results obtained with them. The automatic reduction to master integrals using integration-by-parts and Lorentz-invariance identities corresponds to the solution of a linear system of equations containing more identities than unknowns. The existence of a solution to this system provides therefore already a check on the self-consistency of the identities. In computing the master integrals from differential equations, one integrates one of the three differential equations in the external invariants, such that the result can be checked by inserting it in the remaining two differential equations.
In short, this paper demonstrates how techniques developed for multi-loop calculation of two-point functions can be extended towards integrals with a larger number of external legs. As a first example of the application of these tools in practice, we computed some up to now unknown two-loop four-point functions, relevant for jet calculus beyond the next-to-leading order. The most important potential application of these tools is the yet outstanding derivation of two-loop virtual corrections to exclusive quantities, such as jet observables.
Acknowledgements {#acknowledgements .unnumbered}
================
We are grateful to Jos Vermaseren for his assistance in the use of the algebraic program FORM. One of the authors (E.R.) wants to thank the Alexander-von-Humboldt Stiftung for supporting his stay at the Institut für Theoretische Teilchenphysik of the University of Karlsruhe. The research work presented in this paper was supported in part by the DFG (Forschergruppe “Quantenfeldtheorie, Computeralgebra und Monte-Carlo Simulation”, contract KU 502/8-2).
Special Functions
=================
This appendix summarises the series and integral representations of the hypergeometric functions appearing in the master integrals. The properties of these functions, in particular their region of analyticity, their analytic continuation as well as reduction formulae, can be found in the literature [@glover; @bateman; @grad; @exton].
Hypergeometric functions are sums with coefficients formed from Pochhammer symbols $$\left( a\right)_n \equiv \frac{\Gamma(a+n)}{\Gamma(a)}.$$
The hypergeometric functions of a single variable are given by: $$\begin{aligned}
\;_2F_1 \left(a,b;c;z\right) & = & \sum_{n=0}^{\infty}
\frac{(a)_n(b)_n}{(c)_n} \frac{z^n}{n!}\; ,\\
\;_3F_2 \left(a,b_1,b_2;c_1,c_2;z\right) & = & \sum_{n=0}^{\infty}
\frac{(a)_n(b_1)_n(b_2)_n}
{(c_1)_n(c_2)_n} \frac{z^n}{n!}\;.\end{aligned}$$
Two types of hypergeometric functions of two variables also appear in our results: $$\begin{aligned}
F_1\left(a,b_1,b_2;c;z_1,z_2\right) &=& \sum_{m,n=0}^{\infty}
\frac{(a)_{m+n}(b_1)_m(b_2)_n}
{(c)_{m+n}} \frac{z_1^m}{m!} \frac{z_2^n}{n!}\; ,\\
S_1\left(a_1,a_2,b;c,d;z_1,z_2\right) &=& \sum_{m,n=0}^{\infty}
\frac{(a_1)_{m+n}(a_2)_{m+n}(b)_m}
{(c)_{m+n}(d)_m} \frac{z_1^m}{m!} \frac{z_2^n}{n!}\; .\end{aligned}$$
These functions have the following integral representations: $$\begin{aligned}
\;_2F_1 \left(a,b;c;z\right) & = & \frac{\Gamma(c)}{\Gamma(b)\Gamma(c-b)}
\int_0^1 \d t\; t^{b-1} (1-t)^{c-b-1} (1-tz)^{-a} \nonumber \\
&& \hspace{2cm} \mbox{Re}(b) > 0, \qquad \mbox{Re}(c-b)>0 \\
\;_3F_2 \left(a,b_1,b_2;c_1,c_2;z\right) & = &
\frac{\Gamma(c_1)\Gamma(c_2)}{\Gamma(b_1)\Gamma(c_1-b_1)\Gamma(b_2)
\Gamma(c_2-b_2)}\nonumber \\
&& \int_0^1 \d t_1 \int_0^1 \d t_2\;
t_1^{b_1-1} t_2^{b_2-1} (1-t_1)^{c_1-b_1-1}(1-t_2)^{c_2-b_2-1}
(1-t_1t_2z)^{-a}\nonumber \\
&& \hspace{-0.7cm}
\mbox{Re}(b_1) > 0, \qquad \mbox{Re}(c_1-b_1)>0, \qquad
\mbox{Re}(b_2) > 0, \qquad \mbox{Re}(c_2-b_2)>0 \\
F_1\left(a,b_1,b_2;c;z_1,z_2\right)&=&\frac{\Gamma(c)}{\Gamma(a)\Gamma(c-a)}
\int_0^1 \d t\; t^{a-1}(1-t)^{c-a-1} (1-tz_1)^{-b_1}(1-tz_2)^{-b_2}
\nonumber \\
&& \hspace{2cm} \mbox{Re}(a) > 0, \qquad \mbox{Re}(c-a)>0 \\
S_1\left(a_1,a_2,b;c,d;z_1,z_2\right) &=&
\frac{\Gamma(c)\Gamma(d)}{\Gamma(a_1)\Gamma(c-a_1)\Gamma(b)\Gamma(d-b)}
\nonumber \\
&& \int_0^1\d t_1\int_0^1\d t_2 t_1^{a_1-1} t_2^{b-1} (1-t_1)^{c-a_1-1}
(1-t_2)^{d-b-1} \left(1-t_1t_2z_1-t_1z_2\right)^{-a_2} \nonumber\\
&&\mbox{Re}(a_1)>0, \quad \mbox{Re}(c-a_1)>0,\quad
\mbox{Re}(b)>0, \quad \mbox{Re}(d-b)>0\end{aligned}$$
Expansion of Hypergeometric Functions
=====================================
To separate divergent and finite parts of the loop integrals derived in this paper, one has to expand them around the physical number of space-time dimensions in the parameter $\epsilon=(4-d)/2$. We demonstrate in this appendix, that this expansion can, at least for the hypergeometric functions in one variable, be carried out in a mechanical way, giving rise to harmonic polylogarithms (HPL), a generalisation of Nielsen’s polylogarithms [@nielsen] introduced in [@hpl].
Expanding the integral representation of $\,_2F_1$ in $\epsilon$ yields simple powers of $(t,1-t,1-tz)$ times the product of some number of $(\ln t, \ln (1-t), \ln(1-tz))$. The powers of $(t,1-t,1-tz)$ can be integrated by parts until one obtains non-trivial integrals $$\int_0^1 \d t \left(\frac{1}{t},\frac{1}{1-t},\frac{1}{1-tz}\right)
\ln^{n_1}t \ln^{n_2}(1-t) \ln^{n_3}(1-tz)\;.$$ All these integrals are combinations of harmonic polylogarithms $H(\vec{a};z)$, where $\vec{a}$ is a vector of indices with $w=n_1+n_2+n_3+1$ components. $w$ is called the weight of the harmonic polylogartihm. The proof by induction in $w$ is trivial, once the HPL formalism [@hpl] is recalled:
1. Definition of the three HPLs at $w=1$: $$\begin{aligned}
H(1;z) & \equiv & -\ln (1-z)\; ,\nonumber \\
H(0;z) & \equiv & \ln z\; ,\nonumber \\
H(-1;z) & \equiv & \ln (1+z)
\label{eq:levelone}\end{aligned}$$ and the three fractions $$\begin{aligned}
f(1;z) & \equiv & \frac{1}{1-z} \;, \nonumber \\
f(0;z) & \equiv & \frac{1}{z} \;, \nonumber \\
f(-1;z) & \equiv & \frac{1}{1+z} \; ,\end{aligned}$$ such that $$\frac{\partial}{\partial z} H(a;z) = f(a;z)\qquad \mbox{with}\quad
a=+1,0,-1\;.$$
2. For $w>1$: $$\begin{aligned}
H(0,\ldots,0;z) & \equiv & \frac{1}{w!} \ln^w z\; ,\\
H(a,\vec{b};z) & \equiv & \int_0^z \d x f(a;x) H(\vec{b};x)\; , \end{aligned}$$ which results in $$\frac{\partial}{\partial z} H(a,\vec{b};z) = f(a;z) H(\vec{b};z)\;.$$ This last relation is a convenient tool for verifying identities among different HPLs. Such identities can be verified by first checking a special point (typically $z=0$) and subsequently checking the derivatives. If agreement in the derivatives is not obvious, this procedure can be repeated until one arrives at relations involving only HPLs with $w=1$.
3. The HPLs fulfil an algebra (see Section 3 of [@hpl]), such that a product of two HPLs (with weights $w_1$ and $w_2$) of the same argument $z$ is a combination of HPLs of argument $z$ with weight $w=w_1+w_2$.
Using these properties of the HPL, one can show that the integrals appearing in the $\epsilon$-expansion of the hypergeometric function can be reexpressed as $$\int_0^1 \d t \left(\frac{1}{t},\frac{1}{1-t},\frac{1}{1-tz}\right)
\ln^{n_1}t \ln^{n_2}(1-t) \ln^{n_3}(1-tz) \to
\int_0^1 \d t \left(\frac{1}{t},\frac{1}{1-t},\frac{1}{t-1/z}\right)
H(\vec{a},t) H( \vec{b},zt)\;.
\label{eq:NTint}$$ Following the argumentation of Section 7 of [@hpl], one can show that the integral on the right hand side of the above equation yields a linear combination of HPLs of weight $w=w_a+w_b+1$. The proof goes via induction in $w_b$.
For $w_a=w_b=0$ one has $ H(\vec{a};t) H( \vec{b};zt)=1$. The $t$-integral in (\[eq:NTint\]) yields then a combination of HPL of weight $w=1$ (\[eq:levelone\]). Likewise, for $w_b=0$ the right hand side of (\[eq:NTint\]) will yield a linear combination of HPLs of weight $w=w_a+1$ and of argument $z$, as proven in Section 7 of [@hpl].
Considering $$\frac{\partial}{\partial z}
\int_0^1 \d t \left(\frac{1}{t},\frac{1}{1-t},\frac{1}{t-1/z}\right)
H(\vec{a};t) H(B, \vec{b};zt)\;,
\label{eq:induct}$$ we observe that $$\begin{aligned}
\frac{\partial}{\partial z} \frac{\d t}{t-1/z} &=& \frac{1}{z^2}
\, \frac{1}{t-1/z}\, \d t\,
\frac{\partial}{\partial t} + \mbox{boundary terms} \;,
\nonumber \\
\frac{\partial}{\partial z} H(B,\vec{b};zt) &=& t\,f(B;zt)\,H(\vec{b};zt)
\nonumber \;.\end{aligned}$$ Making these replacements in (\[eq:induct\]) and applying partial fractioning to all denominators, we are left with $$\left(\frac{1}{z},\frac{1}{1-z},\frac{1}{1+z}\right)
\int_0^1 \d t \left(\frac{1}{t},\frac{1}{1-t},\frac{1}{t-1/z}\right)
H(\vec{a};t) H(\vec{b};zt)\;,$$ which is a combination of HPLs with argument $z$ and weight $w=w_a+w_b+1$ multiplied with $(1/z,1/(1-z),1/(1+z))$. Integrating (\[eq:induct\]) over $z$ will thus yield a combination of HPLs with argument $z$ and weight $w+1$, which completes the proof by induction.
The $\epsilon$-expansion of $\,_3F_2$, corresponding to a double integral in $t_1$ and $t_2$, is obtained by carrying out the procedure described here twice, again resulting in a combination of HPLs. A systematic $\epsilon$-expansion of $F_1$ and $S_1$, which are functions of two variables $z_1$ and $z_2$, will in general go beyond the harmonic polylogarithms in one variable.
[99]{}
C.G. Bollini and J.J. Giambiagi, Nuovo Cim. [**12B**]{} (1972) 20;\
G.M. Cicuta and E. Montaldi, Nuovo Cim. Lett. [**4**]{} (1972) 329.
G. ’t Hooft and M. Veltman, Nucl. Phys. [**B44**]{} (1972) 189.
F.V. Tkachov, Phys. Lett. [**100B**]{} (1981) 65;\
K.G. Chetyrkin and F.V. Tkachov, Nucl. Phys. [**B192**]{} (1981) 159. K.G. Chetyrkin, J.H. Kühn and A. Kwiatkowski, Phys. Rept. [**277**]{} (1996) 189. S. Laporta and E. Remiddi, Phys. Lett. [**B379**]{} (1996) 283 (hep-ph/9602417). J.A.M. Vermaseren, [*Symbolic Manipulation with FORM*]{}, Version 2, CAN, Amsterdam, 1991. , Copyright 1981-1994 by Waterloo Maple Software and the University of Waterloo. V.A. Smirnov, Phys. Lett. [**B460**]{} (1999) 397 (hep-ph/9905323). J.B. Tausk, Phys. Lett. [**B469**]{} (1999) 225 (hep-ph/9909506). C. Anastasiou, E.W.N. Glover and C. Oleari, Nucl. Phys. [**B565**]{} (2000) 445 (hep-ph/9907523); Nucl. Phys. [**B572**]{} (2000) 307 (hep-ph/9907494). V.A. Smirnov and O.L. Veretin, Nucl. Phys. [**B566**]{} (2000) 469 (hep-ph/9907385). A.V. Kotikov, Phys. Lett. [**B254**]{} (1991) 158. E. Remiddi, Nuovo Cim. [**110A**]{} (1997) 1435 (hep-th/9711188);\
M. Caffo, H. Czyz, S. Laporta and E. Remiddi, Proceedings of Zeuthen Workshop on [*Loops and Legs in Gauge Theories*]{}, Rheinsberg, 1998, Acta Phys. Polonica [**29**]{} (1998) 2627 (hep-th/9807119). M. Caffo, H. Czyz and E. Remiddi, Nuovo Cim. [**111A**]{} (1998) 365 (hep-th/9805118). G. Kramer and B. Lampe, J. Math. Phys. [**28**]{} (1987) 945. E. Kamke, [*Differentialgleichungen: Lösungsmethoden und Lösungen*]{}, Vol. 1 & 2, B.G. Teubner (Stuttgart, 1977). A. Erdelyi, W. Magnus, F. Oberhettinger and F. Tricomi (Bateman Manuscript Project) , [*Higher Transcendental Functions*]{} , Vol. I, McGraw Hill (New York, 1953). I.S. Gradshteyn and I.M. Ryzhik, [*Table of Integrals, Series and Products*]{}, Academic Press (Boston, 1980). H. Exton, [*Multiple Hypergeometric Functions*]{}, Ellis Horwood (Chichester, 1976). I.N. Bronstein and K.A. Semendjajew, [*Taschenbuch der Mathematik*]{}, BSB B.G. Teubner Verlagsgesellschaft (Leipzig, 1979). R.K. Ellis, D.A. Ross and A.E. Terrano, Nucl. Phys. [**B178**]{} (1981) 421. S. Catani, Phys. Lett. [**B427**]{} (1998) 161. N. Nielsen, Nova Acta Leopoldiana (Halle) [**90**]{} (1909) 123;\
L. Lewin, [*Polylogarithms and Associated Functions*]{}, North Holland (Amsterdam, 1981). E. Remiddi and J.A.M. Vermaseren, Int. J. Mod. Phys. [**A15**]{} (2000) 725 (hep-ph/9905237).
[^1]: Supported by Alexander-von-Humboldt Stiftung, permanent address: Dipartimento di Fisica, Università di Bologna, I-40126 Bologna, Italy
[^2]: The importance of this fact was first pointed out by S. Laporta and exploited in [@laporta].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Counterfactual reasoning and contextuality is defined and critically evaluated with regard to its nonempirical content. To this end, a uniqueness property of states, explosion views and link observables are introduced. If only a single context associated with a particular maximum set of observables can be operationalized, then a context translation principle resolves measurements of different contexts.'
author:
- Karl Svozil
title: On Counterfactuals and Contextuality
---
[address=[Institute of Theoretical Physics, Vienna University of Technology, Wiedner Hauptstraße 8-10/136, A-1040 Vienna, Austria]{},email=[[email protected]]{}]{}
Counterfactuals
===============
With the rise of quantum mechanics [@schrodinger; @reich-44; @jammer:66; @jammer1] physics proper entered an ancient and sometimes fierce debate in theology and philosophy: the controversy between realism versus idealism. Whereas realism has been subsumed by the proposition that [@stace] [*“some entities sometimes exist without being experienced by any finite mind,”*]{} idealism put forward that [*“we have not the faintest reason for believing in the existence of unexperienced entities. \[\[Realism\]\] has been adopted $\ldots$ solely because it simplifies our view of the universe.”*]{} And whereas these issues can be considered nonoperational and thus metaphysical or even ideological, it is also true that they have inspired a great number of minds, to the effect of stimulating new approaches to quantum mechanics, revealing many theoretical details, quantum phenomena and quantum technologies.
The Kochen-Specker theorem [@kochen1], for example, was motivated from the onset by scholasticism, as in an early programmatic article [@specker-60] Ernst Specker related the discussion on the foundations of quantum mechanics to scholastic speculations about the existence of [*infuturabilities.*]{} The scholastic issue was whether or not the omniscience (comprehensive knowledge) of God extends to what nowadays are called [*counterfactuals.*]{} And if so, can all events be pasted together to form a consistent whole?
Informally, counterfactuals will be defined as follows. By counterfactual events we mean [*events which would have occurred if something had happened which did not happen.*]{} The associated counterfactual proposition is henceforth regarded “true” if it states the occurrence of an event which would have occurred if something had happened which did not happen.
Classically, at least in principle, it makes no difference whether or not a particular observable is measured. It is assumed to possess a definite value, irrespective of any measurement. Thus classical counterfactuals do not present a conceptual challenge.
Quantum mechanically, the situation appears to be very different, and the use of counterfactuals for quantized systems is problematic and unresolved. Let us briefly mention two novel, nonrealistic features of quantum mechanics which challenge the sort of realism suggested by classical physics. Note, however, that also in classical times a certain uneasiness with the prevailing realistic perception remained; e.g., in Hertz’s perception of the formalisms of classical mechanics Ref. [@hertz-94].
[*Complementarity*]{} and the [*uncertainty principle*]{} limit the precision of co-measurements of certain entities. However, nondistributive propositional structures characteristic for complementarity not necessarily imply total abandonment of nonclassicality; e.g., in automaton logic [@svozil-ql Chapter 10] or generalized urn models [@wright; @svozil-2001-eua].
[*Value indefiniteness*]{} manifests itself through the scarcity of two-valued, nondispersive states or probability measures for particular logics or propositional structures occurring in quantum mechanics. There are “not enough” two-valued states to allow a faithful embedding into a Boolean algebra. Such two-valued states can be logically interpreted as truth assignments. Quantum logics [@pulmannova-91; @svozil-ql] has developed and characterised classes of “scarcities,” ranging from nonunital to nonseparable set of two-valued states to the nonexistence of two-valued states. The Kochen-Specker theorem [@kochen1] (see also [@specker-60; @kamber64; @kamber65; @ZirlSchl-65; @bell-66; @Alda; @Alda2; @peres; @mermin-93; @svozil-tkadlec; @tkadlec-00]) is a finitistic argument against the existence of any consistent global truth assignment for quantized systems associated with Hilbert spaces of dimension higher that two.
The question remains whether the set of probability measures increases or decreases as the set of two-valued measures increases or decreases. In classical physics and even for nondistributive logics with a separating set of two-valued states, the answer appears to be straightforward: By the Minkowsky-Weyl representation theorem [@ziegler p.29], the set of probability measures is the convex hull of the set of two-valued states representable by vertices of the associated polytope. Due to convexity, the set of probability measures can only increase as the number of vertices increases. But more general, e.g., quantum, probabilities are based on different assumptions.
Uniqueness property of states
=============================
In what follows the possibility of testing certain assumptions related to counterfactuals, in particular contextuality, will be reviewed. For the sake of the argument, certain properties of states will be defined, and the operationalization of explosion views of theoretical arguments will be discussed.
Definition
----------
A multiquantum state will be said to satisfy the [*uniqueness property*]{} if knowledge of a property of one quantum entails the certainty that, if this property were measured on the other quantum (or quanta) as well, the outcome of the measurement would be a unique function of the outcome of the measurement performed. This uniqueness property could be experimentally tested by performing the associated experiments for co-measurable observables of different quanta.
Einstein, Podolsky and Rosen [@epr] have proposed a way of counterfactual interference of particle properties by using certain entangled two-quantum states with the uniqueness property. Suppose one is willing to accept counterfactuals. Then one may pretend to obtain knowledge of noncommeasurable observables referring to a single quantum by measurement of one observable per quantum in a multiquantum state satisfying the uniqueness property at a time, and by subsequent counterfactual inference. This assumption (to be falsified) is also used in the Kochen-Specker [*reductio ad absurdum*]{} proof.
This method could in principle be generalized to $n$ quanta, such that the measurement of the state of one quantum in an $n$-quantum system fixes the states of all the others as well.
Example states
--------------
Apparently, the uniqueness property can be satisfied for multipartite states for which the number of terms contributing to the coherent sum of the joint amplitudes does not exceed the dimension of the single particle Hilbert space. Indeed, for nontrivial configurations, the number of terms should be [*identical*]{} to the single particle Hilbert space dimension. For if the number of terms exceeds the dimension, then at least for one quantum there are more than one possibilities of counterfactual existence. Consider, for example, three spin one quanta. Their only singlet state is (see also [@kok-02]) $$\vert \Psi_3 \rangle
= {1\over \sqrt{6}}(
\vert - + 0\rangle
-
\vert - 0 +\rangle
+
\vert + 0 - \rangle
-
\vert + - 0\rangle
+
\vert 0 - + \rangle
-
\vert 0 + - \rangle
).
\label{2004-qnc-e1}$$ Now, suppose the outcome of a spin measurement on the first quantum is “$-$;” serving as a filter and reducing $\vert \Psi_3 \rangle$ to the first two terms $$({1/ \sqrt{2}})(
\vert - + 0\rangle
-
\vert - 0 +\rangle
)$$ in the coherent sum of Eq. (\[2004-qnc-e1\]). Thereby, two possibilities “$0$” and “$+$” remain for the state of every one of the other quanta. This ambiguity in the counterfactual argument results in nonuniqueness.
With regards to uniqueness, the situation gets worse for singlet states of four or more spin one quanta; e.g., the three states $\vert \Psi_4^1\rangle , \vert \Psi_4^2\rangle , \vert \Psi_4^3\rangle$ of four spin one quanta [@schimpf-svozil] $$\begin{aligned}
\vert \Psi_4^1\rangle &=&
\frac{1}{\sqrt{5}}
\big[
\frac{2}{3}
\vert0,0,0,0\rangle
+
\vert -1,-1,1,1\rangle +\vert 1,1,-1,-1\rangle
\nonumber \\
&&\quad
- \frac{1}{2}
\big(
\vert-1,0,0,1\rangle
+\vert0,-1,0,1\rangle
+\vert-1,0,1,0\rangle
+\vert0,-1,1,0\rangle \nonumber \\
&&\quad
\qquad
+\vert0,1,-1,0\rangle
+\vert 1,0,-1,0\rangle
+\vert0,1,0,-1\rangle
+\vert1,0,0,-1\rangle
\big)\nonumber \\
&&\quad
+\frac{1}{3}
\big(
\vert0,0,-1,1\rangle
+\vert -1,1,0,0\rangle
+\vert1,-1,0,0\rangle
+\vert0,0,1,-1\rangle
\big)\nonumber \\
&&\quad
+
\frac{1}{6}
\big(
\vert-1,1,-1,1\rangle
+\vert1,-1,-1,1\rangle
+\vert-1,1,1,-1\rangle
+\vert1,-1,1,-1\rangle
\big)
\big] ,
\\
\vert \Psi_4^2\rangle &=&\frac{1}{2{\sqrt{3}}}
\big(
\vert -1,0,0,1\rangle
- \vert 0,-1,0,1\rangle
- \vert 0,1,0,-1\rangle
+\vert 1,0,0,-1\rangle \nonumber \\
&&\quad
- \vert -1,0,1,0\rangle
+\vert 0,-1,1,0\rangle
+\vert 0,1,-1,0\rangle
- \vert 1,0,-1,0\rangle \nonumber \\
&&\quad
+\vert -1,1,1,-1\rangle
- \vert 1,-1,1,-1\rangle
- \vert -1,1,-1,1\rangle
+\vert 1,-1,-1,1\rangle
\big) ,
\\
\vert \Psi_4^3\rangle &=& \frac{1}{3}
\big(
\vert 0,0,0,0\rangle
-
\vert 0,0,-1,1\rangle
-
\vert -1,1,0,0\rangle
-
\vert 1,-1,0,0\rangle
-
\vert 0,0,1,-1\rangle
\nonumber \\
&& \quad
+
\vert 1,-1,1,-1\rangle
+
\vert -1,1,-1,1\rangle
+
\vert 1,-1,-1,1\rangle
+
\vert -1,1,1,-1\rangle
\big) ,\end{aligned}$$ do not hold a uniqueness property, as can be readily verified with the same argument as above.
For singlet states of quanta of three dimensions, the uniqueness limit is reached for two particles $$\vert \Psi_2 \rangle
= ({1/ \sqrt{3}})(
\vert + -\rangle
+
\vert - +\rangle
-
\vert 0 0\rangle
).$$ For less terms the configuration may be effectively lower dimensional.
As already stated, for measuring entangled particles at different contexts, the uniqueness property must hold for [*every*]{} such context. This additional assumption is satisfied for singlet states, which are form invariant under identical unitary transformations of the single quantum Hilbert spaces. This is not true for the Greenberger, Horne and Zeilinger three spin one half quanta state in the form proposed by Mermin [@mermin1] $$\vert \Psi_{GHZM} \rangle
= ({1/ \sqrt{2}})(
\vert z+ z+ z+\rangle
+
\vert z- z- z-\rangle
).$$ Here $z\pm$ stands for the outcome $\pm$ of a spin measurement measured along the $z$-axis. A careful calculation (e.g., Eq. (3) of Ref. [@krenn1]) shows that $\vert \Psi_{GHZM}\rangle $ satisfies the uniqueness property only along a single direction, the $z$-axis, of spin state measurements; otherwise $\vert \Psi_{GHZM}\rangle $ contains eight summands.
Whether or not nontrivial states (in the sense mentioned above) exist which satisfy the uniqueness property in “sufficiently many” spin state measurement directions to make them useful for conterfactual reasoning remains an open question Alas, the lack of uniqueness may be the reason why inconsistencies such as the ones derived in the Kochen-Specker type proofs cannot be directly operationalized; such as in the “explosion type” nonclassical setups discussed below.
Explosion views
===============
If the quantum state also satisfies the uniqueness property when transformed to different, complementary measurements, then different, complementary, observables on other quanta could be measured, for which a similar uniqueness property holds. In that way, one may pretend to obtain knowledge of all these noncommeasurable observables referring to a single quantum by measurement of one context per quantum at a time, and by subsequent counterfactual inference. Of course, only one of these properties would actually be obtained by direct measurement on the quantum; all the other properties are merely counterfactually inferred. In principle, this method can be applied to an arbitrary number of contexts and quanta as long as the uniqueness property holds. This kind of setups will be referred to as of the [*“explosion view”*]{} type.
The advantage of explosion views is that they do not require different terms which refer to different detector settings and thus to measurements performed at different times. On the contrary, all standard Bell-Clauser-Horne-Shimony-Halt type or Greenberger-Horne-Zeilinger type measurements [@panbdwz] known today involve the summation or consideration of terms which are not co-measurable. The inevitable time delay of consecutive measurements of the relevant noncomeasurable terms makes these arguments vulnerable to a critique put forward by Hess and Philipp . Furthermore, explosion views offer the possibility to directly measure contextuality.
In this counterfactual sense, the measurement of observables associated with arbitrary operators becomes feasible, since formally, any matrix $A$ can be decomposed into two self-adjoint components $A_1, A_2$ as follows: $$\begin{array}{lllllll}
A&=&A_1+iA_2 \\
A_1&=&{1\over 2}(A+A^\dagger) =:\Re A,\; \;
A_2=-{i\over 2}(A-A^\dagger)=:\Im A.
\end{array} \label{e-decom1}$$ By assuming the uniqueness property, $A_1, A_2$ can be measured along two different entangled particles, respectively, and subsequently counterfactually “completed.”
Contextuality
=============
There exist various notions of contextuality. In what follows, the term contextuality will be used as envisioned by Bell and Redhead [@bell-66; @hey-red; @redhead]. Already Bohr [@bohr-1949] mentioned [*“the impossibility of any sharp separation between the behavior of atomic objects and the interaction with the measuring instruments which serve to define the conditions under which the phenomena appear.”*]{} Bell (Ref. [@bell-66], Sec. 5) stated that the [*“$\ldots$ result of an observation may reasonably depend not only on the state of the system $\ldots$ but also on the complete disposition of the apparatus.”*]{} That is, the outcome of the measurement of an observable $A$ might depend on which other observables from systems of maximal observables are measured alongside with $A$.
This concept was mainly introduced to maintain a certain amount of realism in view of the challenges of the theorems by Gleason [@Gleason] and Kochen and Specker; i.e., the “scarcity” of two valued states mentioned above. (Other attempts towards this goal have assumed nonconstructive measure theory utilizing paradoxical set decomposition [@pitowsky-82; @pitowsky-83], or abandoned the continuity of Hilbert space [@meyer:99; @havlicek-2000].) This section presents a critical evaluation of its empirical content.
Context and link observables
----------------------------
A [*context*]{} can formally be defined as a single (nondegenerate) “maximal” self-adjoint operator ${ C}$. It has a spectral decomposition into some complete set of orthogonal projectors ${ E}_i$ which correspond to propositions in the usual Von Neumann-Birkhoff type sense [@birkhoff-36; @v-neumann-49]. That is, ${ C}=\sum_{i=1}^n e_i { E}_i$ with mutually different $e_i$ and some orthonormal basis $\{{ E}_i{ H} \mid i=1,\ldots n\}$ of $n$-dimensional Hilbert space ${ H}$. In $n$ dimensions, contexts can be viewed as $n$-pods spanned by the $n$ orthogonal projectors ${ E}_1,
{ E}_2, \cdots,{ E}_n$.
In the finite subalgebras considered, an observable belonging to two or more contexts is called [*link observable*]{}.
Contexts can thus be depicted by Greechie (orthogonality) diagrams [@greechie:71], consisting of [*points*]{} which symbolize observables (representable by the spans of vectors in $n$-dimensional Hilbert space). Any $n$ points belonging to a context; i.e., to a maximal set of commeasurable observables (representable as some orthonormal basis of $n$-dimensional Hilbert space), are connected by [*smooth curves*]{}. Two smooth curves may be crossing in common [*link observables*]{}. In three dimensions, smooth curves and the associated points stand for tripods. Still another compact representation is in terms of Tkadlec diagrams, where points represent complete tripods and smooth curves represent single legs interconnecting them. In quantum logic [@pulmannova-91; @svozil-ql; @kalmbach-83], contexts are often referred to as [*subalgebras*]{} or [*blocks.*]{}
Experimental falsification of contextuality in simple configurations: Two contexts in three dimensions
------------------------------------------------------------------------------------------------------
In two dimensional Hilbert space, contextuality does not exist, since every context is fixed by the assumption of one property. The entire context is just this property, together with its negation, which corresponds to the orthogonal ray (which spans a one dimensional subspace) or projection associated with the ray corresponding to the property.
The simplest nontrivial configuration of contexts exist in three dimensional Hilbert space. Consider an arrangement of five observables $A,B,C,D,K$ with two systems of operators $\{A,B,C\}$ and $\{D,K,A\}$ called [*contexts*]{}, which are interconnected by $A$. With a context, the operators commute and the associated observables are commeasurable. For two different contexts, operators outside the link operators do not commute. $A$ will be called a [*link observable*]{}. This propositional structure can be represented in three dimensional Hilbert space by two tripods with a single common leg. Fig. \[2004-qnc-f1\] depicts this configuration in three dimensional real vector space, as well as in the associated Greechie and Tkadlec diagrams.
[ccccc]{} 0.70mm
(40.00,49.67) (15.00,45.00)[(0,-1)[30.00]{}]{} (15.00,15.00)[(-1,-1)[15.00]{}]{} (15.00,15.00)[(1,0)[25.00]{}]{} (15.00,15.00)[(3,-4)[11.00]{}]{} (15.00,15.00)[(5,3)[16.67]{}]{} (3.33,-1.67)[(0,0)\[cc\][$B$]{}]{} (30.00,0.00)[(0,0)\[cc\][$D$]{}]{} (40.00,11.33)[(0,0)\[cc\][$C$]{}]{} (35.00,23.67)[(0,0)\[cc\][$K$]{}]{} (19.33,49.67)[(0,0)\[cc\][$A$]{}]{} (20.00,6.67)[(2,1)[0.2]{}]{} (7.67,6.67)(14.33,2.67)(20.00,6.67) (30.00,23.00)[(-1,2)[0.2]{}]{} (29.67,16.00)(32.33,19.67)(30.00,23.00) (13.33,1.00)[(0,0)\[cc\][$\varphi$]{}]{} (40.00,18.67)[(0,0)\[rc\][$\varphi$]{}]{}
&& 0.80mm
(61.33,36.00) (0.33,35.00)(0.36,-0.12)[84]{}[(1,0)[0.36]{}]{} (30.33,25.00)(0.36,0.12)[84]{}[(1,0)[0.36]{}]{} (30.33,25.00) (45.33,30.00) (60.33,35.00) (0.33,35.00) (15.33,30.00) (60.33,31.00)[(0,0)\[cc\][$K$]{}]{} (45.33,26.00)[(0,0)\[cc\][$D$]{}]{} (30.33,30.00)[(0,0)\[cc\][$A$]{}]{} (15.33,26.00)[(0,0)\[cc\][$C$]{}]{} (0.33,31.00)[(0,0)\[cc\][$B$]{}]{} (0.00,20.00)(0.00,17.33)(3.00,17.33) (3.00,17.33)(10.00,17.00)(10.00,17.00) (10.00,17.00)(15.00,16.00)(15.00,13.33) (30.00,20.00)(30.00,17.33)(27.00,17.33) (27.00,17.33)(20.00,17.00)(20.00,17.00) (20.00,17.00)(15.00,16.00)(15.00,13.33) (15.00,5.33)[(0,0)\[cc\][$\{B,C,A\}$]{}]{} (60.00,20.00)(60.00,17.33)(57.00,17.33) (57.00,17.33)(50.00,17.00)(50.00,17.00) (50.00,17.00)(45.00,16.00)(45.00,13.33) (30.00,20.00)(30.00,17.33)(33.00,17.33) (33.00,17.33)(40.00,17.00)(40.00,17.00) (40.00,17.00)(45.00,16.00)(45.00,13.33) (45.00,5.33)[(0,0)\[cc\][$\{A,D,K\}$]{}]{}
&& 0.80mm
(51.37,10.00) (0.00,25.00) (30.00,25.00) (1.33,25.00)[(1,0)[27.33]{}]{} (15.00,30.00)[(0,0)\[cc\][$A$]{}]{} (30.00,15.00)[(0,0)\[cc\][$\{A,D,K\}$]{}]{} (0.00,15.00)[(0,0)\[cc\][$\{B,C,A\}$]{}]{}
\
a)&&b)&&c)\
The operators $B,C,A$ and $D,K,A$ can be identified with the projectors corresponding to the two bases $$\begin{array}{lcl}
B_{B-C-A}&=&
\{
(1,0,0)^T,
(0,1,0)^T,
(0,0,1)^T
\}
,
\\
B_{D-K-A}&=&
\{
(\cos \varphi , \sin \varphi ,0)^T,
(-\sin \varphi ,\cos \varphi , 0)^T,
(0,0,1)^T
\},
\end{array}
\label{e-vaxjo1}$$ (the superscript “$T$” indicates transposition). Their matrix representation is the dyadic product of every vector with itself.
Physically, the union of contexts $\{B,C,A\}$ and $\{D,K,A\}$ interlinked along $A$ does not have any direct operational meaning; only a single context can be measured along a single quantum at a time; the other being irretrievably lost if no reconstruction of the original state is possible. Thus, in a direct way, testing the value of observable $A$ against different contexts $\{B,C,A\}$ and $\{D,K,A\}$ is metaphysical.
It is, however, possible to counterfactually retrieve information about the two different contexts of a single quantum indirectly by considering a singlet state $
\vert \Psi_2 \rangle
= ({1/ \sqrt{3}})(
\vert + -\rangle
+
\vert - +\rangle
-
\vert 0 0\rangle
)$ via the “explosion view” Einsten-Podolsky-Rosen type of argument discussed above. Since the state is form invariant with respect to variations of the measurement angle and at the same time satisfies the uniqueness property, one may retrieve the first context $\{B,C,A\}$ from the first quantum and the second context $\{D,K,A\}$ from the second quantum. (This is a standard procedure in Bell type arguments with two spin one-half quanta.)
In this indirect, counterfactual sense, contextuality becomes measurable. From an experimental point of view, this amounts to performing two tasks.
- In the preparation stage, a singlet state of two spin one quanta must be realized. This has become feasible recently by engineering entangled states in any arbitrary dimensional Hilbert space [@mvwz-2001; @vwz-2002; @gisin-2002-d; @tdtm-2003] (see also generalized beam splitter setups [@rzbb; @reck-94; @zukowski-97; @svozil-2004-analog] for proofs of principle).
- In the analyzing stage, the context structure $\{B,C,A\}$ and $\{D,K,A\}$ interlinked at $A$ must be realized.
Let the matrix $[{ v}^T{ v}]$ stand for the dyadic product of the vector ${ v}$ with itself. The operators associated with the geometrical configuration enumerated in Eq. (\[e-vaxjo1\]) depicted in Fig. \[2004-qnc-f1\]a) are given by $$\begin{array}{lcl}
C_{B-C-A}&=& \sum_{i=1,2,3} e_i [B_{B-C-A,i}^T B_{B-C-A,i}]=
\left(
\begin{array}{ccc}
e_1&0&0\\
0&e_2&0\\
0&0&e_3\\
\end{array}
\right)
,
\\
C_{D-K-A}
&=& \sum_{i=1,2,3} e_i' [B_{D-K-A,i}^T B_{D-K-A,i}]=
\left(
\begin{array}{ccc}
e_1' \cos^2 \varphi + e_2'\sin^2 \varphi&(e_1'-e_2')\sin \varphi \cos \varphi &0\\
(e_1'-e_2')\sin \varphi \cos \varphi &e_2' \cos^2 \varphi + e_1'\sin^2 \varphi &0\\
0&0&e_3'\\
\end{array}
\right).
\end{array}
\label{e-vaxjo2}$$ As mentioned before, $e_1,e_2,e_3$ as well as $e_1',e_2',e_3'$ must be mutually distinct.
Contextuality would imply that the outcome of a measurement of $A$ depends on which other operators are measured alongside with it. If contextuality were taken seriously for the two context configuration discussed above, the measured value of $A$ would be different for the first and for the second quantum in the entangled two particle singlet state $\vert \Psi_3 \rangle$. Note that the context structure discussed here not necessarily implies that $A$ corresponds to the joint event associated with “$00$,” as the labelling applied to the state preparation has nothing to do with the logico-algebraic propositional structure realized by the contexts.
Despite the necessity to falsify contextuality for the interlinked two-context structure experimentally, there can hardly be any doubt about the outcome of the verdict against contextuality. This is due to the symmetry of the prepared singlet state, which is an expression of the conservation laws of quantities such as angular momentum. A conceivable option to save contextuality would be to assume that only in the nununique state cases contextuality unfolds. However, this would make contextuality a metaphysical property which cannot be measured at all and which has no physical meaning whatsoever.
Three contexts in three dimensions
----------------------------------
A next step further would be the logico-algebraic propositional structure with three interlinked contexts such as $\{A,B,C\}$, $\{A,D,K\}$ and $\{K,L,M\}$ interconnected at $A$ and $K$. This configuration is depicted in Fig. \[2004-vaxjo-f2\]a). Here, for the first time, the Einstein-Podolsky-Rosen “explosion view” type of setup encounters the problem of nonuniqueness: for the three quantum singlet state $
\vert \Psi_3 \rangle
$ enumerated in Eq. (\[2004-qnc-e1\]) the uniqueness property does not hold.
Note also that too tightly interlinked contexts are not realizable in Hilbert space. The interconnected “triangular” system of contexts $\{A,B,C\}$, $\{A,D,K\}$ and $\{K,L,C\}$ drawn in Fig. \[2004-vaxjo-f2\]b) has no representation as operators in Hilbert space. Likewise, no system of three tripods exist such that every tripods is interlinked with the other two tripods in two different legs.
[ccccc]{} 0.70mm
(91.34,36.00) (0.33,35.00)(0.36,-0.12)[84]{}[(1,0)[0.36]{}]{} (30.33,25.00) (45.33,25.00) (0.33,35.00) (15.33,30.00) (0.00,20.00)(0.00,17.33)(3.00,17.33) (3.00,17.33)(10.00,17.00)(10.00,17.00) (10.00,17.00)(15.00,16.00)(15.00,13.33) (30.00,20.00)(30.00,17.33)(27.00,17.33) (27.00,17.33)(20.00,17.00)(20.00,17.00) (20.00,17.00)(15.00,16.00)(15.00,13.33) (15.00,5.33)[(0,0)\[cc\][$\{B,C,A\}$]{}]{} (60.67,20.00)(60.67,17.33)(57.67,17.33) (57.00,17.33)(50.00,17.00)(50.00,17.00) (50.00,17.00)(45.00,16.00)(45.00,13.33) (30.00,20.00)(30.00,17.33)(33.00,17.33) (33.00,17.33)(40.00,17.00)(40.00,17.00) (40.00,17.00)(45.00,16.00)(45.00,13.33) (45.00,5.33)[(0,0)\[cc\][$\{A,D,K\}$]{}]{} (30.33,25.00)[(1,0)[30.00]{}]{} (90.34,35.00)(-0.36,-0.12)[84]{}[(-1,0)[0.36]{}]{} (60.34,25.00) (90.34,35.00) (75.34,30.00) (90.67,20.00)(90.67,17.33)(87.67,17.33) (87.67,17.33)(80.67,17.00)(80.67,17.00) (80.67,17.00)(75.67,16.00)(75.67,13.33) (60.67,20.00)(60.67,17.33)(63.67,17.33) (63.67,17.33)(70.67,17.00)(70.67,17.00) (70.67,17.00)(75.67,16.00)(75.67,13.33) (75.67,5.33)[(0,0)\[cc\][$\{K,L,M\}$]{}]{}
& & 0.80mm
(40.52,42.02) (0.00,9.98)[(1,0)[40.02]{}]{} (40.02,9.98)[(-4,5)[20.20]{}]{} (19.82,35.23)[(-4,-5)[20.20]{}]{} (0.00,4.99)[(0,0)\[cc\][$A$]{}]{} (19.96,4.99)[(0,0)\[cc\][$B$]{}]{} (40.02,4.99)[(0,0)\[cc\][$C$]{}]{} (35.03,24.95)[(0,0)\[cc\][$L$]{}]{} (19.96,42.02)[(0,0)\[cc\][$K$]{}]{} (4.99,24.95)[(0,0)\[cc\][$D$]{}]{} (19.96,9.98) (19.86,35.13) (40.02,9.98) (30.04,22.46) (9.68,22.46) (0.00,9.98)
\
a)& &b)
Explosion views of more elaborate contexts in three dimensions
--------------------------------------------------------------
The same nonuniqueness problems plaguing already the three quanta singlet state get worse for logico-algebraic configurations with a higher number of interlinked contexts. This makes impossible a “direct” counterfactual inference of the Einsten-Podolsky-Rosen type in the case of Kochen-Specker type proofs; i.e., for configurations supporting only nununital, nonseparating or even nonexisting sets of two-valued probability measures corresponding to classical truth assignments. All these nonclassical configurations lack essential classical properties. Note that, in general, the nonexistence of any two-valued state is a very strong property requiring “a lot more” contexts than more subtle nonclassical features such as nununitality or nonseparability [@svozil-ql].
For example, the logico-algebraic structure of observables depicted in Fig. \[2004-qnc-f2\] are representable by quanta in three dimensional Hilbert space. It has the nonclassical feature of a nonseparating set of two-valued probability measures: For all two valued probability measures $P(x)\in\{0,1\}$, $P(a)=P(b)=1$, there is no probability measure separating $a$ from $b$ through $P(a)\neq P(b)$.
That is, if $P(a=a_0=a_9') = 1$ for any two-valued probability measure $P$, then $P(a_8) =0$. Furthermore, $P(a_7)=0$, since by a similar argument $P(a)=1$ implies $P(a_7)=0$. Therefore, $P(b=a_9=a_0') = 1$. Symmetry requires that the reverse implication is also fulfilled, and therefore $P(b) =
P(a)$ for every two-valued probability measure $P$.
An explosion requires 16 contexts and could in principle be realized with some singlet state of 16 spin-one quanta; a state which contains by far too many terms to satisfy the uniqueness property. Here, nonuniquess seems to serve as a kind of [*protection principle*]{} in the case of nonclassical features of two-valued probability measures, the extreme case being the nonexistence of any such measure (cf. Graph $\Gamma_2$ of Ref. [@kochen1]).
0.50mm
(190.67,109.67) (165.67,19.67)(-0.12,0.12)[167]{}[(-1,0)[0.12]{}]{} (145.67,39.67)[(0,1)[40.00]{}]{} (145.67,79.67)(0.12,0.12)[167]{}[(0,1)[0.12]{}]{} (165.67,99.67)(0.12,-0.12)[167]{}[(0,-1)[0.12]{}]{} (185.67,79.67)[(0,-1)[40.00]{}]{} (185.67,39.67)(-0.12,-0.12)[167]{}[(-1,0)[0.12]{}]{} (185.34,59.67)[(-1,0)[39.67]{}]{} (185.67,39.67) (185.67,59.67) (185.67,79.67) (145.67,39.67) (145.67,59.67) (145.67,79.67) (165.67,19.67) (165.67,99.67) (165.67,19.67)(-0.21,0.12)[334]{}[(-1,0)[0.21]{}]{} (95.67,59.67)(0.21,0.12)[334]{}[(1,0)[0.21]{}]{} (95.67,59.67) (95.67,74.67)[(0,0)\[cc\][$a_8=a_8'$]{}]{} (165.67,109.67)[(0,0)\[cc\][$b=a_9=a_0'$]{}]{} (140.67,79.67)[(0,0)\[cc\][$a_2'$]{}]{} (140.67,59.67)[(0,0)\[cc\][$a_6'$]{}]{} (140.67,39.67)[(0,0)\[cc\][$a_4'$]{}]{} (190.67,39.67)[(0,0)\[cc\][$a_3'$]{}]{} (190.67,59.67)[(0,0)\[cc\][$a_5'$]{}]{} (190.67,80.00)[(0,0)\[cc\][$a_1'$]{}]{} (165.67,9.67)[(0,0)\[cc\][$a_7'$]{}]{} (25.00,19.67)(0.12,0.12)[167]{}[(1,0)[0.12]{}]{} (45.00,39.67)[(0,1)[40.00]{}]{} (45.00,79.67)(-0.12,0.12)[167]{}[(0,1)[0.12]{}]{} (25.00,99.67)(-0.12,-0.12)[167]{}[(0,-1)[0.12]{}]{} (5.00,79.67)[(0,-1)[40.00]{}]{} (5.00,39.67)(0.12,-0.12)[167]{}[(1,0)[0.12]{}]{} (5.33,59.67)[(1,0)[39.67]{}]{} (5.00,39.67) (5.00,59.67) (5.00,79.67) (45.00,39.67) (45.00,59.67) (45.00,79.67) (25.00,19.67) (25.00,99.67) (25.00,19.67)(0.21,0.12)[334]{}[(1,0)[0.21]{}]{} (95.00,59.67)(-0.21,0.12)[334]{}[(-1,0)[0.21]{}]{} (25.00,109.67)[(0,0)\[cc\][$a=a_0=a_9'$]{}]{} (50.00,79.67)[(0,0)\[cc\][$a_2$]{}]{} (50.00,59.67)[(0,0)\[cc\][$a_6$]{}]{} (50.00,39.67)[(0,0)\[cc\][$a_4$]{}]{} (-0.00,39.67)[(0,0)\[cc\][$a_3$]{}]{} (-0.00,59.67)[(0,0)\[cc\][$a_5$]{}]{} (-0.00,80.00)[(0,0)\[cc\][$a_1$]{}]{} (25.00,9.67)[(0,0)\[cc\][$a_7$]{}]{} (15.00,89.67) (35.33,89.34) (25.00,59.67) (15.00,29.67) (35.00,29.67) (155.67,89.67) (176.00,89.34) (165.67,59.67) (155.67,29.67) (175.67,29.67)
Principle of context translation
================================
Suppose it is not unreasonable to speculate about the following three assumptions [@svozil-2003-garda]:
- [*Unique context preparation:*]{} It is possible to encode into quanta a certain finite amount of information by preparing them in a single context. This amount is determined by the dimension of the associated Hilbert space.
- [*Nonexistence of different contexts:*]{} (Counterfactual) Elements of physical reality which go beyond that single context do not exist.
- [*Context translation principle:*]{} If quanta are measured along a different context, that context may be translated by the measurement apparatus into the context the quanta have been originally prepared for. The capability of the measurement apparatus to translate the context may depend on certain parameters, such as temperature.
As strange as the first assumption may appear, it amounts to the everyday experience that no agent, deterministic or other, can be prepared to render answers to every conceivable question. A “silly” example for this feature would be the attempt of a person to enter the question [*“is there enough oil in the car’s engine?”*]{} at the command prompt of a desktop computer. Most likely, the machine would respond with some sort of statement that this input is not recognized as a command, executable, or batch file. Within the standalone desktop context, the question makes no sense at all. Yet nobody would come up with the suggestion that something strange, bordering to the mysterious, or “mindboggling” is going on. Compare also Zeilinger’s [*“Foundational Principle”*]{} stating that $n$ elementary two-state systems carry $n$ bits [@zeil-99; @zeil-bruk-02].
When considering quantized spin systems, one often has classical angular momentum models in mind; an observable which can be defined precisely in all directions. Quantum mechanically, this is no longer the case. Hence, some classical properties have to be given up. Spin or quantized angular momentum cannot be conceive as something being precisely defined in all directions simultaneously; it can be only defined precisely in a singular direction.
Nevertheless, spin measurements in different directions [*do*]{} give results, albeit randomized ones. Here, the context translation principle (iii) might be assumed, stating that any measurement apparatus capable of measuring different contexts from the one the quantum was originally prepared for, performs some kind of translation between the contexts. This translation may be thought of as brought about by intrinsic microscopic processes in the measurement apparatus. Take, as an analogy, linearly polarized light along a particular direction, and a linear polarization measurement in a different direction. Due to the dynamics of the oriented macromolecules of the polarization filters, there is light leaving the measurement filter (if it is not oriented perpendicular to the original polarization direction), and its polarization direction is changed to the orientation of the measurement device.
Assumptions (i) and (ii) amount to a subtle form of realism, since some “maximal” property—the context in which the quantum has been prepared for—is assumed to exist even without being observed by any finite mind. Yet other properties, associated with different contexts, are assumed to not exist at all. The possibility to measure these nonexisting properties presents an illusion, which is mediated by the ability of the measurement apparatus to translate the measurement context into the prepared context. There is no general method to obtain knowledge of an unknown specific preparation context for individual quanta. Neither is it possible to obtain [*a posteriori*]{} knowledge of such a context.
In this regard, quanta behave like little universal automata capable of storing a multitude of properties; yet only a single property at a time. The realm of this property is defined by the dimension of Hilbert space associated with this quantum.
Summary and open questions
==========================
This paper contains two main threads: a critical evaluation of counterfactuals, in particular contextuality, and a discussion of a context translation principle. Although there are strong connections, both issues could also be perceived independently.
The experimental and theoretical status can be summarized as follows. Measurements on the context (in)dependence of two-context two quanta (in three dimensions per quantum) configurations are feasible but still need to be done.
It remains an open theoretical question whether or not nonsinglet states exist which satisfy the nuniqueness property for sufficiently many different measurement “directions” or setups to allow for “explosion views” of more than two contexts by Einstein-Podolsky-Rosen type “explosion view” multipartite configurations. There maybe some higherdimensional states which satisfy a uniqueness property in some lowerdimensional subspace. If, as the author suspects, no such states exist, then the case against contextuality is quite firm. For, insofar as contextuality could be operationalized, it is likely to be falsified; and in more nontrivial cases it could not be operationalized. To state it pointedly, contextuality might turn out to be a [*red herring.*]{}
Finally, the entire issue of context translation remains speculative and theoretically and experimentally unsettled. Recall that quanta can only be measured and prepared in a single complete context associated with a maximal operator (per context). If the preparation and measurement context coincides, then ideally the measurement will just reveal the preparation with 100% certainty.
Suppose that the two contexts do not coincide. In this case, for any measurement to take place without a null result, it could be assumed that the measurement context may be translated by the measurement apparatus into the context the quanta have been originally prepared for. In this scenario, the context translation is carried out by the measurement apparatus alone. That is, the capability of the measurement apparatus to translate the context may depend on certain parameters, such as temperature. Alas, at the moment no models exist which could predict the exact mechanism and performance of context translation.
[10]{} \[1\][“\#1”]{} url \#1[`#1`]{}urlprefix
Schr[ö]{}dinger, E., *Naturwissenschaften*, **23**, 807–812, 823–828, 844–849 (1935), <http://www.emr.hibu.no/lars/eng/cat/>, [E]{}nglish translation in [@trimmer] and [@wheeler-Zurek:83 pp. 152-167]; http://www.emr.hibu.no/lars/eng/cat/.
Reichenbach, H., *Philosophic foundations of quantum mechanics*, University of California Press, Berkeley, 1944.
Jammer, M., *The Conceptual Development of Quantum Mechanics*, McGraw-Hill Book Company, New York, 1966.
Jammer, M., *The Philosophy of Quantum Mechanics*, John Wiley & Sons, New York, 1974.
Stace, W. T., “The Refutation of Realism,” in *Readings in philosophical analysis*, edited by H. Feigl and W. Sellars, Appleton–Century–Crofts, New York, 1949, previously published in [*Mind*]{} [**53**]{}, 1934.
Kochen, S., and Specker, E. P., *Journal of Mathematics and Mechanics*, **17**, 59–87 (1967), reprinted in [@specker-ges pp. 235–263].
Specker, E., *Dialectica*, **14**, 175–182 (1960), reprinted in [@specker-ges pp. 175–182]; [E]{}nglish translation: [*The logic of propositions which are not simultaneously decidable*]{}, reprinted in [@hooker pp. 135-140].
Hertz, H., *[P]{}rinzipien der [M]{}echanik*, Barth, Leipzig, 1894.
Svozil, K., *Quantum Logic*, Springer, Singapore, 1998.
Wright, R., *Foundations of Physics*, **20**, 881–903 (1990).
Svozil, K., *International Journal of Theoretical Physics*, **44**, in print (2005).
Pt[á]{}k, P., and Pulmannov[á]{}, S., *Orthomodular Structures as Quantum Logics*, Kluwer Academic Publishers, Dordrecht, 1991.
Kamber, F., *Nachr. Akad. Wiss. G[ö]{}ttingen*, **10**, 103–124 (1964).
Kamber, F., *Mathematische Annalen*, **158**, 158–196 (1965).
Zierler, N., and Schlessinger, M., *Duke Mathematical Journal*, **32**, 251–262 (1965).
Bell, J. S., *Reviews of Modern Physics*, **38**, 447–452 (1966), <http://dx.doi.org/10.1103/RevModPhys.38.447>, reprinted in [@bell-87 pp. 1-13].
Alda, V., *Aplik. mate.*, **25**, 373–374 (1980).
Alda, V., *Aplik. mate.*, **26**, 57–58 (1981).
Peres, A., *Quantum Theory: Concepts and Methods*, Kluwer Academic Publishers, Dordrecht, 1993.
Mermin, N. D., *Reviews of Modern Physics*, **65**, 803–815 (1993), <http://dx.doi.org/10.1103/RevModPhys.65.803>.
Svozil, K., and Tkadlec, J., *Journal of Mathematical Physics*, **37**, 5380–5401 (1996), <http://dx.doi.org/10.1063/1.531710>.
Tkadlec, J., *International Journal of Theoretical Physics*, **39**, 921–926 (2000), <http://dx.doi.org/10.1023/A:1003695317353>.
Ziegler, G. M., *Lectures on Polytopes*, Springer, New York, 1994.
Einstein, A., Podolsky, B., and Rosen, N., *Physical Review*, **47**, 777–780 (1935), <http://dx.doi.org/10.1103/PhysRev.47.777>.
Kok, P., Nemoto, K., and Munro, W. J., Properties of multi-partite dark states (2002), <http://arxiv.org/abs/quant-ph/0201138>.
Schimpf, M., and Svozil, K., All the singlet states (2004), in preparation.
Mermin, N. D., *American Journal of Physics*, **58**, 731 (1990).
Krenn, G., and A.Zeilinger, *Physical Review A*, **54**, 1793–1797 (1996), <http://dx.doi.org/10.1103/PhysRevA.54.1793>.
Pan, J.-W., Bouwmeester, D., Daniell, M., Weinfurter, H., and Zeilinger, A., *Nature*, **403**, 515–519 (2000).
Hess, K., and Philipp, W., *Europhysics Letters*, **57**, 775–781 (2002).
Heywood, P., and Redhead, M. L. G., *Foundations of Physics*, **13**, 481–499 (1983).
Redhead, M., *Incompleteness, Nonlocality, and Realism: A Prolegomenon to the Philosophy of Quantum Mechanics*, Clarendon Press, Oxford, 1990.
Bohr, N., “Discussion with [E]{}instein on epistemological problems in atomic physics,” in *[A]{}lbert [E]{}instein: Philosopher-Scientist*, edited by P. A. Schilpp, The Library of Living Philosophers, Evanston, Ill., 1949, pp. 200–241, <http://www.emr.hibu.no/lars/eng/schilpp/Default.html>.
Gleason, A. M., *Journal of Mathematics and Mechanics*, **6**, 885–893 (1957).
Pitowsky, I., *Physical Review Letters*, **48**, 1299–1302 (1982), <http://dx.doi.org/10.1103/PhysRevLett.48.1299>, cf. N. D. Mermin, [*Physical Review Letters*]{} [**49**]{}, 1214 (1982); A. L. Macdonald, [*Physical Review Letters*]{} [**49**]{}, 1215 (1982); Itamar Pitowsky, [ *Physical Review Letters*]{} [**49**]{}, 1216 (1982).
Pitowsky, I., *Physical Review D*, **27**, 2316–2326 (1983), <http://dx.doi.org/10.1103/PhysRevD.27.2316>.
Meyer, D. A., *Physical Review Letters*, **83**, 3751–3754 (1999), <http://dx.doi.org/10.1103/PhysRevLett.83.3751>.
Havlicek, H., Krenn, G., Summhammer, J., and Svozil, K., *J. Phys. A: Math. Gen.*, **34**, 3071–3077 (2001), <http://dx.doi.org/10.1088/0305-4470/34/14/312>.
Birkhoff, G., and von Neumann, J., *Annals of Mathematics*, **37**, 823–843 (1936).
von Neumann, J., *Mathematische Grundlagen der Quantenmechanik*, Springer, Berlin, 1932, [E]{}nglish translation: [*Mathematical Foundations of Quantum Mechanics*]{}, Princeton University Press, Princeton, 1955.
Greechie, J. R., *Journal of Combinatorial Theory*, **10**, 119–132 (1971).
Kalmbach, G., *Orthomodular Lattices*, Academic Press, New York, 1983.
Mair, A., Vaziri, A., Weihs, G., , and Zeilinger, A., *Nature*, **412**, 313–316 (2001), <http://dx.doi.org/10.1038/35085529>.
Vaziri, A., Weihs, G., , and Zeilinger, A., *Physical Review Letters*, **89**, 240401 (2002), <http://dx.doi.org/10.1103/PhysRevLett.89.240401>.
Riedmatten, H. D., Marcikic, I., Zbinden, H., and Gisin, N., *Quantum Information and Computing*, **2**, 425–433 (2002), [http://www.gap-optique.unige.ch/Publications/Pdf/QICfinale.pd%
f](http://www.gap-optique.unige.ch/Publications/Pdf/QICfinale.pd%
f).
Torres, J. P., Deyanova, Y., Torner, L., and Molina-Terriza, G., *Physical Review A*, **67**, 052313 (2003), <http://dx.doi.org/10.1103/PhysRevA.67.052313>.
Reck, M., Zeilinger, A., Bernstein, H. J., and Bertani, P., *Physical Review Letters*, **73**, 58–61 (1994), <http://dx.doi.org/10.1103/PhysRevLett.73.58>.
Reck, M., and Zeilinger, A., “Quantum phase tracing of correlated photons in optical multiports,” in *Quantum Interferometry*, edited by F. D. Martini, G. Denardo, and A. Zeilinger, World Scientific, Singapore, 1994, pp. 170–177.
Zukowski, M., Zeilinger, A., and Horne, M. A., *Physical Review A*, **55**, 2564–2579 (1997), <http://dx.doi.org/10.1103/PhysRevA.55.2564>.
Svozil, K., Single particle interferometric analogues of multipartite entanglement (2004), <http://arxiv.org/abs/quant-ph/0401113>.
Svozil, K., *Journal of Modern Optics*, **51**, 811–819 (2004).
Zeilinger, A., *Foundations of Physics*, **29**, 631–643 (1999).
Brukner, [Č]{}., and Zeilinger, A., “Information and fundamental elements of the structure of quantum theory,” in *Time, Quantum, Information*, edited by L. Castell and O. Ischebek, Springer, 2003.
Trimmer, J. D., *Proc. Am. Phil. Soc.*, **124**, 323–338 (1980), reprinted in [@wheeler-Zurek:83 pp. 152-167].
Wheeler, J. A., and Zurek, W. H., *Quantum Theory and Measurement*, Princeton University Press, Princeton, 1983.
Specker, E., *Selecta*, Birkh[ä]{}user Verlag, Basel, 1990.
Hooker, C. A., *The Logico-Algebraic Approach to Quantum Mechanics. [V]{}olume [I]{}: Historical Evolution*, Reidel, Dordrecht, 1975.
Bell, J. S., *Speakable and Unspeakable in Quantum Mechanics*, Cambridge University Press, Cambridge, 1987.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Using the effective Lagrangian approach, we clarify general issues about Nambu–Goldstone bosons without Lorentz invariance. We show how to count their number and study their dispersion relations. Their number is less than the number of broken generators when some of them form canonically conjugate pairs. The pairing occurs when the generators have a nonzero expectation value of their commutator. For non-semi-simple algebras, central extensions are possible. The underlying geometry of the coset space in general is partially symplectic.'
author:
- Haruki Watanabe
- Hitoshi Murayama
bibliography:
- 'references.bib'
title: 'Unified Description of Nambu–Goldstone Bosons without Lorentz Invariance'
---
#### Introduction.
—Spontaneous symmetry breaking (SSB) is ubiquitous in nature. The examples include magnets, superfluids, phonons, Bose–Einstein condensates (BECs), neutron stars, and cosmological phase transitions. When continuous and global symmetries are spontaneously broken, the Nambu–Goldstone theorem [@Nambu:1961; @Goldstone:1961; @Goldstone:1962] ensures the existence of gapless excitation modes, i.e., Nambu–Goldstone bosons (NGBs). Since the long-distance behavior of systems with SSB is dominated by NGBs, it is clearly important to have general theorems on their number of degrees of freedom and dispersion relations.
In Lorentz-invariant systems, the number of NGBs $n_{\mathrm{NGB}}$ is always equal to the number of broken generators $n_{\mathrm{BG}}$. All of them have the identical linear dispersion $\omega=c|k|$. However, once we discard the Lorentz invariance, the situation varies from one system to another.
Until recently, systematic studies on NGBs without Lorentz invariance have been limited. (See Ref. [@Brauner:2010] for a recent review.) Nielsen and Chadha [@Nielsen:1975] classified NGBs into two types: type I (II) NGBs have dispersion relations proportional to odd (even) powers of their momenta in the long-wavelength limit. They proved $n_\mathrm{I}+2n_\mathrm{II}\geq n_\mathrm{BG}$, where $n_\mathrm{I}$ ($n_\mathrm{II}$) is the number of type I (II) NGBs. Schäfer *et al.* [@Schafer:2001] showed that $n_\mathrm{NGB}$ is exactly equal to $n_\mathrm{BG}$ if $\langle0|[Q_i,Q_j]|0\rangle$ vanishes for all pairs of the symmetry generators $Q_i$. A similar observation is given in Ref. [@Nambu:2004]. Given these results, Brauner and one of us (H. W.) [@Watanabe:2011] conjectured $$\begin{gathered}
n_\mathrm{BG}-n_\mathrm{NGB}=\frac12\mathrm{rank}\,\rho, \label{eq:conjecture}\\
\rho_{ij}\equiv\lim_{\Omega\to\infty}\frac{-i}{\Omega}\langle0|[Q_i,Q_j]|0\rangle,
\label{eq:rho}\end{gathered}$$ where $\Omega$ is the spatial volume of the system.
In this Letter, we clarify these long-standing questions about the NGBs in Lorentz-non-invariant systems by proving the conjecture and showing the equality in the Nielsen-Chadha theorem with an improved definition using effective Lagrangians $\mathcal{L}_{\mathrm{eff}}$. We also clarify how the central extension of the Lie algebra makes a contribution to $\rho$ [^1].
#### Coset space.
—When a symmetry group $G$ is spontaneously broken into its subgroup $H$, the space of ground states form the coset space $G/H$ where two elements of $G$ are identified if $g_1 = g_2 h$ for $^{\exists} h \in H$. Every point on this space is equivalent under the action of $G$, and we pick one as the origin. The unbroken group $H$ leaves the origin fixed, while the broken symmetries move the origin to any other point. The infinitesimal action of $G$ is given in terms of vector fields ${\bold
h}_i = h_i^{\phantom{i}a}\partial_a$ $(i =
1,\ldots,\mathrm{dim}\,G)$ on $G/H$, where $\partial_a=\frac{\partial}{\partial \pi^a}$ with the local coordinate system $\{\pi^a\}$ $(a =
1,\ldots,n_{\mathrm{BG}}=\mathrm{dim}\,G-\mathrm{dim}\,H)$ around the origin. The infinitesimal transformations ${\bold h}_i$ satisfy the Lie algebra $[{\bold h}_i, {\bold h}_j] = f^k_{\phantom{k}ij} {\bold
h}_k$. We can always pick the coordinate system such that $\pi^a$’s transform linearly under $H$, namely, that ${\bold h}_i =\pi^bR^p(T_i)_b^{\phantom{b}a}\partial_a$, where $R^p(T_i)$ is a representation of $H$ [@Coleman:1969]. On the other hand, the broken generators are realized nonlinearly, ${\bold h}_b=
h_b^{\phantom{b}a} (\pi)\partial_a$ with $h_b^{\phantom{b}a} (0)\equiv X_b^{\phantom{b}a}$. Since broken generators form a basis of the tangent space at the origin, the matrix $X$ must be full-rank and hence invertible.
The long-distance excitations are described by the NGB fields $\pi^a(x)$ that map the space-time into $G/H$. We now write down its $\mathcal{L}_{\mathrm{eff}}$ in a systematic expansion in powers of derivatives, because higher derivative terms are less important at long distances.
#### Effective Lagrangians without Lorentz invariance.
—We discuss the $\mathcal{L}_{\mathrm{eff}}$ for the NGB degrees of freedom following Refs. [@Leutwyler:1994nonrel; @Leutwyler:1994rel]. Under *global* symmetry $G$, the NGBs transform as $\delta \pi^a = \theta^i h_i^{\phantom{i}a}$ where $\theta^i$ are infinitesimal parameters. However, we do not make $\theta^i$ local (gauge) unlike in these papers because it puts unnecessary restrictions on possible types of symmetries and their realizations, as we will see below.
It is well known that a symmetry transformation can change the Lagrangian density by a total derivative. The examples include space-time translations, supersymmetry, and gauge symmetry in the Chern–Simons theory [@Murayama:1989we]. We allow for this possibility in the $\mathcal{L}_{\mathrm{eff}}$ of the NGB fields. We assume spatial translational invariance and rotational invariance at sufficiently long distances in the continuum limit, while we can still discuss their SSB.
If Lorentz invariant, the $\mathcal{L}_{\mathrm{eff}}$ is highly constrained, $$\begin{aligned}
\mathcal{L}_\mathrm{eff} &=&
\frac{1}{2}g_{ab}(\pi)\partial_\mu\pi^a\partial^\mu\pi^b
+O(\partial_\mu^4).\end{aligned}$$ The invariance of the Lagrangian under $G$ requires that $g_{ab}$ is a $G$-invariant metric on $G/H$, namely $\partial_cg_{ab}h_i^{\phantom{i}c} + g_{ac}\partial_bh_i^{\phantom{i}c}
+g_{cb}\partial_ah_i^{\phantom{i}c}=0$. When the coordinates $\pi^a$ are reducible under $H$, the metric $g$ is a direct sum of irreducible components $g_{ab}= \sum_p F_p^2 \delta^p_{ab}$ where $\delta^p_{ab}$ vanishes outside the irreducible representation $p$ with arbitrary constants $F_p$ for each of them.
On the other hand, once we drop Lorentz invariance, the general $\mathcal{L}_{\mathrm{eff}}$ has substantially more freedom, $$\begin{aligned}
\mathcal{L}_\mathrm{eff} &=& c_a(\pi)\dot{\pi}^a
+\frac{1}{2}\bar{g}_{ab}(\pi)\dot{\pi}^a\dot{\pi}^b
-\frac{1}{2}g_{ab}(\pi)\partial_r\pi^a\partial_r\pi^b \nonumber \\
& & +O(\partial_t^3,\partial_t\partial_r^2,\partial_r^4),
\label{eq:effective}\end{aligned}$$ where $\bar{g}_{ab}$ is also $G$ invariant. Here and hereafter, $r=1,\ldots,d$ refers to spatial directions.
Note that the spatial isotropy does not allow terms with first derivatives in space in the $\mathcal{L}_{\mathrm{eff}}$. Therefore, the spatial derivatives always start with at least the second power $O(\partial_r^2)$. (Actually, it is not critical for us whether there are terms of $O(\partial_r^2)$; it may as well start at $O(\partial_r^4)$ without affecting our results, as we will see below.)
The Lagrangian density changes by a total derivative under the infinitesimal transformation $\delta \pi^a = \theta^i
h_i^{\phantom{i}a}$ *iff* $$\begin{gathered}
\left(\partial_b c_a-\partial_a
c_b\right) h_i^{\phantom{k}b} = \partial_ae_i.
\label{eq:hw}\end{gathered}$$ The functions $e_i(\pi)$ introduced in this way are actually related to the charge densities of the system. By paying attention to the variation of the Lagrangian by the surface term $$\delta \mathcal{L}_{\mathrm{eff}} = \theta^i \partial_t (c_a h_i^{\phantom{i}a}+ e_i),
\label{eq:totalderivative}$$ we can derive the Noether current for the global symmetry $j^0_i = e_i -
\bar{g}_{ab} h_i^{\phantom{i}a} \dot{\pi}^b$. Since the ground state is time independent $\dot{\pi}^b = 0$, $$e_i(0)=\langle 0|j_i^0(x)|0\rangle.
\label{eq:ei}$$ It must vanish in the Lorentz-invariant case, which can be understood as the special situation where $c_a$ and $e_i$ vanish, and $g_{ab}=c^2 \bar{g}_{ab}$.
Before presenting the proof, we explain the advantage in not gauging the symmetry. A tedious calculation verifies $\partial_b(h_i^{\phantom{i}a}\partial_ae_j-f^k_{\phantom{k}ij} e_k)=0$, with a general solution, $$h_i^{\phantom{i}a}\partial_a e_j=f^k_{\phantom{k}ij} e_k+ c_{ij}.
\label{eq:he}$$ Therefore, $e_i(\pi)$’s transform as the adjoint representation under $G$, up to possible integration constants $c_{ij} = - c_{ji}$. These constants play important roles as seen below.
In the presence of such constants, the global symmetry cannot be gauged [@Leutwyler:1994nonrel]. This is reminiscent of the Wess–Zumino term that also changes by a surface term under a global symmetry and produces an anomaly upon gauging [@Wess:1971; @Witten:1983]. It is known that the constants can be chosen to vanish with suitable definitions of $e_i$ for semisimple Lie algebras, while a nontrivial second cohomology of the Lie algebra presents an obstruction [@Woodhouse:1992de].
#### Proof of the conjecture.
—The basic point to show is that when $\rho_{ij} \neq 0$, the NGB fields for the generators $i$ and $j$ are *canonically conjugate* to each other.
From Eq. and the assumed translational symmetry, the formula for $\rho$ in Eq. is reduced to $$\rho_{ij} =-i \langle 0 | [Q_i, j^0_j] | 0 \rangle
= h_i^{\phantom{i}a}\partial_a e_j\big|_{\pi=0}.$$ Obviously, this must vanish for unbroken generators by definition. Combining this with Eq. , we have $$h_i^{\phantom{i}a}h_j^{\phantom{i}b}
\left(\partial_bc_a-\partial_a c_b\right)\big|_{\pi=0}=\rho_{ij}.
\label{eq:de}$$ We now solve this differential equation around the origin. The Taylor expansion of $c_a(\pi)$ can be written as $c_a(\pi)=c_a(0)+(S_{ab}+A_{ab})\pi^b+O(\pi^2)$, where $S_{ab}$ and $A_{ab}$ stand for the symmetric and antisymmetric parts of the derivative $\partial_b c_a|_{\pi=0}$. Obviously $c_a(0)$ and $S_{ab}$ lead to only total derivative terms in the $\mathcal{L}_{\mathrm{eff}}$ and thus will be dropped later: $$c_a(\pi)\dot{\pi}^a
=A_{ab}\dot{\pi}^a\pi^b
+\partial_t
\left[c_a(0)\pi^a+\frac{1}{2}S_{ab}\pi^a\pi^b\right]+O(\pi^3).$$ The equation for the antisymmetric part $2X_c^{\phantom{b}a}X_d^{\phantom{b}b}A_{ab}=\rho_{cd}$ has a unique solution which gives $$c_a(\pi)\dot{\pi}^a=\frac{1}{2}\rho_{ab}\dot{\tilde{\pi}}^a\tilde{\pi}^b+O(\tilde{\pi}^3),$$ where $\tilde{\pi}^a\equiv\pi^b(X^{-1})_b^{\,\,a}$. Since the matrix $\rho$ is real and antisymmetric, we can always transform it into the following form by a suitable orthogonal transformation $\tilde{Q}_i=O_{ij}Q_j$: $$\rho=
\begin{pmatrix}
M_1&&&&&\\
&\ddots&&&&\\
&&M_m&&&\\
&&&0&&\\
&&&&\ddots&\\
&&&&&0
\end{pmatrix}
,\,\,
M_\alpha=
\begin{pmatrix}
0&\lambda_\alpha\\
-\lambda_\alpha&0
\end{pmatrix}.$$ Here, $\lambda_\alpha\neq0$ for $\alpha=1,\ldots,m=\frac{1}{2}\mathrm{rank}\,\rho$, while the remaining elements identically vanish.
The most important step in the proof is to write down the explicit expression of the $\mathcal{L}_{\mathrm{eff}}$ in Eq. , $$c_a(\pi)\dot{\pi}^a
=\sum_{\alpha=1}^m\dfrac{1}{2}\lambda_{\alpha}
(\tilde{\pi}^{2\alpha}\dot{\tilde{\pi}}^{2\alpha-1}
-\dot{\tilde{\pi}}^{2\alpha}\tilde{\pi}^{2\alpha-1}),
\label{eq:c}$$ which is in the familiar form of the Lagrangian on the phase space $L=p_i\dot{q}^i-H$ [^2]. Namely, $\tilde{\pi}^{2\alpha-1}$ and $\tilde{\pi}^{2\alpha}$ are canonically conjugate variables, and they together represent one degree of freedom rather than two degrees of freedom. Hereafter we call the first set of $\tilde{\pi}^a$’s $(a=1,
\ldots 2m)$ type B, and the rest type A. Hence, $n_\mathrm{A}+2n_\mathrm{B}=n_\mathrm{BG}$ with $n_\mathrm{A}=n_\mathrm{BG}-2m$ and $n_\mathrm{B}=m$. Thus we proved the conjecture Eq. .
The definition of a degree of freedom here is the conventional one in physics; i.e., one needs to specify both the instantaneous value and its time derivative for each degree of freedom as initial conditions. This definition does not depend on the terms with spatial derivatives in the Lagrangian.
Now we are in the position to prove that the equality is satisfied in the Nielsen–Chadha theorem if the term with two spatial derivatives exists with a nondegenerate metric $g_{ab}$. Then Eq. implies that the type A NGB fields have linear dispersion relations $\omega\propto k$, while the type B NGB fields have quadratic dispersions $\omega\propto k^2$. In this case, our type A (B) coincides with their type I (II), respectively, and the Nielsen–Chadha inequality is saturated.
On the other hand, if we allow the second-order term $O(\partial_r^2)$ to vanish accidentally but the fourth-order term $O(\partial_r^4)$ to exist, the unpaired (type A) NGBs happen to have a quadratic dispersion ($\omega^2 \propto k^4$, and hence type II) yet count as independent degrees of freedom each [@Watanabe:2011]. Therefore, the Nielsen–Chadha theorem is still an inequality in general. In contrast, our distinction between type A and type B NGBs is clearly determined by the first two time derivatives, and defines the number of degrees of freedom unambiguously. Therefore, the classification between odd and even powers in the dispersion relation is not an essential one, and our theorem is stronger than that by Nielsen and Chadha.
Note that the Lagrangian formalism is mandatory in our discussion, because the presence of the first-order derivatives in time essentially affects the definition of the canonical momentum, while a Hamiltonian is written with a fixed definition of the canonical momentum.
#### Examples.
—The simplest and most famous example of a type B NGB is the Heisenberg ferromagnet $H=-J\sum_{\langle i,j\rangle}
\bm{s}_i \cdot \bm{s}_j$ with $J>0$ on a $d$-dimensional square lattice ($d>1$). In this case, the original symmetry group $\mathrm{O}(3)$ is spontaneously broken down into the subgroup $\mathrm{O}(2)$. The coset space is $\mathrm{O}(3)/\mathrm{O}(2)\cong
S^2$. We assume that the ground state has all spins lined up along the positive $z$ direction without a lack of generality. Even though there are two broken generators, there is only one NGB with the quadratic dispersion relation $\omega \propto k^2$.
The coset space can be parametrized as $(n_x,n_y,n_z)=\big(\pi^1,\pi^2,\sqrt{1-(\pi^1)^2-(\pi^2)^2}\big)$. The $\mathrm{O}(3)$ transformation $h_{i}^{\phantom{i}a}=\epsilon_{iaj}n_j$ ($i,j=x,y,z$; $a=1,2$) is realized linearly for the unbroken generator $h_z^{\phantom{z}a}(\pi)=\epsilon_{ab}\pi^b$, while nonlinearly for broken ones $X_a^{\phantom{a}b}=\epsilon_{ab}$. One can show that the $\mathcal{L}_{\mathrm{eff}}$ consistent with the $\mathrm{O}(3)$ symmetry up to $O(\partial^2$) is $$\mathcal{L}_\mathrm{eff} = m\frac{
n_y\dot{n}_x-n_x\dot{n}_y}{1+n_z}
+\frac{1}{2}\bar{F}^2 \dot{\bm{n}}^2
-\frac{1}{2}F^2\partial_r\bm{n}\partial_r\bm{n} .
\label{eq:ferro}$$ Comparing to Eqs. and , we can read off $c_a$ and $e_i$ as $c_1=\frac{mn_y}{1+n_z}$, $c_2=-\frac{mn_x}{1+n_z}$ and $e_i=m n_i$. Hence $m=\langle j^0_z\rangle$ represents the magnetization of the ground state. It is clear that there is only one type B NGB because $\pi^1$ and $\pi^2$ are canonically conjugate to each other, with a quadratic dispersion $\omega \propto k^2$.
However, for an antiferromagnet, $J<0$, the overall magnetization cancels between sublattices, and therefore $e_i = 0$, which in turn requires $c_a = 0$. As a consequence, the lowest order term in the time derivative expansion has two powers, and we find that both $\pi^1$ and $\pi^2$ represent independent type A NGBs with linear dispersions $\omega
\propto |k|$. The generalization to the ferrimagnetic case is straightforward.
Another example is the spontaneously broken translational invariance that leads to acoustic phonons in an isotropic medium [@Landau:1986]. The displacement vector $\bm{u} (x)$ represents the NGBs under the spatial translation $\bm{u} \rightarrow \bm{u} + \bm{\theta}$, hence $G={\mathbb R}^3$ and $H=0$. Then with $\mathrm{O}(3)$ symmetry of spatial rotations, the most general form of the continuum $\mathcal{L}_{\mathrm{eff}}$ is $$\mathcal{L}_{\mathrm{eff}} = \frac{1}{2} \dot{\bm{u}}^2
- \frac{c_\ell^2}{2} (\bm{\nabla} \cdot \bm{u})^2
- \frac{c_t^2}{2} (\bm{\nabla} \times \bm{u})^2.$$ We recover the usual result of one longitudinal and two transverse phonons with linear dispersions $\omega=c_\ell k$ and $\omega=c_t k$, respectively (type A). When the $\mathrm{O}(3)$ symmetry is reduced to $\mathrm{SO}(2)\times
{\mathbb Z}_2$ for rotation in the $xy$ plane and the reflection $z
\rightarrow -z$, there are considerably more terms one can write down. Using the notation $\psi = u_x + i u_y$, $\partial=\frac{1}{2}(\partial_x - i \partial_y)$, and $\bar{\psi}$ and $\bar{\partial}$ for their complex conjugates, the most general $\mathcal{L}_{\mathrm{eff}}$ is $$\begin{aligned}
\lefteqn{
{\cal L}_{\rm eff}
= \dfrac{ic_{xy}}{2} \bar{\psi} \dot{\psi}
+ \frac{1}{2} \dot{u}_z^2 + \dot{\bar{\psi}} \dot{\psi}
- F_{0}^2 (\bar{\partial} \psi) (\partial \bar{\psi})
- \frac{1}{2} F_{1}^2 (\partial_z u_z)^2
} \nonumber \\
& &
- (\bar{\partial} u_z, \partial_z \psi)
\left(
\begin{array}{cc}
F_{2}^2 & F_{3}^2 \\ F_{3}^{2} & F_{4}^2
\end{array}
\right)
\left(
\begin{array}{c}
\partial u_z \\ \partial_z \bar{\psi}
\end{array}
\right)
- \frac{1}{2} ( F_{5}^2 (\partial \psi)^2 + c.c.) . \nonumber \\\end{aligned}$$ With $c_{xy}\neq 0$, we find there is one type A NGB with a linear dispersion, and one type B NGB with a quadratic dispersion. The first term $\frac{ic_{xy}}{2} \bar{\psi} \dot{\psi} =\frac12 c_{xy} (u_y \dot{u}_x-u_x \dot{u}_y)$ implies $$\rho_{xy} =-i\langle 0 | [P_x, j_y^0] | 0\rangle
= c_{xy} \neq 0.$$ Namely, this Lie algebra is a *central extension* of the Abelian algebra of the translation generators, i.e., $[P_i,
P_j] = c_{ij}\Omega$. As pointed out in Ref. [@Watanabe:2012], when the medium is electrically charged, an external magnetic field along the $z$ axis precisely leads to this behavior with $c_{xy}=2\omega_c$ (the cyclotron frequency), because the gauge-invariant translations in a magnetic field are generated by $P_i = -i\hbar \partial_i - \frac{e}{c} A_i$, which satisfy $\langle 0|[P_x , P_y]|0\rangle = i\frac{\hbar e}{c} B_z
N$ with the number of particles $N$. This would not be possible with the gauged $\mathcal{L}_{\mathrm{eff}}$ in Ref. [@Leutwyler:1994nonrel] that does not allow for the central extension.
As a more nontrivial example, let us consider a spinor BEC with $F=1$. The symmetry group is $G=\mathrm{SO}(3) \times \mathrm{U}(1)$, where $\mathrm{SO}(3)$ rotates three components of $F=1$ states, while $\mathrm{U}(1)$ symmetry gives the number conservation. The Lagrangian is written using a three-component complex Schrödinger field $\psi$, $$\mathcal{L}=i\hbar\psi^\dagger\dot{\psi}
-\dfrac{\hbar^2}{2m}\partial_r\psi^\dagger\partial_r\psi
+\mu\psi^\dagger\psi-\dfrac{\lambda}{4}(\psi^\dagger\psi)^2
-\dfrac{\kappa}{4}|\psi^T \psi|^2.
\label{eq:BEC}$$ Since the potential reads $\frac{\lambda+\kappa}{4}\hat{n}^2-\mu \hat{n}-\frac{\kappa}{4}\hat{\bm{S}}^2$ ($\hat{n}\equiv \psi^{\dagger}\psi$, $\hat{\bm{S}}\equiv\psi^{\dagger}\bm{S}\psi$ and $\bm{S}$ is the 3 by 3 spin matrix), we identify two possibilities for condensates $$\psi =v_p(0,0,1)^T\quad
\mbox{or} \quad \frac{v_f}{\sqrt{2}}(1,i,0)^T$$ for “polar” $(-\lambda < \kappa < 0)$ or “ferromagnetic” $(\kappa>0)$ states, where $v_p=\sqrt{\frac{2\mu}{\lambda+\kappa}}$ and $v_f= \sqrt{\frac{2\mu}{\lambda}}$ [@Ho:1998zz]. The magnetization density is given by $e_i(0)=\langle j_i^0 \rangle=- i \hbar \epsilon_{ijk} \psi_j^*
\psi_k$. In the polar case, there is no net magnetization, and the symmetry is broken into $H=\mathrm{SO}(2) \subset \mathrm{SO}(3)$. For the ferromagnetic case, there is a net magnetization $e_z(0)=\hbar v_f^2$, and the symmetry is broken into the diagonal subgroup $H$ of $\mathrm{U}(1)$ and $\mathrm{SO}(2) \subset \mathrm{SO}(3)$. Therefore, the unbroken symmetry is the same for both cases \[$H=\mathrm{SO}(2)=\mathrm{U}(1)$\], yet we see three type A NGBs for the polar case while one type A and one type B NGB for the ferromagnetic case as shown below.
For the polar case, we parameterize $\psi$ as $$\psi = (v_p + h) e^{i\theta} (\vec{n} + i \vec{\chi}), \quad
\vec{n}^2 = 1, \quad \vec{\chi} \perp \vec{n}.$$ After integrating out the gapped modes $h$ and $\vec{\chi}$, the Lagrangian (\[eq:BEC\]) becomes $$\mathcal{L}_{\mathrm{eff}} = \frac{\hbar^2}{\lambda+\kappa} \dot{\theta}^2
+ \frac{\hbar^2}{|\kappa|} \dot{\vec{n}}^2
- \frac{\hbar^2 v_p^2}{2m} \left[ (\partial_r \theta)^2
+ (\partial_r \vec{n})^2
\right].$$ We do find three type A NGBs with linear dispersions.
For the ferromagnetic case, we parameterize $\psi$ as $$\psi = (v_f+h)
\dfrac{e^{i\theta}}{\sqrt{2}(1+z^* z)}\begin{pmatrix}1-z^2\\i(1+z^2)\\2z\end{pmatrix}.
\label{eq:RP3}$$ After integrating out $h$, we find $$\begin{aligned}
\lefteqn{
\mathcal{L}_{\mathrm{eff}} = \hbar v^2_f i
\dfrac{z^* \dot{z} - \dot{z}^* z}{1+z^* z}
+ \frac{\hbar^2}{\lambda} \left( \dot{\theta} -i \dfrac{z^* \dot{z}-\dot{z}^*
z}{1+z^* z}\right)^2 } \nonumber \\
& & -\frac{\hbar^2 v^2_f}{2m} \left[
\left(\partial_r \theta -i \dfrac{z^* \partial_r{z}-\partial_r{z}^*
z}{1+z^* z}\right)^2
+ \dfrac{2\partial_r z^* \partial_r z}{(1+z^* z)^2}\right].\quad
\label{eq:LeffRP3}\end{aligned}$$ Clearly, $z$ and $z^*$ are canonically conjugate to each other, representing one type B NGB with a quadratic dispersion, while $\theta$ represents one type A NGB with a linear dispersion.
#### Underlying geometry.
—Having demonstrated our theorem Eq. at work in very different examples, we now study the underlying geometry. Usually, canonically conjugate pairs in mechanics (such as type B NGBs) imply a *symplectic structure* mathematically, which requires an even-dimensional manifold $M$, and if closed, a nontrivial second de Rham cohomology $H^2(M) \neq 0$. However, we have seen in the last two examples that type A and type B NGBs can coexist on an odd-dimensional $M$ with $H^2(M)=0$. This puzzle can be solved as follows.
The time integral of the first term in Eq. defines a one-form $c= c_a d \pi^a$ on the coset space, and its exterior derivative gives a closed two-form $\sigma=dc$. Using the coordinates in Eq. , $\sigma
=\sum_{\alpha=1}^m\lambda_\alpha
d\tilde{\pi}^{2\alpha}\wedge d\tilde{\pi}^{2\alpha-1}$ for $m$ type B NGBs, which resembles a symplectic two-form. However, type A NGB fields for the remaining $n_{\mathrm{BG}}-2m$ broken generators do not have terms with first order in time derivatives, and hence do not take part in $\sigma$. Therefore, $\sigma$ has a constant rank but is degenerate, and hence is not a symplectic structure in the usual sense.
This kind of a partially symplectic (or *presymplectic* [@Woodhouse:1992de]) structure is possible on a coset space by considering the following fiber bundle, $F\hookrightarrow G/H\stackrel{\pi}{\rightarrow}B$, where the base space $B = G/(H\times F)$ is symplectic. The fiber $F$ is a subgroup of $G$ that commutes with $H$. The symplectic structure $\omega$ on $B$ is pulled back to $G/H$ as $\sigma=\pi^* \omega$. Since $d\omega=0$ on $B$ implies $d\sigma = 0$ on $G/H$, we can always find a one-form $c$ such that $dc =
\sigma$ locally on $G/H$, which appears in $\mathcal{L}_{\mathrm{eff}}$. Type B NGBs live on the symplectic base manifold $B$, whose coordinates form canonically conjugate pairs, while the type A NGBs live on the fiber $F$, each coordinate representing an independent NGB. The type A and type B NGBs can coexist on $G/H$ in this fashion.
The Heisenberg ferromagnetic model has the coset space $S^2 = {\mathbb
C}{\mathrm P}^1$ which is Kähler and hence symplectic, with one type B NGB. On the other hand, the spinor BEC example in its ferromagnetic state has $G/H={\mathbb R}{\mathrm P}^3$ which is not symplectic. The last term in Eq. is nothing but the Fubini–Study metric on $S^2 = \mathbb{C}\mathrm{P}^{1}$ which is Kähler and hence symplectic. The first term in Eq. defines the one-form $c$ whose exterior derivative ${\mathrm d}c$ gives precisely the Kähler form associated with the metric up to normalization. However $\theta$ is an orthogonal direction with no connection to the symplectic structure. We can define the projection $\pi:{\mathbb R}{\mathrm P}^3 \rightarrow
S^2$ simply by eliminating the $\theta$ coordinate. It shows the structure of a fiber bundle $\mathrm{U}(1) \hookrightarrow {\mathbb R}{\mathrm
P}^3 \stackrel{\pi}{\rightarrow} \mathbb{C}\mathrm{P}^{1}$, which is the well-known Hopf fibration (the difference between $S^3$ and ${\mathbb R}{\mathrm P}^3 = S^3/{\mathbb Z}_2$ is not essential here). The phonons in the magnetic field also show a partially symplectic structure.
In fact, it is always possible to find such a symplectic manifold $B$ if $G$ is compact and semisimple, thanks to Borel’s theorem [@Borel:1954]. Generalizations to non-semi-simple groups would be an interesting future direction in mathematics.
#### Final remarks.
—In this Letter, we exclusively focused on true NGBs. We do not regard pseudo-NGBs [@Weinberg:1972] as NGBs, since they do not correspond to the broken symmetries and tend to acquire mass corrections. Also, we assumed that there are no gapless excitations other than NGBs; especially, this assumption fails when there is a Fermi surface on the ground state. Taking such degrees of freedom into account would be another interesting future direction.
We thank T. Brauner, D. Stamper-Kurn, M. Ueda, and K. Hori for useful discussions. H. W. is also grateful to T. Hayata and H. M. to T. Milanov. The work of H. M. was supported in part by the U.S. DOE under Contract No. DEAC03-76SF00098, by the NSF under Grant No. PHY-1002399, by the JSPS Grant (C) No. 23540289, by the FIRST program Subaru Measurements of Images and Redshifts (SuMIRe), CSTP, and by WPI, MEXT, Japan.
[^1]: For more details and extensive discussions, see H. Watanabe and H. Murayama (to be published).
[^2]: Darboux’s theorem says one can choose a local coordinate system such that higher order terms $O(\tilde{\pi}^3)$ vanish.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We have studied the transport process in the two dimensional electron gas (2DEG) in presence of a magnetic field and a dissipative environment at temperature T. By means of imaginary time series functional integral method we calculate the decay rates at finite temperature and in the presence of dissipation. We have studied decay rates for wide range of temperatures —- from the thermally activated region to very low temperature region where the system decays by quantum tunneling. We have shown that dissipation and impurity helps the tunneling. We have also shown that tunneling is strongly affected by the magnetic field. We have demonstrated analytical results for all the cases mentioned above.'
author:
- Malay Bandyopadhyay
title: 'Dissipative Tunneling in 2 DEG: Effect of Magnetic Field, Impurity and Temperature.'
---
-0.5cm
The decay of a metastable state received an extensive study recently [@legg1; @legg2] because of its important role played in physics, chemistry as well as transport in biomolecules [@vau; @chan; @marc]. The tunneling of two dimensional electron gas (2DEG) is a general phenomena and is widely encountered in mesoscopic systems [@sav; @pao1; @jain; @pao2], in semiconductor heterostructures [@butch; @lee], in quantum Hall systems [@zheng]. Here our focus is on tunneling of 2DEG in the presence of magnetic field, impurity potential and in a dissipative environment.\
In 2DEG the system can be thought of a collection of noninteracting charged particles moving in 2D. So one can visualize the system in question as if a charged particle of mass M moving in a metastable potential U(q) while coupled to a heat bath environment which is assumed here as a collection of harmonic oscillators. In macroscopic systems, the tunneling probability is strongly influenced by the interaction with the environment at finite temperature [@grab1; @grab2]. The problem of tunneling in the presence of a coupling to many degrees of freedom such as phonons, magnons, gives rise to quantum friction which strongly affects the tunneling rate [@grab1; @grab2; @grab3; @lang]. But all the above studies were in the absence of magnetic field. But here we shall study the tunneling process of 2DEG in presence of magnetic field. Besides dissipation we have special attention to the effects of impurity potential on the tunneling rate.\
Here we are actually discussing about the decay of a charged particle from a metastable potential well. Now at high temperatures the decay is thermally activated and the rate follows the famous Arrhenius law: $$\Gamma_{cl} = f_{cl}exp(-\frac{v_b}{k_BT})$$ where $V_b$ is the barrier height and $f_{cl}$ is the attempt frequency. As the temperature is lowered the classical escape rate decreases exponentially and it can decay only via quantum tunneling.\
Our method of analysis is based on pure thermodynamic equilibrium method which was pioneered by Langer [@lang]. In this approach the quantity of interest is the free energy of the metastable system. Because of states of lower energy on the other side of the barrier, the partition function can only b defined by means of an analytic continuation from a stable potential to the metastable situation depicted in Fig. 1. The procedure of analytic continuation leads to a unique Imaginary part of the free energy of the metastable state. This quantity is then related to the decay probability of the system. The thermal quantum rate of decay is related to $Im F$ by the following relation: $$K = -\frac{2}{\hbar}Im F$$ The method we employ to calculate the decay rate is a functional-integral approach which is convenient because it can easily take into account the dissipation as a nonlocal term. Beside this, we are dealing with a problem in a higher dimensional space: the particle moves in 2-D, and the effective dimension of the dissipative environment is infinite. The most convenient method for such problem is the path integral method which we use here. The Langer’s method [@lang] for the calculation of the quantum decay rate was developed through the work of Miller, Stone, Coleman Callan [@miller]. They use the term “bounce Euclidean action" $S_B$ and defined the quantum decay rate as $$\Gamma = \omega_q exp(-\frac{S_B}{\hbar}),$$ where quantum mechanical preexponential factor $\omega_q$ is related with fluctuations about the bounce. Dissipation was first incorporated into the bounce technique by Caldeira and Leggett [@cal]. Also Grabert [*et al*]{} [@grab1; @grab2; @grab3] and Larkin [*et al*]{} [@larkin] used this method as an effective scheme to calculate decay rates in the entire range of temperature. In the present paper we extend these study by adding a couple of new features in addition to the usual dissipative effect:(a) the effect of the magnetic field, (b) the effect of the impurity potential. With the preceding background I organize the paper as follows. In $Sec. II$ I discuss about the free energy and Euclidean action of the system. We use the path integral method to reduce the infinite dimensional problem to an effective one dimensional problem. In $Sec. III$ I discuss about the thermodynamic method for the calculation of decay rates in the high temperature region. In $Sec. IV$ I talk about the quantum correction near the crossover region. Section V is our main focus. Below the crossover temperature the metastable state predominantly decay via quantum tunneling. Here we discuss about the effect of dissipation, impurity potential and magnetic field on the quantum tunneling rate of 2DEG in great details. Finally I summarize and conclude the paper in $Sec. VI$.\
Free Energy of the Damped 2DEG
==============================
I am specifically dealing with two dimensional electron gas (2 DEG) which confined in the x-y plane. The magnetic field is applied perpendicular to the x-y plane. Now following Caldeira and Leggett [@legg1; @legg2] the system is assumed to be coupled linearly to its environment which can be represented as a collection of harmonic oscillators. Following Feynmann-Vernon [@feyn] one can integrates out all the environmental degrees of freedom and other irrelevant degrees of freedom. So the problem effectively comes out to be 1-D problem. Our starting point is the total Hamiltonian of the Global system i.e. the system and the reservoir which are coupled linearly: $${\cal H} = {\cal H_S} + {\cal H_B} + {\cal H_I},$$ where subscripts $S$, $B$, and $I$ stand for ‘system’,‘bath’ and ‘interaction’ term respectively. $$\begin{aligned}
{\cal H}& = & \sum_{l=x,y}\Big[\frac{1}{2M}\Big(p_{l} + \frac{e}{c}{A_{l}}({q_{l}})\Big)^{2}+U(q_{l})\Big]\\ & & +\sum_{j,l=x,y}\Big[\frac{1}{2m_j}p_{j,l}^{2} + \frac{1}{2}m_j\omega_{j}^{2}q_{j,l}^{2}\Big]\\ & & + \sum_{j,l=x,y}\frac{c_{j,l}^{2}}{2m_j\omega_{j}^{2}}q_{l}^{2} - \sum_{j,l=x,y}c_{j,l}q_{j,l}q_l.\end{aligned}$$ or, $$\begin{aligned}
{\cal H} & = & \sum_{l=x,y}\Big[\frac{1}{2M}\Big(p_{l} + \frac{e}{c}{A_{l}}({q_{l}})\Big)^{2}+U(q_{l})\Big] \\ \nonumber & & +\sum_{j,l=x,y}\Big[\frac{1}{2m_j}p_{j,l}^{2} + \frac{1}{2}m_j\omega_{j}^{2}(q_{j,l}-\frac{c_{j,l}}{m_j\omega_{j}^{2}}q_{l})^{2}\Big],\end{aligned}$$ where $q_l = (q_x,q_y)= \vec{q}$ are the co-ordinate of charged particle and $q_{j,l} = (q_{j,x},q_{j,y})$ are environmental degrees of freedom. Now all quantities characterizing the environment may be expressed in terms of the spectral density of bath oscillators $$J_l(\omega) = \pi \sum_{j=1}{N} \frac{c_{j,l}^{2}}{2m_{j}\omega_{j}}\delta(\omega - \omega_{j}).$$ Sometimes it is needed to introduce cut-off frequency $\omega_c$ [@pao1; @pao2]. Thus $$J_l(\omega) = \eta_{l}\omega^{s_l}exp(-\frac{\omega}{\omega_c}).$$ In the present paper we follow the second definition of spectral density. The partition function of the charged Brownian particle can be written as a functional integral [@cal] over periodic paths where path probability is weighted according to the Euclidean action [@legg1; @grab2; @cal; @hanggi]. $$\begin{aligned}
S^{E}[\vec{q}(\tau),\dot{\vec{q}}(\tau)] & = & \int_{0}^{\hbar\beta}d\tau\Big[\frac{M}{2}\dot{\vec{q}}^{2}(\tau) + U(\vec{q}(\tau))\\ \nonumber & & - iM\omega_c(\vec{q}(\tau)\times\dot{\vec {q}}(\tau))_{z}\Big]\\ \nonumber & & + \sum_{j,l=x,y}\Big[\frac{1}{2m_j}p_{j,l}^{2} + \frac{1}{2}m_j\omega_{j}^{2}(q_{j,l}-\frac{c_{j,l}}{m_j\omega_{j}^{2}}q_{l})^{2}\Big]\end{aligned}$$ The above action can be transformed into an effective action $$\begin{aligned}
S^{E}[\vec{q}(\tau),\dot{\vec{q}}(\tau)] & = & \int_{0}^{\hbar\beta}d\tau\Big[\frac{M}{2}\dot{\vec{q}}^{2}(\tau) + U(\vec{q}(\tau))\\ \nonumber & & - iM\omega_c(\vec{q}(\tau)\times\dot{\vec {q}}(\tau))_{z}\Big]\\ \nonumber & & +\frac{1}{2}\int_{0}^{\hbar\beta}d\tau\int_{0}^{\hbar\beta}d\tau^{\prime}{\cal K}(\tau-\tau^{\prime})q(\tau)q(\tau^{\prime}),\end{aligned}$$ where $\beta = \frac{1}{k_BT}$. The first term in Eq. (9) represents the conservative or reversible motion of the particle whereas the second term introduces the dissipation. The damping kernel ${\cal K}(\tau)$ may be expressed as a Fourier series $${\cal K}(\tau) = \frac{M}{\hbar\beta}\sum_{n=-\infty}^{+\infty}{\cal K}(\nu_n)exp(i\nu_n\tau),$$ where Matsubara frequencies $\nu_n = \frac{2\pi nk_BT}{\hbar}$ and the Fourier coefficients are given by ${\cal K}(\nu_n) = \sum_{i}\frac{c_i^{2}}{m_i\omega_i^{2}}\frac{\nu_n^{2}}{\nu_n^{2}+\omega_i^{2}}$. The partition function is defined as $${\cal Z} = \int D[q]exp\Big[-\frac{1}{\hbar}S^{E}[q]\Big],$$ where the functional integral is over all periodic paths with period $\hbar\beta$. The free energy is given by $$F = -\frac{1}{\beta}ln{\cal Z}.$$\
Thermally Activated Decay
=========================
Temperature dependence of quantum decay rates in dissipative systems was first studied by Grabert [*et al*]{} [@grab1; @grab2; @grab3]. They define crossover temperature $T_0$ at which the transition between thermal hopping and quantum tunneling occurs. For temperature above $T_{0}$, the main contribution in the functional integral (11) arises from the vicinity of the time-independent trajectories $q(\tau) = 0$, where the particle sits on top of the potential barrier of the inverted potential, and $q(\tau) = q_b$, where it sits at the bottom of the well as shown in Fig. 2. a periodic path near $q(\tau) = 0$ can be expanded in terms of Matsubara frequencies $(\nu_n)$ as follows $$x^{\prime}(\tau) = \sum_{n = -\infty}^{+\infty}X_n^{\prime}exp(i\nu_n\tau).$$
When we put Eq. (13) into Eq. (9) one obtains $$S[X_n^{\prime}] = \frac{M\hbar\beta}{2}\sum_{n = -\infty}^{+\infty}\lambda_n^{0}X_n^{\prime}X_{-n}^{\prime},$$ where, $\lambda_n^{0} = \nu_{n}^{2}+\omega_0^{2}+2\omega_c\nu_n+|\nu_n|\hat{\gamma}(|\nu_n|)$, with well frequency $\omega_0 = [\frac{U^{\prime\prime}(0)}{M}]^{\frac{1}{2}}$ and cyclotron frequency $\omega_c = \frac{eB}{Mc}$. Now the partition function ${\cal Z}_0$ can be obtained by performing the Gaussian integrals over the amplitudes $X_n$. A periodic path around $q(\tau) = q_b$ can be written as $$y^{\prime}(\tau) = q_b +\sum_{n = -\infty}^{+\infty}Y_n^{\prime}exp(i\nu_n\tau).$$ Now the action becomes $$S[Y_n^{\prime}] = \hbar\beta U_b + \frac{M\hbar\beta}{2}\sum_{n = -\infty}^{+\infty}\lambda_n^{b}Y_n^{\prime}Y_{-n}^{\prime},$$ where, $\lambda_n^{b} = \nu_{n}^{2}-\omega_b^{2}+2\omega_c\nu_n+|\nu_n|\hat{\gamma}(|\nu_n|)$, with frequency $\omega_b = [\frac{U^{\prime\prime}(b)}{M}]^{\frac{1}{2}}$. Since $\lambda_{0}^{b} = - \omega_{b}^{2}$; hence the integral over the amplitude $Y_{0}$ diverges. This leads to an imaginary part of free energy. The imaginary part of the free energy is given by $$\begin{aligned}
Im F & = & -\frac{1}{\beta{\cal Z}_{0}}Im {\cal Z}_{b}\\ \nonumber
& = & -\frac{1}{2\beta}\Big[\frac{D_0}{D_b}\Big]^{\frac{1}{2}}exp(-\beta U_b),\end{aligned}$$ where, $$D_0 = \prod_{n=-\infty}^{\infty}\lambda_n^{0};\ \ \ \ \ D_b = \prod_{n=-\infty}^{\infty}\lambda_n^{b}.$$ Affleck [@aff] has shown that above $T_0$ the decay rate given by Eq. (2) is modified as follows $$\Gamma = - \frac{2}{\hbar}\Big(\frac{T_0}{T}\Big)Im F.$$ Now using the fact that $\lambda_0^{0}=\omega_0^{2}$ and $\lambda_0^{b}=-\omega_b^{2}$ and combining Eq. (17) with Eq. (19) we obtain $$\Gamma = \frac{\omega_0}{2\pi}\frac{\omega_R}{\omega_b}\Big(\prod_{n=1}^{\infty}\frac{\nu_{n}^{2}+\omega_0^{2}+2\omega_c\nu_n+|\nu_n|\hat{\gamma}(|\nu_n|)}{\nu_{n}^{2}-\omega_b^{2}+2\omega_c\nu_n+|\nu_n|\hat{\gamma}(|\nu_n|)}\Big)exp(-\frac{U_b}{k_BT}),$$ where we made use the definition of crossover temperature $T_0 = \frac{\hbar\omega_R}{2\pi k_B}$ [@grab1; @hanggi]. $\omega_R$ is the largest positive root of the equation $\omega_R^{2}+\omega_R(\hat{\gamma}(\omega_R)+2\omega_c) = \omega_b^{2}$. Now at high temperature $T>>T_0$ the product term in the bracket (...) becomes unity, thus we obtain $$\Gamma_{cl} = \frac{\omega_0}{2\pi}\frac{\omega_R}{\omega_b}exp(-\frac{U_b}{k_BT}),$$ which is the famous Arrhenius law. The influence of dissipation and magnetic field on the classical rate arises through memory renormalized barrier frequency $\omega_R$.\
As the temperature is lowered, the thermally activated decay rate is modified by the quantum correction factor. Quantum fluctuations enhances the decay probability because they increase the average energy of the Brownian charged particle. Now following Hanggi [*et al*]{} [@hanggi] we obtain the resulting leading quantum corrections are found to be given by the formula $$f_q = exp\Big[\frac{\hbar^{2}}{24}\frac{(\omega_0^{2}+\omega_b^{2})}{(k_BT)^{2}}\Big]$$ This $f_q$ is independent of dissipative and magnetic field mechanism. Let us now consider frequency independent damping case i.e. ohmic damping case. For the ohmic damping case the product in Eq. (20) can be evaluated in terms of $\Gamma$ functions as follows [@woly] $$f_q = \frac{\Gamma(1-\frac{\lambda_b^{+}}{\nu})\Gamma(1-\frac{\lambda_b^{-}}{\nu})}{\Gamma(1-\frac{\lambda_0^{+}}{\nu})\Gamma(1-\frac{\lambda_0^{-}}{\nu})},$$ where $\nu = \frac{2\pi k_BT}{\hbar}$ and $$\begin{aligned}
\lambda_b^{\pm}& = &\frac{-(\gamma+2\omega_c)\pm\sqrt{(\gamma+2\omega_c)^{2}+4\omega_b^{2}}}{2};\\ \nonumber \lambda_0^{\pm}& = &\frac{-(\gamma+2\omega_c)\pm\sqrt{(\gamma+2\omega_c)^{2}-4\omega_0^{2}}}{2}.\end{aligned}$$ For strongly damped system $\gamma>>\omega_0,\omega_b$ and $T>>T_0$ the ohmic quantum correction factor Eq. (23) becomes $$f_q = exp\Big\lbrace\frac{T_0}{T}\Big[1+\frac{\omega_0^{2}}{\omega_b^{2}}\Big]\Big[\Psi\Big[1+4\alpha^{2}\frac{T_0}{T}\Big]-\Psi(1)\Big]\Big\rbrace,$$ where $\Psi(x)$ is the digamma function and $\alpha = \frac{\gamma+2\omega_c}{2\omega_b}$. On the other hand, for intermediate temperatures $T_0<<T<<4\alpha^{2}T_0$ we obtain the quantum correction factor [@grab1] $$f_q = \Big(\frac{4\alpha^{2}T_0}{T}\Big)^{\frac{(\omega_b^{2}+\omega_0^{2})T_0}{\omega_b^{2}T}}.$$ Now in the third case where $\gamma<<\omega_0,\omega_b$ we have [@grab1] $$f_q = \frac{\omega_b}{\omega_0}\frac{sinh\Big(\frac{\hbar\omega_0}{2k_BT}\Big)}{sin\Big(\frac{\hbar\omega_b}{2k_BT}\Big)}exp[D\alpha+O(\alpha^{2})],$$ where $\alpha = \frac{\gamma+2\omega_c}{2\omega_b}$ and $$D = \frac{\omega_b}{\nu}\Big[\Psi(1+\frac{\omega_b}{\nu})+\Psi(1-\frac{\omega_b}{\nu})-\Psi(1+i\frac{\omega_0}{\nu})-\Psi(1-i\frac{\omega_0}{\nu})\Big].$$\
NEAR CROSSOVER REGION
=====================
In the above section we have seen that the quantum corrections become very much important as we approach the crossover temperature $T_0$. In the vicinity of $T_0$ the above mentioned semiclassical approximation fails because the eigenvalue $\lambda_1^{b}=\lambda_{-1}^{b}=\nu^{2}-\omega_b^{2}+(2\omega_c+\hat{\gamma}(\nu))\nu$ vanishes for $T=T_0$. Hence the Gaussian integral over the amplitude $Y_1,Y_{-1}$ diverges. Now to regularize this divergent integral I add the higher order terms in amplitudes $Y_1,Y_{-1}$ [@grab1]. I expand the potential U(q) about the barrier top $U(q) = U_b - \frac{1}{2}M\omega_b^{2}(q-q_b)^2 + \sum_{k=3}^{\infty}\frac{1}{k}Mc_k(q-q_b)^k$. Now following Grabert [*et al*]{} I obtain the quantum correction factor in the crossover region as follows: $$f_q = \Big(\frac{\lambda_1^{0}}{\Lambda_1}\Big)f_R,$$ where $$f_R = \prod_{n=2}^{\infty}\Big(\frac{\lambda_n^{0}}{\lambda_n^{b}}\Big)\nonumber,$$ $\lambda_n^{0} = \nu_n^{2}+\omega_0^{2}+(2\omega_c+\hat{\gamma}(\nu_n))\nu_n$,\
$\lambda_n^{b} = \nu_n^{2}-\omega_b^{2}+(2\omega_c+\hat{\gamma}(\nu_n))\nu_n$,\
$\Lambda_1 = \Big[\Big(\frac{\pi M\beta}{2B}\Big)^{\frac{1}{2}}erfc\Big[\lambda_1^{b}\Big(\frac{M\beta}{2B}\Big)^{\frac{1}{2}}\Big]exp\Big[(\lambda_1^{b})^{2}(\frac{M\beta}{2B})\Big]\Big]^{-1}$ and $B = \frac{4c_3^{2}}{\omega_b^{2}+\omega_c^{2}}-\frac{2c_2^{2}}{\lambda_2^{b}} + 3c_4$.\
A more detailed analysis on the crossover region has been discussed by many others [@grab1; @weiss; @grab2; @grab3], although they did not consider the magnetic field effect which I have considered here. Now I switch over to our main focus i.e. the quantum tunneling region.\
QUANTUM TUNNELING REGION
========================
Below the crossover temperature $T_0$, the tunneling through the barrier becomes more probable and the decay rate is mainly given by the tunneling rate. In this region the most probable escape path is the bounce trajectory $q_B(\tau)$. For temperature below $T_0$ the imaginary part of the free energy comes from the so-called bounce trajectory [@woly]. The decay rate is given by [@grab1; @grab2; @grab3]: $$\Gamma = \Big(\frac{S_0}{2\pi\hbar}\Big)\Big(\frac{D_0}{D_B^{\prime}}\Big)^{\frac{1}{2}}exp\Big(-\frac{S_B}{\hbar}\Big),$$ where $S_B$ is the action (9) evaluated along the bounce trajectory $q_B(\tau)$, $S_0$ is the zero-mode normalization factor $$S_0 = M\int_{0}^{\hbar\beta}d\tau \dot{q_B}^{2}(\tau),$$ $D_0$ is the product of eigenvalues of fluctuation modes about the metastable minimum, and $D_B^{\prime}$ is the product of eigenvalues of fluctuation modes about the bounce trajectory with the zero eigenvalue omitted.\
In my case I am interested in the 2D electron gas tunneling out of the metastable state $q_x = 0$ in the x direction. The electron gas is moving in the x-y plane and magnetic field is applied perpendicular to the x-y plane. The 2DEG is coupled with a dissipative environment. so far, we have considered arbitrary forms of the metastable potential. Now I consider the practically important case of a cubic potential, $U_1(q_x) = \frac{1}{2}M\omega_x^{2}\Big(q_x^{2}-\frac{q_x^{3}}{q_0}\Big)$,which is relevant for SQUID or a current-biased Josephson junction. I also approximate the impurity potential in the y direction by the harmonic potential. Thus $U(q_x,q_y) = \frac{1}{2}M\omega_x^{2}\Big(q_x^{2}-\frac{q_x^{3}}{q_0}\Big) + \frac{1}{2}M\omega_y^{2}q_y^{2}$. For my convenience I use the notation $\vec{x} = (x,y)$ instead of $\vec{q}_l = (q_x,q_y)$ . The tunneling is described by the Euclidean action $$\begin{aligned}
S^{E}& = & \int_{0}^{\hbar\beta}d\tau\Big[\frac{1}{2}M(\dot{x}^{2}+\dot{y}^{2})+iM\omega_c\dot{x}y+U_1(x)\Big]\\ \nonumber & & +\frac{1}{2}M\omega_y^{2}y^{2} +\sum_{j}\Big[\frac{1}{2m_j}\vec{p}_{j}^{2} + \frac{1}{2}m_j\omega_{j}^{2}(\vec{q}_{j}-\frac{c_{j}}{m_j\omega_{j}^{2}}\vec{x})^{2}\Big],\end{aligned}$$ where I have used Landau gauge. After the tunneling the other degrees of freedoms, $y$ and ${q_j}$ can take any allowed values. Hence one can easily integrate out the environmental degrees of freedom ${q_j}$ and $y$ coordinates. Before doing all these integration over ${q_j}$ and $y$ coordinates, we perform a Fourier transformation in the time interval $[0,\hbar\beta]$: $$\begin{aligned}
\Big(x(\tau),y(\tau)\Big)& = & \frac{1}{\hbar\beta}\sum_{n=-\infty}^{+\infty}(x_n,y_n)e^{i\nu_n\tau} \\
q_j &=& \frac{1}{\hbar\beta}\sum_{n=-\infty}^{+\infty} q_{j,n}e^{i\nu_n\tau}\end{aligned}$$ Now the action can be rewritten in terms of Fourier components as follows: $$\begin{aligned}
S^{E}& = & \frac{1}{\hbar\beta}\sum_{n=-\infty}^{+\infty}\Big[\frac{1}{2}M\nu_n^{2}x_nx_{- n}+\frac{1}{2}(M\nu_n^{2} + M\omega_y^{2})y_ny_{- n}\Big]\\
\nonumber
& & +U_1(x_n) + \frac{1}{\hbar\beta}\sum_{n=-\infty}^{+\infty}M\omega_cx_ny_{- n}\nu_{- n}\\
\nonumber
& & +\frac{1}{\hbar\beta}\sum_{n=-\infty}^{+\infty}\sum_{j}\Big[\frac{1}{2}m_j\nu_n^{2}\vec{q}_{j,n}\vec{q}_{j,- n}\\
\nonumber
& & + \frac{1}{2}m_j\omega_j^{2}\Big(\vec{q}_{j,n}-\frac{c_j}{m_j\omega_j^{2}}\vec{x}_{n}\Big)\Big(\vec{q}_{j,- n}-\frac{c_j}{m_j\omega_j^{2}}\vec{x}_{- n}\Big)\Big],\end{aligned}$$ with $U_1(x_n) = \hbar\beta\int_{0}^{\hbar\beta}d\tau U(x(\tau).$ After doing integration over the environmental degrees of freedom $q_{j,n}$, I obtain $$\begin{aligned}
S^{E}& = & \frac{1}{\hbar\beta}\sum_{n=-\infty}^{+\infty}\frac{1}{2}(M\nu_n^{2}+\xi_{x,n})x_nx_{- n}\\ & & +\frac{1}{\hbar\beta}\sum_{n=-\infty}^{+\infty}\frac{1}{2}(M\nu_n^{2} + M\omega_y^{2}+\xi_{y,n})y_ny_{- n}\\
& & +U_1(x_n) + \frac{1}{\hbar\beta}\sum_{n=-\infty}^{+\infty}M\omega_cx_ny_{- n}\nu_{- n},\end{aligned}$$ with $ \xi_{l,n} = \frac{1}{\pi}\int_{0}^{\infty}d\omega \frac{J_{l}(\omega)}{\omega}\frac{2\nu_n^{2}}{\nu_n^{2}+\omega^{2}};\ \ \ \ \ l=x,y.$ Here we are considering tunneling along x-direction, hence I can integrate out the y variables. After doing the integration over y variables I obtain the effective one dimensional action as follows: $$\begin{aligned}
S^{E}& = & \frac{1}{\hbar\beta}\sum_{n=-\infty}^{+\infty}\frac{1}{2}(M\nu_n^{2}+\xi_{x,n})x_nx_{- n} + U_1(x_n)\\ \nonumber & & -\frac{1}{\hbar\beta}\sum_{n=-\infty}^{+\infty}\frac{(M\omega_c)^{2}\nu_n\nu_{- n}}{M\nu_n^2+M\omega_y^{2}+\xi_{y,n}}x_nx_{- n}\end{aligned}$$ The above effective 1D action can be rewritten as shown below: $$\begin{aligned}
S_{eff}^{E}[x(\tau)] & = & \int_{0}^{\hbar\beta}d\tau \Big[\frac{1}{2}M\dot{x}^{2}(\tau)+U_1(x(\tau))\Big]\\ \nonumber & &
+\frac{1}{2}\int_{0}^{\hbar\beta}d\tau\int_{0}^{\hbar\beta}d\tau^{\prime}\Big[{\cal K}(\tau-\tau^{\prime})+g(\tau-\tau^{\prime})\Big]\\ \nonumber & &\ \ \ \ \
\Big(x(\tau)-x(\tau^{\prime})\Big)^{2},\end{aligned}$$ where the normal damping kernel is given by $${\cal K}(\tau) = \frac{1}{\pi}\int_{0}d\omega J(\omega) \frac{cosh[\omega(\frac{\hbar\beta}{2}-\tau)]}{sinh[\frac{\omega\hbar\beta}{2}]},$$ and the anomalous damping kernel is given by $$\begin{aligned}
g(\tau) & = & -\frac{1}{\hbar\beta}\sum_{n=-\infty}^{+\infty}\frac{(M\omega_c)^{2}\nu_n^{2}}{M\nu_n^2+M\omega_y^{2}+\xi_{y,n}}e^{i\nu_n\tau}\\ \nonumber & = &
-\frac{1}{\hbar\beta}(M\omega_c)^{2}\Big[\sum_{n=-\infty}^{+\infty}e^{i\nu_n\tau}\\ \nonumber & & -\sum_{n=-\infty}^{+\infty}\frac{M\omega_y^{2}+\xi_{y,n}}{M\nu_n^2+M\omega_y^{2}+\xi_{y,n}}e^{i\nu_n\tau}\Big].\end{aligned}$$ The first term is the delta function and has no contribution to the semiclassical action, thus finally I obtain $$g(\tau) = \frac{1}{\hbar\beta}(M\omega_c)^{2}\sum_{n=-\infty}^{+\infty}\frac{M\omega_y^{2}+\xi_{y,n}}{M\nu_n^2+M\omega_y^{2}+\xi_{y,n}}e^{i\nu_n\tau}.$$ To derive the normal damping kernel I use the identity [@weiss] $\sum_n\frac{\nu_n^2}{\nu_n^2+\omega^2}e^{i\nu_n\tau}=\frac{\hbar\beta}{2}\frac{cosh[\omega(\frac{\hbar\beta}{2}-\tau)]}{sinh[\frac{\omega\hbar\beta}{2}]}.$ Thus we have obtained an effective one dimensional problem. The physics due to the influence of the dissipative environment and magnetic field is underlined in the normal and anomalous damping kernel respectively. Now I shall discuss different limiting cases.
[(i)****]{}
In this limiting condition the normal damping kernel ${\cal K}(\tau) = 0$ and the anomalous damping kernel $g(\tau) = \frac{(M\omega_c)^2}{\hbar\beta}$. Thus the effective Euclidean action becomes $$\begin{aligned}
S_{eff}^{E}& = &\int_{0}^{\hbar\beta}d\tau\Big[\frac{1}{2}M\dot{x}^2(\tau)+U_1(x(\tau))\Big]\\ \nonumber & & +\frac{(M\omega_c)^2}{\hbar\beta}\int_{0}^{\hbar\beta}d\tau\int_{0}^{\hbar\beta}d\tau^{\prime}[x(\tau)-x(\tau^{\prime})]^2.\end{aligned}$$ $S_{eff}^{E}$ depends on magnetic field through $\omega_c$ and this dependence is proportional to $B^2$. Thus for a strong magnetic field the tunneling rate \[Eq. (30)\] vanishes at very low temperature. At zero temperature the effective action Eq. (40) can be written as $$S_{eff}^{E} = \int_{0}^{\hbar\beta}d\tau\Big[\frac{1}{2}M\dot{x}^2(\tau)+U_1(x(\tau))+\frac{1}{2}M\omega_c^{2}x^{2}(\tau)\Big].$$ Equation (41) shows that the magnetic field renormalizes the potential $U_1(x(\tau)$ such that the local minima at $ x=0$ is now more stable. Because the renormalized potential is now a double well potential and thus the second local minima may be lower than at $ x = 0$.
[(ii)****]{}
Since there is no dissipation, hence the normal damping kernel ${\cal K}(\tau)=0$ and the anomalous damping kernel is given by $$\begin{aligned}
g(\tau) & = & -\frac{1}{\hbar\beta}(M\omega_c)^{2}\sum_{n=-\infty}^{+\infty}\frac{\nu_n^{2}}{M\nu_n^2+M\omega_y^{2}+\xi_{y,n}}e^{i\nu_n\tau}\\ & = &-\frac{1}{\hbar\beta}(M\omega_c)^{2}\frac{1}{M}\sum_{n=-\infty}^{+\infty}\frac{\nu_n^{2}}{\nu_n^2+\omega_y^{2}}\\ & = & -\frac{1}{\hbar\beta}(M\omega_c)^{2}\frac{1}{M}\frac{\hbar\beta}{2}\frac{cosh[\omega_y(\frac{\hbar\beta}{2}-\tau)]}{sinh[\frac{\omega_y\hbar\beta}{2}]} \end{aligned}$$ Thus the effective euclidean action becomes $$\begin{aligned}
S_{eff}^{E}& = &\int_{0}^{\hbar\beta}d\tau \Big[\frac{1}{2}M\dot{x}^2(\tau)+U_1(x(\tau))\Big]\\ \nonumber & & -\frac{1}{4M}(M\omega_c)^{2}\int_{0}^{\hbar\beta}d\tau\int_{0}^{\hbar\beta}d\tau^{\prime}\frac{cosh[\omega_y(\frac{\hbar\beta}{2}-|\tau-\tau^{\prime}|)]}{sinh[\frac{\omega_y\hbar\beta}{2}]}\\ \nonumber & & \ \ \ \ \ [x(\tau)-x(\tau^{\prime})]^2\end{aligned}$$ Now one can easily be able to understand that the tunneling rate is finite for any value of magnetic field. Thus the impurity acts in opposite way as that of magnetic field. So impurity suppresses the effect of magnetic field and it actually enhances the tunneling rate.
[(i)****]{}
I set the impurity potential i.e. $\omega_y^{2} = 0$. We know that the quantum tunneling predominates at very low temperature, i.e. at very large imaginary time limit $(\hbar\beta\rightarrow\infty)$. So we have to take large time limit values of the normal and anomalous damping kernel [@pao1; @pao2]. In the large $\tau$ limit the normal damping kernel is given by [@pao1; @pao2] ${\cal K}(\tau) = \frac{1}{\pi}\eta \frac{1}{\tau^{s+1}}$ and the anomalous damping kernel becomes $g(\tau) = \frac{(M\omega_c)^2}{2\pi M}\frac{1}{\eta^{\prime}}\frac{1}{\tau^{2-s+1}}$, where $\eta^{\prime} = \frac{\eta}{2M}\frac{2}{\pi}\int_0^{\infty}dz\frac{z^{s-1}}{z^2+1}$. Thus the effective action for the ohmic damping (s=1) case becomes $$\begin{aligned}
S_{eff}^{E}& = &\int_{0}^{\infty}d\tau\Big[\frac{1}{2}M\dot{x}^2(\tau)+U_1(x(\tau))\Big]\\ \nonumber & &
+\frac{1}{2}\int_{0}^{\infty}d\tau\int_{0}^{\infty}d\tau^{\prime}\Big[\frac{1}{\pi}\eta\frac{1}{|(\tau-\tau^{\prime})|^{1+1}}\\ \nonumber & & +\frac{(M\omega_c)^2}{2\pi M}\frac{1}{|(\tau-\tau^{\prime})|^{2}}\frac{1}{\frac{\eta}{2M}\frac{2}{\pi}\int_0^{\infty}\frac{dz}{z^2+1}}\Big][x(\tau)-x(\tau^{\prime})]^2
, \end{aligned}$$ or, $$\begin{aligned}
S_{eff}^{E} & = & \int_{0}^{\infty}d\tau\Big[\frac{1}{2}M\dot{x}^2(\tau)+U_1(x(\tau))\Big]\\ \nonumber & & +\frac{\Big(\eta+\frac{(M\omega_c)^{2}}{\eta}\Big)}{2\pi}\int_{0}^{\infty}d\tau\int_{0}^{\infty}d\tau^{\prime}\frac{1}{|(\tau-\tau^{\prime})|^{2}}\\& &[x(\tau)-x(\tau^{\prime})]^2.\end{aligned}$$ Finally I obtain $$\begin{aligned}
S_{eff}^{E} & = & \int_{0}^{\infty}d\tau\Big[\frac{1}{2}M\dot{x}^2(\tau)+U_1(x(\tau))\Big]\\ \nonumber & &+ \frac{\eta_{eff}}{2\pi}\int_{0}^{\infty}d\tau\int_{0}^{\infty}d\tau^{\prime}\frac{1}{|\tau-\tau^{\prime}|^2}[x(\tau)-x(\tau^{\prime})]^2,\end{aligned}$$ where $\eta_{eff} = \eta + \frac{(M\omega_c)^2}{\eta}$. From the Eq. (43) it is well understood that the effect of dissipation is to suppress the effect of magnetic field. Because of the factor $\frac{1}{|\tau-\tau^{\prime}|^2}$ we should get finite tunneling at any amount of magnetic field strength. So dissipative environment helps in tunneling of 2DEG in presence of magnetic field.
[(i)****]{}
Since both impurity and dissipation helps in tunneling of 2DEG in presence of magnetic field, so the total effect of them will be the same. The effective action is $$\begin{aligned}
S_{eff}^{E} & = & \int_{0}^{\hbar\beta}d\tau\Big[\frac{1}{2}M\dot{x}^2(\tau)+U_1(x(\tau))\Big]\\ \nonumber & &+\frac{1}{2}\int_{0}^{\hbar\beta}d\tau\int_{0}^{\hbar\beta}d\tau^{\prime}[{\cal K}(|\tau-\tau^{\prime}|)+g(\tau-\tau^{\prime})]\\ & & [x(\tau)-x(\tau^{\prime})]^2.\end{aligned}$$ Now in case of strong impurity the anomalous damping kernel is negligible compared to the normal damping kernel, because in this case the anomalous superohmic kernel decays rapidly compare to ohmic normal damping kernel. Hence the effective action becomes: $$\begin{aligned}
S_{eff}^{E} & = & \int_{0}^{\hbar\beta}d\tau\Big[\frac{1}{2}M\dot{x}^2(\tau)+U_1(x(\tau))\Big]\\ \nonumber & & + \frac{\eta}{2\pi}\int_{0}^{\infty}d\tau\int_{0}^{\infty}d\tau^{\prime}\frac{1}{|\tau-\tau^{\prime}|^2}[x(\tau)-x(\tau^{\prime})]^2.\end{aligned}$$ Equation (44) suggests that dirty sample enhances the tunneling rate of 2DEG in presence of magnetic field. On the other hand for strong damping case one can neglect the superohmic normal damping kernel compared to ohmic anomalous damping kernel. Thus the effective action now reads as: $$\begin{aligned}
S_{eff}^{E} & = & \int_{0}^{\hbar\beta}d\tau\Big[\frac{1}{2}M\dot{x}^2(\tau)+U_1(x(\tau))\Big]\\ \nonumber & & + \frac{(M\omega_c)^2}{2\pi\eta}\int_{0}^{\infty}d\tau\int_{0}^{\infty}d\tau^{\prime}\frac{1}{|\tau-\tau^{\prime}|^2}[x(\tau)-x(\tau^{\prime})]^2.\end{aligned}$$ From the above expression it is well understood that dissipation opposes the magnetic field effect.
[(i)****]{}
Under very strong magnetic field the kinetic energy of the electron gas freezes. Thus one obtain the effective action as follows $$\begin{aligned}
S_{eff}^{E} & = & \int_{0}^{\hbar\beta}d\tau\Big[iM\omega_c(\dot{x}(\tau)y(\tau)) + U_1(x(\tau))+\frac{1}{2}M\omega_y^2 y^2]\\ \nonumber & & +\frac{1}{2}\int_{0}^{\infty}d\tau\int_{0}^{\infty}d\tau^{\prime}{\cal K}(\tau -\tau^{\prime})[x(\tau)-x(\tau^{\prime})]^2.\end{aligned}$$ Following Jain [*et al*]{} [@jain] I obtain $$\begin{aligned}
S_{cl}^{E} & = & -i\int_{0}^{x_t}dx iM\omega_c(\dot{x}(\tau)y(\tau))\\ & & + \frac{1}{2}\int_{0}^{\infty}d\tau\int_{0}^{\infty}d\tau^{\prime}\frac{1}{|\tau-\tau^{\prime}|^2}[x_{cl}(\tau)-x_{cl}(\tau^{\prime})]^2.\end{aligned}$$ Finally I obtain $$\begin{aligned}
S_{cl} & = & M\omega_c\int_{0}^{x_t}dx \Big[\frac{2U_1(x)}{M\omega_y^2}\Big]^{\frac{1}{2}} \\ \nonumber & & + \frac{1}{2}\int_{0}^{\infty}d\tau\int_{0}^{\infty}d\tau^{\prime}\frac{1}{|\tau-\tau^{\prime}|^2}[x_{cl}(\tau)-x_{cl}(\tau^{\prime})]^2.\end{aligned}$$ Thus as we increase the impurity potential along y-direction $S_{cl}$ decreases and the tunneling rate of the 2DEG along x-direction increases.\
CONCLUSION AND SUMMARY
======================
In this paper I have presented the imaginary time functional integral approach to reaction rates calculation. I have provided a complete description of the decay of a metastable state extending from the classical regime to quantum region. I have covered a wide range of temperature region — extending from the high temperature thermally activated region to a very low temperature quantum tunneling regime. It has been seen that well above the crossover temperature $T_0$, the decay rate follows the famous Arrhenius law. Here the preexponential factor is affected by the damping and magnetic field through the renormalized frequency $\omega_R$. Now as we approach to the crossover temperature the quantum correction factor to the decay rate becomes very much important. Now well below the crossover temperature the system decays predominantly by quantum tunneling process. I have shown that at zero temperature the tunneling rate of 2DEG from the metastable state vanishes in a strong magnetic field and in absence of both dissipation and impurity. The effect of impurity potential and the dissipation have been discussed in great details in Sec. IV. Both the impurity and dissipation subdue the effect of magnetic field and enhances the tunneling rate. I have discussed this tunneling rate in context of a very relevant cubic potential. So my results are also applicable to the physically interesting problem of quantum tunneling in SQUID or current-biased Josephson junctions.
Acknowledgments {#acknowledgments .unnumbered}
===============
I wish to express my gratitude to Prof. Sushanta Dattagupta for his constant encouragement. Financial support from CSIR is gratefully acknowledged.\
[99]{} A. O. Caldeira and A. J. Leggett, Ann. Phys. (NY),[*149, 374*]{}(1983); [*153, 445*]{} (1984). A. J. Leggett, S. Chakravarty, A. T. Dorsey, M. P. A. Fisher, A. Garg, W. Zwerger, Rev. Mod. Phys. [*59, 1*]{} (1987); [*67, 725(E)*]{}(1995). D. De Vault, Q. Rev. Biophys. [*13, 387*]{} (1980). Tunneling in Biological Systems, edited by B. Chance [*et al*]{}(Academic, New York, 1977). R. A. Marcus, J. Chem. Phys. [*24, 966*]{} (1956);[*43, 679*]{} (1963). G. F. Saville, J. M. Goodkind and P. M. Platzman, Phys. Rev. Lett. [*70, 1517*]{} (1993). P. Ao, Ph.D. Thesis, University of Illinois at Urbana Champaign, 1991. J. K. Jain, Steven Kivelson; Phys. Rev B [*37 4111*]{}(1988); S. Chakravarty, A. J. Leggett, Phys. Rev. Lett. [*53, 5*]{} (1984). Ping Ao and David J. Thouless, Phys. Rev. Lett [*70, 2158*]{} (1993); D. J. Thouless, P. Ao and Q. Niu, Physica A [*200, 42*]{} (1993); Ping Ao, arXiv: Cond-mat/9609021 V1. Physics of Low-Dimensional Semiconductor Structures, edited by P. Butcher, N. H. March, and M. P. Tosi (Plenum, New York, 1993). Mesoscopic Phenomena in Solids, edited by B. L. Altershuler, P. A. Lee, and R. A. Webb (North-Holland, Amsterdam,1991). L. Zheng and A. H. MacDonald, Phys. Rev. B [*47, 10619*]{} (1993); S. He, P. M. Platzman, and B. I. Halperin, Phys. Rev. Lett. [*71, 777*]{} (1993). Hermann Grabert, Peter Olschowski and Ulrich Weiss, Phys. Rev. B [*36, 1931*]{} (1987). Hermann Grabert, Ulrich Weiss and Peter Hanggi, Phys. Rev. Lett. [*52, 2193*]{} (1984). Hermann Grabert, Peter Olschowski and Ulrich Weiss, Phys. Rev. B [*32, 3348*]{} (1985). J. S. Langer, Ann. Phys. (NY) [*41, 108*]{} (1967). W. H. Miller, J. Chem. Phys. [*62, 1899*]{} (1975); Adv. Chem. Phys. [*25, 69*]{} (1974); M. Stone, Phys. Lett. [*67B, 186*]{} (1977); C. G. Callan and S. Coleman, Phys. Rev. D [*16, 1762*]{} (1977); S. Coleman, in The ways of subnuclear Physics, edited by A. Zichichi (Plenum, New York, 1979). A. O. Caldeira and A. J. Leggett , Phys. Rev. Lett. [*46, 211*]{} (1981); A. J. Leggett, S. Chakravarty, A. T. Dorsey, M. P. A. Fisher, A. Garg, and W. Zwerger Rev. Mod. Phys. [*59, 1*]{} (1987): H. Grabert and Ulrich Weiss, Phys. Rev. Lett. [54, 1605]{} (1985); U. Weiss, H. Grabert and S. Linkwitz,J. Low Temp. Phys. [*68, 213*]{} (1987); H. Grabert and U. Weiss, Phys. Lett [*108A, 63*]{} (1985). A. I. Larkin and Yu. N. Ovchinnikov, So. Phys.-JETP Lett. [*37, 382*]{} (1983) R. P. Feynmann, F. L. Vernon,Jr.: Ann. Phys (N. Y.) [*24, 118*]{} (1963). H. Grabert, U. Weiss and P. Hanggi, Z. Phys. B [*56, 171*]{} (1984). I. K. Affleck, Phys. Rev. Lett. [*46, 388*]{} (1981). P. G. Wolynes, Phys. Rev. Lett. [*47, 968*]{} (1981). Quantum Dissipative Systems, edited by Ulrich Weiss (World Scientific, Singapore, 1999).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'D. Batcheldor'
bibliography:
- 'bbl.bib'
title: ' The $M_\bullet - \sigma_\ast$ Project '
---
Introduction
============
The discovery of the relationship between supermassive black hole (SMBH) mass ($M_\bullet$) and the stellar velocity dispersion ($\sigma_\ast$) of the host bulge [@2000ApJ...539L...9F; @2000ApJ...539L..13G] signaled a new and fundamental understanding on the nature of black hole and galaxy formation and evolution; the $M_\bullet-\sigma_\ast$ relation intimately links the most basic characteristic of a SMBH with an underlying dynamical property of the surrounding galaxy. Consequently, the $M_\bullet-\sigma_\ast$ relation, that is typically described by $\log M_\bullet = \alpha + \beta\log(\sigma_\ast)$, has received an extraordinary amount of attention.
The $M_\bullet-\sigma_\ast$ relation is a member of a family of scaling relations all linking $M_\bullet$ to various properties of the host bulge, and possibly the entire gravitational mass of the host galaxy. For a comprehensive review of all the scaling relations see [@2005SSRv..116..523F]. Despite all these other relations, $M_\bullet-\sigma_\ast$ has traditionally been the preferred as, at its time of discovery, it showed a remarkably small scatter; a smaller scatter than for the other scaling relations and a scatter consistent with the measurement errors alone. Since then, however, [@2003ApJ...589L..21M] and [@2007MNRAS.379..711G] have shown that with enough care, scaling relation scatters can be comparable especially when using near-infrared luminosities. Regardless of which scaling relation is “best”, it is important to fully investigate them all as the intrinsic scatters themselves contain pivotal information on the precise inter-play between SMBHs and their hosts.
Reproducing the features of the $M_\bullet-\sigma_\ast$ relation has become an integral part of SMBH and galaxy formation and evolution models. The relation also has several other significant applications. For example, the $M_\bullet-\sigma_\ast$ relation may be extrapolated to smaller and larger values of $M_\bullet$, and be used to estimate $M_\bullet$ in systems where the black hole sphere of influence radius ($r_h$) cannot be resolved. It is therefore especially important to know the values of $\alpha$ and $\beta$, and their associated scatters, to high accuracy. To have doubt in either parameter is to introduce large uncertainties in extrapolated values of $M_\bullet$. This in turn has a fundamental impact on our understanding of both black hole and galaxy formation and evolution.
The $M_\bullet-\sigma_\ast$ Relation {#proj}
====================================
The $M_\bullet-\sigma_\ast$ relation has had many attempts to fit its slope, zero-point and scatter to the whole galaxy population [@2000ApJ...539L...9F; @2000ApJ...539L..13G; @2001ApJ...547..140M; @2002ApJ...574..740T] in addition to individual families of galaxies that potentially lie off the relation . So far, these individual fits have produced considerably different results ($\beta$ ranges from 3.8 to 4.9, for example). To demonstrate what affect these uncertainties have on the importance of the $M_\bullet-\sigma_\ast$ relation, we have used the direct $M_\bullet$ and $\sigma_\ast$ data compiled recently by [@2005SSRv..116..523F Table 2] to construct Figure \[fig:1\]. The fit of [@2002ApJ...574..740T] is also plotted as these values are typically used by many authors. Taking $\sigma_\ast=20{\rm{km~s^{-1}}}$ (a large globular cluster) the two slopes predict $\log M_\bullet=4.11M_\odot$ or $\log M_\bullet=3.36M_\odot$. Taking $\sigma_\ast=470{\rm{km~s^{-1}}}$ (a brightest cluster galaxy) the two slopes predict $\log M_\bullet=9.62M_\odot$ or $\log M_\bullet=10.02M_\odot$.
The aim of the $M_\bullet-\sigma_\ast$ Project is to produce the most reliable estimate of the $M_\bullet-\sigma_\ast$ relation to date. It will couple ground-based integral field spectroscopy (IFS) with high spatial and spectral resolution data, from the STIS archive, to estimate $M_\bullet$ in a large sample of galaxies using the latest three-integral, axisymmetric, orbit-based stellar dynamical models. The project will also examine the accurate determination of $\sigma_\ast$ within the host bulges. In addition, the project will identify and address several key issues that surround the uncertainties in determining $\alpha$ and $\beta$.
Estimating $M_\bullet$
----------------------
Four methods have been used to [*directly*]{} estimate the values of $M_\bullet$ defining the $M_\bullet-\sigma_\ast$ relation. Proper motions have been used in the Milky Way [@1997MNRAS.291..219G; @2005ApJ...620..744G] and $\rm H_20$ masers have been used to estimate $M_\bullet$ in NGC 4258 [@1995Natur.373..127M]. All other estimates have been made using gas and stellar dynamics. Gas dynamics [e.g., @1996ApJ...470..444F; @2001ApJ...549..915M], are relatively easy to model as long as the nuclear disk is completely dominated by the gravitational potential of the SMBH. Stellar dynamics [e.g., @1994MNRAS.270..271V; @2002MNRAS.335..517V; @2003ApJ...583...92G], does not suffer from this problem, but is a more complex method that requires a high sensitivity instrument. It is difficult to judge the consistency between estimates made by each method as very few galaxies have been modeled by more than one method.
Given a sufficiently sensitive instrument, and a sufficiently elaborate model, then stellar dynamics is the preferred method for determining $M_\bullet$. Indeed, to date 57% of direct $M_\bullet$ estimations have been made using stellar dynamics; it is the dominant method defining the $M_\bullet-\sigma_\ast$ relation. [@2004ApJ...602...66V] provide a detailed description of an orbit-based model designed to recover the gravitational potential of a system using stellar kinematics. They carry out detailed analytical testing of this model as applied to a dataset from M32. [@2005ApJ...628..137V] apply the same model to constrain $M_\bullet$ in NGC 205. It is found that a large range in $M_\bullet$ gives equally acceptable fits. The degeneracy in the solution is a function of the number of orbits used in the models and the number of constraining data points.
As M32 and NGC 205 are the only datasets for which $\chi^2$ space in stellar dynamical models has been explored in detail, it is clear that the dominant mass estimates within the current $M_\bullet-\sigma_\ast$ relation may include un-investigated issues that could have a fundamental impact on our understanding of the relation itself. The $M_\bullet-\sigma_\ast$ project will address these issues in two key ways. Firstly, it will use a wide range of constraining data points by employing IFS. Secondly, it will serially reduce a large family of dynamical models, that have a wide range of total orbit numbers, on a high performance computing cluster with a large node number.
It has been repeated by many authors [@2002ApJ...578...90F; @2003ApJ...589L..21M; @2004ApJ...602...66V] that resolving $r_h$ (given by $GM_{\bullet}/{\sigma_\ast^2}$) is a minimum requirement for detecting a SMBH. If $r_h$ is un-resolved then estimates of $M_\bullet$ can be tarnished by uncertain contributions from stellar masses. The derivation of $r_h$ was first attempted by [@1972ApJ...178..371P] based on the ad hoc assumption that in a steady state the stellar density is exponentially linked to the total gravitational potential and that $\sigma_\ast$ is constant. The black hole removes stars at small radii (via accretion or ejection, for example) and disrupts this equilibrium. The steady state is restored via an increase in $\sigma_\ast$ of stars brought in from larger radii.
Although the requirement to resolve $r_h$ has been thoroughly argued it must be noted that this creates a somewhat degenerate problem; you are forced to make an assumption about SMBH demographics (i.e., that they follow the $M_\bullet-\sigma_\ast$ relation) in order to deduce whether you can detect that SMBH in the first place. Furthermore, you are using the very parameters in the $M_\bullet-\sigma_\ast$ relation in the calculation of $r_h$. Nevertheless, with the assumption that SMBHs do follow the $M_\bullet-\sigma_\ast$ relation, we can plot $r_h$ within the $M_\bullet-\sigma_\ast$ plane. In addition, we can determine the ability of [*HST*]{}, from which most $M_\bullet$ estimates are derived, to resolve $r_h$ at a specific distance (Fig \[fig:2\]).
Two key points are associated with Figure \[fig:2\]. Firstly, it does not include the sensitivity, and therefore the SMBH detection efficiency, of the telescope. The relatively small aperture of [*HST*]{} ensures that many STIS orbits are required before a sufficiently large continuum signal-to-noise is reached for stellar dynamical studies. Secondly, it shows that even at 16 Mpc [*HST*]{} is unable to explore a significant area of the $M_\bullet-\sigma_\ast$ plane (below the dashed lines). If there were a population of bona-fide under-massive black holes, as suggested by [@2007ApJ...663L...5V] for example, [*HST*]{} would not be able to detect them; they would not be modeled and the $M_\bullet-\sigma_\ast$ relation would still remain to appear universal. In short, [*HST*]{} has a SMBH “blind spot”. To unequivocally define the $M_\bullet-\sigma_\ast$ relation the entire $M_\bullet-\sigma_\ast$ plane needs to be detectable over a large enough distance to include a large population of host galaxies.
Estimating $\sigma_\ast$
------------------------
One of the problems associated with the fitting of the $M_\bullet-\sigma_\ast$ relation concerns the values of $\sigma_\ast$ used. [@2000ApJ...539L...9F] transformed $\sigma_\ast$ to an aperture of $r_e/8$, where $r_e$ is the host effective radius, while [@2000ApJ...539L..13G] use a value weighted by the host luminosity. The use of different values is understandable considering there is, as yet, no clearly defined aperture size through which $\sigma_\ast$ should be determined.
When considering the nature of a LOSVD, the second order of which is usually described as $\sigma_\ast$, one can see that the size and nature of the aperture through which the LOSVD is collected has a fundamental bearing on its observed shape. In addition to the intrinsic variation of $\sigma_\ast$ with aperture size [@2005ApJS..160...76B], a velocity field across the aperture will rotationally broaden the LOSVD as a function of the aperture diameter. This will produce an over estimate of the intrinsic $\sigma_\ast$. In the case of an axisymmetric velocity field, the aperture shape and orientation will also have a bearing; a different value of $\sigma_\ast$ will be derived from, for example, a long-slit aligned along the major or minor axis of rotation. IFS can be used to address these concerns.
Summary
=======
The $M_\bullet-\sigma_\ast$ relation is one of the most valuable correlations in modern galaxy and SMBH formation and evolutionary models. There are, however, concerns in the ways both $\sigma_\ast$ and many $M_\bullet$ estimates have been determined. The latest stellar dynamical models have shown that great care must be used to fully explore $\chi^2$ space around $M_\bullet$. It is essential that $r_h$ is fully resolved, that there are large number of high quality constraints, and that a large variety of orbit families are employed in the models. The most appropriate values of $\sigma_\ast$ may not have been determined as it has yet to be clearly defined. STIS long-slit, and integral field spectroscopy can be used to address all these issues and will lead to the most accurate form of the $M_\bullet-\sigma_\ast$ relation to date.
The author gratefully acknowledges useful discussions with D. J. Axon, A. Marconi, D. Merritt and M. Valluri. Support for Proposal number HST-AR-10935.01 was provided by NASA through a grant from the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Incorporated, under NASA contract NAS5-26555.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We discuss the relationship between the recurrence coefficients of orthogonal polynomials with respect to a generalized Freud weight $$\ww{x}=|x|^{2\la+1}\exp\left(-x^4+tx^2\right),\qquad x\in\R,$$ with parameters $\la>-1$ and $t\in\R$, and classical solutions of the fourth equation. We show that the coefficients in these recurrence relations can be expressed in terms of Wronskians of parabolic cylinder functions that arise in the description of special function solutions of the fourth equation. Further we derive a second-order linear ordinary differential equation and a differential-difference equation satisfied by the generalized Freud polynomials.'
author:
- |
Peter A. Clarkson\
School of Mathematics, Statistics and Actuarial Science,\
University of Kent, Canterbury, CT2 7NF, UK\
`[email protected]`\
Kerstin Jordaan and Abey Kelil\
Department of Mathematics and Applied Mathematics,\
University of Pretoria, Pretoria, 0002, South Africa\
`[email protected]`, `[email protected]`
title: A Generalized Freud Weight
---
Introduction
============
Let $\mu$ be a positive Borel measure defined on the real line for which all the moments $$\mu_n=\int_{\R} x^n\, d\mu(x), \quad n\in\N,$$ exist. The Hilbert space $L^2(\mu)$ contains the set of polynomials and hence Gram-Schmidt orthogonalization applied to the set $\{1,x,x^2,\dots\}$ yields a set of monic orthogonal polynomials on the real line defined by $$\int_{\R} P_n(x) P_m(x) \,d\mu(x)=h_n\delta_{mn}, \qquad m,n\in\N,$$ with $h_n>0$, $\delta_{mn}$ the Kronecker delta and $P_n(x)$ is a polynomial of exact degree $n$.
A family $\{P_n(x)\}_{n=0}^{\infty}$ of monic orthogonal polynomials satisfies a three-term recurrence relation of the form $$\label{3trr1} xP_n(x) = P_{n+1}(x) + \a_n P_n(x) + \b_n P_{n-1}(x), \quad n\in\N,$$ with $P_{-1}(x)=0$ and $P_0(x)=1$. Respectively multiplying by $P_n(x)$ and $P_{n-1}(x)$ and integrating gives $$\begin{aligned}
\a_n=\frac{1}{h_n}\int_{\R} x \,P_n^2(x)\,d\mu(x),\qquad
\b_n&=\frac{1}{h_{n-1}}\int_{\R} x\,P_n(x)P_{n-1}(x)\,d\mu(x).\end{aligned}$$ The converse statement, known as the spectral theorem for orthogonal polynomials, is often attributed to Favard [@refFavard] but was seemingly discovered independently, around the same time, by both Shohat [@refShohat36; @refShohat38] and Natanson [@refNatanson]. It is interesting to note that the result can be traced back to earlier work on continued fractions with a rudimentary form given already in 1894 by Stieltjes [@refStieltjes], see [@refChihara], also [@refMarcellan; @refwalter]. The result also appears in the books by Stone [@refStone] and Wintner [@refWintner], see [@refIsmail]. A modern proof of the result is given by Beardon [@refBeardon].
If a family of polynomials satisfies a three-term recurrence relation of the form $$xP_n(x) = P_{n+1}(x) + \a_n P_n(x) + \b_n P_{n-1}(x),$$ with initial conditions $P_{-1}(x)=0$ and $P_0(x)=1$, where $\a_n,\b_n\in\R$, then there exists a measure $\mu$ on the real line such that these polynomials are monic orthogonal polynomials satisfying $$\int_{\R} P_n(x) P_m(x)\,d\mu(x)=h_n\delta_{mn}, \qquad m,n \in\N,$$ with $h_n>0$.
See, for example, [@refChihara; @refIsmail].
Two important problems arise in the study of orthogonal polynomials.
1. *Direct problem:* Given a measure $\mu(x)$, what can be deduced about the recurrence coefficients $\{\a_n,\b_n\}$, $n\in\N$?
2. *Inverse problem:* Given recurrence coefficients $\{\a_n,\b_n\}$, $n\in\N$, what can be deduced about the orthogonality measure $\mu(x)$?
Given an orthogonality measure $\mu(x)$, several characteristic properties of the sequence $\{P_n(x)\}_{n=0}^{\infty}$ are determined by the nature of the measure. Extracting this information from the measure is an extension of the direct problem and one of the interesting and challenging problems in the study of systems of orthogonal polynomials. Properties typically studied include the Hankel determinants, the coefficients of the three-term recurrence relation, the coefficients of the differential-difference equation and the differential equation satisfied by the polynomials. For instance, the recurrence coefficients can be expressed in terms of Hankel determinants comprising the moments of the orthogonality measure. We refer to [@refChihara; @refIsmail; @refSzego] for further information about orthogonal polynomials.
For classical orthogonal polynomials, namely the Hermite, Laguerre and Jacobi polynomials, the measure $\mu(x)$ is absolutely continuous and can be expressed in terms of a weight function $w(x)$ which is non-negative, with support on some interval $[a,b] \in\R$ (where $a=-\infty$ and $b=\infty$ are allowed), i.e. $d\mu(x)=w(x)\,dx$. The properties that these orthogonal polynomials satisfy are well known and include the fact that (cf. [@refIsmail]):
- their derivatives also form orthogonal polynomial sets;
- they satisfy a Rodrigues’ type formula $$P_n(x)=\frac{1}{\kappa_n w(x)}{\frac{\d^{n}{}}{{\d{x}}^{n}}}\left\{w(x)\sigma^n(x)\right\},$$ where $w(x)$ is the weight function, $\sigma(x)$ is a polynomial in $x$ independent of $n$ and $\kappa_n$ a constant;
- they satisfy a non-linear equation of the form $$\begin{split}\sigma(x){\frac{\d^{}{}}{{\d{x}}^{}}}\big[{P_n(x)P_{n-1}(x)}\big]=(a_n x+b_n) &P_n(x) P_{n-1}(x)+c_n P_n^2(x) +d_n P_{n-1}^2 (x),\end{split}$$ where $a_n$, $b_n$, $c_n$ and $d_n$ are independent of $x$;
- they satisfy a second order linear differential equation of the Sturm-Liouville type $$\sigma(x){\frac{\d^{2}{P_n}}{{\d{x}}^{2}}}+\tau(x){\frac{\d^{}{P_n}}{{\d{x}}^{}}}=\la_nP_n(x),$$ where $\sigma(x)$ is a polynomial of degree $\leq2$, $\tau(x)$ is a linear polynomial, both independent of $n$, and $\la_n$ is independent of $x$;
- they satisfy the differential-difference equation $$\pi(x){\frac{\d^{}{P_n}}{{\d{x}}^{}}}=(a_n x+b_n)P_{n}(x)+c_nP_{n-1}(x).\label{d}$$
The converse, that any polynomial set which satisfies any one of the above properties must necessarily be one of the classical orthogonal polynomial sets, also holds. In particular, Al-Salam and Chihara [@ASC] showed that orthogonal polynomial sets satisfying must be either Hermite, Laguerre or Jacobi polynomials.
Askey raised the more general question of what orthogonal polynomial sets have the property that their derivatives satisfy a relation of the form $$\pi(x){\frac{\d^{}{P_n}}{{\d{x}}^{}}}=\sum_{k=n-t}^{n+s}a_{n,k}P_k(x).$$ This problem was solved by Shohat [@refShohat39] and later, independently, by Freud [@refFreud], as well as Bonan and Nevai [@refBNevai]. Maroni [@Maroni1; @Maroni2] stated the problem in a different way, trying to find all orthogonal polynomial sets whose derivatives are quasi-orthogonal, and called such orthogonal polynomial sets semi-classical.
A useful characterization of classical orthogonal polynomials is the Pearson equation $$\label{eq:Pearson}{\frac{\d^{}{}}{{\d{x}}^{}}}[\sigma(x)w(x)]=\tau(x)w(x),$$ where $\sigma(x)$ and $\tau(x)$ are polynomials with deg$(\sigma)\leq 2$ and deg$(\tau)=1$.
Semi-classical orthogonal polynomials are defined as orthogonal polynomials for which the weight function satisfies a Pearson equation with deg$(\sigma)\geq 2$ or deg$(\tau)\neq 1$; see Hendriksen and van Rossum [@Hendriksen].
The relationship between semi-classical orthogonal polynomials and integrable equations dates back to Shohat [@refShohat39] and Freud [@refFreud], see also Bonan and Nevai [@refBNevai]. However it was not until the work of Fokas, Its and Kitaev [@refFIKa; @refFIKb] that these equations were identified as discrete equations. The relationship between semi-classical orthogonal polynomials and the equations was discussed by Magnus [@refMagnus95] who, for example, showed that the coefficients in the three-term recurrence relation for the Freud weight [@refFreud] $$\ww{x}=\exp\left(-x^4+tx^2\right),\qquad x\in\R,\label{Freud}$$ with $t\in\R$ a parameter, can be expressed in terms of simultaneous solutions of the discrete equation $$\label{eq:dPIf}
q_n(q_{n-1}+q_n+q_{n+1})+2tq_n=n,$$ which is [discrete ]{} (), as shown by Bonan and Nevai [@refBNevai], and the differential equation $$\label{eq:PIVf}
{\frac{\d^{2}{q_n}}{{\d{z}}^{2}}}= \frac{1}{2q_n}\left({\frac{\d^{}{q_n}}{{\d{z}}^{}}}\right)^{2} + \frac{3}{2}q_n^3 + 4z q_n^2 + 2(z^2 +\tfrac12n)q_n- \frac{n^2}{2q_n},$$ which is a special case of the fourth equation – see equation below – with $n\in\N$. This connection between the recurrence coefficients for the Freud weight and simultaneous solutions of and is due to Kitaev, see [@refFIKa; @refFIZ]. The discrete equation is also known as the “Freud equation" or the “string equation".
It is known (cf. [@refIsmail]) that polynomials orthogonal with respect to exponential weights $w(x)=\exp\{-Q(x)\}$ on $\R$ for polynomial $Q(x)$ satisfy structural relations of the form $${\frac{\d^{}{P_n}}{{\d{x}}^{}}}(x) = A_n(x) P_{n-1}(x) + B_n(x) P_n(x).$$ Such structural relations and the three-term recurrence relation reveal that the orthogonal polynomials satisfy a second order differential equation.
It had been generally accepted that explicit expressions for the associated coefficients in the three-term recurrence relation and orthogonal polynomials were nonexistent for weights such as the Freud weight . To quote from the [*[Digital Library of Mathematical Functions]{}*]{} [@DLMF §18.32]:
> “A [*[Freud weight]{}*]{} is a weight function of the form $$w(x)=\exp\{-Q(x)\},\quad -\infty<x<\infty,$$ where $Q(x)$ is real, even, nonnegative, and continuously differentiable. Of special interest are the cases $Q(x)=x^{2m}$, $m=1,2,\dots$. No explicit expressions for the corresponding OP’s are available. However, for asymptotic approximations in terms of elementary functions for the OP’s, and also for their largest zeros, see Levin and Lubinsky [@LevinLubinsky] and Nevai [@Nevai]. For a uniform asymptotic expansion in terms of Airy functions for the OP’s in the case $Q(x)=x^4$ see Bo and Wong [@refBoWong]. For asymptotic approximations to OP’s that correspond to Freud weights with more general functions $Q(x)$ see Deift *et al*. [@refDKMVZa; @refDKMVZb], Bleher and Its [@BleherIts], and Kriecherbauer and McLaughlin [@refKrMcL].”
In [@refCJ], the direct problem was studied for semi-classical Laguerre polynomials orthogonal with respect to the semi-classical Laguerre weight $$\label{Lw}
\ww{x} = x^{\la}\, \exp (-x^2 +tx), \qquad x\in \R^{+},\quad \la>-1,$$ with $t\in\R$ a parameter, and it was shown that the coefficients in the three-term recurrence relation of these polynomials can be explicitly expressed in terms of Wronskians of parabolic cylinder functions which also arise in the description of special function solutions of the fourth equation () $$\label{eq:PIV}
{\frac{\d^{2}{q}}{{\d{z}}^{2}}}= \frac{1}{2q}\left({\frac{\d^{}{q}}{{\d{z}}^{}}}\right)^{2} + \frac{3}{2}q^3 + 4z q^2 + 2(z^2 - A)q + \frac{B}{q},$$ where $A$ and $B$ are constants, and the second degree, second order equation satisfied by the associated Hamiltonian function, see equation .
Polynomials orthogonal with respect to a symmetric measure can be generated via quadratic transformation from the classical orthogonal polynomials, cf. [@refChihara]. For example, Laguerre polynomials generate a class of generalized Hermite polynomials while Jacobi polynomials give rise to a class of generalized Ultraspherical polynomials.
In this paper we are concerned with the positive, even weight function on the real line arising from a symmetrization of the semi-classical Laguerre weight , namely the generalized Freud weight $$\ww{x}=|x|^{2\la+1}\exp\left(-x^4+tx^2\right),\qquad x\in\R\label{genFreud},$$ with parameters $\la>-1$ and $t\in\R$.
We use two different methods to derive the differential-difference equation satisfied by generalized Freud polynomials $\{S_n(x;t)\}_{n=0}^{\infty}$, orthogonal with respect to the generalized Freud weight . In §\[sec:scOPs\] we modify the ladder operator method (cf. [@refChenFeigin; @refChenIsmail; @refIsmail]) to derive a general formula for the coefficients of the differential-difference as well as the second order differential equation satisfied by polynomials orthogonal with respect to a generalized Freud type weight which vanishes at a point. It is important to note that the second order differential equation is linear with coefficients that are rational functions of $x$ with parameters $t$ and $\la$ involving parabolic cylinder functions. In §\[sec:scLag\] we derive the generalized Freud weight through a symmetrization of the semi-classical Laguerre weight . In §\[sec:genFpoly\] we are concerned with specific results for generalized Freud weight . We show that the coefficient $\b_n(t;\la)$ in the three-term recurrence relation $$xS_n(x;t)=S_{n+1}(x;t)+\b_{n}(t;\la)S_{n-1}(x;t),$$ satisfied by the polynomials associated with the weight can be expressed in terms of Wronskians that arise in the description of special function solutions of which are expressed in terms of parabolic cylinder functions. Further we apply the results of §\[sec:scOPs\] to the weight to derive the differential-difference equation and the linear, second order differential equations satisfied by generalized Freud polynomials $\{S_n(x;t)\}_{n=0}^{\infty}$. A method due to Shohat [@refShohat39], based on the concept of quasi-orthogonality and applied to the weight is discussed in §\[sho\].
Semi-classical orthogonal polynomials {#sec:scOPs}
=====================================
The coefficients $A_n(x;t)$ and $B_n(x;t)$ in the relation $$\label{ddee}{\frac{\d^{}{P_n}}{{\d{x}}^{}}}(x;t)=-B_n(x;t)P_n(x;t)+A_n(x;t)P_{n-1}(x;t),$$ satisfied by semi-classical orthogonal polynomials are of interest since differentiation of this differential-difference equation yields the second order differential equation satisfied by the orthogonal polynomials. Shohat [@refShohat39] gave a procedure using quasi-orthogonality to derive for weights $\ww{x}$ such that $\displaystyle{w'(x;t)}/{\ww{x}}$ is a rational function, which we apply to later. This technique was rediscovered by several authors including Bonan, Freud, Mhaskar and Nevai approximately 40 years later, see [@Nevai p. 126–132] and the references therein for more detail. The method of ladder operators was introduced by Chen and Ismail in [@refChenIsmail]. Related work by various authors can be found in, for example, [@refChenIts; @refChenZhang; @refFvAZ; @mhaskar] and a good summary of the technique is provided in [@refIsmail Theorem 3.2.1].
In [@refChenFeigin], Chen and Feigin adapt the method of ladder operators to the situation where the weight function vanishes at one point. Our next result generalizes the work in [@refChenFeigin], giving a more explicit expression for the coefficients in when the weight function is positive on the real line except for one point.
\[Thm:ABn\] Let $$\label{gft}\ww{x}=|x-k|^\ga \exp\{-\v(x;t)\},\qquad x,\,t,\,k\in\R,$$ where $\v(x;t)$ is a continuously differentiable function on $\R$. Assume that the polynomials $\{P_n(x;t)\}_{n=0}^{\infty}$ satisfy the orthogonality relation $$\int_{-\infty}^{\infty}P_n(x;t)P_m(x;t)\ww{x}\,dx=h_n\delta_{mn}.$$ Then, for ${\ga\geq1}$, $P_{n}(x;t)$ satisfy the differential-difference equation $$\label{dde}
(x-k)\,{\frac{\d^{}{P_n}}{{\d{x}}^{}}}(x;t)=-B_n(x;t)P_n(x;t)+A_n(x;t)P_{n-1}(x;t),$$ where
\[ABn\] $$\begin{aligned}
\label{An}
A_n(x;t)&=\frac{x-k}{h_{n-1}}\int_{-\infty}^{\infty}P_n^2(y;t)\vvxy\ww{y}\,dy+a_n(x;t),\\
B_n(x;t)&= \frac{x-k}{h_{n-1}}\int_{-\infty}^{\infty}P_n(y;t)P_{n-1}(y;t)\vvxy\ww{y}\,dy+b_n(x;t),\label{Bn}\end{aligned}$$
where $$\vvxy=\frac{\v'(x;t)-\v'(y;t)}{x-y},\label{def:Kxy}$$ with
\[abn\] $$\begin{aligned}
\label{Aa}a_n(x;t)&=\frac{\ga}{h_{n-1}}\int_{-\infty}^{\infty}\frac{P_n^2(y;t)}{y-k}\,\ww{y}\,dy,\\
\label{Ab}b_n(x;t)&=\frac{\ga}{h_{n-1}} \int_{-\infty}^{\infty}\frac{P_n(y;t)P_{n-1}(y;t)}{y-k}\,\ww{y}\,dy.\end{aligned}$$
Consider the generalized Freud-type weight . Since $\displaystyle{\frac{\d^{}{P_n}}{{\d{x}}^{}}}(x;t)$ is a polynomial of degree $n-1$ in $x$, then it can be expressed in terms of the orthogonal basis as $$\label{eq:orthexp}{\frac{\d^{}{P_n}}{{\d{x}}^{}}}(x;t)=\sum_{j=0}^{n-1}c_{n,j}P_j(x;t).$$ Applying the orthogonality relation and integrating by parts, we obtain $$\begin{aligned}
c_{n,j}\,h_j&=\int_{-\infty}^{\infty}{\frac{\d^{}{P_n}}{{\d{y}}^{}}}(y;t)P_j(y;t)\ww{y}\,dy\\
&=\Bigl[P_n(x;t)P_j(x;t)\ww{x}\Bigr]_{-\infty}^{\infty} -\int_{-\infty}^{\infty}P_n(y;t)\left\{{\frac{\d^{}{P_j}}{{\d{y}}^{}}}(y;t)\ww{y}+P_j(y;t){\frac{\d^{}{w}}{{\d{y}}^{}}}(y;t)\right\}dy\\
&=-\int_{-\infty}^{\infty}P_n(y;t)P_j(y;t){\frac{\d^{}{w}}{{\d{y}}^{}}}(y;t)\,dy\\
&=\int_{-\infty}^{\infty}P_n(y;t)P_j(y;t)\left[\vvy-\frac{\ga}{y-k}\right]\ww{y}\,dy,\end{aligned}$$ provided that ${\ga\geq 1}$.
Now, from , we can write $$\begin{aligned}
{\frac{\d^{}{P_n}}{{\d{x}}^{}}}(x;t)
&=\sum_{j=0}^{n-1}\frac{1}{h_j}\left\{\int_{-\infty}^{\infty}P_n(y;t)P_j(y;t)\left[\vvy-\frac{\ga}{y-k}\right]\ww{y}\,dy\right\}P_j(x;t)\\
&=\int_{-\infty}^{\infty}P_n(y;t)\sum_{j=0}^{n-1}\frac{P_j(y;t)P_j(x;t)}{h_j}\left[\vvy-\frac{\ga}{y-k}\right]\ww{y}\,dy\\
&=\int_{-\infty}^{\infty}P_n(y;t)\sum_{j=0}^{n-1}\frac{P_j(y;t)P_j(x;t)}{h_j}\left[\vvy-\vvx\right]\ww{y}\,dy\\
&\qquad\qquad+\vvx\int_{-\infty}^{\infty}P_n(y;t)\sum_{j=0}^{n-1}\frac{P_j(y;t)P_j(x;t)}{h_j}\,\ww{y}\,dy\\
&\qquad\qquad-\ga\int_{-\infty}^{\infty}\frac{P_n(y;t)}{y-k}\sum_{j=0}^{n-1}\frac{P_j(y;t)P_j(x;t)}{h_j}\,\ww{y}\,dy
\\
&=\int_{-\infty}^{\infty}P_n(y;t)\sum_{j=0}^{n-1}\frac{P_j(y;t)P_j(x;t)}{h_j}\left[\vvy-\vvx\right]\ww{y}\,dy\\
&\qquad\qquad-\ga\int_{-\infty}^{\infty}\frac{P_n(y;t)}{y-k}\sum_{j=0}^{n-1}\frac{P_j(y;t)P_j(x;t)}{h_j}\,\ww{y}\,dy.\end{aligned}$$ Next, using the orthogonality relation again, we obtain $$\begin{aligned}
{(x-k) {\frac{\d^{}{P_n}}{{\d{x}}^{}}}(x;t)}
&=(x-k) \int_{-\infty}^{\infty}P_n(y;t)\sum_{j=0}^{n-1}\frac{P_j(y;t)P_j(x;t)}{h_j}\left[\vvy-\vvx\right]\ww{y}\,dy\\
&\qquad\qquad-\ga\int_{-\infty}^{\infty}P_n(y;t)\left(\frac{x-k}{y-k}\right)\sum_{j=0}^{n-1}\frac{P_j(y;t)P_j(x;t)}{h_j}\,\ww{y}\,dy.\\
&=(x-k) \int_{-\infty}^{\infty}P_n(y;t)\sum_{j=0}^{n-1}\frac{P_j(y;t)P_j(x;t)}{h_j}\left[\vvy-\vvx\right]\ww{y}\,dy\\
&\qquad\qquad-\ga\int_{-\infty}^{\infty}P_n(y;t)\frac{x-y}{y-k}\sum_{j=0}^{n-1}\frac{P_j(y;t)P_j(x;t)}{h_j}\,\ww{y}\,dy.\end{aligned}$$ It now follows from the Christoffel-Darboux formula (cf. [@refIsmail]) $$\sum_{j=0}^{n-1}\frac{P_j(y;t)P_j(x;t)}{h_j}=\frac{P_n(x;t)P_{n-1}(y;t)-P_n(y;t)P_{n-1}(x;t)}{(x-y)h_{n-1}},$$ that $$\begin{aligned}
A_n(x;t)&=\frac{x-k}{h_{n-1}}\int_{-\infty}^{\infty}P_n^2(y;t)\vvxy\ww{y}\,dy+a_n(x;t)\\
B_n(x;t)&= \frac{x-k}{h_{n-1}}\int_{-\infty}^{\infty}P_n(y;t)P_{n-1}(y;t)\vvxy\ww{y}\,dy +b_n(x;t),\end{aligned}$$ with $a_n(x;t)$ and $b_n(x;t)$ given by .
\[lemmaeven\] Consider the weight defined by and assume that $\v(x;t)$ is an even, continuously differentiable function on $\R$. Assume that the polynomials $\{P_n(x;t)\}_{n=0}^{\infty}$ satisfy the orthogonality relation $$\int_{-\infty}^{\infty}P_n(x;t)P_m(x;t)\ww{x}\,dx=h_n\delta_{mn}.$$ and the three-term recurrence relation $$\label{3trr}
P_{n+1}(x;t)=xP_{n}(x;t)-\b_n(t;\la)P_{n-1}(x;t),$$ with $P_0=1$ and $P_1=x$. Then the polynomials $P_n(x;t)$ satisfy $$\begin{aligned}
\int_{-\infty}^{\infty}\frac{P_n^2(y;t)}{y-k}\,\ww{y}\,dy&=0,\\
\int_{-\infty}^\infty \frac{P_n(y;t)P_{n-1}(y;t)}{y-k}\,\ww{y}\,dy&= \tfrac12[1-(-1)^n]\,h_{n-1},
\label{int22}\end{aligned}$$ where $n\in\N$ and $$h_n = \int_{-\infty}^\infty {P_n^2(y;t)\ww{y}}\,dy.$$
Since $\ww{x}$ is even when $\v(x;t)$ is assumed to be even, the integrand in is odd and hence $a_n(x;t)=0$.
Furthermore, the monic orthogonal polynomials $P_{n}(x;t)$ satisfy the three-term recurrence relation , with $P_0=1$ and $P_1=x$, hence $$\begin{aligned}
\int_{-\infty}^\infty \frac{P_n(y;t)P_{n-1}(y;t)}{y-k}\,\ww{y}\,dy
&=\int_{-\infty}^\infty \frac{\big[ yP_{n-1}(y;t)-\b_{n-1}P_{n-2}(y;t)\big]P_{n-1}(y;t)}{y-k}\,\ww{y} \,dy\\
&=\int_{-\infty}^\infty P_{n-1}^2(y;t)\ww{y}\,dy +k\int_{-\infty}^{\infty}\frac{P_{n-1}^2(y;t)}{y-k}\, \ww{y}\,dy\\
&\qquad- \b_{n-1}\int_{-\infty}^\infty \frac{P_{n-1}(y;t)P_{n-2}(y;t)}{y-k}\,\ww{y}\,dy\\
&= h_{n-1} - \b_{n-1}\int_{-\infty}^\infty \frac{P_{n-1}(y;t)P_{n-2}(y;t)}{y-k}\,\ww{y}\,dy,\end{aligned}$$ since the integrand in the second integral is odd. Hence, if we define $$I_n=\int_{-\infty}^\infty \frac{P_n(y;t)P_{n-1}(y;t)}{y-k}\,\ww{y}\,dy,$$ then $I_n$ satisfies the recurrence relation $$I_n = h_{n-1} - \b_{n-1}I_{n-1}= h_{n-1} -\frac{h_{n-1}}{h_{n-2}} I_{n-1},$$ since $\b_n=h_{n}/h_{n-1}$. Iterating this gives $$\begin{aligned}
I_n &= \frac{h_{n-1}}{h_{n-3}} I_{n-2}
= h_{n-1}-\frac{h_{n-1}}{h_{n-4}} I_{n-3}
= \frac{h_{n-1}}{h_{n-5}} I_{n-4}
= h_{n-1}-\frac{h_{n-1}}{h_{n-6}} I_{n-5}, \end{aligned}$$ and so on. Hence, by induction, $$\begin{aligned}
I_{2N} =\frac{h_{2N-1}}{h_{1}} I_{2},\qquad I_{2N+1} =h_{2N}-\frac{h_{2N}}{h_{1}} I_{2},\end{aligned}$$ and so, since $$\begin{aligned}
I_2&=\int_{-\infty}^\infty \frac{P_2(y;t)P_{1}(y;t)}{y-k}\,\ww{y}\,dy\\
& =\int_{-\infty}^\infty P_2(y;t)\ww{y}\,dy+k\int_{-\infty}^{\infty}\frac{P_2(y;t)}{y-k}\,\ww{y}\,dy=0,\end{aligned}$$ we have that $$I_{2N} =0,\qquad I_{2N+1} =h_{2N},$$ as required.
\[cor:ABn\] Let $$\ww{x}=|x-k|^\ga \exp\{-\v(x;t)\},\qquad x,\,t,\,k\in\R,$$ where $\v(x;t)$ is an even, continuously differentiable function on $\R$. Assume that the polynomials $\{P_n(x;t)\}_{n=0}^{\infty}$ satisfy the orthogonality relation $$\int_{-\infty}^{\infty}P_n(x;t)P_m(x;t)\ww{x}\,dx=h_n\delta_{mn}.$$ Then, for ${\ga\geq1}$, $P_{n}(x;t)$ satisfy the differential-difference equation $$(x-k)\,{\frac{\d^{}{P_n}}{{\d{x}}^{}}}(x;t)=-B_n(x;t)P_n(x;t)+A_n(x;t)P_{n-1}(x;t),$$ where
\[ABn2\] $$\begin{aligned}
A_n(x;t)&=\frac{x-k}{h_{n-1}}\int_{-\infty}^{\infty}P_n^2(y;t)\vvxy\ww{y}\,dy,\\
B_n(x;t)&= \frac{x-k}{h_{n-1}}\int_{-\infty}^{\infty}P_n(y;t)P_{n-1}(y;t)\vvxy\ww{y}\,dy+\tfrac12{\ga}[1-(-1)^n].\end{aligned}$$
The result is an immediate consequence of Theorem \[Thm:ABn\] and Lemma \[lemmaeven\].
\[lem22\] Let $\ww{x}$ be the weight defined by with $\v(x;t)$ an even, continuously differentiable function on $\R$ and $A_n(x;t)$ and $B_n(x;t)$ defined by . Then $$B_n(x;t)+B_{n+1}(x;t)=\frac{x A_n(x;t)}{\b_n}+\ga-(x-k)\vvx,$$ when $\ga\geq1$.
From , and the fact that $h_n=h_{n-1}\b_n$ we have $$\begin{aligned}
\nonumber
B_n(x;t)+B_{n+1}(x;t)&=\frac{x-k}{h_{n-1}}\int_{-\infty}^{\infty}P_n(y;t)P_{n-1}(y;t)\vvxy\ww{y}\,dy\nonumber\\
&\qquad+\frac{x-k}{h_{n}}\int_{-\infty}^{\infty}P_n(y;t)P_{n+1}(y;t)\vvxy\ww{y}\,dy\nonumber\\
&\qquad+\tfrac12{\ga}[1-(-1)^{n}]+\tfrac12{\ga}[1-(-1)^{n+1}]\nonumber\\
&=\frac{x-k}{h_{n}}\left\{\int_{-\infty}^{\infty}P_{n}(y;t)\left[\b_nP_{n-1}(y;t)+P_{n+1}(y;t)\right]
\vvxy\ww{y}\,dy\right\}+\ga\nonumber\\
&=\frac{x-k}{h_{n}}\int_{-\infty}^{\infty}yP^2_{n}(y;t)\vvxy\ww{y}\,dy+\ga\nonumber\\
&=\frac{x-k}{h_{n}}\int_{-\infty}^{\infty}P_n^2(y;t)\left[\vvy-\vvx\right]\ww{y}\,dy\nonumber\\
&\qquad+\frac{x(x-k)}{h_n}\int_{-\infty}^{\infty}P_n^2(y;t)\vvxy\ww{y}\,dy+\ga\nonumber\\
&=\frac{x-k}{h_n}\int_{-\infty}^{\infty}P_n^2(y;t)\vvy\ww{y}\,dy\nonumber\\
&\qquad-\frac{x-k}{h_n}\vvx\int_{-\infty}^{\infty}P_n^2(y;t)\ww{y}\,dy+\frac{x}{h_n}h_{n-1}A_n(x;t)+\ga\nonumber\\
&=\frac{\gamma(x-k)}{h_n}\int_{-\infty}^{\infty}P_n^2(y;t)\frac{\ww{y}}{y-k}\,dy-\frac{x-k}{h_n}\int_{-\infty}^{\infty}P_n^2(y;t){\frac{\d^{}{w}}{{\d{y}}^{}}}(y;t)\,dy\nonumber\\
&\qquad-(x-k)\vvx+\frac{xA_n(x;t)}{\b_n}+\ga,\label{bet}\end{aligned}$$ since $$\ds{{\frac{\d^{}{w}}{{\d{y}}^{}}}(y;t)=\left[-\vvy+\frac{\ga}{y-k}\right]\ww{y}}.$$ The first integral in vanishes since the integrand is odd, hence it follows, using integration by parts, that $$\begin{aligned}
B_n(x;t)+B_{n+1}(x;t)&=-\frac{x-k}{h_n}\Big\{\Bigr[P_n^2(x;t)\ww{x}\Bigr]_{-\infty}^{\infty}-\int_{-\infty}^{\infty}2P_n(y;t){\frac{\d^{}{P_n}}{{\d{y}}^{}}}(y;t)\ww{y}\,dy\Bigr\}\\
&\qquad -(x-k)\vvx+\frac{x A_n(x;t)}{\b_n}+\ga,\end{aligned}$$ and the result follows from the orthogonality of $P_n$.
Let $\ww{x}$ be the weight defined by with $\v(x;t)$ an even, continuously differentiable function on $\R$ and $A_n(x;t)$ and $B_n(x;t)$ defined by . Then $$\begin{aligned}
\left\{-(x-k){\frac{\d^{}{}}{{\d{x}}^{}}}+B_n(x;t)+(x-k)\vvx-\ga\right\}&P_{n-1}(x;t)\label{dde2} =\frac{A_{n-1}(x;t)}{\b_{n-1}}P_n(x;t).\end{aligned}$$
From the differential-difference equation we have $$\begin{aligned}
(x-k){\frac{\d^{}{P_{n-1}}}{{\d{x}}^{}}}(x;t)&=-B_{n-1}(x;t)P_{n-1}(x;t)+A_{n-1}(x;t)P_{n-2}(x;t)\\
&=-B_{n-1}(x;t)P_{n-1}(x;t) +\frac{A_{n-1}(x;t)}{\b_{n-1}}\left[xP_{n-1}(x;t)-P_n(x;t)\right],\end{aligned}$$ using the recurrence relation . Hence, using Lemma \[lem22\], $$\begin{aligned}
\frac{A_{n-1}(x;t)}{\b_{n-1}}P_n(x;t)
&=-(x-k){\frac{\d^{}{P_{n-1}}}{{\d{x}}^{}}}(x;t)-B_{n-1}(x;t)P_{n-1}(x;t) +\frac{x A_{n-1}(x;t)}{\b_{n-1}}P_{n-1}(x;t)\\
&=-(x-k){\frac{\d^{}{P_{n-1}}}{{\d{x}}^{}}}(x;t)-B_{n-1}(x;t)P_{n-1}(x;t)\\ &\qquad
+\left\{B_n(x;t)+B_{n-1}(x;t)-\ga+(x-k)\vvx\right\}P_{n-1}(x;t)\\
&=-(x-k){\frac{\d^{}{P_{n-1}}}{{\d{x}}^{}}}(x;t)+B_{n}(x;t)P_{n-1}(x;t) +\left[(x-k)\vvx-\ga\right]P_{n-1}(x;t)\\
&=\left\{-(x-k){\frac{\d^{}{}}{{\d{x}}^{}}}+B_n(x;t)+(x-k)\vvx-\ga\right\}P_{n-1}(x;t).\end{aligned}$$
\[thm:gende\]Let $\ww{x}=|x-k|^\ga \exp\{-\v(x;t)\}$, for $x,\,t,\,k\in\R$, with $\v(x;t)$ an even, continuously differentiable function on $\R$. Then
$$\label{eq:gende}
(x-k){\frac{\d^{2}{P_n}}{{\d{x}}^{2}}}(x;t)+R_n(x;t){\frac{\d^{}{P_n}}{{\d{x}}^{}}}(x;t)+T_n(x;t)P_n(x;t)=0,$$
where $$\begin{aligned}
\label{coef1}
R_n(x;t)&=\ga-(x-k)\vvx-\frac{x-k}{A_n(x;t)}\,{\frac{\d^{}{A_n}}{{\d{x}}^{}}}(x;t)+1,\\[5pt]
\label{coef2}T_n(x;t)&=\frac{A_n(x;t)A_{n-1}(x;t)}{(x-k)\b_{n-1}}+{\frac{\d^{}{B_n}}{{\d{x}}^{}}}(x;t)+\frac{\ga B_n(x;t)}{x-k}\nonumber\\
&\qquad-B_n(x;t)\left[\vvx +\frac{B_n(x;t)}{x-k}\right]-\frac{B_n(x;t)}{A_n(x)}{\frac{\d^{}{A_n}}{{\d{x}}^{}}}(x;t),\end{aligned}$$ with $$\begin{aligned}
A_n(x;t)&=\frac{x-k}{h_{n-1}}\int_{-\infty}^{\infty}P_n^2(y;t)\vvxy\ww{y}\,dy\\
B_n(x;t)&= \frac{x-k}{h_{n-1}}\int_{-\infty}^{\infty}P_n(y;t)P_{n-1}(y;t)\vvxy\ww{y}\,dy
+\tfrac12{\ga}[1-(-1)^n].\end{aligned}$$
Differentiating both sides of with respect to $x$ we obtain $$\begin{aligned}
\label{eq:dif1}
(x-k){\frac{\d^{2}{P_n}}{{\d{x}}^{2}}}(x;t)&=-[B_n(x;t)+1]{\frac{\d^{}{P_n}}{{\d{x}}^{}}}x,t)+{\frac{\d^{}{A_n}}{{\d{x}}^{}}}(x;t)P_{n-1}(x;t)\nonumber\\
&\qquad-{\frac{\d^{}{B_n}}{{\d{x}}^{}}}(x;t)P_n(x;t)+A_n(x;t)P_{n-1}(x;t).\end{aligned}$$ Substituting into yields $$\begin{aligned}
(x-k){\frac{\d^{2}{P_n}}{{\d{x}}^{2}}}(x;t)&=\nonumber-[B_n(x;t)+1]{\frac{\d^{}{P_n}}{{\d{x}}^{}}}(x;t)-\left[{\frac{\d^{}{B_n}}{{\d{x}}^{}}}(x;t)+\frac{A_n(x;t)A_{n-1}(x;t)}{\b_{n-1}(x-k)}\right]P_n(x;t)\\&\qquad+\left\{{\frac{\d^{}{A_n}}{{\d{x}}^{}}}(x;t)+A_n(x;t)\left[\frac{B_n(x;t)}{x-k}+\vvx-\frac{\ga}{x-k}\right]\right\}P_{n-1}(x;t),\label{eq:fin}\end{aligned}$$ and the results follows by substituting $P_{n-1}(x;t)$ in using .
Semi-classical Laguerre weight {#sec:scLag}
==============================
We recall that the semi-classical Laguerre weight [@refCJ; @Hendriksen] is given by $$\begin{aligned}
\label{laguerreweight}
\ww{x} = x^{\la}\, \exp (-x^2 +tx), \quad x\in \R^{+},\end{aligned}$$ where $\la > -1$. The weight function satisfies the differential equation $$x{\frac{\d^{}{w}}{{\d{x}}^{}}}(x;t) + (2x^2-tx-\la) \ww{x} = 0,$$ which is the Pearson equation with $\sigma(x) =x$ and $ \tau(x;t) = -2x^2+tx+\la +1$.
Explicit expressions for the recurrence coefficients $\a_n(t;\la)$ and $\b_n(t;\la)$ in the three-term recurrence relation $$xP_n(x;t)=P_{n+1}(x;t)+\a_n(t;\la)P_n(x;t)+\b_n(t;\la)P_{n-1}(x;t),\label{eq:scLag2}$$ associated with the semi-classical Laguerre weight were obtained in [@refCJ], and are given in the following theorem.
Suppose $\Psi_{n,\la}(z)$ is given by $$\label{eq:PT.SF.eq411}
\Psi_{n,\la}(z)=\W\left(\psi_{\la},{\frac{\d^{}{\psi_{\la}}}{{\d{z}}^{}}},\ldots,{\frac{\d^{n-1}{\psi_{\la}}}{{\d{z}}^{n-1}}}\right),\qquad \Psi_{0,\la}(z)=1,$$ where $$\label{eq:PT.SF.eq413}
\psi_{\la}(z)=\begin{cases} D_{-\la-1}\big(-\sqrt2\,z\big)\exp\big(\tfrac12z^2\big),\quad
&\mbox{\rm if}\quad \la\not\in\N,\\[2.5pt]
\ds {\frac{\d^{m}{}}{{\d{z}}^{m}}}\left\{\big[1+\erf(z)\big]\exp(z^2)\right\}, &\mbox{\rm if}\quad \la=m\in\N,
\end{cases}$$ with $D_{\nu}(\zeta)$ is the [parabolic cylinder function]{} and $\erfc(z)$ the complementary error function $$\label{def:erfc}
\erfc (z)=\frac{2}{\sqrt{\pi}} \int_z^\infty \exp (-t^2) \,dt,$$ Then coefficients $\a_n(t;\la)$ and $\b_n(t;\la)$ in the recurrence relation associated with the semi-classical Laguerre weight are given by
\[eq:scLag31ab\]$$\begin{aligned}
\label{eq:scLag31a}
\a_n(t;\la)&=\tfrac12q_n(z)+\tfrac12t,\\ \label{eq:scLag31b}
\b_n(t;\la)&=-\tfrac18{\frac{\d^{}{q_n}}{{\d{z}}^{}}}-\tfrac18 q_n^2(z) -\tfrac14zq_n(z)+\tfrac14\la+\tfrac12n,\end{aligned}$$
with $z=\tfrac12t$, where $$\label{eq:PT.SF.eq410}q_n(z)=-2z+{\frac{\d^{}{}}{{\d{z}}^{}}}\ln\frac{\Psi_{n+1,\la}(z)}{\Psi_{n,\la}(z)},$$ which satisfies with parameters $(A,B)=(2n+\la+1,-2\la^2)$.
See [@refCJ].
Symmetrization of semi-classical Laguerre weight
------------------------------------------------
In this section we show that symmetrizing the semi-classical Laguerre weight gives rise to the generalized Freud weight .
Let $\lbrace L^{(\la)}_n(x;t)\rbrace$ denote the monic semi-classical Laguerre polynomials orthogonal with respect to semi-classical Laguerre weight with $$\begin{aligned}
\int_0^{\infty} L^{(\la)}_m(x;t) L^{(\la)}_n(x;t)
\, x^{\la}\exp(-x^2+tx) \,d{x} = K_n(t)\, \delta_{mn}.\end{aligned}$$
Define $$\begin{aligned}
\label{quadratic}
S_{2m}(x;t) =L^{(\la)}_m(x^2;t); \quad S_{2m+1}(x;t) =xQ^{(\la)}_n(x^2;t),\end{aligned}$$ where $$\begin{aligned}
\label{Kernelsym}
Q^{(\la)}_n(x;t)= \frac{1}{x}\left[ L^{(\la)}_{n+1}(x;t) - \dfrac{L^{(\la)}_{n+1}({0;t})}{L^{(\la)}_n({0;t})} L^{(\la)}_n(x;t) \right],\end{aligned}$$ are also monic and of degree $n$, for $x\neq0$.
Then $$\begin{aligned}
\int_0^{\infty} L^{(\la)}_m(x;t) L^{(\la)}_n(x;t) x^{\la}\exp(-x^2+tx) \,d{x}
& = 2\int_0^{\infty} L^{(\la)}_m(x^2;t) L^{(\la)}_n(x^2;t) x^{2\la+1} \exp\left(-x^4+tx^2\right) d{x}
\\& = \int_{-\infty}^{\infty} L^{(\la)}_m(x^2;t) L^{(\la)}_n(x^2;t) |x|^{2\la+1} \exp\left(-x^4+tx^2\right) d{x}
\\&= \int_{-\infty}^{\infty}S_{2m}(x;t) S_{2n}(x;t) |x|^{2\la+1}\exp\left(-x^4+tx^2\right) d{x}\\&=K_n(t) \delta_{mn}.\end{aligned}$$ which implies that $\lbrace S_{2m}(x;t) \rbrace_{m=0}^{\infty}$ is a symmetric orthogonal sequence with respect to the even weight $$\ww{x} =|x|^{2\la+1}\exp (-x^4 +tx^2),$$ on $\R$, i.e. the generalized Freud weight .
It is proved in [@refChihara Thm 7.1] that the kernel polynomials $Q_{m}^L(x;t)$ are orthogonal with respect to $x \ww{x}= x^{\la+1}\exp(-x^2+tx)$. Hence $$\begin{aligned}
K_n(t) \delta_{mn}&=
\int_0^{\infty} Q^{(\la)}_m(x;t) Q^{(\la)}_n(x;t) x^{\la+1}\exp(-x^2+tx) \,d{x} \\&=
2\int_0^{\infty} Q^{(\la)}_m(x^2;t) Q^{(\la)}_n(x^2;t) x^{2\la+3}\exp\left(-x^4+tx^2\right) d{x}
\\&=
\int_{-\infty}^{\infty} \left[ xQ^{(\la)}_m(x^2;t)\right)]\left[ xQ^{(\la)}_n(x^2;t)\right] |x|^{2\la+1}\exp\left(-x^4+tx^2\right) d{x}
\nonumber\\&=
\int_{-\infty}^{\infty} S_{2m+1}(x;t) S_{2n+1}(x;t) |x|^{2\la+1}\exp\left(-x^4+tx^2\right)d{x}.\end{aligned}$$ Lastly, since in each case the integrand is odd, we have that $$\begin{aligned}
\int_{-\infty}^{\infty} S_{2m+1}(x;t) S_{2n}(x;t) |x|^{2\la+1}\exp\left(-x^4+tx^2\right) d{x}=0,\qquad m,n\in\N,\end{aligned}$$ and so we conclude that $\lbrace S_{n}(x;t) \rbrace_{n=0}^{\infty}$ is a sequence of polynomials orthogonal with respect to the even weight on $\R.$
The generalized Freud weight {#sec:genFpoly}
============================
The moments of the generalized Freud weight
-------------------------------------------
It is shown in [@refCJ] that the moments of the semi-classical weight provide the link between the weight and the associated equation. The recurrence coefficients in the three-term recurrence relations associated with semi-classical orthogonal polynomials can often be expressed in terms of solutions of the equations and associated discrete equations.
For the generalized Freud weight , the first moment, $\mu_0(t;\la)$, can be obtained using the integral representation of a parabolic cylinder (Hermite-Weber) function ${D_{v}}(\xi)$. By definition $$\begin{split}
\mu_0(t;\la)&=\int_{-\infty}^{\infty}|x|^{2\la+1}\exp\left(-x^4+tx^2\right)dx\\
&=2\int_{0}^{\infty}x^{2\la+1}\exp\left(-x^4+tx^2\right)dx\\
&=\int_{0}^{\infty}y^{\la}\exp\left(-y^2+ty\right)dy\\
&= \frac{\Gamma(\la+1)}{2^{(\la+1)/2}}\,\exp\left(\tfrac18t^2\right){D_{-\la-1}}\big(-\tfrac12\sqrt2\,t\big).
\end{split}$$ since the parabolic cylinder function ${D_{v}}(\xi)$ has the integral representation [@DLMF §12.5(i)] $${D_{v}}(\xi) = \dfrac{\exp(-\tfrac{1}{4}\xi^2)}{\Gamma(-v)} \int_{0}^{\infty} s^{-v-1} \, \exp\left( -\tfrac{1}{2}s^2 -\xi s \right) d{s},\qquad \Re(\nu)<0.$$ The even moments are $$\begin{aligned}
\mu_{2n}(t;\la) &= \int_{-\infty}^{\infty} x^{2n}\, |x|^{2\la+1}\exp\left(-x^4+tx^2\right) d{x} \\&=
{\frac{\d^{n}{}}{{\d{t}}^{n}}}\left(\int_{-\infty}^{\infty}|x|^{2\la+1}\exp\left(-x^4+tx^2\right) d{x} \right),\nonumber\\&
= {\frac{\d^{n}{}}{{\d{t}}^{n}}}\mu_0(t;\la), \quad n\geq1,\end{aligned}$$ whilst the odd ones are $$\begin{aligned}
\mu_{2n+1}(t;\la) &= \int_{-\infty}^{\infty} x^{2n+1}\, |x|^{2\la+1}\exp\left(-x^4+tx^2\right) d{x} =0, \quad n\in\N,\end{aligned}$$ since the integrand is odd.
We note that $$\begin{aligned}
\mu_{2n}(t;\la) &= \int_{-\infty}^{\infty} x^{2n}\, |x|^{2\la+1}\exp\left(-x^4+tx^2\right) d{x}\\
&= \int_{-\infty}^{\infty} |x|^{2\la+2n+1}\exp\left(-x^4+tx^2\right) d{x}\\
&= \mu_0(t;\la+n).\end{aligned}$$ Also, when $\la=n\in\N$, then $${D_{-n-1}}\big(-\tfrac12\sqrt2\,t\big)=\tfrac12\sqrt{2\pi}\,{\frac{\d^{n}{}}{{\d{t}}^{n}}}\left\{\left[1+\erf\left(\tfrac12t\right)\right]\exp\left(\tfrac18t^2\right)\right\},$$ where $\erf(z)$ is the error function [@DLMF §12.7(ii)].
Recurrence coefficients of the generalized Freud polynomials
------------------------------------------------------------
Monic orthogonal polynomials with respect to the symmetric generalized Freud weight satisfy the three-term recurrence relation $$xS_n(x;t)=S_{n+1}(x;t)+\b_{n}(t;\la)S_{n-1}(x;t),\label{eq:gfrr}$$ where $S_{-1}(x;t)=0$ and $S_0(x;t)=1.$
Our interest is to determine explicit expressions for the recurrence coefficients $\b_n(t;\la)$ in the three-term recurrence relation . First we relate them to solutions of .
[The recurrence coefficients $\b_n(t;\la)$ in satisfy the equation $$\label{eq:bn}
{\frac{\d^{2}{\b_n}}{{\d{t}}^{2}}}=\frac{1}{2\b_n}\left({\frac{\d^{}{\b_n}}{{\d{t}}^{}}}\right)^2+\tfrac32\b_n^3-t\b_n^2+(\tfrac18 t^2-\tfrac12A_n)\b_n+\frac{B_n}{16\b_n},$$ where the parameters $A_n$ and $B_n$ are given by $$\begin{array}{l@{\quad}l}A_{2n} = -2\la-n-1, & A_{2n+1}=\la-n,\\
B_{2n} = -2n^2, & B_{2n+1}=-2(\la+n+1)^2.\end{array}$$ Further $\b_n(t;\la)$ satisfies the nonlinear difference equation $$\label{eq:dPI}
\b_{n+1}+\b_n+\b_{n-1}=\tfrac12t+\frac{2n+(2\la+1)[1-(-1)^n]}{8\b_n},$$ which is known as discrete (). ]{}
See, for example, [@refFvAZ Theorem 6.1].
1. Equation is equivalent to through the transformation $\b_n(t;\la)=\tfrac12 w(z)$, with $z=-\tfrac12t$. Hence
\[bnp4\] $$\begin{aligned}
\b_{2n}(t;\la)&=\tfrac12 w\big(z;-2\la-n-1,-2n^2\big),\\
\b_{2n+1}(t;\la)&=\tfrac12 w\big(z;\la-n,-2(\la+n+1)^2\big),\end{aligned}$$
with $z=-\tfrac12t$, where $w(z;A,B)$ satisfies . The conditions on the parameters in are precisely those for which has has solutions expressible in terms of the parabolic cylinder function $$\psi(z)=\mu_0(-2z;\la)=\frac{\Gamma(\la+1)}{2^{(\la+1)/2}}\,\exp\left(\tfrac12z^2\right){D_{-\la-1}}\big(\sqrt2\,z\big),$$ [@refFW01; @refOkamotoPIIPIV]; see also [@refCJ Theorem 3.5].
2. The link between the differential equation and the difference equation is given by the Bäcklund transformations
$$\begin{aligned}
\b_{n+1}&=\frac{1}{2\b_n}{\frac{\d^{}{\b_n}}{{\d{t}}^{}}}-\tfrac12\b_n+\tfrac14t+\frac{\ga_n}{4\b_n},\label{def:bnp}\\
\b_{n-1}&=-\frac{1}{2\b_n}{\frac{\d^{}{\b_n}}{{\d{t}}^{}}}-\tfrac12\b_n+\tfrac14t+\frac{\ga_n}{4\b_n},\label{def:bnm}\end{aligned}$$
with $\ga_n=\tfrac12n+\tfrac14(2\la+1)[1-(-1)^n]$. Letting $n\to n+1$ in and substituting gives the differential equation , whilst eliminating the derivative yields the difference equation .
The recurrence coefficients $\b_n(t;\la)$ in are given by
$$\begin{aligned}
\b_{2n}(t;\la) &={\frac{\d^{}{}}{{\d{t}}^{}}}\ln\frac{\tau_n(t;\la+1)}{\tau_n(t;\la)},\\
\b_{2n+1}(t;\la) &={\frac{\d^{}{}}{{\d{t}}^{}}}\ln\frac{\tau_{n+1}(t;\la)}{\tau_n(t;\la+1)},\end{aligned}$$
where $\tau_n(t;\la)$ is the Wronskian given by $$\begin{aligned}
\label{taun}
\tau_n(t;\la)&=\mathcal{W}\left(\phi_{\la},{\frac{\d^{}{\phi_{\la}}}{{\d{t}}^{}}},\ldots,{\frac{\d^{n-1}{\phi_{\la}}}{{\d{t}}^{n-1}}}\right),\end{aligned}$$ with $$\phi_{\la}(t)=\mu_0(t;\la)=\frac{\Gamma(\la+1)}{2^{(\la+1)/2}}\,\exp\left(\tfrac18t^2\right){D_{-\la-1}}\big(-\tfrac12\sqrt2\,t\big).$$
From the parabolic cylinder solutions of given in [@refCJ Theorem 3.5], it is easily shown that the equation $$\label{eqb}
{\frac{\d^{2}{y}}{{\d{t}}^{2}}}=\frac{1}{2y}\left({\frac{\d^{}{y}}{{\d{t}}^{}}}\right)^2+\tfrac32y^3-ty^2+(\tfrac18 t^2-\tfrac12A)y+\frac{B}{16y},$$ has the solutions $\left\{y_{n}^{[j]}\big(t;A_{n}^{[j]},B_{n}^{[j]}\big)\right\}_{j=1,2,3}$,
\[eqbsol\] $$\begin{aligned}
&y_{n}^{[1]}\big(t;\la+2n-1, -2\la^2\big)= \tfrac12t+{\frac{\d^{}{}}{{\d{t}}^{}}}\ln\frac{\tau_{n-1}(t;\la)}{\tau_{n}(t;\la)},\\
&y_{n}^{[2]}\big(t;-2\la-n-1, -2n^2\big) ={\frac{\d^{}{}}{{\d{t}}^{}}}\ln\frac{\tau_n(t;\la+1)}{\tau_n(t;\la)},\\
&y_{n}^{[3]}\big(t;\la-n+1, -2(\la+n)^2\big) ={\frac{\d^{}{}}{{\d{t}}^{}}}\ln\frac{\tau_{n}(t;\la)}{\tau_{n-1}(t;\la+1)},\end{aligned}$$
where $\tau_n(t;\la)$ is the Wronskian . Comparing and gives the desired result.
The first few recurrence coefficients $\b_n(t;\la)$ are given by $$\begin{split}
\b_1(t;\la)&=\Ph,\\
\b_2(t;\la)&=-\frac{2\Ph^2-t\Ph-\la-1}{2\Ph},\\
\b_3(t;\la)&=-\frac{\Ph}{2\Ph^2-t\Ph-\la-1}-\frac{\la+1}{2\Ph},\\
\b_4(t;\la)&=\frac{t}{2(\la+2)}+\frac{\Ph}{2\Ph^2-t\Ph-\la-1} +\frac{(\la+1)(t^2+2\la+4)\Ph+(\la+1)^2t}{2(\la+2)[2(\la+2)\Ph^2-(\la+1) t\Ph-(\la+1)^2]},
\end{split}$$ where $$\begin{aligned}
\Ph(t)&={\frac{\d^{}{}}{{\d{t}}^{}}}\ln\left\{{D_{-\la-1}}\big(-\tfrac12\sqrt2\,t\big)\exp\left(\tfrac18t^2\right)\right\}
=\tfrac12t+\tfrac12\sqrt{2}\,\frac{{D_{-\la}}\big(-\tfrac12\sqrt2\,t\big)}{{D_{-\la-1}}\big(-\tfrac12\sqrt2\,t\big)}.
\label{def:Phi}\end{aligned}$$ We note that $$\Ph(t)={\frac{\d^{}{}}{{\d{t}}^{}}}\ln\phi_\la(t),$$ where $\phi_\la(t)=\exp(\tfrac18t^2)D_{-\la-1}\left(-\tfrac12\sqrt{2}\,t\right)$ satisfies $${\frac{\d^{2}{\phi_\la}}{{\d{t}}^{2}}}-\tfrac12t{\frac{\d^{}{\phi_\la}}{{\d{t}}^{}}}-\tfrac12(\la+1)\phi_\la=0,$$ and $\Ph(t)$ satisfies the Riccati equation $${\frac{\d^{}{\Ph}}{{\d{t}}^{}}}=-\Ph^2+\tfrac12t\Ph+\tfrac12(\la+1)=0,$$ see Lemma \[lem42\].
Using the recurrence relation , the first few polynomials are $$\begin{split}
S_1(x;t,\la)&={x},\\
S_2(x;t,\la)&=x^2-\Ph,\\
S_3(x;t,\la)&=x^3+\frac {t\Ph +\la+1}{2\Ph}\,x,\\
S_4(x;t,\la)&=x^4+{\frac {2t\Ph^{2}-({t}^{2}+2)\Ph -(\la+1) t}{ 2(2\Ph^{2}-t\Ph -\la-1)}}\,x^2 -{\frac {2(\la+2)\Ph^2-(\la+1) t\Ph-(\la+1)^2}{2(2\Ph^{2}-t\Ph -\la-1)}},\\
S_5(x;t,\la)&=x^5-{\frac {2(\la+3)t\Ph^2-(\la+1) (t^2-2)\Ph-(\la+1)^2t}{4(\la+2)\Ph^2-2(\la+1) t\Ph-2(\la+1)^2}}\,x^3\\
&\qquad -{\frac {\big[2(\la+2)^2-t^2\big]\Ph^{2}-(\la+1)(\la+4)t\Ph -(\la+1)^2(\la+3)}{4(\la+2)\Ph^2-2(\la+1) t\Ph-2(\la+1)^2}}\,x.\\
\end{split}$$
[\[lem42\]The function $\Ph(t)$ defined by satisfies the Riccati equation $$\label{eq:Phi}{\frac{\d^{}{\Ph}}{{\d{t}}^{}}}=-\Ph^2+\tfrac12t\Ph+\tfrac12(\la+1),$$ and has the asymptotic expansion as $t\to\infty$ $$\label{Phi:expan}\Ph(t)=\tfrac12t+\sum_{n=1}^{\infty}\frac{a_n}{t^{2n-1}},$$ where the constants $a_n$ are given by the nonlinear recurrence relation $$a_{n+1}=2(2n-1)a_{n}-2\sum_{j=1}^{n}a_ja_{n+1-j},$$ with $a_1=\la$. In particular, as $t\to\infty$ $$\label{Phasyn}
\Ph(t)=\tfrac12 t+\frac{\la}{t} +\frac{2\la(1-\la)}{t^3}+\frac{4\la(\la-1)(2\la-3)}{t^5}+\mathcal{O}\big(t^{-7}\big).$$ ]{}
Letting $\ds\Ph(t)={\frac{\d^{}{}}{{\d{t}}^{}}}\ln \phi_{\la}(t)$ in yields $${\frac{\d^{2}{\phi_{\la}}}{{\d{t}}^{2}}}-\tfrac12t{\frac{\d^{}{\phi_{\la}}}{{\d{t}}^{}}}-\tfrac12(\la+1)\phi_{\la}=0,$$ which has solution $$\phi_{\la}(t)=\left\{C_1{D_{-\la-1}}\big(-\tfrac12\sqrt2\,t\big)+C_2{D_{-\la-1}}\big(\tfrac12\sqrt2\,t\big)\right\}\exp(\tfrac18t^2),$$ with $C_1$ and $C_2$ arbitrary constants. Hence setting $C_1=1$ and $C_2=0$ gives the solution and shows that $\Ph(t)$ satisfies .
Substituting into gives $$\begin{aligned}
\sum_{n=1}^{\infty}\frac{(2n-1)a_n}{t^{2n}} &=\tfrac12t \sum_{n=1}^{\infty}\frac{a_n}{t^{2n-1}}
+\left(\sum_{n=1}^{\infty}\frac{a_n}{t^{2n-1}}\right)^2-\tfrac12\la\\
&=a_1-\la+\tfrac12\sum_{n=1}^{\infty}\frac{a_{n+1}}{t^{2n}}+\sum_{n=1}^{\infty}\frac{1}{t^{2n}}\sum_{j=1}^{n}a_ja_{n+1-j},\end{aligned}$$ hence, comparing coefficients of powers of $t$ gives $a_1=\la$ and $$a_{n+1}=2(2n-1)a_{n}-2\sum_{j=1}^{n}a_ja_{n+1-j},$$ as required. Hence $$\begin{split}a_1&=\la,\qquad a_2=-2\la( \la-1),\qquad a_3=4\la(\la-1)(2\la-3),
\end{split}$$ which gives as required.
[Let $h_n(t;\la)$ be defined by $$\label{def:Hn}
H_n(t;\la) = {\frac{\d^{}{}}{{\d{t}}^{}}}\ln\tau_n(t;\la),$$ where $\tau_n(t;\la)$ is the Wronskian given by $$\tau_n(t;\la)=\mathcal{W}\left(\phi_{\la},{\frac{\d^{}{\phi_{\la}}}{{\d{t}}^{}}},\ldots,{\frac{\d^{n-1}{\phi_{\la}}}{{\d{t}}^{n-1}}}\right),$$ with $$\phi_{\la}(t)=\frac{\Gamma(\la+1)}{2^{(\la+1)/2}}\,\exp\left(\tfrac18t^2\right){D_{-\la-1}}\big(-\tfrac12\sqrt2\,t\big).$$ Then $H_n(t;\la)$ satisfies the second-order, second-degree equation $$\begin{aligned}
\left({\frac{\d^{2}{H_n}}{{\d{t}}^{2}}}\right)^2&-\tfrac14\left(t{\frac{\d^{}{H_n}}{{\d{t}}^{}}}-H_n\right)^2+{\frac{\d^{}{H_n}}{{\d{t}}^{}}}\left(2{\frac{\d^{}{H_n}}{{\d{t}}^{}}}-n\right)\left(2{\frac{\d^{}{H_n}}{{\d{t}}^{}}}-n-\la\right)=0.\label{eq:Hn}\end{aligned}$$ ]{}
[Equation is equivalent to , the $\sigma$-equation $$\label{eq:sPIV}
\left({\frac{\d^{2}{\sigma}}{{\d{z}}^{2}}}\right)^{2} - 4\left(z{\frac{\d^{}{\sigma}}{{\d{z}}^{}}}-\sigma\right)^{2} +4{\frac{\d^{}{\sigma}}{{\d{z}}^{}}}\left({\frac{\d^{}{\sigma}}{{\d{z}}^{}}}+2\theta_0\right)
\left({\frac{\d^{}{\sigma}}{{\d{z}}^{}}}+2\theta_{\infty}\right)=0,$$ as shown in [@refCJ Theorem 4.11]. Equation is the same as equation (4.15) in [@refCJ]. Special function solutions of in terms of parabolic cylinder functions have been classified in [@refFW01; @refOkamotoPIIPIV]; see also [@refCJ Theorem 3.5]. ]{}
We remark that equation , and hence also equation , is equivalent to equation SD-I.c in the classification of second order, second-degree differential equations with the property by Cosgrove and Scoufis [@refCS], an equation first derived and solved by Chazy [@refChazy11].
As $t\to\infty$, the recurrence coefficient $\b_n(t)$ has the asymptotic expansion
$$\begin{aligned}
\b_{2n}(t;\la)&=\frac{n}{t}-\frac{2n(2\la-n+1)}{t^3}+\mathcal{O}\big(t^{-5}\big),\\
\b_{2n+1}(t;\la)&=\frac{t}{2}+\frac{\la-n}{t}-\frac{2(\la^2-4\la n+n^2-\la-n)}{t^3}+\mathcal{O}\big(t^{-5}\big),\end{aligned}$$
for $n\in\N$.
In terms of the function $H_n(t;\la)$ defined by , the recurrence coefficients are given as
\[bn:asym\] $$\begin{aligned}
\b_{2n}(t;\la) &= H_n(t;\la+1)-H_n(t;\la),\\ \b_{2n+1}(t;\la) &= H_{n+1}(t;\la)-H_n(t;\la+1).\end{aligned}$$
As $t\to\infty$, $H_n(t;\la)$ has the asymptotic expansion $$\label{asym:Hn}
H_n(t;\la)=\frac{nt}{2}+{\frac {n\la}{t}}+{\frac {2n\la(n- \la)}{{t}^{3}}}+\mathcal{O}\big(t^{-5}\big),$$ for $n\in\N$, see [@refCJ Lemma 5.2]; note that the functions $\Delta_n(t)$ and $S_n(t)$ in [@refCJ] are the same as our functions $\tau_n(t;\la)$ and $H_n(t;\la)$, respectively. Substituting in immediately gives the result.
The differential-difference equation satisfied by generalized Freud polynomials {#sec:diff-differential}
-------------------------------------------------------------------------------
In this section we derive a differential-difference equation satisfied by generalized Freud polynomials using our results in §\[sec:scOPs\].
[\[lem31\]For the generalized Freud weight the monic orthogonal polynomials $S_{n}(x;t)$ satisfy $$\begin{aligned}
\int_{-\infty}^\infty \vvxy&{S_n^2(y;t)}\,\ww{y}\,dy
=4\big(x^2-\tfrac12t +\b_n+\b_{n+1}\big)h_n, \label{int1a}\\
\int_{-\infty}^\infty \vvxy& {S_n(y;t)S_{n-1}(y;t)}\,\ww{y}\,dy = 4xh_n,\label{int1b}\end{aligned}$$ where $$\vvxy=\frac{\vx-\vy}{x-y},$$ with $v(x;t)=x^4-tx^2$ and $$\label{def:hn} h_n = \int_{-\infty}^\infty {S_n^2(y;t)\ww{y}}\,dy.$$]{}
For the generalized Freud weight we have $$\ww{x}=|x|^{2\la+1}\exp\left(-x^4+tx^2\right),$$ i.e.$v(x;t)=x^4-tx^2$, and so $$\vvxy= 4x^2+4xy+4y^2-2t.$$ Hence for $$\begin{aligned}
\int_{-\infty}^\infty \vvxy {S_n^2(y;t)}\,\ww{y}\,dy &=
(4x^2-2t)\int_{-\infty}^\infty {S_n^2(y;t)}\,\ww{y}\,dy \\&\qquad + 4x\int_{-\infty}^\infty {yS_n^2(y;t)}\,\ww{y}\,dy + 4\int_{-\infty}^\infty {y^2S_n^2(y;t)}\,\ww{y}\,dy\\
&= (4x^2-2t)h_n +4x\int_{-\infty}^\infty {S_n(y;t)}\big[S_{n+1}(y;t)+\b_nS_{n-1}(y;t)\big]\ww{y}\,dy\\ &\qquad
+ 4\int_{-\infty}^\infty \big[S_{n+1}(y;t) + \b_nS_{n-1}(y;t)\big]^2\ww{y}\,dy\\
&= (4x^2-2t)h_n + 4h_{n+1}+ 4\b_n^2h_{n-1}\\
&= 4\big(x^2-\tfrac12t +\b_n+\b_{n+1}\big)h_n,\end{aligned}$$ since $\b_n=h_n/h_{n-1}$, the monic orthogonal polynomials $S_{n}(x;t)$ satisfy the three-term recurrence relation , and are orthogonal, i.e. $$\int_{-\infty}^\infty {S_m(y;t)}S_{n}(y;t)\ww{y}\,dy=0,\qquad{\rm if}\quad m\neq n.\label{Pnorth}$$ Also for $$\begin{aligned}
\int_{-\infty}^\infty \vvxy &{S_n(y;t)S_{n-1}(y;t)}\,\ww{y}\,dy\nonumber\\ &=
(4x^2-2t) \int_{-\infty}^\infty {S_n(y;t)S_{n-1}(y;t)}\,\ww{y}\,dy + 4x \int_{-\infty}^\infty {yS_n(y;t)S_{n-1}(y;t)}\,\ww{y}\,dy \\ &\qquad
+ 4 \int_{-\infty}^\infty {y^2S_n(y;t)S_{n-1}(y;t)}\,\ww{y}\,dy \\
&= 4x \int_{-\infty}^\infty S_{n}(y;t)\big[S_{n}(y;t)+\b_{n-1}S_{n-2}(y;t)\big] \ww{y}\,dy \\
&\qquad +4 \int_{-\infty}^\infty \big[S_{n+1}(y;t)+\b_nS_{n-1}(y;t)\big] \big[S_{n}(y;t)+\b_{n-2}S_{n-2}(y;t)\big]\ww{y}\,dy\\
&=4xh_n,\end{aligned}$$ using the recurrence relation and orthogonality .
For the generalized Freud weight the monic orthogonal polynomials $S_{n}(x;t)$ satisfy the differential-difference equation $$\label{eq:Snddeq}
x{\frac{\d^{}{S_n}}{{\d{x}}^{}}}(x;t)=-B_n(x;t)S_n(x;t)+A_n(x;t)S_{n-1}(x;t),$$ where
\[AnBn\]$$\begin{aligned}
A_n(x;t)&=4x\b_n(x^2-\tfrac12t+\b_n+\b_{n+1}),\\
B_n(x;t)&=4x^2\b_n+(\la+\tfrac12)[1-(-1)^n],\end{aligned}$$
with $\b_n$ the recurrence coefficient in the three-term recurrence relation .
Corollary \[cor:ABn\] shows that monic orthogonal polynomials $S_{n}(x;t)$ with respect to the weight $$\ww{x}=|x|^{2\la+1}\exp\{-v(x;t)\},$$ satisfy the differential-difference equation , where
\[def:AnBn\] $$\begin{aligned}
A_n(x;t)&=\frac{x}{h_{n-1}}\int_{-\infty}^\infty \vvxy {S_n^2(y;t)}w(y;t)\,dy,\\
B_n(x;t)&=\frac{x}{h_{n-1}}\int_{-\infty}^\infty \vvxy {S_n(y;t)S_{n-1}(y;t)} w(y;t)\,dy +(\la+\tfrac12)[1+(-1)^n] .\end{aligned}$$
For the generalized Freud weight , using Lemma \[lem31\] yields the result.
The differential equation satisfied by generalized Freud polynomials {#sec:diffeqn}
--------------------------------------------------------------------
Now we derive a differential equation satisfied by generalized Freud polynomials.
\[thm43\]For the generalized Freud weight the monic orthogonal polynomials $S_{n}(x;t)$ satisfy the differential equation $$\label{eq:Snode}
x{\frac{\d^{2}{S_n}}{{\d{x}}^{2}}}(x;t)+R_n(x;t){\frac{\d^{}{S_n}}{{\d{x}}^{}}}(x;t)+T_n(x;t)S_n(x;t)=0,$$ where
$$\begin{aligned}
R_n(x;t) &=-4x^4+2tx^2+{2\la+1}-\frac{2x^2}{x^2-\tfrac12t+\b_n+\b_{n+1}},\\[5pt]
T_n(x;t) &= 4nx^3+16x\b_n(\b_n+\b_{n+1}-\tfrac12t)(\b_n+\b_{n-1}-\tfrac12t)\nonumber\\
&\qquad+4x[1+(2\la+1)(-1)^n]\b_n -\frac{8\b_n x^3+(2\la+1)[1-(-1)^n]x}{x^2-\tfrac12t+\b_n+\b_{n+1}}\nonumber\\
&\qquad+(2\la+1)[1-(-1)^n]x\left(t-\frac{1}{2x^2}\right).\end{aligned}$$
In Theorem \[thm:gende\] we proved that the coefficients in the differential equation satisfied by polynomials orthogonal with respect to the weight $$\ww{x}=|x-k|^{\ga}\exp\{-\v(x;t)\},$$ are given by and . For the generalized Freud weight we use and with $k=0$, $\ga=2\la+1$, $\v(x;t)=x^4-tx^2$, and $A_n$ and $B_n$ given by to obtain the stated result.
We note that if $\{P_n(x)\}_{n=0}^{\infty}$, is a sequence of *classical* orthogonal polynomials (such as Hermite, Laguerre and Jacobi polynomials), then $P_n(x)$ satisfies second-order ordinary differential equation $$\label{eq:Pn}
\sigma(x){\frac{\d^{2}{P_n}}{{\d{x}}^{2}}}+\tau(x){\frac{\d^{}{P_n}}{{\d{x}}^{}}}=\la_nP_n,$$ where $\sigma(x)$ is a monic polynomial with deg$(\sigma)\leq2$, $\tau(x)$ is a polynomial with deg$(\tau)=1$, and $\la_n$ is a real number which depends on the degree of the polynomial solution, see Bochner [@refBochner]. For [classical]{} orthogonal polynomials, the polynomials $\sigma(x)$ and $\tau(x)$ are the same as in the associated Pearson equation . In contrast the coefficients in second-order ordinary differential equation satisfied by the polynomials for the generalized Freud weight given in Theorem \[thm43\] are not the same as the polynomials $\sigma(x)=x$ and $\tau(x)=-4x^4+2tx^2+2\la+1$ in the Pearson equation satisfied by the generalized Freud polynomials since $$\nonumber
\dfrac{w'(x;t)}{w(x;t)} =\frac{\tau(x)-\sigma'(x)}{\sigma(x)}=-4x^{3}+2tx + \dfrac{2\la+1}{x}, \quad x\in \R\backslash \lbrace 0\rbrace.$$
An alternative method due to Shohat {#sho}
-----------------------------------
It is shown in [@Nevai] and [@refShohat39] that the monic orthogonal polynomials $S_{n}(x;t)$ orthogonal with respect to the generalized Freud weight are quasi-orthogonal of order $m=5$ and hence we can write $$\begin{aligned}
\label{quasi}
x{\frac{\d^{}{S_n}}{{\d{x}}^{}}}(x;t) &= \sum_{k=n-4}^{n}c_{n,k} S_k(x;t),\end{aligned}$$ where the coefficient $c_{n,k}$ is given by $$\begin{aligned}
\label{Ccoeff}
c_{n,k} &= \frac{1}{h_{k}} \int_{-\infty}^{\infty} x{\frac{\d^{}{S_n}}{{\d{x}}^{}}}(x;t) S_k(x;t) \ww{x} \,d{x},\end{aligned}$$ for $n-4 \leq k\leq n$ and $h_k\neq 0$.
Integrating by parts, we obtain for $n-4 \leq k\leq n-1$, $$\begin{aligned}
\nonumber
h_k c_{n,k} &=\Big[ xS_k(x;t) S_n(x;t) w(x;t)\Big]_{-\infty}^{\infty}
- \int_{-\infty}^\infty {\frac{\d^{}{}}{{\d{x}}^{}}}\left[ x S_k (x;t) \ww{x} \right]S_n(x;t) \,d{x} \nonumber\\
&=- \int_{-\infty}^{\infty} \left[S_n(x;t) S_k(x;t)
+ xS_n(x;t) {\frac{\d^{}{S_k}}{{\d{x}}^{}}}(x;t)\right] w(x;t) \,d{x}
- \int_{-\infty}^{\infty} xS_n(x;t) S_k(x;t) {\frac{\d^{}{w}}{{\d{x}}^{}}}(x;t) \,d{x},
\nonumber\\
&= \nonumber - \int_{-\infty}^{\infty} { xS_n(x;t) S_k(x;t) {\frac{\d^{}{w}}{{\d{x}}^{}}}(x;t)}\,d{x}\\
&= -\int_{-\infty}^{\infty} {S_n(x;t) S_k(x;t) \left( -4x^{4} +2tx^2 +2\la +1\right) } \ww{x}\,d{x}\nonumber
\\&=
\int_{-\infty}^{\infty} \big(4x^4-2tx^2\big) S_n(x;t) S_k(x;t) \ww{x} \,d{x},\label{maineq}
\end{aligned}$$ since $$x{\frac{\d^{}{w}}{{\d{x}}^{}}}(x;t) =\big(-4x^4+2x^2 + 2\la+1\big)\ww{x}.$$ Iterating the three-term recurrence relation , the following relations are obtained
\[recurrence13\] $$\begin{aligned}
x^2 S_n = S_{n+2} &+( \b_{n} + \b_{n+1}) S_{n} + \b_{n} \b_{n-1} S_{n-2},\label{recurrence1}\\
x^4 S_n = S_{n+4} &+ ( \b_{n}+ \b_{n+1} + \b_{n+2} +\b_{n+3}) S_{n+2}
\nonumber\\&
+ \big[\b_{n} ( \b_{n-1} + \b_{n} + \b_{n+1}) +\b_{n+1} ( \b_{n} + \b_{n+1} + \b_{n+2} )\big] S_{n}
\nonumber\\& +\b_{n} \b_{n-1}( \b_{n-2} +\b_{n-1} + \b_{n} + \b_{n+1} ) S_{n-2}
+ ( \b_{n} \b_{n-1} \b_{n-2} \b_{n-3} ) S_{n-4}.\label{recurrence3}\end{aligned}$$
Substituting into yields the coefficients $\lbrace c_{n,k} \rbrace_{k=n-4}^{n-1}$ in .
\[Aau\] $$\begin{aligned}
c_{n,n-4} & = 4 \b_{n} \b_{n-1} \b_{n-2} \b_{n-3},\\
c_{n,n-3} & =0,\\
c_{n,n-2} &= 4\b_{n} \b_{n-1} (\b_{n-2} + \b_{n-1} + \b_{n} + \b_{n+1}- \tfrac12t),\\
c_{n,n-1} & =0.
\end{aligned}$$
Lastly we consider the case when $k=n$. Integration by parts in yields $$\begin{aligned}
h_n c_{n,n} &=\int_{-\infty}^{\infty} x{\frac{\d^{}{S_n}}{{\d{x}}^{}}}(x;t) S_n(x;t) \ww{x} \,d{x}\nonumber\\
&=-\tfrac12\int_{-\infty}^{\infty} S_n^2(x;t) \left[\ww{x} + x{\frac{\d^{}{w}}{{\d{x}}^{}}}(x;t)\right] d{x}
\nonumber\\&= - \tfrac12h_n
+\int_{-\infty}^{\infty} S_n^2(x;t)\big(2x^{4} -tx^2 -\la -\tfrac12\big)\,\ww{x} \,d{x}
\nonumber\\&= \int_{-\infty}^{\infty} \big(2x^4-tx^2\big) S_n^2(x;t)\,\ww{x} \,d{x} - (\la+1) h_{n}. \label{eq:cnn1}
\end{aligned}$$ From the three-term recurrence relation , we have $$\begin{aligned}
x^2 S_n^2&=(S_{n+1}+\b_n S_{n-1})^2 \\
&= S_{n+1}^2+2\b_nS_{n+1}S_{n-1}+\b_n^2S_{n-1}^2\\
x^4 S_n^2&= x^2\big(S_{n+1}^2+2\b_nS_{n+1}S_{n-1}+\b_n^2S_{n-1}^2)\\
&= x^2 S_{n+1}^2 +2\b_n(xS_{n+1})(xS_{n-1})+\b_n^2x^2S_{n-1}^2\\
&= \big(S_{n+2}+\b_{n+1}S_{n}\big)^2 + 2\b_n\big(S_{n+2}+\b_{n+1}S_{n}\big)\big(S_{n}+\b_{n-1}S_{n-2}\big)
+\b_n^2\big(S_{n}+\b_{n-1}S_{n-2}\big) ^2\\
&= S_{n+2}^2+2(\b_{n+1}+\b_n)S_{n+2}S_{n}+(\b_{n+1}+\b_n)^2 S_{n}^2+2\b_n\b_{n-1}S_{n+2}S_{n-2}
\\ &\qquad\quad
+2\b_n\b_{n-1}(\b_n+\b_{n+1})S_{n}S_{n-2}+\b_n^2\b_{n-1}^2S_{n-2}^2
\end{aligned}$$ and so by orthogonality $$\begin{aligned}
\int_{-\infty}^{\infty} x^2 S_n^2(x;t)\,\ww{x} \,d{x} \
&= h_{n+1}+\b_n^2 h_{n-1} = (\b_{n+1}+\b_{n})h_{n},\label{eq:cnn2} \\
\int_{-\infty}^{\infty} x^4 S_n^2(x;t)\ww{x} \,d{x} &= h_{n+2}+(\b_{n+1}+\b_n)^2h_n +\b_n^2\b_{n-1}^2h_{n-2}\nonumber\\
&= \b_{n+2}\b_{n+1}h_n+(\b_{n+1}+\b_n)^2h_n +\b_n\b_{n-1}h_n \nonumber\\
&=\big[(\b_{n+1}+\b_{n}+\b_{n-1})\b_n + (\b_{n+2}+\b_{n+1}+\b_n)\b_{n+1}\big]h_n\nonumber\\
&=\tfrac12 \big[t(\b_{n+1}+\b_{n})+n+\la+1\big]h_{n},\label{eq:cnn3}
\end{aligned}$$ using $h_{n+1}=\b_{n+1}h_n$ and . Hence from , and we have $$\begin{aligned}
c_{n,n}&= t(\b_{n+1}+\b_{n})+n+\la+1-t(\b_{n+1}+\b_{n})- (\la+1)\nonumber\\
&=n. \label{eq:cnn4}\end{aligned}$$
Combining with , we write $$\begin{aligned}
\label{AQquasi}
x{\frac{\d^{}{S_n}}{{\d{x}}^{}}}(x;t) = c_{n,n-4} S_{n-4}(x;t) + c_{n,n-2} S_{n-2}(x;t) + c_{n,n} S_{n}(x;t).\end{aligned}$$ In order to express $S_{n-4}$ and $S_{n-2}$ in in terms of $S_n$ and $S_{n-1}$, we iterate to obtain $$\begin{aligned}
\label{recoa}
S_{n-2}&= \dfrac{xS_{n-1}-S_n}{\b_{n-1}}, \\
S_{n-3}&= \dfrac{xS_{n-2}-S_{n-1}}{\b_{n-2}} = \dfrac{x^2-\b_{n-1}}{\b_{n-1}\b_{n-2}}\, S_{n-1} - \dfrac{x}{\b_{n-1}\b_{n-2}}\, S_{n},\\ S_{n-4}
&= \dfrac{xS_{n-3} - S_{n-2}}{\b_{n-3}} = \label{recoab} \dfrac{x^3- ( \b_{n-1}+\b_{n-2})x}{\b_{n-1}\b_{n-2}\b_{n-3}}\, S_{n-1} -
\dfrac{x^2-\b_{n-2}}{ \b_{n-1}\b_{n-2}\b_{n-3}} \, S_{n}.
\end{aligned}$$ Substituting , , and into yields $$\begin{aligned}
\label{Aomega}
x{\frac{\d^{}{S_n}}{{\d{x}}^{}}}(x;t)= -B_n(x) S_n(x;t) + A_n(x;t) S_{n-1}(x;t),\end{aligned}$$ where $A_n(x)$ and $B_n(x)$ are given by .
Conclusion {#sec:concl}
==========
In this paper, for the generalized Freud weight we have obtained explicit expressions for the coefficients of the three-term recurrence relation and differential-difference equation satisfied by generalized Freud polynomials. We also proved that the generalized Freud polynomials satisfy a linear ordinary differential equation. We note that the closed form expressions for the coefficients provided allow investigation of other properties, including properties of the zeros such as monotonicity, convexity and inequalities satisfied by the zeros. However, although the expressions for the coefficients given in this paper are explicit, they are rather complicated and given in terms of special function solutions of the fourth equation which does not necessarily lead to elegant results in applications. For this reason, a natural extension of this work would be an investigation of asymptotic properties using limiting relations satisfied by the polynomials as the parameters $t$ and/or $\la$ tend to $\infty$.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank the London Mathematical Society for support through a “Research in Pairs" grant. PAC thanks Alexander Its, Ana Loureiro, and Walter van Assche for their helpful comments and illuminating discussions. KJ thanks the School of Mathematics, Statistics & Actuarial Science at the University of Kent for their hospitality during her visit when some of this research was done. The research by KJ was partially supported by the National Research Foundation of South Africa.
\#1\#2\#3\#4\#5\#6\#7[ , , *\#4*, **\#5** (\#7) \#6.]{}
\#1\#2\#3\#4\#5\#6\#7[ , , *\#4*, \#6.]{} \#1\#2\#3\#4\#5[ , *\#3*, \#4, \#5.]{}
\#1\#2\#3\#4\#5\#6[ , , in *\#4*, [\#5]{}, \#6.]{}
[99]{}
\[Reprinted by Dover Publications, 2011.\]
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We show that a linear algebraic group is the Galois group of a parameterized Picard-Vessiot extension of $k(x), \ x''=1$, for certain differential fields $k$, if and only if its identity component has no one dimensional quotient as a linear algebraic group.'
author:
- 'Michael F. Singer[^1]'
date:
-
-
title: 'Linear Algebraic Groups as Parameterized Picard-Vessiot Galois Groups'
---
\[section\] \[section\] \[thm\][Lemma]{} \[thm\][Corollary]{} \[thm\][Proposition]{} \[thm\][Definition]{} \[thm\][Definitions]{} \[thm\][Remark]{} \[thm\][Remarks]{} \[thm\][Example]{} \[thm\][Examples]{}
Introduction
============
In the usual Galois theory of polynomial equations, one starts with a polynomial having coefficients in a field[^2] $k$, forms a splitting field $K$ of this polynomial and then defines the Galois group of this equation to be the group of field automorphisms of $K$ that leave $k$ element-wise fixed. A natural inverse question then arises: [*Given the field $k$, which groups can occur as Galois groups*]{}. For example, if $k= C(x), C$ an algebraically closed field and $x$ transcendental over $C$, any finite group occurs as a Galois group (Corollary 7.10,[@volklein]). In the Galois theory of linear differential equations, one starts with a homogeneous linear differential equation with coefficients in a differential field $k$ with algebraically closed constants $C$, forms a Picard-Vessiot extension $K$ (the analogue of a splitting field) and defines the Galois group of the linear differential equation to be the differential automorphisms of $K$ that leave $k$ element-wise fixed. This Galois group is a linear algebraic group defined over $C$ and one can again seek to determine which groups occur as the Galois group of a homogeneous linear differential equation over a given differential field. For example, if $k = C(x), C$ an algebraically closed field, $x' = 1$ and $c'=0$ for all $c \in C$, then any linear algebraic group occurs as Galois group of a Picard-Vessiot extension of $k$ ([@hartmann2005; @hartmann2007] for proofs of this as well as references to earlier work). Besides putting the Picard-Vessiot theory on a firm modern footing, Kolchin developed a generalization of Picard-Vessiot extensions called [*strongly normal extensions*]{} and developed a Galois theory for these fields (see [@DAAG] for an exposition and references to the original articles and [@kovacic03] for a reworking of this theory in terms of differential schemes). The Galois groups of these extensions can be arbitrary algebraic groups. Kovacic [@kovacic69; @kovacic71] studied the general inverse problem in the context of strongly normal extensions and showed that this problem can be reduced to the inverse problem for linear algebraic groups and for abelian varieties. If $k = C(x)$ as above, Kovacic showed that any abelian variety can be realized and, combining this with the solution for linear algebraic groups described above, one sees that any algebraic group defined over $C$ can be realized as a Galois group of a strongly normal extension of $C(x)$ (Kovacic also solved the inverse problem for connected solvable linear algebraic groups and laid out a general plan for attacking the inverse problem for linear groups over arbitrary fields).\
In [@Landesman], Landesman developed a new Galois theory generalizing Kolchin’s theory of strongly normal extension to include, for example, certain differential equations that contain parameters. The Galois groups appearing here are differential algebraic groups (as in [@kolchin_groups]). A special case was developed in [@CaSi] where the authors consider parameterized [*linear*]{} differential equations and discuss various properties of the associated Galois groups, named parameterized Picard-Vessiot groups or PPV-groups for short. These latter groups are linear differential algebraic groups in the sense of Cassidy [@cassidy1], that is, groups of matrices whose entries belong to a differential field and satisfy a fixed set of differential equations. The inverse problem in these theories is not well understood. Landesman showed that any connected differential algebraic group is a Galois group in his theory over [*some*]{} differential field that may depend on the given differential algebraic group (Theorem 3.66, [@Landesman]). The analogue of the field $C(x)$ mentioned above is a field $k_0(x)$ with commuting derivations $\Delta = \{\d_x,\d_1, \ldots, \d_m\}, \ m\geq 1,$ where $k_0$ is a differentially closed (see the definition below) $\Pi = \{\d_1, \ldots ,\d_m\}$-differential field, $x$ is transcendental over $k_0$, $\d_i(x) = 0$ for $i=1, \ldots ,m$ and $\d_x$ is defined on $k$ by setting $\d_x(a) = 0$ for all $a \in k_0$ and $\d_x(x) = 1$. It is not known, in general, which differential algebraic groups appear as Galois groups in Landesman’s theory over this field. In [@Landesman] and [@CaSi], it is shown that the additive group $\Ga(k_0)$ cannot appear while any proper subgroup of these groups does appear as a Galois group (the same situation for $\Gm(k_0)$ is also described in [@Landesman])[^3].\
In this paper, further progress is made on the inverse problem for the parameterized Picard-Vessiot theory over the field $k=k_0(x)$ described in the previous paragraph. In the following, I characterize those linear [*algebraic*]{} groups, considered as linear [*differential*]{} algebraic groups, that can occur as PPV-groups of PPV-extensions of $k$ (under suitable hypotheses concerning $k$). Before I state the main result of this paper, I will recall some definitions. Although these definitions may be stated in more generality, I will state them relative to the field $k$ defined above.\
The parameterized Picard-Vessiot theory (PPV-theory) considers linear differential equations of the form $$\begin{aligned}
\label{ppveqn}\d_xY &=& AY\end{aligned}$$ where $A \in \gl_n(k)$. In analogy to classical Galois theory and Picard-Vessiot theory, we consider fields, called [*PPV-extensions of $k$*]{}, that act as “splitting fields” for such equations. A PPV-extension $K$ of $k$ for (\[ppveqn\]) is a $\Delta$-field $K$ such that
1. $K = k\langle Z \rangle$, the $\Delta$-field generated by the entries of a matrix $Z \in \gl_n(K)$ satisfying $\d_xZ = AZ, \ \det(Z) \neq 0$.
2. $K^{\d_x} = k^{\d_x} = k_0$, where for any $\Delta$-extension $F$ of $k$, $F^{\d_x} = \{c \in F \ | \ \d_xc=0\}$.
A $\Pi$-field $E$ is said to be [*differentially closed*]{} (also called [*constrainedly closed*]{}, see, for example §9.1 of [@CaSi]) if for any $n$ and any set $\{P_1(y_1, \ldots, y_n), \ldots, P_r(y_1, \ldots ,y_n), \linebreak Q(y_1, \ldots
,y_n)\} \subset E\{y_1, \ldots , y_n\}$, the ring of differential polynomials in $n$ variables, if the system $$\{P_1(y_1, \ldots, y_n)=0, \ldots ,
P_r(y_1, \ldots ,y_n)=0, Q(y_1, \ldots ,y_n)\neq 0\}$$ has a solution in some differential field $F$ containing $E$, then it has a solution in $E$. In [@CaSi] (and more generally in [@HaSi08]), it is shown that under the assumption that $k_0$ is differentially closed, then PPV-extensions exist and are unique up to $\Delta$-$k$-isomorphisms. This hypothesis has been weakened to non-differentially closed $k_0$ in [@GGO] and [@wibmer]. In these papers the authors give conditions weaker than differential closure for the existence and uniqueness of PPV-extensions and and discuss the corresponding Galois theory. Although some of our results remain valid under these weaker hypotheses, we will assume in this paper that $k_0$ is $\Pi$-differentially closed. The set of field-theoretic automorphisms of $K$ that leave $k$ elementwise fixed and commute with the elements of $\Delta$ forms a group $G$ called the [*parameterized Picard-Vessiot group* ]{} (PPV-group) of (\[ppveqn\]). One can show that for any $\sigma \in G$, there exists a matrix $M_\sigma \in \GL_n(k_0)$ such that $\sigma(Z) =(\sigma(z_{i,j})) = ZM_\sigma$. Note that $\d_x$ applied to an entry of such an $ M_\sigma$ is $0$ since these entries are elements of $k_0$ but that such an entry need not be constant with respect to the elements of $\Pi$. In [@CaSi], the authors show that the map $\sigma \mapsto M_\sigma$ is an isomorphism whose image is furthermore a [*linear differential algebraic group*]{}, that is, a group of invertible matrices whose entries satisfies some fixed set of polynomial [*differential*]{} equations (with respect to the derivations $\Pi = \{\d_1, \ldots, \d_m\}$) in $n^2$ variables. We say that a set $X \subset \GL_n(k_0)$ is [*Kolchin-closed*]{} if it is the zero set of such a set of polynomial differential equations. One can show that the Kolchin-closed sets form the closed sets of a topology, called the [*Kolchin topology*]{} on $\GL_n(k_0)$ ([*cf.*]{} [@cassidy1; @cassidy6; @CaSi; @kolchin_groups]).\
One more definition is needed before stating the main result of this paper, Theorem \[thm1\]. A $\Pi$-field $F$ is a [*$\Pi$-universal field* ]{} if for any $\Pi$-field $E \subset F$, finitely differentially generated over $\QX$, any $\Pi$-finitely generated extension of $E$ can be differentially embedded over $E$ into $F$ ([@DAAG], p. 133). Note that a universal field is differentially closed.
\[thm1\]Let $k$ be as above and $G(k_0)$ the group of $k_0$-points of a linear algebraic group $G$ defined over $k_0$.
1. If $k_0$ is $\Pi$-differentially closed and $G(k_0)$ is a PPV-group of a PPV-extension of $k$, then the identity component of $G$ has no quotient (as an algebraic group) isomorphic to the additive group $\Ga$ or the multiplicative group $\Gm$.
2. If $k_0$ is a $\Pi$-universal field and the identity component of $G$ has no quotient (as an algebraic group) isomorphic to the additive group $\Ga$ or the multiplicative group $\Gm$, then $G(k_0)$ is a PPV-group of a PPV-extension of $k$.
The remainder of this paper is organized as follows. Section \[sec2\] contains the proof of Theorem 1.1. In fact, I prove the stronger result (Proposition \[prop1\]) that if $G$ is a linear [*differential*]{} algebraic group such that $G(k_0)$ is the PPV-group of a PPV-extension of $k$, then the identity component of $G(k_0)$ has no quotient (as a linear differential algebraic group) isomorphic to $\Ga(k_0)$ or $\Gm(k_0)$. In Section \[sec3\], I show that a linear algebraic group contains a Kolchin-dense finitely generated subgroup if and only if it has no quotient (as an algebraic group) isomorphic to the additive group $\Ga$ or the multiplicative group $\Gm$. Theorem 1.2 then follows from the fact that a linear differential algebraic group containing a Kolchin-dense finitely generated subgroup is a PPV-group of a PPV-extension of the field $k$ when $k_0$ is a $\Pi$-universal field. This latter fact is proven in [@MiSi11] using analytic tools. In Section \[sec4\], I show that Theorem 1.2 does not hold for general linear differential algebraic groups, that is, there is a connected linear differential algebraic group $G$ having neither $\Ga$ nor $\Gm$ as a quotient with the further property that $G$ is not the PPV-group of any PPV-extension of $k$. Section \[sec5\] contains some final comments.\
The author wishes to thank Phyllis Cassidy for helpful discussions concerning the content of this paper.
Proof of Theorem 1.1 {#sec2}
====================
In this section we will prove the stronger result (Proposition \[prop1\]) that if $G$ is a linear [*differential*]{} algebraic group such that $G(k_0)$ is the PPV-group of a PPV-extension of $k$, then the identity component of $G(k_0)$ has no quotient (as a linear differential algebraic group) isomorphic to $\Ga(k_0)$ or $\Gm(k_0)$. Let $ k_0$ be a $\Pi$-differentially closed field and let $G$ be a linear differential group defined over $k_0$ and let $G^0$ be its identity component in the Kolchin topology.
\[lem0\] The homomorphism $l\d_1:\Gm(k_0) \rightarrow \Ga(k_0)$ where $l\d_1(u) = \d_1(u)/u$ maps $\Gm(k_0)$ onto $\Ga(k_0)$.
Since $k_0$ is differentially closed, we need only show that for any $u \in k_0$, there is a $\Pi$-differential extension $F$ of $k_0$ such that $\d_1y = u y $ has a solution $y \neq 0$ in $F$. Let $\Pi_1 = \{\d_2, \ldots , \d_m\}$ and let $F$ be the $\Pi_1$-field $k_0\langle v \rangle$, where $v$ is a $\Pi_1$-differentially transcendental element. We extend the derivation $\d_1$ from $k_0$ to $F$ by setting $\d_1 v = uv,$ and $\d_1(\d_2^{i_2}\ldots \d_m^{i_m} v) = \d_2^{i_2}\ldots \d_m^{i_m}(\d_1v) = \d_2^{i_2}\ldots \d_m^{i_m}(uv)$. With these definitions, $F$ becomes a $\Pi$-differential extension of $k_0$ and $y=v$ satisfies $\d_1y = u y $.
\[lem1\] If $G^0(k_0)$ has $\Gm(k_0)$ or $\Ga(k_0)$ as a homomorphic image (under a differential algebraic homomorphism) and $G(k_0)$ is a PPV-group of a PPV-extension of $k_0(x)$, then $\Ga(k_0)$ is a PPV-group of a PPV-extension of a finite algebraic extension $E$ of $k_0(x)$.
I will show that this result follows from the Galois theory of parameterized linear differential equations ([@Landesman],[@CaSi]). Let $K$ be a PPV-extension of $k_0(x)$ having $G$ as its PPV-group. The fixed field $E$ of $G^0$ is a finite algebraic extension of $k_0(x)$. If $G^0$ has $\Gm(k_0)$ as a homomorphic image under a differential homomorphism then composing composing this homomorphism with $l\d_1:\Gm(k_0) \rightarrow \Ga(k_0)$ where $l\d_1(u) = \d_1(u)/u$, Lemma \[lem0\] implies that $\Ga(k_0)$ would also be a homomorphic image of $G^0(k_0)$ under a differential homomorphism. Therefore we shall only deal with this latter case. Let $\phi: G^0(k_0) \rightarrow \Ga(k_0)$ be a surjective differential algebraic homomorphism and let $H$ be its kernel. The Galois theory (Theorem 9.5, [@CaSi]) implies that the fixed field of $H$ is a PPV-extension $F$ of $E$ whose PPV-group over $E$ is differentially isomorphic to $\Ga(k_0)$.
The following lemma is the key to showing that $\Ga(k_0)$ is not a PPV-group over a finite algebraic extension of $k_0(x)$.
\[lem2\] Let $E$ be a finite algebraic extension of $k_0(x)$ and $f \in E$. Let $K$ be the PPV-extension of $k_0(x)$ corresponding to the equation $$\d_x y = f.$$ Let $z \in K$ satisfy $\d_xz = f$. Then there exists a nonzero linear differential operator $L \in k_0[\d_1]$ and an element $g \in E$ such that $$L(z) = g.$$
The proof of this lemma is a slight modification of Manin’s construction of the Picard-Fuchs equations (see Section 3, pp. 64-65 of the English translation of [@manin58]). We shall use (as does Manin) ideas and results that appear in [@chevalley51]. In Ch. VI, §7 of [@chevalley51], Chevalley shows that $\d_1$ can be used to define a map $D$ on differentials of $E$ satisfying $D(ydx) = (\d_1y)dx$. Furthermore, Theorem 13 of Ch. VI, §7 of [@chevalley51] states that for any differential $\omega$ and any place $P$ of $E$, we have $\res_PD(\omega) = \d_1(\res_P\omega)$ (where $\res_P$ denotes the residue at $P$). Let $\alpha_1, \ldots , \alpha_m$ be the non-zero residues of $fdx$ and let $R \in k[\d_1]$ be a nonzero linear differential operator such that $R(\alpha_i) =0, i = 1, \ldots , m$[^4]. We then have that for any place $P$, $\res_P(R(f)dx) = R(\res_P(fdx)) = 0$. Therefore $R(f)dx$ has residue $0$ at all places, that is, it is a differential of the second kind. Note that $ \d_1^i(R(f))dx$ is also a differential of the second kind for any $i\geq1$. The factor space of differentials of the second kind by the space of exact differentials has dimension $2g$ over $k$, where $g$ is the genus of $E$ (Corollary 1, Ch. VI, §8,[@chevalley51]). Therefore there exist $v_{2g}, \ldots v_0 \in k_0$ such that $$v_{2g}\d_1^{2g}(R(f))dx + \ldots +v_0R(f)dx = d\tilde{g} = \d_x\tilde{g} dx$$ for some $\tilde{g} \in E$. This implies that there exists a linear differential operator $L\in k_0[\d_1]$ such that $$L(f) = \d_x\tilde{g}.$$ Furthermore, $\d_x(L(z)) = L(\d_xz) = L(f) = \d_x\tilde{g}$. Therefore $L(z) = g$ where $g = \tilde{g} +c$ for some $c \in k_0$.
\[prop1\] If $G$ is a linear differential algebraic group defined over $k_0$ such that $G^0(k_0)$ has $\Gm(k_0)$ or $\Ga(k_0)$ as a quotient (as a linear differential group), then $G(k_0)$ cannot be a PPV-group of a PPV-extension of $k_0(x)$.
Assume that $G(k_0)$ is a PPV-group of a PPV-extension of $k_0(x)$. Lemma \[lem1\] implies that, in this case, $\Ga(k_0)$ is a PPV-group of a PPV-extension $K$ of $E$, where $E$ is a finite algebraic extension of $k_0(x)$. From Proposition 9.12 of [@CaSi], $K$ is the function field of a $\Ga(k_0)$-principal homogeneous space. The corollary to Theorem 4 of (Ch. VII,§3, [@kolchin_groups]) implies that this principal homogeneous space is the trivial principal homogeneous space and so $K = E\langle z \rangle$ where for any $\sigma \in \Ga(k_0)$ there exists a $c_\sigma \in k_0$ such that $\sigma(z) = z +c_\sigma$. In particular, $\sigma(\d_x z) =
\d_x z$ for all $\sigma \in \Ga(k_0)$ and so $\d_x z = f \in E$. Lemma \[lem2\] implies that there exists a linear differential operator $L \in k_0[\d_1]$ and an element $g \in E$ such that $L(z) = g$. For any $\sigma \in \Ga(k_0)$, we have $g = \sigma(g) = \sigma(L(z)) = L(\sigma(z)) = L(z+ c_\sigma) = g + L(c_\sigma)$ so $L(c_\sigma) = 0$. This implies that the PPV-group of $K$ over $E$ is a proper subgroup of $\Ga(k_0)$, a contradiction.
Theorem 1.1 follows from Propostion \[prop1\] by noting that a linear algebraic group is [*a fortiori*]{} a linear differential algebraic group.
Proof of Theorem 1.2 {#sec3}
====================
The proof of Theorem 1.2 is inspired by [@tretkoff79]. In this latter paper, the authors mix analytic and algebraic facts to show that any linear algebraic group defined over $\CX$ is the Galois group of a Picard-Vessiot extension of $\CX(x)$. Their proof is based on the following facts:
1. Any linear algebraic group contains a finitely generated Zariski-dense subgroup.
2. Let $\PX^1(\CX)$ be the Riemann Sphere and $x_0, x_1, \ldots , x_n$ be distinct points of $\PX^1(\CX)$. If $\rho:\pi_1(\PX^1(\CX)\backslash\{x_1, \ldots , x_m\}, x_0) \rightarrow \GL_n(\CX)$ is a representation of the first homotopy group of the Riemann sphere with $m$ punctures, then there exists a linear differential equation $$\begin{aligned}
\label{eqn1}\frac{dY}{dx} &= &AY, \mbox{ $A$ an $n\times n$ matrix with entries in $\CX(x)$}\end{aligned}$$ with only regular singularities having $\rho$ as its monodromy representation (for some choice of fundamental solution matrix).
3. If equation \[eqn1\] has only regular singular points, then for some choice of fundamental solution matrix, the Galois group of this equation is the smallest linear algebraic group containing the image of the monodromy representation.
With these facts, the authors of [@tretkoff79] proceed as follows. Let $G \subset \GL_n(\CX)$ be a linear algebraic group. Using (1), there exist elements $g_1, \ldots g_m \in G$ that generate a Zarski-dense subgroup of $G$. We can furthermore assume that the $g_i$ have been chosen so that $\prod_{i=1}^mg_i = 1$. Let $x_0, x_1, \ldots , x_n$ be distinct points of $\PX^1(\CX)$ and let $\gamma_i, i=1, \ldots , m$, be the obvious loops, starting and ending at $x_0$ that each enclose a unique $x_i$. The map $\rho:\gamma_i \mapsto g_i$ defines a homomorphism $\rho:\pi_1(\PX^1(\CX)\backslash\{x_1, \ldots , x_m\}, x_0) \rightarrow G \subset \GL_n(\CX)$. From (2), we can conclude that there is a linear differential equation (\[eqn1\]) with only regular singular points having $\rho$ as its monodromy representation. From (3), we conclude that $G$ is the Galois group of this equation.\
When one tries to mimic this proof in the parameterized case, one immediately is confronted with the fact that there are linear differential algebraic groups that have no finitely generated Kolchin-dense subgroups, that is, the analogue of (3) is no longer true. For example, as we have seen in the proof of Lemma \[lem2\], if $g_1, \ldots , g_m$ are any elements in $\Ga(k_0)$, there exists a linear differential operator $L \in k_0[\d_1]$ such that $L(g_i) = 0$ for $i =1, \ldots ,m$. This implies that any finitely generated subgroup of $\Ga(k_0)$ is contained in a proper Kolchin-closed subset and so cannot be Kolchin-dense in $\Ga(k_0)$. Nonetheless, analogues of facts (2) and (3) can be proven in the context of parameterized linear differential equations (see [@MiSi11]) and one can conclude the following (Corollary 5.2, [@MiSi11]). Let $k_0$ be a $\Pi$-universal differential field and let $k_0(x)$ be a differential field as defined in the introduction.
\[mono\]Let $G$ be a linear differential algebraic group defined over $k_0$ and assume that $G(k_0)$ contains a finitely generated subgroup that is Kolchin-dense in $G(k_0)$. Then $G(k_0)$ is the PPV-group of a PPV-extension of $k_0(x)$.
The assumption that $k_0$ is universal is forced on us, at present, because the analytic techniques used to prove this result do not let us control the algebraic nature of the coefficients appearing in the differential equation defining the PPV-extension. This forces us to assume that $k_0$ is “sufficiently large”.\
Therefore to prove Theorem 1.2 we need to show that under the stated hypotheses, $G$ contains a Kolchin-dense finitely generated subgroup. In fact, we show the following. Note that for the rest of this section, $k_0$ will denote a $\Pi$-differentially closed field.
\[fingen\] Let $G \subset \GL_n$ be a linear algebraic group defined over $k_0$. The group $G(k_0)$ contains a Kolchin-dense finitely generated subgroup if and only if the identity component $G^0(k_0)$ has no quotient isomorphic (as an algebraic group) to $\Ga(k_0)$ or $\Gm(k_0)$.
To prove this result we will need the following three lemmas.
\[lem3\] Let $G$ be a linear algebraic group defined over $k_0$ and $G^0$ be its identity component. $G(k_0)$ contains a Kolchin-dense finitely generated group if and only if $G^0(k_0)$ contains a Kolchin-dense finitely generated group.
Assume that $G^0(k_0)$ contains a Kolchin-dense group generated by $g_1, \ldots, g_s$. Let $\{h_1, \ldots , h_t\}$ be a subset of $G(k_0)$ mapping surjectively onto $G(k_0)/G^0(k_0)$. The set $\{g_1, \ldots , g_s,h_1, \ldots , h_t\}$ generates a group that is Kolchin-dense in $G(k_0)$.\
Assume that $G(k_0)$ contains elements $g_1, \ldots , g_s$ that generate a Kolchin-dense subgroup. From ([@wehrfritz], p.142) or ([@borel-serre], lemme 5.11, p.152), one knows that any linear algebraic group $G(k_0)$, $k_0$ algebraically closed, is of the form $HG^0(k_0)$ where $H$ is a finite subgroup of $G(k_0)$. Therefore we may write each $g_i$ as a product of an element of $H$ and an element of $G^0(k_0)$ and so we may assume that there is a finite set $S=\{\tilde{g}_1, \ldots, \tilde{g}_t\} \subset G^0(k_0)$ such that the group generated by $S$ and $H$ is Kolchin-dense in $G(k_0)$. Extending $S$ if necessary, we may assume that $S$ is stable under conjugation by elements of $H$ and therefore that the group generated by $S$ is stable under conjugation by the elements of $H$. An elementary topological argument shows that the Kolchin-closure $G'$ of the group generated by $S$ is also stable under conjugation by $H$. Therefore $H\cdot G'$ forms a group. It is a finite union of Kolchin-closed sets, so it is also Kolchin-closed. It contains $H$ and $S$ so it must be all of $G(k_0)$. Finally $G'$ is normal and of finite index in $G(k_0)$ so it must contain $G^0(k_0)$. Clearly $ G'\subset G^0(k_0)$ so $G^0(k_0) = G'$ and this shows that $G^0(k_0)$ is finitely generated.
\[lem4\]Let $P\subset \GL_n$ be a connected semisimple linear algebraic group defined over $k_0$. Then $P(k_0)$ contains a finitely generated Kolchin-dense subgroup.
From Proposition 1 of [@tretkoff79] or Lemma 5.13 of [@PuSi2003], we know that a linear algebraic group contains a Zariski-dense finitely generated subgroup $H$. We also know that $P$ contains a maximal torus $T$ of positive dimension. After conjugation, we may assume that $T$ is diagonal and that the projection onto the first diagonal entry is a homomorphism of $T$ onto $k_0^* = k_0\backslash \{0\}$. Since $k_0$ is differentially closed, the derivations $\d_1, \ldots, \d_m$ are linearly independent so there exist nonzero elements $x_1, \ldots , x_m \in k_0$ such that $\det(\d_ix_j)_{1 \leq i,j\leq m} \neq 0$ (Theorem 2, p. 96, [@DAAG]). For each $i = 1 ,\ldots , m$, let $g_i \in T$ be an element whose first diagonal entry is $x_i$. Let $P'$ be the Kolchin-closure of the group generated by $H$ and $\{g_1, \ldots , g_m\}$. I claim $P'=P$.\
To see this note that since $P'$ contains $H$, $P'$ is Zariski-dense in $P$. If $P' \neq P$, then results of [@cassidy6] imply that there exist a nonempty subset $\Sigma \subset k_0\Pi$, the $k_0$ span of $\Pi$, such that $P'$ is conjugate to a group of the form $P''(C)$ where $P''$ is a semisimple algebraic group defined over $\QX$ and $C = \{ c \in k_0 \ | \ \d c = 0 \mbox{ for all } \d \in \Sigma\}$. This implies that each element of $G$ has eigenvalues in $C$ and so, for each $x_i, \ \d(x_i) = 0$ for all $\d \in \Sigma$. Yet, if $\d = \sum_{j=1}^m a_j \d_j$, not all $a_j$ zero and $\d(x_i) = 0$ for $i = 1, \ldots , m$, then $(a_1, \ldots a_m)X = (0, \ldots , 0)$ where $X = (\d_ix_j)_{1 \leq i,j\leq m}$. This contradicts the fact that $\det X \neq 0$. Therefore $P' = P$.
\[lem5\] Let $G(k) = P(k_0)\ltimes U(k_0)$ be a connected linear algebraic group where $P(k_0)$ is a semisimple linear algebraic group and $U(k_0)$ is a commutative unipotent group, both defined over $k_0$. If $G(k_0)$ has no quotient isomorphic to $\Ga(k_0)$, then $G(k_0)$ contains a Kolchin-dense finitely generated subgroup.
Note that $U(k_0)$ is isomorphic to $k_0^m$ for some $m$. Since $P$ acts on $U$ by conjugation, we may write $U = \oplus_{i=1}^m U_i$ where each $U_i$ is an irreducible $P$-module. Furthermore, if the action of $P$ on some $U_j$ is trivial, then this $U_j$ would be of the form $\Ga(k_0)$ and we could write $P\ltimes U = (P\ltimes \oplus_{i\neq j}U_i)\times \Ga(k_0)$. This would imply that there is an algebraic morphism of $G(k_0)$ onto $\Ga(k_0)$, a contradiction. Therefore we may assume the action of $P$ on each $U_i$ is nontrivial. Let $B$ be a Borel subgroup of $P$. From the representation theory of semisimple algebraic groups (Ch.13.3, [@humphreys]), we know that each $U_i$ contains a unique $B$-stable one-dimensional subspace corresponding to a weight $\lambda_i:B \rightarrow \Gm(k_0)$ (the highest weight of $U_i$). For each $i$, let $u_i$ span this one-dimensional space. We claim that the $P(k_0)$-orbit of $u_i$ generates a group that equals $U_i(k_0)$. Note that since $B(k_0)$ is connected and $\lambda_i$ is not trivial, we have the $P(k_0)$-orbit of $u_i$ contains $\Gm(k_0)u_i$. Since $U_i$ is an irreducible $P(k_0)$-module, there exist $g_1, \ldots , g_s \in P(k)$, such that $g_1u_i g_1^{-1}, \ldots g_su_i g_s^{-1}$ span $U_i$. Since $g_j(\Gm(k_0)u_i)g_j^{-1} = \Gm(k_0)(g_ju_ig_j^{-1})$ for $j = 1, \ldots , s$, we have that the $P(k_0)$-orbit of $u_i$ generates all of $U_i$.\
Now Lemma \[lem4\] asserts that there exists a finite set $S\subset P(k_0)$ that generates a Kolchin-dense subgroup of $P(k_0)$. We then have that $S\cup\{u_i\}_{i=1}^m$ generates a Kolchin-dense subgroup of $G(k_0)$.
[*Proof of Proposition \[fingen\].*]{} Assume that $G(k_0)$ contains a Kolchin-dense finitely generated subgroup. Lemma \[lem3\] implies that $G^0(k_0)$ also contains a Kolchin-dense finitely generated subgroup. If there is an algebraic morphism of $G^0(k_0)$ onto $\Gm(k_0)$ or $\Ga(k_0)$ then, in the first case, composing this with the differential algebraic morphism $l\d_1:\Gm(k_0) \rightarrow \Ga(k_0)$ where $l\d_1(u) = \d_1(u)/u$, we would have a differential algebraic morphism of $G^0(k_0)$ onto $\Ga(k_0)$. Therefore, in either case we have a differential homomorphism of $G^0(k_0)$ onto $\Ga(k_0)$. This implies that $\Ga(k_0)$ would contain a Kolchin-dense finitely generated subgroup. On the other hand, any finite set of elements of $\Ga(k_0)$ satisfy a linear differential equation and so could not generate a Kolchin-dense subgroup. Therefore there is no algebraic morphism of $G^0(k_0)$ onto $\Gm(k_0)$ or $\Ga(k_0)$.\
Assume that there is no algebraic morphism of $G^0(k_0)$ onto $\Gm(k_0)$ or $\Ga(k_0)$. Lemma \[lem3\] implies that it is enough to show that $G^0(k_0)$ contains a Kolchin-dense finitely generated group. We may write $G^0 = P \ltimes R_u$ where $P$ is a Levi subgroup and $R_u$ is the unipotent radical of $G$ ([@humphreys], Ch.30.2).\
We first claim that $P$ must be semisimple. We may write $P = (P,P) Z(P)$ where $(P,P)$ is the derived subgroup of $P$ and $Z(P)$ is the center of $P$. Furthermore, $Z(P)^0$ is a torus ([@humphreys], Ch.27.5). We therefore have a composition of surjective morphisms $$G^0 \rightarrow G^0/R_u \simeq P \rightarrow P/(P,P) \simeq Z(P)/(Z(P) \cap (P,P)).$$ Since $G$ is connected, its image lies in the image of $Z^0(P)$ in $Z(P)/(Z(P) \cap (P,P)$ and therefore is a torus. This torus, if not trivial, has a quotient isomorphic to $\Gm$. This would yield a homomorphism of $G^0(k_0)$ onto $\Gm(k_0)$ and, by assumption, this is not possible. Therefore $Z^0(P)$ is trivial. Since $G^0$ is connected we must have $Z(P) \subset (P,P)$. Therefore $P = (P,P)$ and is therefore semisimple.\
We shall now show that it suffices to prove that $G^0(k_0)$ contains a Kolchin-dense finitely generated subgroup under the assumption that $R_u$ is commutative. In [@kovacic69], Kovacic shows ([@kovacic69],Lemma 2): [*Let $G$ be an abstract group, $H$ a subgroup and $N$ a nilpotent normal subgroup of $G$. Suppose $H\cdot (N,N) = G$. Then $H = G$.*]{} Therefore, if we can find a Kolchin-dense finitely generated subgroup of the $k_0$-points of $G^0/(R_u,R_u) \simeq P\ltimes (R_u/(R_u,R_u))$, then the preimage of this group under the homomorphism $G^0 \rightarrow G^0/(R_u,R_u)$ generates a Kolchin-dense subgroup of $G(k_0)$.\
Therefore, we need only consider connected groups satisfying the hypotheses of Proposition \[fingen\] and of the form $P(k_0)\ltimes U(k_0)$, where $P$ is semisimple and $U$ is a commutative unipotent group. Lemma \[lem5\] guarantees that such a group has a finitely generated Kolchin-dense subgroup. $\QED$\
Proposition \[mono\] and Proposition \[fingen\] imply that Theorem 1.2 is true.\
It would be of interest to find a purely algebraic proof of Theorem 1.2 that would perhaps also show that this result is true when we weaken the hypotheses to assume that $k_0$ is only differentially closed (or even just algebraically closed). Furthermore, the relation between the conditions that a group contains a Kolchin-dense finitely generated subgroup and that the group appears as a PPV-group of a PPV-extension of $k_0(x)$ should be further studied. I know of no example of a linear differential algebraic group that is a PPV-group of a PPV-extension of $k_0(x)$ and that does not contain a Kolchin-dense finitely generated subgroup. See Section \[sec5\] for further discussion concering this.
An Example {#sec4}
==========
In this section we give an example that shows that Theorem 1.2 and Proposition \[fingen\] are not true for linear differential algebraic groups in general.\
Let $k_0$ be an ordinary differentially closed field with derivation $\d_1$ and let $$G = \{ \left(\begin{array}{cc} 1& 0 \\ a & b \end{array} \right) \ | a, b \in k_0 , \ b\neq 0, \ \d_1 b = 0 \} \simeq G_1 \rtimes G_2$$ where $$\begin{aligned}
G_1&= & \{ \left(\begin{array}{cc} 1& 0 \\ a& 1 \end{array} \right) \ | a \in k_0 \} \simeq \Ga(k_0)\\
G_2& = & \{ \left(\begin{array}{cc} 1& 0 \\ 0 & b \end{array} \right) \ | b \in k_0 , \ b\neq 0, \ \d_1b = 0 \}
\simeq \Gm(C) \end{aligned}$$ where $C = \{c \in k_0 \ | \ \d_1c = 0\}$. Let $k = k_0(x)$ be a $\Delta = \{ \d_x, \d_1\}$-field as in the introduction.\
If $k_0$ is a universal field then Proposition \[mono\] implies that statement 1. of the following proposition follows from statement 4. We do not make this assumption.
\[exprop\] $\mbox{ } $
1. $G(k_0)$ contains no Kolchin-dense finitely generated subgroup.
2. $G$ is Kolchin-connected.
3. There is no surjective differential algebraic homomorphism of $G(k_0)$ onto $\Ga(k_0)$ or $\Gm(k_0)$.
4. $G(k_0)$ is not a PPV-group of a PPV-extension of $k_0(x)$.
1. To see that $G(k_0)$ contains no Kolchin-dense finitely generated subgroup, note that any element of $G(k_0)$ can be written as a product of an element of $\Ga(k_0)$ and $\Gm(C)$. Therefore it is enough to show that any set of elements of the form $$\left(\begin{array}{cc} 1& 0 \\ a_1 & 1 \end{array} \right), \ldots , \left(\begin{array}{cc} 1& 0 \\ a_n & 1 \end{array} \right),\left(\begin{array}{cc} 1& 0 \\ 0 & b_1 \end{array} \right), \ldots , \left(\begin{array}{cc} 1& 0 \\ 0 & b_m \end{array} \right)$$ with the $a_i \in k_0$ and the $b_i \in C$ do not generate a Kolchin-dense subgroup of $G$. Let $H$ be the group generated by these elements and let $L \in k[\d_1]$ be a nonzero differential operator such that $L(a_i) = 0$ for all $i = 1, \ldots , n$. A calculation shows that any element of $H$ is of the form $$\left(\begin{array}{cc} 1& 0 \\ c_1a_1+ \ldots +c_na_n & b \end{array} \right)$$ with $b \in k_0$ and the $c_i \in C$. Therefore $H$ is a subgroup of $$\{\left(\begin{array}{cc} 1& 0 \\ a & b \end{array} \right) \ | \ L(a) = 0, \d_1 b = 0, b\neq 0\}$$ which is a proper Kolchin-closed subgroup of $G$.\
2. Assume that $G$ is not Kolchin-connected and let $G^0$ be the identity component in the Kolchin topology. One has that $G^0$ is normal and of finite index in $G$. Furthermore, $G/G^0$ is again a linear differential algebraic group and $\pi:G \rightarrow G/G^0$ is a differential algebraic homomorphism. Since $G_1 \simeq \Ga(k_0)$ is a divisible group and any homomorphism of a divisible group into a finite group is trivial, we have that $G_1$ is contained in the kernel of $\pi$. This implies that $\pi$ induces a differential algebraic homomorphism $\pi^*$ from $G_2\simeq \Gm(C)$ to a finite group. Since the elements of $\Gm(C)$ have constant entries, $\pi^*$ is really an algebraic homomorphism of $\Gm(C)$ into a finite group. Since $\Gm(C)$ is connected in the Zariski topology, this homomorphism is trivial. Therefore $G = G^0$.\
3. Since $\Ga(k_0)$ is a differential homomorphic image of $\Gm(k_0)$, it suffices to show that there is no surjective differential algebraic homomorphism of $G(k_0)$ onto $\Ga(k_0)$. Assume not and let $\phi:G(k_0)\rightarrow \Ga(k_0)$. Restricting $\phi$ to $G_2$ yields an [*algebraic*]{} homomorphism of $\Gm(C)$ into $\Ga(k_0)$. Since algebraic homomorphisms preserve the property of being semisimple, we must have that $G_2 \subset \ker \phi$. Therefore for any $a \in k$ and any $b \in C^*$, we have $$\left(\begin{array}{cc} 1& 0 \\ a & 1 \end{array} \right)\left(\begin{array}{cc} 1& 0 \\ 0 & b \end{array} \right)\left(\begin{array}{cc} 1& 0 \\ -a & 1 \end{array} \right) = \left(\begin{array}{cc} 1& 0 \\ a-ba & b \end{array} \right)\in \ker \phi .$$ For any $\tilde{a} \in k_0$ and $1\neq b \in C$ there exists a $a \in k_0$ such that $a-ba = \tilde{a}$, so $\ker \phi$ contains all elements of the form $$\left(\begin{array}{cc} 1& 0 \\ \tilde{a} & b \end{array} \right)$$ $a \in k_0$, $1\neq b \in C$. Since $G_2 \subset \ker \phi$ as well, we have that $G \subset \ker \phi$, a contradiction.\
4. Assume $G$ is a PPV-group of a PPV-extension $K$ of $k=k_0(x)$. The field $K$ is then a PPV-extension corresponding to a second order linear differential equation, in $\d_x$, $L(Y) = 0$. There are elements $u,v \in K$ forming a $k_0$-basis of the solutions space such that for any $\sigma \in G, \ \sigma = \left(\begin{array}{cc} 1& 0 \\ a & b \end{array} \right)$, we have $\sigma(u) = u+av$ and $\sigma(v) = bv$. We have $K = k<u,v>_\Delta$, the $\Delta =\{\d_x, \d_1\}$-differential field generated by $u$ and $v$. I will describe the structure of $K$ in more detail.
1. Let $E = k(v)$. Since for any $\sigma \in G$ there exists a $b_\sigma \in C$ such that $\sigma(v) = b_\sigma v$, we have that $\d_xv/v \in k$ and $\d_1 v/v \in k$. Furthermore, $E$ is the fixed field of $G_1$ and the PPV-group of $E$ over $F$ is $G/G_1 \simeq G_2$.
2. We may write $K = E<w>_\Delta$ where $w = \frac{u}{v}$. For any $\sigma = \left(\begin{array}{cc} 1& 0 \\ a & 1 \end{array} \right) \in G_1$ we have $\sigma (w) = w + a$. Therefore $\sigma(\d_x w) =\d_x w$, so $\d_x w \in E$.
3. For any $\sigma = \left(\begin{array}{cc} 1& 0 \\ 0 & b \end{array} \right) \in G_2$ we have $\sigma(\d_x w) = \d_x(\sigma(w)) = \d_x(\frac{1}{b}w) = \frac{1}{b} \d_x w$. Since $\sigma(v) = bv$, we have $\sigma(\d_xw \cdot v) = \d_1 w \cdot v$. This implies $\d_1 w = r/v$ for some $r \in k$. In particular, we may write $E = k(\d_x w)$ and that $\d_x(\d_xw) /\d_x w = A \in k$ and $\d_1(\d_xw) /\d_x w = B \in k$.
Summarizing, we have $$k = k_0(x) \subset E =k(\d_x w) \subset E<w>_\Delta = K$$ where $$\frac{\d_x(\d_xw)}{\d_x w }= A \in k\mbox{ and } \frac{\d_1(\d_xw)}{\d_x w} = B \in k.$$ I now claim that there exists an element $h \in k_0(x)$ and a nonzero operator $$L = \sum_{i=0}^M \alpha_i \d_1^i\in k_0[\d_1]$$ such that $$L(\d_x w) = (\d_x h + hA)\d_xw.$$ Let us assume that this last claim is true. We then would have that $$\d_x(L(w) - h\d_xw) = L(\d_x w) - \d_xh \d_xw - hA\d_xw = 0.$$ Therefore $L(w) = h\d_xw + c \in E$ for some $c \in k_0$. In particular $L(w)$ is left fixed by all $\sigma \in H \simeq \Ga(k_0)$. This means that $$L(w) = \sigma(L(w)) = L(w + a_\sigma) = L(w) + L(a_{\sigma})$$ where $\sigma = \left(\begin{array}{cc} 1& 0 \\ a_\sigma & 1 \end{array} \right)$ and so $L(a) = 0$ for all $a \in k_0$, a contradiction.\
We shall now show that the claimed $L$ and $h$ exist. If $A \in k_0$, then we let $L = A$ and $h = 1$. This yields $L(\d_xw) = A\d_xw$ as desired. Therefore we may assume $A \in k_0(x)$ but $A \notin k_0$ and so $A$ has poles. Let $x_1, \ldots , x_p \in \PX^1(k_0)$ include all the poles of $A$ and $B$ (including $x_p = \infty$). Select $n \in \NX$ so that $A$ and $B$ have poles of orders at most $n$ at each of these points. Select integers $M$ and $N$ such that $$M > np \mbox{ and } N> n(M-1)$$ (the reason for this choice will be apparent later). Let $$\begin{aligned}
L & = & \sum_{i=0}^M \alpha_i\d_1^i\\
h & = &\beta_{p,0} + \beta_{p,1}x + \ldots + \beta_{p,N}x^N + \sum_{i=1}^{p-1}\sum_{j=1}^N \frac{\beta_{i,j}}{(x-x_i)^j} \end{aligned}$$ where the $\alpha_i$ and the $\beta_{i,j}$ are indeterminates. We shall show that the condition “$L(\d_xw) = (\d_x h + hA)\d_xw$" forces these indeterminates to satisfy an undertedermined system of linear equations over $k_0$ and so there will always be a way to select elements of $k_0$ (not all zero) satisfying this system. Furthermore, we will show that not all the $\alpha_i$ can be zero. To see this, we shall look at the expressions $L(\d_x w)$ and $(\d_x h + hA)\d_1w$ separately.\
$\underline{L(\d_xw):}$ Note that $$\begin{aligned}
\d_1(\d_x w) & = & B \d_xw\\
\d_1^2(\d_x w) & = & \d_1 B \d_xw + B^2 \d_xw\\
\vdots \ \ \ & \vdots & \ \ \ \vdots \\
\d_1^i(\d_x w) & = & R_i \d_xw\end{aligned}$$ where $R_i \in k_0(x)$ and $R_0 = 1, R_1 = B, R_{i+1} = \d_1 R_i + BR_i$. We furthermore have $$L(\d_x w) = (\sum_{i=0}^M \alpha_iR_i) \d_x w .$$ Each $R_i$ has poles only at $x_1, \ldots , x_p$. Furthermore the order of a poles of each $R_i$ is at most $in$. Therefore the partial fraction decomposition of $\sum_{i=0}^M \alpha_iR_i$ is of the form $$\begin{aligned}
\label{gammaeqn}
\Gamma_{p,0} + \Gamma_{p,1}x + \ldots + \Gamma_{p,Mn}x^{Mn} + \sum_{i=1}^{p-1}\sum_{j=1}^{Mn} \frac{\Gamma_{i,j}}{(x-x_i)^j}& &\end{aligned}$$ where the $\Gamma_{i,j}$ are linear forms in the $\{\alpha_r\}_{r=0}^M$.\
$\underline{(\d_x h + Ah)\d_xw:}$ Once again $\d_x h + Ah$ has poles only at $x_1, \ldots , x_p$. The order at any of these poles is at most $n + N$. Therefore the partial fraction decomposition of $\d_x h + Ah$ is of the form: $$\begin{aligned}
\label{lambdaeqn}
\Lambda_{p,0} + \Lambda_{p,1}x + \ldots + \Lambda_{p,Mn}x^{n+N} + \sum_{i=1}^{p-1}\sum_{j=1}^{n+N} \frac{\Lambda_{i,j}}{(x-x_i)^j}&&\end{aligned}$$ where the $\Lambda_{i,j}$ are linear forms in the $\{\beta_{r,s}\}$ with coefficients in $k_0$.\
The equation $L(\d_xw) = (\d_x h + hA)\d_xw$ implies that expression (\[gammaeqn\]) equals expression (\[lambdaeqn\]). When we equate coefficients of powers of $x$ in these two expressions, we will produce $n+N+ 1$ homogeneous linear equations in the variables $\{\alpha_r\}$ and $\{\beta_{r,s}\}$ (note that our assumption $N > n(M-1)$ implies that $N+n>Mn$). For each $x_i, \ i=1, \ldots, p-1$, equating the coefficients of powers of $x-x_i$ in these two expressions yields $N+n$ homogeneous linear equations in the variables $\{\alpha_r\}$ and $\{\beta_{r,s}\}$. In total, equating coefficients of like terms yields $p(n+N) + 1$ homogenous linear equations in the $\{\alpha_r\}$ and $\{\beta_{r,s}\}$. The total number of the $\{\alpha_r\}$ and $\{\beta_{r,s}\}$ is $M+1 + Np+1$. Because we have selected $M >np$, we have $$M+1 + Np + 1 > np + 1 + Np + 1 = p(n+N) + 2$$ and this exceeds the number of equations. Therefore we can find $\{\alpha_r\}$ and $\{\beta_{r,s}\}$ in $k_0$, not all zero, that satisfy these equations.\
We shall now show that in any such choice, not all the $\alpha_r$ are zero. Assume all the $\alpha_r$ are zero. In this case we would have $\d_xh + Ah = 0$. This implies that $\d_x(h\d_x w ) = 0$ and so $\d_x w \in k_0(x)$. This contradicts the fact that the field $E$ in our tower $k \subset E \subset K$ is a proper extension of $k$.
It is interesting to contrast the group $G$ above with the slightly larger group $$\begin{aligned}
G'&= & \{ \left(\begin{array}{cc} 1& 0 \\ a & b \end{array} \right) \ | a, b \in k_0 , \ b\neq 0, \ \d_1(\frac{\d_1 b}{b}) = 0 \}\\
& = & G_1' \rtimes G_2'
\end{aligned}$$ where $$\begin{aligned}
G_1'&= & \{ \left(\begin{array}{cc} 1& 0 \\ a& 1 \end{array} \right) \ | a, \in k_0 \} \simeq \Ga(k_0)\\
G_2'& = & \{ \left(\begin{array}{cc} 1& 0 \\ 0 & b \end{array} \right) \ | b \in k_0 , \ b\neq 0, \ \d_1(\frac{\d_1 b}{b}) = 0 \}
\simeq \{b\in \Gm(k_0) \ | \d_1(\frac{\d_1 b}{b}) = 0 \}\end{aligned}$$ Note that $G \subset G'$. I will show that, $G'$ contains a Kolchin-dense finitely generated subgroup. To see this let $$g = \left(\begin{array}{cc} 1& 0 \\ 1 & 1 \end{array} \right) \ \ \ \ \ \mbox{ and } \ \ \ \ \ h= \left(\begin{array}{cc} 1& 0 \\ 0 & e\end{array} \right)$$ where $0 \neq e \in k_0$ satisfies $\d e = e$. I will show that the Kolchin closure of the group generated by these two elements is $G'$.\
For $n \in \ZX$, $$h^{n}gh^{-n} = \left(\begin{array}{cc} 1& 0 \\ e^n & 1 \end{array} \right)$$ lies in $G_1'$. Since $S = \{e^n \ | \ n \in \ZX \}$ is a set of elements linearly independent over the $\d_1$-constants of $k_0$, $S$ cannot be a subset of the solution space of a nonzero homogeneous linear differential equation with coefficients in $k_0$. Therefore the Kolchin-closure of the group generated by $g$ and $h$ contains $G_1'$.\
Let $H$ be the Kolchin-closure of the group generated by $h$. We will identify this with a subgroup of $\{b\in \Gm(k_0) \ | \d_1(\frac{\d_1 b}{b}) = 0 \} \subset \Gm(k_0)$. One sees that $H$ is Zariski-dense in $\Gm(k_0)$ and so, by (Corollary 2, p. 938,[@cassidy1]), there exists a set $\calL$ of linear operators in $\d_1$ with coefficients in $k_0$ such that $H = \{y \in \Gm(k_0) \ | \ L(\frac{\d_1 y}{y}) = 0, \forall \ L \in \calL\}$. We may assume that $\calL$ is a left ideal in the ring $k_0[D]$ of linear operators. Since this ring is a right euclidean domain, we have that $\calL$ is generated by a single element $R$. Since $H \subset G_1'$ we have $R$ divides the operator $D$ on the right. Therefore $R$ must equal $D$ and so $H = G_1'$. Therefore the Kolchin-closure of the group generated by $g$ and $h$ contains $G_1'$ and $G_2'$ and so must equal $G'$.\
When $k_0$ is a universal field then Proposition \[mono\] implies that $G'$ is a PPV-group of some PPV-extension of $k$. In fact, as noted in Section 7 of [@CaSi], when $k_0$ is the differential closure of $\CX(t)$ where $\d_1 = \frac{\d}{\d t}$, then $G'$ is the PPV-group of $$\begin{aligned}
\label{gammafnceqn}
\d_x^2 y -\frac{t-1-x}{x}\d_xy &= & 0 . \end{aligned}$$ Knowing that $G'$ is the PPV-group of equation (\[gammafnceqn\]), one can also deduce from the results of [@dreyfus] (see the final comments below) that $G'$ contains a Kolchin-dense finitely generated subgroup.
Final Comments {#sec5}
==============
The results of Section \[sec2\] and Section \[sec3\] imply that if $k_0$ is a universal field and $G$ is a linear algebraic group defined over $k_0$, then $G(k_0)$ is a PPV-groups of a PPV- extension of $k$ if and only if $G(k_0)$ contains a Kolchin-dense finitely generated subgroup. One can ask if this result is true for arbitrary linear differential algebraic groups. I know of no counterexamples. Of course, the implication in one direction is true since Proposition \[mono\] states that a sufficient condition for $G(k_0)$ to be a PPV-group of a PPV-extension of $K$ is that $G(k_0)$ contains a Kolchin-dense finitely generated subgroup. A recent result of Dreyfus [@dreyfus] bears on the implication in the other direction. In [@dreyfus], the author proves a generalization of the Ramis Density Theorem [@martinet_ramis] by showing that the local parameterized Picard-Vessiot group of a parameterized linear differential equation with a fixed singular point is the Kolchin-closure of the group generated by an element called the formal parameterized monodromy, a finite number of elements called the (parameterized) Stokes operators and the constant points of a linear [*algebraic*]{} group called the parameterized exponential torus. As is pointed out in [@dreyfus], this latter group is finitely generated. Using this result, one can further conclude that at least in the case of parameterized equations with fixed singular points, one has that the PPV-group of a parameterized linear differential equation contains a Kolchin-dense finitely generated subgroup.
[10]{}
A. Borel and J.-P. Serre, Théorèmes de finitude en cohomologie galoisienne, , 39:111–164, 1964-65.
P. J. Cassidy, Differential algebraic groups, , 94:891-954, 1972.
P. J. Cassidy, The classification of the semisimple differential algebraic groups and the linear semisimple differential algebraic [L]{}ie algebras, . 121(1):169–238, 1989.
P. J. Cassidy, M. F. Singer, Galois Theory of Parameterized Differential Equations and Linear Differential Algebraic Groups, , D. Bertrand et. al., eds., IRMA Lectures in Mathematics and Theoretical Physics, 9:113-157,2006.
C. Chevalley, , , American Mathematical Society, New York, 1951.
T. Dreyfus, A density theorem for parameterized differential Galois theory, Preprint, Institut de mathématiques de Jussieu, Paris, July 22, 2011.
H. Gillet, S. Gorchinskiy, A. Ovchinnikov, Parameterized Picard-Vessiot extensions and Atiyah extensions, Preprint, May, 2011.
C. Hardouin, M. F. Singer, Differential [G]{}alois theory of linear difference equations, , 342(2):333–377, 2008.
J. Hartmann, On the inverse problem in differential [G]{}alois theory, , 586:21-44, 2005.
J. Hartmann, Patching and differential Galois groups (joint work with David Harbater), . 4(2):1490-1493, 2007.
J. E. Humphreys, , Graduate Texts in Mathematics. Springer-Verlag, New York, 1975.
E. R. Kolchin, Differential Algebra and Algebraic Groups, Academic Press, New York, 1976.
E. R. Kolchin, Differential Algebraic Groups, Academic Press, Orlando, 1985.
J. Kovacic, The inverse problem in the [Galois]{} theory of differential fields, , 89:583–608, 1969.
J. Kovacic, On the inverse problem in the [Galois]{} theory of differential fields, , 93:269-284, 1971.
J. Kovacic, The differential [G]{}alois theory of strongly normal extensions, , 355(11):4475-4522, 2003.
P. Landesman, Generalized differential Galois theory, , 360(8):4441–4495, 2008.
B. Malgrange, On nonlinear differential [G]{}alois theory, ,World Sci. Publ., River Edge, NJ, 185-196, 2004.
Ju. I. Manin, Algebraic curves over fields with differentiation, , 22:737–756, 1958, An English translation appears in Transl. Amer. Math. Soc. Ser. 2, 37:59-78, 1964.
D. Marker, A. Pillay, Differential [G]{}alois theory [III]{}. [S]{}ome inverse problems, 41(3):453-561, 1997.
J. Martinet, J.-P. Ramis, Théorie de [Galois]{} différentielle et resommation, in [*Computer Algebra and Differential Equations*]{}, E. Tournier, ed., Academic Press, 115-214, 1989.
C. Mitschi, M. F. Singer, Monodromy Groups of Parameterized Linear Differential Equations with Regular Singular Points, :1106.2664v1 \[math.CA\] 14 June 2011.
A. Pillay, Differential [G]{}alois theory [II]{}, . 88(2-3):181-191, 1997.
A. Pillay, Differential [G]{}alois theory [I]{}, :42(4):678-699, 1998.
M. van der Put, M. F. Singer, , volume 328 of [*Grundlehren der mathematischen Wissenshaften*]{}, Springer, Heidelberg, 2003.
C. Tretkoff, M. Tretkoff, Solution of the Inverse Problem in Differential [Galois]{} Theory in the Classical Case, , 101:1327-1332, 1979.
H. Umemura, Invitation to [G]{}alois theory, , D. Bertrand et. al., eds., IRMA Lectures in Mathematics and Theoretical Physics, 9:269-289, 2007.
H. Volklein, , Cambridge Studies in Advanced Mathematics, 53. Cambridge University Press, Cambridge, 1996.
B. A. F. Wehrfritz, , Ergebnisse der Mathematik. Springer-Verlag, Berlin, 1973.
M. Wibmer, Existence of $\d$-parameterized Picard-Vessiot extensions over fields with algebraically closed constants, :1104.3514v1 \[math.AC\] 18 April 2011.
[^1]: Department of Mathematics, North Carolina State University, Box 8205, Raleigh, North Carolina 27695-8205, email:`[email protected]`. The author was partially supported by NSF Grant CCF-1017217. 2010 Mathematics Subject Classification: 12H05, 20G15, 34M03, 34M15, 34M50
[^2]: In this paper, all fields considered are of characteristic zero.
[^3]: There are other Galois theories of differential equations due primarily to Malgrange [@malgrange_galois_04], Pillay [@pillaygalois2; @pillaygalois1; @pillaygalois3] and Umemura [@umemura_invitation]. In particular, inverse problems are addressed in [@pillaygalois1]. We will not consider these theories here.
[^4]: Let $C$ be the $\d_1$-constants of $k_0$ and $\beta_1, \ldots ,\beta_s$ a $C$-basis of the $C$-span of the $\alpha_i$’s. Let $R(Y) = wr(Y,\beta_1, \ldots , \beta_s)$ where $wr$ denotes the wronskian determinant. $R(Y)$ is a linear differential polynomial yielding the desired $R \in k_0[\d_1]$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We discuss an incentivizing market and model-based approach to design the energy management and control systems which realize high-quality ancillary services in dynamic power grids. Under the electricity liberalization, such incentivizing market should secure a high speed market-clearing by using the market players’ private information well. Inspired by contract theory in microeconomics field, we propose a novel design method of such incentivizing market based on the integration of the economics models and the dynamic grid model. The conventional contract problems are analyzed for static systems or dynamical systems with control inputs directly operated by the principal. The analysis is, however, in discord with the incentivizing market. The main challenge of our approach is to reformulate the contract problems adapted to the market from the system and control perspective. We first establish the fundamental formulas for optimal design, and clarify the basic properties of the designed market. We also discuss possibilities, limitation and some challenges in the direction of our approach and general market-based approaches.'
author:
- 'Yasuaki Wasa, Kenji Hirata and Kenko Uchida [^1]'
nocite: '[@*]'
title: |
Contract Theory Approach to\
Incentivizing Market and Control Design
---
Introduction {#sec:1}
============
Achieving a quality assurance of electric energy, called the ancillary service, is a key target of next-generation energy management and control systems for dynamic electric smart grids where electricity liberalization is fully enforced and renewable energy is highly penetrated [@bib01]. Frequency, voltage and power controls, which are typical contents of the ancillary service, have been technical requirements for the electric energy supplier (e.g., see [@bib02; @bib03]). Since the electricity liberalization starts, such ancillary control services have been investigated and realized in competitive electricity markets [@bib04; @bib05; @bib06]. In view these, future energy management and control systems should include ancillary service markets with some incentive mechanisms, as core elements, which provide high-quality and fast-response control services to the extent of the primary level. Moreover, if we need ancillary control services of transient state, ancillary service markets should include physical models of dynamic power grids. In this article, we propose an incentivizing market-based approach to design the energy management and control systems which realize high-quality ancillary services in such dynamic power grids. Using this approach, we develop a design method of such incentivizing market based on the integration of the economics models and the dynamic grid model, and provide fundamental conditions and formulas for the incentivizing market design.
Our approach is developed under the assumption that an energy dispatch scheduling on a future time interval has been finished in a spot energy market at the tertiary control level [@bib07; @bib08], e.g., for one hour future interval, and that each agent has a linearized model of his/her own system along the scheduled trajectory over the future time interval. For this linear time-varying model, we formulate a design problem of energy management and control systems based on a real-time regulation market, called the ancillary market, at the secondary and primary control levels [@bib07; @bib08]. Participants in the dynamic electric smart grid are consumers, suppliers or prosumers, called agents, who control their physical system selfishly according to their own criterion, and utility (independent public commission), who integrates economically all the controls of agents into a high-quality power demand and supply. In the integration, a market mechanism is adopted inevitably in order to secure selfish behaviors of agents in electricity liberalization; that is, each agent bids his/her certain private information in response to a market-clearing price, while utility (auctioneer) clears the market based on the bidding and decides the prices, in real-time.
The market model in our approach is characterized by two terminologies: *private information* and *incentivizing market*. An iterative market-clearing model so-called the tâtonnement model does not need rigorous models, but does not generally guarantee the convergence to a specified equilibrium. Moreover, if it converges, the tâtonnement model takes a long time to converge at a market clearing equilibrium. To overcome the issues, we propose a novel model-based and market-based approach that designs first some incentives for the agents to report their private information (including their own model information) to the utility in the market, and makes it possible to realize a high speed market-clearing. This approach needs incentivizing costs, and the resulting optimization process can be recognized as an intermediate model (the second best model) between two extremal models, namely the tâtonnement model and the so-called supply/demand function equilibrium model (the first best model) which uses for free all agents’ rigorous models, i.e., agents’ private information. We provide this approach with fundamental formulas and tools to design the incentivizing mechanism in the market, and discuss the basic properties of the designed market. We also discuss the relationships of our incentivizing mechanism with the Lagrange multiplier based integration/decomposition mechanism and the mechanism design.
This article has been organized as follows: Section \[sec:2\] introduces a dynamic power grid model and a model-based incentivizing market model. We next derive some theoretical results on a general reward design problem in Section \[sec:3\]. In Section \[sec:4\], we show the relationship between the private information and the incentives and discuss possibilities and limitation of our approach through three typical scenarios. In Section \[sec:5\], we summarize the results.
Grid Model and Incentivizing Market Model {#sec:2}
=========================================
In this paper, we consider the two level architecture with the two layers market; spot energy market and real-time regulation market. The well-known temporally-separated architecture [@bib07; @bib08] motivated by the conventional power system control is divided into the primary control level (voltage and frequency stabilization), the secondary control level (quasi-stationary power imbalance control) and the tertiary control level (economic dispatch). The two layers market reorganizes the conventional three-level architecture according to the functions of the markets. Our approach is developed under the assumption that an energy dispatch scheduling on a future time interval has been finished in a spot energy market (at the tertiary control level), and that each agent has a linearized model of his/her own system along the scheduled trajectory over the future time interval. For this linear time-varying model, we formulate a design problem of energy management and control systems to realize ancillary services based on a real-time regulation market (at the secondary and primary control levels).
Linearized Grid Model
---------------------
Let us first consider the linearized time-varying model used in the ancillary market. This paper considers one of the standard grid models, the average system frequency model [@bib09], as a generic model of high speed response for ancillary service control problems with two area power networks and with two kinds of players: *Utility* and *Agents*. Here we present a linearized model of each player’s own system along the scheduled trajectory over a future time interval during when an energy dispatch scheduling has been finished in a spot energy market.
The utility dynamics, which describes the deviation of the power and/or frequency balance and other deviations from physical constraints as well, obeys the following equation: $$\begin{aligned}
dx_{0t}
=
\left( A_{00}(t) x_{0t} + A_{01}(t) x_{1t} + A_{02}(t) x_{2t} \right) dt
+ D_0(t) d\beta_t, \quad
t_0 \leq t \leq t_f,
\label{eq:u_dyn}\end{aligned}$$ and is evaluated by the utility’s revenue functional: $$J_0(t,x;u) = {\mathbb{E}}_{t,x}
\left[ \varphi_0(t_f, x_{t_f} ) + \int_{t}^{t_f} l_0( \tau, x_\tau, u_\tau ) d\tau \right]$$ where $x=(x_0^\top, x_1^\top, x_2^\top)^\top\in\mathbb{R}^n$ is the collection of the states of the utility dynamics $x_0\in\mathbb{R}^{n_0}$ and the agents’ dynamics $x_i\in\mathbb{R}^{n_i}$, $i=1, 2$, at time $t\in[t_0,t_f]$, and $u=( u_1^\top, u_2^\top )^\top$ is the local control inputs, respectively; $\beta_t$ is the disturbance modeled by a standard Wiener process on $[t_0,t_f]$; ${\mathbb{E}}_{t,x}$ indicates an expectation given initial data $(t,x)$; we use an abbreviation like $x_{0t}=x_0(t)$, $x_t=x(t)$. The dynamics of the agent $i$ ($i=1,2$) obeys $$dx_{it} = ( A_i(t) x_{it} + B_i(t) u_{it} )dt + D_i(t) d\beta_t,
\quad t_0\leq t\leq t_f, \ i=1,2,
\label{eq:a_dyn}$$ and is evaluated by the agent’s revenue functional: $$J_i(t,x;u) = {\mathbb{E}}_{t,x}
\left[ \varphi_i(t_f, x_{t_f} ) + \int_{t}^{t_f} l_i( \tau, x_\tau, u_\tau ) d\tau \right], \quad i=1,2.$$ The agent’s state $x_i$ indicates typically the deviation of power generation or consumption from the scheduled trajectory; the control $u_i$ compensates the deviation. An admissible control of agent $i$, denoted as $u_i\in\Gamma_i$, is a state feedback $u_{it}=u_i(t,x)$ denoted by $u_i:[t_0,t_f]\times \mathbb{R}^n \to U_i \subset \mathbb{R}^{m_i}$ is continuous at $t\in[t_0,t_f]$ and Lipschitz continuous at $x\in\mathbb{R}^n$, $i=1,2$. To simplify the description in the following, let us describe the grid dynamics by combining the utility dunamics (\[eq:u\_dyn\]) and the agent’s dynamics (\[eq:a\_dyn\]) as follows: $$d x_t
= f(t,x_t,u_t) dt + D(t) d\beta_t
:= \left( f_0(t,x_t) + f_1(t,x_{1t},u_{1t}) + f_2(t,x_{2t},u_{2t}) \right) dt
+ D(t) d\beta_t$$ where $$f_0 =
\left( \begin{array}{c}
A_{00} x_0 + A_{01} x_1 + A_{02} x_2 \\
{\boldsymbol{0}}\\
{\boldsymbol{0}}\end{array} \right),
\
f_1 =
\left( \begin{array}{c}
{\boldsymbol{0}}\\
A_{1} x_1 + B_{1} u_1 \\
{\boldsymbol{0}}\end{array} \right), \
f_2 =
\left( \begin{array}{c}
{\boldsymbol{0}}\\
{\boldsymbol{0}}\\
A_{2} x_2 + B_{2} u_2
\end{array} \right), \
D =
\left( \begin{array}{c}
D_0 \\
D_1 \\
D_2
\end{array} \right).$$
We need the following assumptions, which make the discussions in this paper mathematically regorous. The notations $\nabla_t=\partial/\partial t$, $\nabla_x=(\partial/\partial x_0, \partial/\partial x_1, \partial/\partial x_2)$ and $\nabla_x^2=[ \partial^2/\partial x_i \partial x_j ]$ are used.
(A1)
: Each element of matrices $A_{00}(t)$, $A_{0i}(t)$, $A_i(t)$, $B_i(t)$, $D_0(t)$, $D_i(t)$, $i=1,2$, are continuous at $t\in[t_0,t_f]$, and $D(t)D(t)^\top > 0$ for all $t\in[t_0,t_f]$.
(A2)
: The set $U_i$, $i=1,2$, are compact and convex.
(A3)
: The function $\varphi_0(t_f,\cdot):\mathbb{R}^n\to \mathbb{R}$ is of class ${{\mathcal{C}}}^2$ and $\nabla_x \varphi_0(t_f,\cdot)$ is polynomial growth. The function $l_0:[t_0,t_f]\times \mathbb{R}^n\times U_1 \times U_2 \to \mathbb{R}$ is ${{\mathcal{C}}}^1$ at $(t,u_1,u_2)\in[t_0,t_f]\times U_1 \times U_2$ and ${{\mathcal{C}}}^2$ at $x\in\mathbb{R}^n$, and $\nabla_x l_0$, $\nabla_{u_i} l_0$, $i=1,2$, are polynomial growth at $(x,u_1,u_2)\in\mathbb{R}^n\times U_1 \times U_2$.
(A4)
: The function $\varphi_i(t_f,\cdot):\mathbb{R}^{n_i}\to \mathbb{R}$ is of class ${{\mathcal{C}}}^2$ and $\nabla_x \varphi_i(t_f,\cdot)$ is polynomial growth. The function $l_i:[t_0,t_f]\times \mathbb{R}^{n_i}\times U_i \to \mathbb{R}$ is ${{\mathcal{C}}}^1$ at $t\in[t_0,t_f]$ and ${{\mathcal{C}}}^2$ at $(x_i,u_i)\in\mathbb{R}^{n_i}\times U_i$, and $\nabla_{x_i} l_i$, $\nabla_{u_i} l_i$ are polynomial growth at $(x_i,u_i)\in\mathbb{R}^{n_i}\times U_i$ and $\nabla_{u_i}^2 l_i < 0$, $i=1,2$.
We formulated the grid model with the evaluation functionals on the finite time interval $[t_0,t_f]$. For simplicity, from now on, we consider the state feedback strategies $u=( u_1^\top, u_2^\top )^\top$ derived by dynamic programming. We will discuss possibilities of the other options in Section \[sec:4\]. To achieve the objective, we reformulate our problems on the future time interval from the current time $t$ to the final time $t_f$ based on the time-consistency property.
To describe formulas concisely, we adopt the continuous-time model in this article; we can develop in parallel the same results in the discrete-time model. On the other hand, to develop our discussion in the continuous-time model in a mathematically sound way, we need some technical assumptions as stated above and in the later discussion; however, the assumptions except that on convexity (or concavity) are for assuring an appropriate smoothness and boundness of the variables appearing in the discussions, but not essential for developing our key ideas.
Incentivizing Market Model
--------------------------
To describe market mechanism, we need to specify participant’s private information. Private information of agent $i=1,2$ consists of model information $\Xi_i=(f_i,\varphi_i,l_i)$ and on-line information $Z_{it} \subset \{ x_{it}^{t_f}, u_{it}^{t_f} \}$, where $x_{it}^{t_f} := \{ x_{i\tau}, t\leq \tau \leq t_f \}$ and $u_{it}^{t_f} := \{ u_{i\tau}, t\leq \tau \leq t_f \}$.
To incentivize agent’s behavior in market model, we (or a market planner) use a reward (salary) functional of the following form. The reward (salary) functional: $$\begin{aligned}
W_i^w(t,x_t^{t_f};u)
=
w_{if}(t_f,x_{t_f}) + w_{i0}(t,x)
+ \int_{t}^{t_f} w_{i1}(\tau,x_\tau) d\tau
+ \int_{t}^{t_f} w_{i2}(\tau,x_\tau) dx_\tau,
\quad
t_0 \leq t \leq t_f, \ i=1,2, \end{aligned}$$ are defined along with the grid dynamics $ d x_\tau = f(\tau,x_\tau,u_\tau) d\tau + D(\tau) d\beta_\tau$ where $w=(w_1,w_2)$ and $w_i=(w_{if},w_{i0},w_{i1},w_{i2})$. Admissible parameters of the reward functional, denoted as $w=(w_1,w_2)\in\Pi\times \Pi$, are defiend by: $w_{if}(t_f,\cdot): \mathbb{R}^n\to \mathbb{R}$ is of class ${{\mathcal{C}}}^2$ and $\nabla_x w_{if}(t_f,\cdot)$ is polynomial growth; $w_{i0}:[t_0,t_f]\times\mathbb{R}^n\to \mathbb{R}$ is continuous at $(t,x)\in[t_0,t_f]\times\mathbb{R}^n$; $w_{i1}:[t_0,t_f]\times\mathbb{R}^n\to \mathbb{R}$ are ${{\mathcal{C}}}^1$ at $t\in[t_0,t_f]$ and ${{\mathcal{C}}}^2$ at $x\in\mathbb{R}^n$, and $\nabla_x w_{i1}$ is polynomial growth at $x\in\mathbb{R}^n$; $w_{i2}:[t_0,t_f]\times\mathbb{R}^n\to \mathbb{R}^{1\times n}$ is of class ${{\mathcal{C}}}^1$ and polynomial growth at $x\in\mathbb{R}^n$, and $\nabla_x w_{i2}$ is bounded. We use the notation $W_i^w$ so as to emphasize the dependence of $W_i$ on the choice of the parameter $w$. In the following discussion, we often use the same notation to show such parameter dependence. We try to express the parameter $w$ with another parameter $h$, which we call the price, so that the reward functional depends on the choice of the price $h$; then such dependence is also denoted as $W_i^h$.
The reward functionals together with the utility’s revenue functional and the agent’s revenue functional define the social welfare functional as: $$I^w(t,x;u) = J_0(t,x;u)
- {\mathbb{E}}_{t,x} \left[ W_1^w(t,x_t^{t_f};u) \right]
- {\mathbb{E}}_{t,x} \left[ W_2^w(t,x_t^{t_f};u) \right]$$ and the agent’s profit functional as: $$\begin{aligned}
I_i^w(t,x;u) = J_i(t,x;u)
+ {\mathbb{E}}_{t,x} \left[ W_i^w(t,x_t^{t_f};u) \right],
\quad
i=1,2.\end{aligned}$$
A market planner designs a market mechanism with incentivizing structures and makes auction rules as well, based on the evaluation functionals and the grid model information introduced so far; the auction is performed in the following five steps:
Step 1:
: Utility announces the auction system, and agents decide participation.
Step 2:
: Agent offers his/her bid based on his/her own private information.
Step 3:
: Based on agents’ bids, price is determined so as to maximize social welfare.
Step 4:
: Agent decides his/her control to maximize his/her own profit based on price.
Step 5:
: Utility pay rewards to agents.
Note that Steps 2, 3 and 4 will be performed continuously over a finite time interval.
Model-based One-shot Market Mechanism {#sec:3}
=====================================
Reward Design for Incentivizing {#sec:3.1}
-------------------------------
Components of our market model and their general interplay have been described in the previous section. To complete our market model, we need to fix a concrete shape of agents’ bidding, and design reward functionals by choosing their characterizing parameter, called the reward parameter, $w=(w_1,w_2)$. First, let us specify agents’ private information to be bidden in the market model discussed here: *Each agent’s model information $\Xi_i=(f_i,\varphi_i,l_i)$ is sent a priori to utility, and each agent’s on-line information to be bidden is just the current state, i.e., $Z_{it}=x_{it}$, which means that utility cannot access control input $u_i$.* Then, the design problem of our market is reduced to a social welfare maximization problem, called the reward design problem, subject to the constraints that provide the market with two incentivizing functions by rewards, which is formulated as follows: $$\begin{aligned}
&&
\max_{u\in\Gamma_1\times\Gamma_2,\ w\in\Pi\times\Pi}
I^w(t,x;u)
\nonumber \\
& \mbox{subject to} &
\nonumber \\
& \mbox{(Constraint 1)} &
I^w_1(t,x;u) = \max_{v_1\in\Gamma_1} I_1^w(t,x;v_1,u_2), \quad
I^w_2(t,x;u) = \max_{v_2\in\Gamma_2} I_2^w(t,x;u_1,v_2),
\nonumber \\
& \mbox{(Constraint 2)} &
I^w_1(t,x;u) \geq k_1(t,x), \qquad\qquad\qquad\
I^w_2(t,x;u) \geq k_2(t,x),
\nonumber \end{aligned}$$ where $k_i:[t_0,t_f]\times\mathbb{R}^n\to \mathbb{R}$ is continuous at $(t,x)\in[t_0,t_f]\times\mathbb{R}^n$. By solving this problem, we obtain the optimal reward functional with two incentive functions and the agents’ optimal controls. Constraint 1 claims that the reward incentivizes each agent’s behavior to adopt the optimal control that maximizes her own profit and, in other words, constitutes a Nash equilibrium together with the other agent’s control. This also implies that, since the utility holds the bidden models, the utility can know the control profile, even if it is not bidden. On the other hand, Constraint 2 assures a prescribed level of each agent’s profit. The above formulation is an application of the moral hazard problem in contract theory [@bib10; @bib11] to our market design problem; using terminology of contract theory, we call Constraint 1 and Constraint 2 the incentive compatibility constraint and the individual rationality constraint, respectively. The conventional contract (moral hazard) problems analyzed for static systems and dynamical systems with control inputs directly operated by the principal [@bib10]. The analysis is, however, in discord with the incentivizing market. The main challenge of our incentivizing market design is to reformulate the moral problems adapted to the market as above and synthesize the proposed market from the system and control perspective.
Solutions for General Reward Design {#sec:3.2}
-----------------------------------
To solve the reward design problem, we start specifying a form of the reward functionals by using Constraints 1 and 2. For a parameter $w=(w_1,w_2)\in\Pi\times\Pi$, let $(u_1^w,u_2^w)$ be a pair of optimal controls (a Nash equilibrium in $\Gamma_1\times\Gamma_2$) defined by $u_i^w=\arg\max_{u_i\in\Gamma_i} I_i^w(t,x;u_i,u_{-i}^w)$, $i=1,2$, so that Constraint 1 is fulfilled, where $u_{-1}:=u_2$ and $u_{-2}:=u_1$. Then, as shown in Appendix, the Hamilton-Jacobi-Bellman (HJB) equations for the value functions $$V_i^w(t,x) = \max_{u_i\in\Gamma_i} I_i^w(t,x;u_i,u_{-i}^w) - w_{i0}(t,x),
\quad i=1,2,
\label{eq:agent_Vfunc}$$ lead the reward functional to the form: $$\begin{aligned}
\!\!\!\!\!\!\!\!\!\!\!\! &&
W_i^w(t,x_t^{t_f};u_i,u_{-i}^w)
= h_{i0}^w(t,x) - \varphi_i(t_f,x_{it_f})
\nonumber \\
\!\!\!\!\!\!\!\!\!\!\!\! && \qquad
- \int_{t}^{t_f} \left[ h_{i1}^w(\tau,x_t)f(\tau,x_\tau,u_1^w(\tau,x_\tau),u_2^w(\tau,x_\tau))
+ l_i(\tau,x_{it},u_i^w(\tau,x_\tau)) \right] d\tau
+ \int_{t}^{t_f} h_{i1}^w(\tau,x_{\tau}) dx_{\tau}
\label{eq:V-W}\end{aligned}$$ along with $dx_{\tau} = f(\tau,x_\tau,u_i(\tau,x_\tau),u_{-i}^w(\tau,x_\tau))d\tau+D(\tau)d\beta_\tau$, $t\leq \tau \leq t_f$, where $h_{i0}^w$ and $h_{i1}^w$ are defined by
$$\begin{aligned}
h_{i0}^w(t,x) &=& V_i^w(t,x) + w_{i0}(t,x)
\label{eq:V-h0}
\\
h_{i1}^w(t,x) &=& \nabla_x V_i^w(t,x) + w_{i2}(t,x)
\label{eq:V-h1}\end{aligned}$$
\[eq:V-h\]
Moreover, $u_i^w$, $i=1,2$, which constitute a Nash equilibirum, must satisfy $$\begin{aligned}
u_i^w(\tau,x)
&=&
\arg\max_{u_i\in U_i}
\left[ h_{i1}^w(\tau,x) f(\tau,x,u_i,u_{-i}^w(\tau,x)) + l_i(\tau,x_i,u_i) + w_{i1}(\tau,x) \right]
\nonumber \\
&=&
\arg\max_{u_i\in U_i}
\left[ h_{i1}^w(\tau,x) f_i(\tau,x_i,u_i) + l_i(\tau,x_i,u_i) \right]
\nonumber \end{aligned}$$ so that a function $\mu_i$ given in Lemma \[lem:01\] provides uniquely $u_i^w$ with the expression of an explicit dependence on $h_{i1}^w$ such that $u_i^w(\tau,x) = \mu_i( \tau, x_i, h_{i1}^w(\tau,x) )$. For simplicity of notation, we will denote sometimes $\mu_i(\tau,x_i,h_{i1}^w(\tau,x))$ by $\mu_i^{h_{i1}^w}(\tau,x)$.
\[lem:01\] There exists a unique function $\mu_i$ that satisfies $$\begin{aligned}
\mu_i(\tau,x_i,p_i) =
\arg\max_{u_i\in U_i} \left[ p_i f_i(\tau,x_i,u_i) + l_i(\tau,x_i,u_i) \right],
\quad i=1,2, \end{aligned}$$ for each $(\tau,x_i,p_i) \in [t_0,t_f] \times \mathbb{R}^{n_i} \times \mathbb{R}^{1 \times n}$ such that $\mu_i$ is continuous at $(\tau,x_i,p_i)$ and Lipschitz continuous at $(x_i,p_i)$.
The continuity at $(\tau,x_i,p_i)$ follows from the uniqueness of the maximum. The Lipschitz continuity is shown by Lemma VI.6.3 in [@bib12].
Now, summarizing the above observation, we see that, in solving the reward design problem, Costraint 1 enables us to limit a search of the optimal reward functional to the class of the form (\[eq:V-W\]). In this form of the reward functional, $h_{i0}^w$ and $h_{i1}^w$ are given by (\[eq:V-h0\]) and (\[eq:V-h1\]), respectively, which implies that they depend on a choice of the parameter $w\in(w_1,w_2)\in\Pi\times\Pi$. We can show that this class of reward functionals is invariant, even if the class of parameters $h_i^w=(h_{i0}^w,h_{i1}^w)$ is generalized to a class where dependence on the parameter $w$ is not necessarily required. For this purpose, let $h_i=(h_{i0},h_{i1})$ and define a class of reward parameters $h=(h_1,h_2)\in H \times H$ such that $h_{i0}:[t_0,t_f]\times\mathbb{R}^n\to\mathbb{R}$ is continuous at $(t,x)\in[t_0,t_f]\times\mathbb{R}^n$; $h_{i1}:[t_0,t_f]\times\mathbb{R}^n\to\mathbb{R}^{1\times n}$ is of class ${{\mathcal{C}}}^1$ and polynomial growth at $x\in\mathbb{R}^n$, and $\nabla_x h_{i1}$ is bounded. Note that $h^w=(h_1^w,h_2^w)\in H\times H$ for any $w=(w_1,w_2)\in\Pi\times\Pi$ if $(V_i^w,\nabla_x V_i^w)$ is in the class $H$.
\[prop:01\] (a) A pair of controls $(u_1^w,u_2^w)$ constitutes a Nash equilibrium satisfying Constraint 1 for a pair of reward functionals $(W_1^w,W_2^w)$ with a parameter $w=(w_1,w_2)\in\Pi\times\Pi$ and the corresponding pair of value functions $(V_1^w,V_2^w)$ satisfies the condition that $(V_i^w,\nabla_x V_i^w)$, $i=1,2$ are in the class $H$, only if there is a parameter $h=(h_1,h_2)\in H \times H$ such that the pair of reward functionals has the form $$\begin{aligned}
\!\!\!\!\!\!\!\!\!\!\!\! &&
W_i^h(t,x_t^{t_f};u_i,\mu_{-i}^{h_{-i1}})
= h_{i0}^w(t,x) - \varphi_i(t_f,x_{it_f})
\nonumber \\
\!\!\!\!\!\!\!\!\!\!\!\! && \qquad
- \int_{t}^{t_f} \left[ h_{i1}(\tau,x_t)f(\tau,x_\tau,\mu_1^{h_{11}}(\tau,x_\tau),\mu_{2}^{h_{21}}(\tau,x_\tau))
+ l_i(\tau,x_{it},\mu_i^{h_{i1}}(\tau,x_\tau)) \right] d\tau
+ \int_{t}^{t_f} h_{i1}(\tau,x_{\tau}) dx_{\tau}
\label{eq:h-W}\end{aligned}$$ along with $ dx_\tau = f(\tau,x_\tau,u_i(\tau,x_\tau),\mu_{-i}^{h_{-i1}}(\tau,x_\tau))d\tau + D(\tau)d\beta_\tau$, $i=1,2$.\
(b) For the reward functionals (\[eq:h-W\]) with a parameter $h=(h_1,h_2)\in H\times H$, a pair of controls $(u_1,u_2)\in\Gamma_1\times\Gamma_2$ is a Nash equilibrium if and only if it has the form $$u_i(\tau,x) = \mu_i( \tau, x_i, h_{i1}(\tau,x) ), \quad i=1,2.
\label{eq:Nash_mu}$$ (c) For the reward functionals (\[eq:h-W\]) with a parameter $h=(h_1,h_2)\in H\times H$ and the Nash equilibrium (\[eq:Nash\_mu\]), Constraint 2 is fulfilled if and only if $h_{i0}$, $i=1,2$, are specified such as $h_{i0}(t,x) \geq k_{i}(t,x)$.
\(a) We have already seen that, for a chosen parameter $w=(w_1,w_2)\in\Pi\times\Pi$, the reward functionals for which $(u_1^w,u_2^w)$ constitutes a Nash equilibrium must have the form (\[eq:V-W\]) with the parameters (\[eq:V-h0\]) and (\[eq:V-h1\]), and the Nash equilibrium must be given as $u_i^w(\tau,x)=\mu_i(\tau,x_i,h_{i1}^w(\tau,x))$, $i=1,2$. Now, let $h=(h_1,h_2)\in H\times H$ be chosen independently of $w$ and set a reward parameter $\bar{w}=(\bar{w}_1,\bar{w}_2)\in\Pi\times\Pi$ as
$$\begin{aligned}
\bar{w}_{if}(t_f,x) &=&
- \varphi_i(t_f,x)
\\
\bar{w}_{i0}(t,x) &=&
h_{i0}(t,x)
\\
\bar{w}_{i1}(t,x) &=&
- h_{i1}(\tau,x) f(\tau,x, \mu_1^{h_{11}}(\tau,x),\mu_{2}^{h_{21}}(\tau,x)) - l_i(\tau,x_i,\mu_i^{h_{i1}}(\tau,x))
\\
\bar{w}_{i2} &=&
h_{i1}(\tau,x).\end{aligned}$$
Then, we can show that, for these reward parameters, the HJB equation (\[eq:DP\_V\]) in Appendix has a unique constant solution of the form $V_i^{\bar{w}}(\tau,x)=0$, so that we have $h_{i0}^{\bar{w}}(t,x) = V_i^{\bar{w}}(t,x) + \bar{w}_{i0}(t,x) = h_{i0}(t,x)$ and $h_{i1}^{\bar{w}}(\tau,x) = \nabla_x V_i^{\bar{w}}(\tau,x) + \bar{w}_{i2}(\tau,x) = h_{i1}(\tau,x)$. This implies that the class of reward functionals given by (\[eq:V-W\]) with (\[eq:V-h0\]) and (\[eq:V-h1\]) is invariant, even if the class of parameters $h^w=(h_1^w,h_2^w)$ depending on $w$ is generalized to $H\times H$, and proves the part (a) of this proposition.\
(b) For reward functionals of the form (\[eq:h-W\]) with a parameter $h=(h_1,h_2)\in H \times H$, profit functionals of the agent $i=1,2$ are represented as [ $$\begin{aligned}
I_i^h(t,x_t^{t_f};u_i,\mu_{-i}^{h_{-i1}})
&=& h_{i0}(t,x)
- \int_t^{t_f}
\left[ h_{i1}(\tau,x_\tau) f(\tau,x_\tau, \mu_1^{h_{11}}(\tau,x_\tau), \mu_{2}^{h_{21}}(\tau,x_\tau))
+ l_i(\tau,x_{i\tau},\mu_i^{h_{i1}}(\tau,x_\tau)) \right] d\tau
\nonumber \\
&&
+ \int_t^{t_f}
\left[ h_{i1}(\tau,x_\tau) f(\tau,x_\tau, u_i(\tau,x_\tau), \mu_{-i}^{h_{-i1}}(\tau,x_\tau))
+ l_i(\tau,x_{i\tau},u_i(\tau,x_\tau)) \right]
d\tau.
\label{eq:I_h}\end{aligned}$$]{} From the definition of $\mu_i$ given in Lemma \[lem:01\], the second (integral) term in the right hand side of the identity above is non-negative, and therefore the pair of controls $(u_1,u_2)\in\Gamma_1\times\Gamma_2$ is a Nash equilibrium if and only if $u_i=\mu_i^{h_{i1}}$, $i=1,2$.\
(c) It is obvious because the identity (\[eq:I\_h\]) guarantees $I_i^h(t,x_t^{t_f};\mu_1^{h_{11}},\mu_{2}^{h_{21}})=h_{i0}(t,x)$, $i=1,2$.
A key message of part (a) in Proposition \[prop:01\] is that the original parameter $w=(w_1,w_2)$ can be replaced with the parameter $h=(h_1,h_2)$. We will see below that the parameter $h$ can be interpreted as a price (vector), and show that it enables us to introduce a dynamic contract, which realizes requisite incentives, in the market model. Another message from the parts (a) and (b) is that we can shift the Nash equilibrium (\[eq:Nash\_mu\]) freely to some extent by selecting the price (vector) $h$. The parameterization of reward (salary) functional with the parameter $h$ would be in itself a new result of interest in contract theory, which is different from the known types based on typically the so-called first order condition [@bib11; @bib13; @bib14] and the other types [@bib15; @bib16] in the contract theory literatures. Finally, note that for proving this proposition we do not use the linearity of the grid model in the state, while we need the linearity and additivity in the controls in the grid model and the convexity (concavity) of the control ranges and the revenue functions in (A2) and (A4) as well.
Now, using Proposition \[prop:01\], we can present an optimal control based approach, in which the parameter $h=(h_1,h_2)\in H \times H$ plays a role of control, to the reward design problem.
\[thm:01\] The reward design problem with the parameter $h=(h_1,h_2)\in H \times H$ is equivalent to an optimal control problem described by $$\max_{(h_1,h_2)\in H\times H}
I^h(t,x;\mu_1^{h_{11}},\mu_2^{h_{21}})$$ subject to $$h_{i0}(t,x) \geq k_{i}(t,x), \quad i=1,2,$$ and the stochastic state equation: $$dx_\tau =
f(\tau,x_\tau, \mu_1^{h_{11}}(\tau,x_\tau),\mu_2^{h_{21}}(\tau,x_\tau))d\tau + D(\tau) d\beta_\tau,
\quad t \leq \tau \leq t_f,$$ where $\mu_i^{h_{i1}}(\tau,x)=\mu_i(\tau,x_i,h_{i1}(\tau,x))$, $i=1,2$.
From (a) of Proposition \[prop:01\] that the social welfare functional $I^h(t,x;\mu_1^{h_{11}},\mu_2^{h_{21}})$ is represented by $$\begin{aligned}
I^h(t,x;\mu_1^{h_{11}},\mu_2^{h_{21}})
&=&
J_0(t,x;\mu_1^{h_{11}},\mu_2^{h_{21}})
- {\mathbb{E}}_{t,x} \left[ \sum_{i=1}^2 W_i^h( t, x_t^{t_f} ; \mu_1^{h_{11}},\mu_2^{h_{21}} ) \right]
\nonumber \\
&=& {\mathbb{E}}_{t,x} \left[
\varphi_0(t_f,x_{t_f}) + \int_t^{t_f} l_0(\tau,x_\tau,\mu_1^{h_{11}}(\tau,x_{\tau}),\mu_2^{h_{21}}(\tau,x_{\tau})) d\tau
\right]
\nonumber \\
&&
+ {\mathbb{E}}_{t,x} \left[
\sum_{i=1}^2 \left( \varphi_i(t_f,x_{it_f}) + \int_t^{t_f} l_i(\tau,x_{i\tau},\mu_i^{h_{i1}}(\tau,x_\tau)) d\tau \right)
\right] - \sum_{i=1}^2 h_{i0}(t,x).
\label{eq:I0_h}\end{aligned}$$ Then, from (b) and (c) of Proposition \[prop:01\], Constraints 1 and 2 are fulfilled, respectively, for any $h=(h_1,h_2)\in H\times H$. Thus we have this theorem.
The optimal solution $h^*=(h_1^*,h_2^*)\in H \times H$ leads to the Nash equilibrium $(\mu_1^{h_{11}^*},\mu_2^{h_{21}^*})$, $i=1,2$. Note that $h_{i0}^*(t,x)=k_{i}(t,x)$, $i=1,2$ follows from the expression (\[eq:I0\_h\]), and then Constraint 2 is fulfilled.
Discussion through Typical Scenarios {#sec:4}
====================================
It is generally difficult to solve the optimal control problem in Theorem \[thm:01\]. Here, focusing on some special cases, we discuss qualitative properties of the parameter $h$ and try to give economic meanings to the parameter and the reward functional.
Let the value function be denoted by $$V(t,x) =
\sup_{(h_1,h_2)\in H\times\times H}
I^h(t,x;\mu_1^{h_{11}},\mu_2^{h_{21}}) + h_{10}(t,x) + h_{20}(t,x).$$ Then, the HJB equation is given by
$$\begin{aligned}
\nabla_t V(t,x) &+& \frac{1}{2} \left[ \nabla_x^2 V(t,x) D(t) D(t)^\top \right]
\nonumber \\
&+&
\sup_{(h_{11},h_{21}) \in \mathbb{R}^{1\times n}\times\mathbb{R}^{1\times n}}
\Big[
\nabla_x V(t,x) f_0(t,x)
+ l_0(t,x_i,\mu_1(t,x_1,h_{11}),\mu_2(t,x_2,h_{21}))
\nonumber \\
&& \qquad
+ \sum_{i=1}^2 \left(
\nabla_x V(t,x) f_i(t,x_i,\mu_i(t,x_i,h_{i1}))
+ l_i(t,x_i,\mu_i(t,x_i,h_{i1}))
\right) \Big]
= 0.
\\
V(t_f,x_{t_f}) &=& \varphi_0(t_f,x_{t_f}) + \varphi_1(t_f,x_{1t_f}) + \varphi_2(t_f,x_{2t_f}).\end{aligned}$$ \[eq:DP\_principal\]
In this section, we discuss the relationship between the private information and the incentives, interpretation and limitation of our approach through three typical cases.
[**(A)**]{} Consider the case when the utility evaluates only the grid state $x$ and does not evaluate the agents’ control inputs $u_i$, $i=1,2$ such that $l_0=l_0(\tau,x)$.
\[coro:01\] In Case (A), if the HJB equation (\[eq:DP\_principal\]) has a solution $V(t,x)$ such that $(V,\nabla_x V)$ is in the class $H$, the optimal parameters $h_i^*=(h_{i0}^*,h_{i1}^*)\in H$, $i=1,2$ are given by $h_{i0}^*(t,x) = k_i(t,x)$ and $h_{i1}^*(t,x)=\nabla_x V(t,x)$, $t_0 \leq t \leq t_f$.
$h_{i0}^*(t,x) = k_i(t,x)$ is already noted. In Case (A), the maximization in the HJB equation (\[eq:DP\_principal\]) becomes $$\begin{aligned}
\sum_{i=1}^2 \sup_{h_{i1}\in\mathbb{R}^{1\times n}}
\left[ \nabla_x V(t,x) f_i(t,x_i,\mu_i(t,x_i,h_{i1}))
+ l_i(t,x,\mu_i(t,x_i,h_{i1})) \right], \end{aligned}$$ and it follows from Lemma \[lem:01\] that the maximum is attained by $h_{i1}^*(t,x)=\nabla_x V(t,x)$, $i=1,2$. Then, the verification theorem [@bib11 Theorem VI.4.1] verifies the optimality of the parameter.
The fact $h_{i1}^*(t,x) = \nabla_x V(t,x)$, $i=1,2$, shown in Corollary \[coro:01\] implies that the reward parameter $h_{11}^*(t,x) (=h_{21}^*(t,x))$ can be regarded as a price of quantity $x$ at time $t$; $\nabla_x V(t,x)$ is actually called the shadow price in economics literatures, and our parametrization of the reward functional could be suitable for the market model. Note that the form of the utility’s revenue function as $l_0=l_0(\tau,x)$ is no so restrictive, since the utility dynamics has no control input.
[**(B)**]{} Consider the case that the utility’s revenue functional is given by $$\begin{aligned}
J_0(t,x;u) =
{\mathbb{E}}_{t,x} \left[
\varphi_0(t_f,x_{t_f}) + \int_{t}^{t_f} l_0(\tau,x_\tau) d\tau
+ \sum_{i=1}^2 \left( \varphi_i(t_f,x_{it_f}) + \int_t^{t_f} l_i(\tau,x_{i\tau},u_{i\tau}) d\tau \right) \right].\end{aligned}$$ That is, the utility’s revenue is the sum of the original utility’s revenue and the agent’s revenues. Assume further that the payment of the rewards for the agents is not liquidated in the social welfare, i.e., the utility’s revenue functional above is identical to the social welfare functional, and the agents ask no profit, i.e., $k_i(t,x) \equiv 0$, $i=1,2$. In this case, as the problem is basically equivalent to that in Case (A) with $k_i(t,x) \equiv 0$, $i=1,2$, repeating the same argument as in Case (A), we can obtain the same result as Corollary \[coro:01\] with $h_{i0}^*(t,x) \equiv 0$, $i=1,2$. The result shows that, if the price vector $\nabla_x V(t,x)$, called the adjoint vector in the optimal control theory, is provided by the utility, each agent can realize his/her optimal control in a decentralized way such as $\mu_i^{h_{i1}}(\tau,x) = \mu_i(\tau,x_i,\nabla_x V(\tau,x))$, $i=1,2$; this result corresponds to the dual decomposition of the static optimization based on “Lagrange multiplier" (price). On the other hand, from the viewpoint of the incentive design, each agent in Case (B) has a zero level of incentive to the participation in the market (the decentralized optimization based on the price), because he/she obtains no profit, whereas, in Case (A), agents have the profits $k_i(t,x)$, $i=1,2$, rewarded by the utility and have the incentives to the participation. We see that the implementation of this decentralized optimization scheme may require additional incentives or legal forces for strategic agents.
[**(C)**]{} The reward design discussed so far incentivizes agents to constitute a Nash equilibrium (Constraint 1) and to participate in the market if the profit level is over his/her expectation (Constraint 2). However, these are assured under the tacit assumption that the agents’ private information consisting of the model data and the on-line data is truthfully sent and bidden; if an agent fictitiously bids his/her private information, for example, the Nash equilibrium shifts or disappears; the *mechanism design* [@bib17; @bib18] provides a solution in such case by using additional incentives. Consider the same setting as in Case (B) where the social welfare functional given as above and does not include the budget for payment of the agents’ rewards, and, on the other hand, let the agent’s profit functional have an additional reward functional as $$\begin{aligned}
I_i^w(t,x;u) = J_i(t,x;u)
+ {\mathbb{E}}_{t,x} \left[ W_i^w(t,x_t^{t_f};u) \right]
+ {\mathbb{E}}_{t,x} \left[ W_{ai}^w(t,x_t^{t_f};u) \right]\end{aligned}$$ where $$\begin{aligned}
{\mathbb{E}}_{t,x}\left[ W_{a1}^w(t,x_t^{t_f};u) \right]
= J_0(t,x;u) + J_{2}(t,x;u), &&
{\mathbb{E}}_{t,x}\left[ W_{a2}^w(t,x_t^{t_f};u) \right]
= J_0(t,x;u) + J_{1}(t,x;u)\end{aligned}$$ and $l_0=l_0(\tau,x)$. In this case, replacing $\varphi_i(t_f,x_{it_f})$ with $\varphi_0(t_f,x_{t_f}) + \sum_{i=1}^2 \varphi_i(t_f,x_{it_f})$ and also $l_i(\tau,x_i,u_i)$ with $l_0(\tau,x) + \sum_{i=1}^2 l_i(\tau,x_i,u_i)$ and repeating the same argument as in Case (A), we have the same conclusion as in Corollary \[coro:01\], so that the agents should constitute a Nash equilibrium and participate in the market. In this case, moreover, the agents should report his/her model information and bid his/her on-line information truthfully; the reason for this is as follows. First, note that the additional reward $W^{ai}$ provides the utility and all the agents with the same revenue, so that the optimal price from the viewpoint of the social welfare is optimal for all the agents. Second, note that the utility calculates the optimal price based on the reported model and the bidden states. Therefore, if an agent sends or bids fictitiously his/her private information to the market, the agent obtains a price which is not optimal for his/her own profit. This incentivizing scheme corresponds to the Groves mechanism [@bib17] in mechanism design literatures. Finally, we point out an issue of this scheme; the rewards $W^{ai}$ should be additionally paid from the social welfare budget.
Conclusions {#sec:5}
===========
On the basis of a genetic model suggested from the average system frequency model [@bib09], we have discussed the incentivizing market and model-based approach to design the energy management and control systems which realize ancillary services in dynamic power grids. The key issue of the approach is to incentivize the agents (areas) to open their private information, which is essential to realize our model-based scheme, to the utility. We have proposed a design method of such incentivizing market by integrating the economics models and tools with the dynamic physical model, and clarified its basic properties of use together with its possibilities and limits for further developments.
Based on the principle of optimality, the value function (\[eq:agent\_Vfunc\]) leads to the HJB equation:
$$\begin{aligned}
\nabla_t V_i^w(t,x) &+& \frac{1}{2} {{\rm tr}}\left[ \nabla_x^2 V_i^w(t,x) D(t)D(t)^\top \right]
\nonumber \\
&=&
- \max_{u_i\in U_i} \left[ \left( \nabla_x V_i^w(t,x) + w_{i2}(t,x) \right)
f(t,x,u_i,u_{-i}^w(t,x)) + l_i(t,x_i,u_i) + w_{i1}(t,x) \right]
\nonumber \\
&=&
- \left[ \left( \nabla_x V_i^w(t,x) + w_{i2}(t,x) \right)
f(t,x,u_i^w(t,x),u_{-i}^w(t,x)) + l_i(t,x_i,u_i^w(t,x)) + w_{i1}(t,x) \right],
\\
V_i^w(t_f,x_{t_f}) &=& \varphi_i(t_f,x_{it_f}) + w_{if}(t_f,x_{t_f}). \end{aligned}$$ \[eq:DP\_V\]
Substituting the above relation into the right hand side of the Ito’s differential equality $$dV_i^w(t,x_t)
=
\left( \nabla_t V_i^w(t,x_t)
+ \frac{1}{2} {{\rm tr}}\left[ \nabla_x^2 V_i^w(t,x_t) D(t)D(t)^\top \right] \right) dt
+ \nabla_x V_i^w(t,x_t) dx_t$$ along with $dx_t = f(t,x_t,u_i(t,x_t),u_{-i}^w(t,x_t))dt + D(t)d\beta_t$, and integrating the both side on $[t,t_f]$, we have the reward functional of the form (\[eq:V-W\]) with (\[eq:V-h0\]) and (\[eq:V-h1\]).
[99]{} M. Amin, A.M. Annaswamy, C.L. DeMarco and T. Samad, [*IEEE vision for smart grid controls: 2030 and beyond*]{}, IEEE Press, 2013.
M.D. Ilic and S.X. Liu, [*Hierarchical Power Systems Control – Its Value in a Changing Industry*]{}, Springer, 1996.
Y.G. Rebours, D.S. Kirschen, M. Trotignon and S. Rossignol, “A survey of frequency and voltage control ancillary services – Part I: Technical features," [*IEEE Trans. Power Systems*]{}, vol. 22, no. 1, pp.350–357, 2007.
A.J. Wood and B.F. Wollenberg, [*Power Generator Operation and Control*]{}, Wiley, 1996.
M.A.B. Zammit, D.J. Hill and R.J. Kaye, “Designing ancillary services markets for power system security," [*IEEE Trans. Power Systems*]{}, vol. 15, no. 2, pp. 675–680, 2000.
E. Ela, V. Gevorgian, A. Tuohy, B. Kirby, M. Milligan and M. O’Malley, “Market designs for the primary frequency response ancillary service – Part I: motivation and design," [*IEEE Trans. Power Systems*]{}, vol. 29, no. 1, pp. 421–431, 2014.
M.D. Ilic, “Toward a unified modeling and control for sustainable and resilient electric energy systems," [*Foundations and Trends in Electric Energy Systems*]{}, vol. 1, no. 1–2, pp. 1–141, 2016.
A. Kiani, A. Annaswamy and T. Samad, “A hierachical transactive control architecture for renewables integration in smart grids: Analytical modeling and stability," [*IEEE Trans. Smart Grid*]{}, vol. 5, no. 4, pp. 2054–2065, 2014.
A.W. Berger and F.C. Schweppe, “Real time pricing to assist in load frequency control," [*IEEE Trans. Power Systems*]{}, vol. 4, no. 3, pp. 920–926, 1989.
P. Bolton and M. Dewatripont, [*Contract Theory*]{}, The MIT Press, 2005.
B. Holmstrom and P. Milgrom, “Aggregation and linearity in the provision of intertemporal incentives," [*Econometrica*]{}, vol. 55, no. 2, pp. 303–328, 1987.
W.H. Fleming and R.W. Rishel, [*Deterministic and Stochastic Optimal Control*]{}, Springer, 1975.
H. Schattler and J. Sung, “The first-order approach to the continuous time principal-agent problem with exponential utility," [*J. Economic Theory*]{}, vol. 61, pp. 331–371, 1993.
H.K. Koo, G. Shim and J. Sung, “Optimal multi-agent performance measures for team contracts," [it Mathematical Finance]{}, vol. 18, no. 4, pp. 649–667, 2008.
Y. Sannikov, “Contracts: The theory of dynamic principal-agent relationships and the continuous-time approach," In: D. Acemoglu, M. Arellano and E. Dekel (Eds.), [*Advances in Economics and Econometrics, 10th World Congress of the Econometric Society*]{}, Cambridge University Press, 2013.
J. Cvitanic and J. Zhang, [*Contract Theory in Continuous-Time Models*]{}, Springer, 2013
M.O. Jackson, “Mechanism theory," In: U. Derigs (Ed.), [*Encyclopedia of Life Support Systems*]{}, EOLSS Publishers, 2003.
Y. Okajima, T. Murao, K. Hirata, and K. Uchida, “A dynamic mechanism for LQG power networks with random type parameters and pricing delay," [*Proc. 52nd IEEE Conf. Decision and Control*]{}, pp. 2384–2390, 2013.
J. Moon and T. Basar, “Linear quadratic risk-sensitive and robust mean field games," [*IEEE Trans. Automatic Control*]{}, vol. 62, no. 3, pp. 1062–1077, 2017.
[^1]: Yasuaki Wasa (corresponding author) and Kenko Uchida are with the Department of Electrical Engineering and Bioscience, Waseda University, Tokyo 169-8555, JAPAN. Kenji Hirata is with the Graduate School of Science and Engineering for Research, University of Toyama, Toyama 930-8555, JAPAN. [[email protected], [email protected], [email protected]]{}.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'A new class of micromechanically motivated chain network models for soft biological tissues is presented. On the microlevel, it is based on the statistics of long chain molecules. A wormlike chain model is applied to capture the behavior of the collagen microfibrils. On the macrolevel, the network of collagen chains is represented by a transversely isotropic eight chain unit cell introducing one characteristic material axis. Biomechanically induced remodeling is captured by allowing for a continuous reorientation of the predominant unit cell axis driven by a biomechanical stimulus. To this end, we adopt the gradual alignment of the unit cell axis with the direction of maximum principal strain. The evolution of the unit cell axis’ orientation is governed by a first-order rate equation. For the temporal discretization of the remodeling rate equation, we suggest an exponential update scheme of Euler-Rodrigues type. For the spatial discretization, a finite element strategy is applied which introduces the current individual cell orientation as an internal variable on the integration point level. Selected model problems are analyzed to illustrate the basic features of the new model. Finally, the presented approach is applied to the biomechanically relevant boundary value problem of an in vitro engineered functional tendon construct.'
author:
- |
Ellen Kuhl\
Chair for Applied Mechanics\
University of Kaiserslautern, D-67653 Kaiserslautern, Germany\
[[email protected]]{}\
Krishna Garikipati\
Department of Mechanical Engineering,\
University of Michigan, Ann Arbor, MI 48109, USA\
[[email protected]]{}\
Ellen M. Arruda\
Department of Mechanical Engineering\
Macromolecular Science and Engineering Program\
University of Michigan, Ann Arbor, MI 48109, USA\
[[email protected]]{}\
Karl Grosh\
Department of Mechanical Engineering,\
University of Michigan, Ann Arbor, MI 48109, USA\
[[email protected]]{}
bibliography:
- 'litra.bib'
title: 'Remodeling of biological tissue: Mechanically induced reorientation of a transversely isotropic chain network'
---
\#1 \#1 \#1 \#1 \#1 \#1 \#1
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
We present a direct spectroscopic observation of a shallow hydrogen-like muonium state in SrTiO$_3$ which confirms the theoretical prediction that interstitial hydrogen may act as a shallow donor in this material. The formation of this muonium state is temperature dependent and appears below $\sim 70$ K. From the temperature dependence we estimate an activation energy of $\sim
50$ meV in the bulk and $\sim 23$ meV near the free surface. The field and directional dependence of the muonium precession frequencies further supports the shallow impurity state with a rare example of a fully anisotropic hyperfine tensor. From these measurements we determine the strength of the hyperfine interaction and propose that the muon occupies an interstitial site near the face of the oxygen octahedron in SrTiO$_3$. The observed shallow donor state provides new insight for tailoring the electronic and optical properties of SrTiO$_{3}$-based oxide interface systems.
author:
- 'Z. Salman'
- 'T. Prokscha'
- 'A. Amato'
- 'E. Morenzoni'
- 'R. Scheuermann'
- 'K. Sedlak'
- 'A. Suter'
title: 'Direct spectroscopic observation of a shallow hydrogen-like donor state in insulating SrTiO$_{3}$'
---
The discovery of a high mobility two-dimensional electron gas (2DEG) at the interface between two insulating Perovskite oxides; TiO$_{2}$-terminated SrTiO$_{3}$ (STO) and LaAlO$_{3}$ (LAO)[@Ohtomo04N; @Thiel06S; @Huijben06NM] has prompted great interest in these oxides. In addition to the 2DEG, this interface was found to be magnetic[@Brinkman07NM; @BenShalom09PRB; @Salman12PRL] and even superconducting below $\sim300$ mK[@Reyren07S]. It is generally agreed that the high carrier densities at the interface are associated with various effects, including doping with electrons or oxygen vacancies [@Thiel06S; @Pentcheva06PRB; @Park06PRB; @Takizawa06PRL; @Kalabukhov07PRB], inter-diffusion [@Takizawa06PRL; @Nakagawa06NM; @Willmott07PRL], and the influence of lattice distortions [@Ahn03N; @Gemming06AM; @Hamann06PRB; @Maurice06PSS; @Okamoto06PRL]. These discoveries provide an interesting prospect for producing interfaces with physical properties not present, nor predictable, from the constituent materials, and may lead to new technological applications. Here we discuss the response of these oxides to impurities, in particular hydrogen, which may play an important role in the discovered phenomena. Moreover, we note that the reported results are extremely relevant for possible applications of STO and related insulating oxides in fuel cells and hydrogen sensors [@Higuchi01PRB; @Higuchi98PRB; @Yukawa99SSI].
Hydrogen is an ubiquitous impurity in device or sample fabrication and can cause unintended modifications of the electronic and structural properties. For example, the properties of the 2DEG at STO/LAO interfaces may be affected if additional free charge carriers are present due to unintentional doping [@Son10JPCM]. The electronic behaviour of interstitial hydrogen can be characterized by the position of the H(+/-) level in the band gap, where the formation energies of H$^+$ and H$^-$ are equal. If this level is close to or intersecting the conduction/valence band, hydrogen will act as a shallow donor/acceptor. In elemental and binary semiconductors, a universal alignment of the H(+/-) level at $\sim$ -4.5 eV with respect to the vacuum level has been found theoretically [@vandeWalle03N] and supported by experiments [@Lichti08PRL]. The model predicted successfully hydrogen shallow donor states in ZnO and InN [@Cox01PRL; @Hofmann02PRL; @Davis03APL]. In this model the H(+/-) level coincides approximately with the host’s charge neutrality level (CNL). In general the CNL is found to be located at relatively constant energies with respect to the vacuum level [@vandeWalle03N; @Xiong07JAP]. The coincidence of the H(+/-) level with the CNL can be understood as follows. In a binary semiconductor H$^+$ and H$^-$ behave similarly, i.e. they break a bond at the anion or cation site, respectively, leaving a dangling bond at the opposite cation or anion site. The H(+/-) level is then located midway in between the energy levels of the dangling bonds, which corresponds to the CNL.
In oxides the situation is different, since H$^+$ tends to form an OH$^-$ antibonding state, without breaking a cation–O bond. H$^-$, on the other hand, causes the formation of an oxygen dangling bond by breaking a metal–O bond. Thus, the H(+/-) level is determined by the average of the oxygen dangling bond and OH$^-$ antibonding levels, which is located at relatively constant energies of about $4.5-5.5$ eV above the valence band maximum [@Xiong07JAP; @Peacock03APL]. Therefore, hydrogen should form a shallow donor state in oxides with a band gap of less than $\sim
4.5$ eV, such as STO which has an indirect band gap of $\sim
3.2$ eV. This prediction has been supported recently by a new theoretical work on hydrogenated vacancies and hidden hydrogen in STO [@Varley14PRB]. In contrast, the universal alignment model described above places H(+/-) deep in the band gap, i.e. $\sim 0.5$ eV below the conduction band minimum, which equals the electron affinity of STO and is located at about $4.0$ eV below the vacuum level. Note, the model in Refs. [@Xiong07JAP; @Peacock03APL] predicts shallow donor states in SnO$_2$, TiO$_2$, ZrO$_2$, and HfO$_2$ which have been observed or inferred in muon spin rotation ($\mu$SR) experiments [@Cox06JPCM2]. In these experiments, positive muons, when implanted into insulators or semiconductors, can capture an electron to form interstitial muonium \[Mu=($\mu^+\rm{e}^-$)\], which can be considered as a light hydrogen isotope and mimics its chemical and electrical interactions. In fact, a large amount of information on the structure and electrical activity of isolated interstitial H states in semiconductors and insulators has been obtained by Mu spectroscopy [@Patterson88RMP; @muSR2010].
In this Letter we present a direct spectroscopic observation of a shallow hydrogen-like Mu state in a STO (100) single crystal, in agreement with the model suggested by Refs. [@Xiong07JAP; @Peacock03APL]. Our results are consistent with muons occupying an interstitial site in the lattice, between two O–O bonds near the face of the oxygen octahedron. We find that up to $\sim
60$% of the implanted muons form a shallow muonium state at 25 K with a relatively small hyperfine coupling. From the field dependence of the Mu characteristic precession frequencies we find that the hyperfine tensor is fully anisotropic and estimate the hyperfine interaction along the principal axes of the tensor $A_X=1.4 \pm 0.1$ MHz, $A_Y=6.7\pm 0.1$ MHz and $A_Z=11.5 \pm 0.1$ MHz. This Mu state is one of the rare clear examples of a fully anisotropic Mu.
The $\mu$SR experiments were performed on the DOLLY, GPS and LEM spectrometers at the Paul Scherrer Institut in Villigen, Switzerland. In these experiments $100 \%$ polarized positive muons are implanted into the sample. Each implanted muon decays (lifetime $\tau_{\mu}=2.2$ $\mu$s) emitting a positron preferentially in the direction of its polarization at the time of decay. Using appropriately positioned detectors, one measures the asymmetry of the muon beta decay along different directions as a function of time, $A(t)$, which is proportional to the time evolution of the muon spin polarization. $A(t)$ depends on the electronic environment of the muon and is used to extract information on the hyperfine interaction between the muon and the electrons in the system. Muonium is spectroscopically identified by its characteristic precession frequencies which allow to determine the Mu hyperfine parameters [@muSR2010; @Patterson88RMP]. In a low energy $\mu$SR (LE-$\mu$SR) experiment, the energy of the implanted muons, $E$, can be tuned (1-30 keV) to perform a measurement of $A(t)$ at depths in the range $\sim 1-200$ nm [@Morenzoni03PB; @Prokscha08NIMA]. The measurements reported here were performed on a $15 \times 15 \times 1$ mm single side epitaxially polished $(100)$ STO single crystal substrate (Crystal GmbH). In the bulk $\mu$SR measurements, the sample was suspended on an aluminized Mylar tape and mounted into a He gas flow cryostat. The muons were implanted with their polarization nominally along $\langle 100
\rangle$ with the field applied perpendicular to it (nominally along $\langle 010 \rangle$). In the LE-$\mu$SR measurements the sample was glued to a cold finger cryostat, with the field applied along $\langle
100 \rangle$ and the polarization of implanted muons perpendicular to it.
![(Color online) The normalized diamagnetic fraction of the $\mu^+$ precession signal at $B=10$ mT as a function of temperature. Circles, squares and triangles are measurements at 4.1 MeV ($B \perp \langle100\rangle$), 14.1 keV and 1.6 keV ($B \parallel \langle100\rangle$), respectively. The drop in asymmetry below $\sim 70$ K is due to formation of muonium. The dashed lines are fits to Eq. \[Activation\].[]{data-label="AsyT"}](AsyT){width="0.8\columnwidth"}
We start by looking at muon spin precession measurements in a field of $B=10$ mT applied perpendicular to the muon’s initial spin direction. At room temperature we find that all muons implanted at 4.1 MeV in bulk STO precess at the Larmor frequency of $\mu^+$, $\nu_0=\gamma_\mu/2\pi~B$, with almost no damping. Here, $\gamma_\mu/2\pi = 135.5$ MHz/T is the muon gyromagnetic ratio. As we decrease the temperature, we find that below $\sim 70$ K the amplitude of the signal precessing at $\nu_0$ (the so called diamagnetic fraction, $f_{\mu^+}$) decreases sharply, reaches a minimum at $\sim
30$ K, and then increase at lower temperatures (Fig. \[AsyT\]). Similar behaviour is observed at lower muon implantation energies. However, in this case the decrease in the $f_{\mu^+}$ is smaller, and at $E=1.6$ keV we observe no increase at low temperatures.
Closer investigation of the measured asymmetries in the bulk below $70$ K reveals that the polarization contains additional components with precession frequencies higher than $\nu_0$. Example asymmetries, measured at $25$ K and applied fields of 1 and 10 mT are shown in Fig. \[AsyField\](a). Note that the additional frequencies are of the order of few MHz \[see Fig. \[AsyField\](b)\], which we attribute to Mu precession frequencies.
![(Color online) (a) The asymmetries at $T=25$ K as a function of time for applied fields 1 and 10 mT ($B \perp
\langle100\rangle$) and (b) the corresponding real Fourier transform showing at least four muonium precession frequencies in both cases. The solid lines are fits to a sum of precessing and damping signals.[]{data-label="AsyField"}](AsyvsB){width="0.8\columnwidth"}
These are directly related to the hyperfine interaction of Mu and confirm its formation at these temperatures. The size of these frequencies ($<10$ MHz) indicates that a “shallow” Mu state is formed, i.e. much smaller than the $\sim 4.4$ GHz frequency observed for vacuum Mu.
The temperature dependence of $f_{\mu^+}$ is related to the Mu fraction, $f_{\rm Mu}=1-f_{\mu^+}$. It can be calculated following a semi-empirical model [@Cox06JPCM2], $$\label{Activation}
f_{\mu^+} = 1 - f_{\rm Mu}= 1 - \frac{f_0}{1+N \exp (-\Delta/T)},$$ where $f_0$ is the normalized maximum Mu fraction, $N$ is a density-of-states parameter and $\Delta$ is the Mu activation energy which can be considered as an effective ionization or binding energy [@Cox06JPCM2]. A fit of $f_{\mu^+}$ in bulk (circles in Fig. \[AsyT\]), between 25 and 200 K, yields an ionization energy of $\Delta=600 \pm 15$ K or $52 \pm 2$ meV. Similar fits for the data at low implantation energies yield $\Delta=515 \pm 15$ K ($45 \pm 2$ meV) and $270\pm 10$ K ($23\pm1$ meV) for 14.1 keV and 1.6 keV, respectively. Similar decrease in shallow muonium ionization energy near the surface has been recently detected in ZnO and CdS. This effect is attributed to the presence of electric fields due to band bending near the free surface [@Prokscha14condmat].
It is worthwhile noting here that the absence of (or weak) energy dependence of the diamagnetic fraction below $\sim 10$ K indicates that the Mu formation at low $T$ does not depend on the number of track electrons, i.e., there is no *delayed* Mu formation below $T \sim 10$ K. Mu can form in two ways: i) *promptly* during charge-exchange collisions with subsequent thermalization as neutral Mu, and ii), *delayed*, where the positive muon thermalizes at an interstitial lattice site, followed by a capture of an electron from its own ionization track. The latter is significantly suppressed if the number of track electrons is $\lesssim 10^3$, corresponding to implantation energies of less than a few keV [@Prokscha07PRL]. It is also suppressed in materials with a large dielectric constant, $\varepsilon$, where the electric field of the muon point charge is effectively screened by the surrounding medium, thus suppressing the probability of capturing a track electron. In STO, $\varepsilon \sim
300$ at room temperature, reaches $\sim 10000$ at $\sim25$ K and then saturates at $20000-25000$ below 10 K [@Mueller79PRB]. The increase of the diamagnetic fraction below $\sim 30$ K in the bulk may be due to the increasing $\varepsilon$ at these temperatures, as *delayed* Mu formation becomes less likely. The low temperature flattening of $f_{\mu^+}$ corresponds well to the saturation of $\varepsilon$.
It is somewhat surprising that a shallow Mu state can be observed in a system with such a large $\varepsilon$. For shallow donors, one usually applies a hydrogenic effective-mass model to estimate the binding energy of the donor to be $E_D \sim R_y
(m^*/m_e)/\varepsilon^2$, where $R_y = 13.6$ eV the Rydberg constant , $m_e$ the electron mass, and $m^*$ the conduction-band effective mass of the electron [@Cox06JPCM2]. This approximation is justified by the large electron wave-function spreading over several lattice sites. This implies that at 25 K, where $\varepsilon \sim 10000$ in STO, the binding energy should be zero, i.e. no Mu should form. Therefore, the observation of a shallow Mu state indicates that it might have a more localized polaronic character [@Cox06JPCM2]. Furthermore, it is known that doping of STO may increase $m^*/m_e$ up to $\sim 20$ [@Ravichandran11PRB], and that an electric field may lower $\varepsilon$ [@Hyun01APL]. Hence, “doping” STO with Mu may cause a local deformation of the lattice with a local modification of $m^*$ and $\varepsilon$.
Now we turn to evaluating the hyperfine interaction tensor of Mu. We consider a general Hamiltonian for a muon (spin $\mathbf{I}$) interacting with an electron (spin $\mathbf{S}$) with a fully anisotropic hyperfine interaction [@Senba00PRA], $$\label{Ham}
{\cal H}=\gamma_{e}BS_{z}-\gamma_\mu BI_z+A_Z S_{Z}I_{Z}+A_Y S_{Y}I_{Y}+A_X S_{X}I_{X}.$$ where $\gamma_{\mu}$ and $\gamma_e$ are the muon and electron gyromagnetic ratios, $z$ is defined by the direction of $B$ and $A_i$ ($i=X,Y,Z$) are the Mu hyperfine interactions along the principal axes.
![(Color online) The asymmetries measured (a) parallel and (b) perpendicular to $\langle100\rangle$ at $T=25$ K and in ZF. (c) and (d) are the corresponding Fourier transform showing clear ZF muonium precession frequencies in both cases. The solid lines are fits to a sum of precessing and damping signals[]{data-label="ZFAsy"}](ZFAsy){width="1.0\columnwidth"}
The coordinates $(X,Y,Z)$/$(x,y,z)$ denote the hyperfine interaction/laboratory frame, such that the components of a spin angular momentum vector seen in the $(X,Y,Z)$ system are expressed in terms of the components of the same vector seen in the $(x,y,z)$ system as $$\left[ \begin{array}{l} S_X \\ S_Y \\ S_Z \end{array} \right] =
D(\alpha,\beta,\gamma)
\left[ \begin{array}{l} S_x \\ S_y \\ S_z \end{array} \right],$$ where $\alpha$, $\beta$,and $\gamma$ are the Euler angles of the three consecutive rotations around $Z$, $Y$, and $Z$ axes of the $(X,Y,Z)$ coordinate system, which initially coincides with the $(x,y,z)$ system. In ZF, the Hamiltonian (\[Ham\]) can be diagonalized analytically to calculate the Mu frequencies, giving a maximum of 6 possible frequencies depending on the values of $A_i$ [@Senba00PRA]. These are the sums and differences of the different $A_i$ parameters. Indeed our ZF measurement, shown in Fig. \[ZFAsy\](a) and (b), with different relative orientations between the implanted muon spin and STO crystal give $\nu=2.4, 2.7,
4.1, 5.1, 6.5$ and $9$ MHz \[Fig. \[ZFAsy\](c) and (d)\]. From these we estimate the hyperfine parameters $A_X=1.4 \pm 0.1$ MHz, $A_Y=6.7\pm 0.1$ MHz and $A_Z=11.5 \pm 0.1$ MHz.
Next, we extract the field dependent Mu precession frequencies from the asymmetries measured in bulk STO at 25 K. Here, we limit ourselves to four/five most visible frequencies for each field, as plotted in Fig. \[FrqB\].
![(Color online) The observed Muonium precession frequencies as a function of applied field (with an angle of $\sim 14^\circ$ between $B$ and $\langle010\rangle$) at $T=25$K. The lines are calculated values with $A_X=1.37$, $A_Y=6.67$ and $A_Z=11.52$ MHz with $(\alpha,\beta,\gamma)=(0^\circ,57^\circ,27^\circ)$ (solid) and $(0^\circ,17^\circ,0^\circ)$ (dashed).[]{data-label="FrqB"}](FrqB){width="0.8\columnwidth"}
Using $A_i$ values we calculate the Mu frequencies in an applied field by numerical diagonalization of ${\cal H}$. The amplitudes or probabilities of precession between different Mu energy states depend on the initial spin direction of the muon relative to the hyperfine principal axes. We find that in order to best model the field dependence in Fig. \[FrqB\], one has to set the $(\alpha,\beta,\gamma)=(0^\circ,57^\circ,27^\circ)$ (solid line in Fig. \[FrqB\]) and $(0^\circ,17^\circ,0^\circ \pm 45^\circ)$ (dashed line in Fig. \[FrqB\]). The agreement with our experimental results is excellent considering the uncertainties and limited resolution in the experimental data. Note, there is a small frequency splitting in the highest frequency branch. This may be attributed to a small amount of mosaicity or the structural distortion in STO at this temperature [@Cowley96PTSL]. Such effect results in different domain orientations and thus a small variation in the local environment of Mu, which can only be resolved at high frequencies. Using the extracted hyperfine parameter to calculate the precessing amplitudes reveals that there is no missing fraction, i.e., $f_{\mu^+} + f_{\rm
Mu}^{\rm S} = 1$, where $f_{\rm Mu}^{\rm S}$ is the fraction of shallow Mu. This excludes the existence of any Mu state deep in the band gap. Although we cannot exclude that the observed neutral muonium centre is metastable, the fact that the summ of the Mu$^+$ and Mu$^0$ fractions accounts for the full muon polarization, suggests an equilibrium balance between the two charge states. The observed hyperfine interaction strength and ionization energy imply a shallow donor state, modelling the analogous hydrogen state [@Cox09RPP].
Note that the angle between the STO cubic axes and the normal to the face of the oxygen octahedron are $\sim 54^\circ$. Therefore, the hyperfine tensor and the first set of angles (solid lines in Fig. \[FrqB\]) could be attributed to a Mu occupying an interstitial site between two O–O bonds and near the face of the oxygen octahedron in the STO crystal. The determination of the second set of angles (dashed lines in Fig. \[FrqB\]) is much less reliable since it shows a much weaker angular dependence. Nevertheless, if we assume that the angle between the applied field and the $\langle 010 \rangle$ is $\sim
14^\circ$, then these angles are also consistent with the same Mu site. Our results are consistent with neutron diffraction results [@Sata96PRB], infrared absorption experiments and theoretical studies on hydrogen defect vibrational modes [@Tarun11JAP; @Thienprasert12PRB] as well as other theoretical work [@Varley14PRB]. However, they disagree with Refs. [@Weber86PRB; @Houde87PRB; @Klauer92PRL; @Villamagua07PS] which place the hydrogen on O–O bond or the face of the cube between corner sharing Sr atoms and the O atoms at the face center. Note also that such sites, which have high symmetry, will result in an axially symmetric Mu hyperfine tensor. Surprisingly, we also find that the implanted muons occupy a different site from that occupied by other implanted impurities in STO such as Li [@Salman04PRB; @Salman06PRL].
In conclusion, we present a direct spectroscopic observation of a shallow hydrogen-like muonium state in STO. This confirms a theoretical prediction that interstitial hydrogen may act as a shallow donor in STO [@Xiong07JAP; @Peacock03APL; @Varley14PRB]. The formation of this muonium state appears below $\sim 70$ K and implies an activation energy of $\sim 50$ meV in bulk which decrease to $\sim
25$ meV near the surface of the crystal. We find that the shallow impurity state has a fully anisotropic hyperfine tensor, with $A_X=1.4
\pm 0.1$ MHz, $A_Y=6.7\pm 0.1$ MHz and $A_Z=11.5 \pm 0.1$ MHz. These results provide strong evidence of the sensitivity of the electronic properties of STO, and in particular its surface region, to impurities. Finally, since hydrogen is an ubiquitous impurity, these findings may prove crucial for interpretation of the variety of observed phenomena at LAO/STO interfaces [@Ohtomo04N; @Thiel06S; @Huijben06NM]. We believe that hydrogen doping effect may be a possible explanation for the excess charge carriers at the interfaces of LAO/STO, and therefore require a more detailed experimental and theoretical consideration. We propose that a systematic study of the transport properties of LAO/STO interfaces as a function of hydrogen doping may provide quantitative information about this effect.
This work was performed at the Swiss Muon Source (S$\mu$S), Paul Scherrer Institute (PSI, Switzerland). We would like to thank Rob Kiefl, Kim Chow and João Campos Gil for useful suggestions and fruitful discussions. We are also grateful to Ekaterina Pomjakushina for her assistance with the allignment of the crystals.
[52]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, , , , , ****, ().
, , , , , , , , , ****, ().
, , , , , , , , , , ****, ().
, , , , , , , , ****, ().
, , , , , , , , , , , ****, ().
, , , , , , , , , , , ****, ().
, ****, ().
, , , ****, ().
, , , , , , , , , , , ****, ().
, , , , , , ****, ().
, , , ****, ().
, , , , , , , , , , , ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, , , , , , , ****, ().
, , , ****, ().
, , , , , , , , , ****, ().
, , , , , , ****, ().
, , , ****, ().
, , , , ****, ().
, ****, ().
, , , ****, ().
, , , , , , , , , , , ****, ().
, , , , , , , , , ****, ().
, , , , ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, , , , , , , , , , , ****, ().
, ****, ().
, ** (, ).
, , , , , , , , , , , ****, ().
, , , , , , , , ****, ().
, , , , , , , , (), .
, , , , , , , , ****, ().
, ****, ().
, , , , , , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , , , ****, ().
, ****, ().
, , , , , ****, ().
, , , ****, ().
, , , , , ****, ().
, ****, ().
, , , , ****, ().
, , , , , , , , , , ****, ().
, , , , , , , , , , , ****, ().
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this paper, we obtain a canonical central element $\nu_H$ for each semi-simple quasi-Hopf algebra $H$ over any field $k$ and prove that $\nu_H$ is invariant under gauge transformations. We show that if $k$ is algebraically closed of characteristic zero then for any irreducible representation of $H$ which affords the character $\chi$, $\chi(\nu_H)$ takes only the values 0, 1 or -1, moreover if $H$ is a Hopf algebra or a twisted quantum double of a finite group then $\chi(\nu_H)$ is the corresponding Frobenius-Schur Indicator. We also prove an analog of a Theorem of Larson-Radford for split semi-simple quasi-Hopf algebra over any field $k$. Using this result, we establish the relationship between the antipode $S$, the values of $\chi(\nu_H)$, and certain associated bilinear forms when the underlying field $k$ is algebraically closed of characteristic zero.'
author:
- 'Geoffrey Mason and Siu-Hung Ng'
title: 'Central Invariants and Frobenius-Schur Indicators for Semisimple Quasi-Hopf Algebras'
---
Introduction
============
In the paper ([@LM00]), Linchenko and Montgomery introduced and studied Frobenius-Schur indicators for irreducible representations of a semi-simple Hopf algebra $H$ over an algebraically closed field of characteristic $p \neq 2$. If $\Lambda$ is the *unique* normalized left integral of $H$, i.e. $\e(\Lambda) = 1$, set $$\label{eq: centralhopfelt}
\nu = \nu_H = \sum_{(\Lambda)} \Lambda_1 \Lambda_2.$$ Here we have used Sweedler notation $ \Delta(\Lambda) =
\sum_{(\Lambda)} \Lambda_1 \otimes \Lambda_2$, so that if $m$ is multiplication in $H$ then $\nu = m \circ \Delta
(\Lambda)$. Then $\nu$ is a central element of $H$ and the Frobenius-Schur indicator $\nu_{\chi}$ of an irreducible $H$-module $M$ with character $\chi$ is defined via $$\label{eq: hopfs}
\nu_{\chi} = \chi(\nu).$$ In case $H$ is a group algebra $k[G], \nu = |G|^{-1}\sum_{g \in
G} g^2$ and $\nu_{\chi} = |G|^{-1}\sum_{g \in G} \chi(g^2)$ reduces to the original definition of Frobenius and Schur (cf.[@CR88] or [@serre77], for example). Generalizing the famous result of Frobenius and Schur for group algebras, Linchenko and Montgomery show that for general semi-simple $H$, $\nu_{\chi}$ can take only the values $0, 1$, or $-1$. Moreover $\nu_{\chi}
\neq 0$ if, and only if, $M \cong M^*$, and in this case $M$ admits a non-degenerate $H$-invariant bilinear form $\<\cdot
,\cdot \>$ satisfying $$\<u, v\> = \nu_{\chi} \<v, u\>$$ for $u, v \in M$. Recall that $\<\cdot ,\cdot \>$ is $H$-invariant if $$\label{eq: invbilform}
\sum_{(h)}\<h_1u, h_2v\> = \epsilon(h)\<u, v\>$$ for $h \in H$ and $u, v \in M$.
In a recent paper ([@KMM02]) the authors showed how one may effectively compute Frobenius-Schur indicators for a certain class of Hopf algebras. Their work applies, in particular, to the case of the *quantum double* $D(G)$ of a finite group $G$, and it was shown (loc. cit.) how the indicators for irreducible modules over $D(G)$ may be given in terms of purely group-theoretic invariants associated to $G$ and its subgroups. The algebra $D(G)$ is of interest in orbifold conformal field theory ([@Mas]), indeed in this context there is a more general object, the *twisted quantum double* $D^\omega(G)$, that arises naturally ([@DPR90]). (Here, $\omega \in Z^3(G,
\BC^\times)$ is a normalized $3$-cocycle about which we shall have more to say below.) The present work originated with a natural problem: understand Frobenius-Schur indicators for twisted quantum doubles.
$D^{\omega}(G)$ is a semi-simple *quasi-Hopf* algebra (over $\BC$, say), but is generally not a Hopf algebra. One of the difficulties this imposes is that the antipode $S$ is not necessarily involutorial (something that is always true for semi-simple Hopf algebras by a Theorem of Larson and Radford ([@LaRa87])), whereas having $S^2 = id$ is fundamental for the Linchenko-Montgomery approach and therefore for the calculations in ([@KMM02]. If it happens that $S^2 = id$ then Theorem 4.4 of (loc. cit.) can be used to obtain indicators given by $$\label{eq: kmmfsind}
\nu_{\chi} = |G|^{-1} \sum_{ x^{-1}gx=g^{-1}} \gamma_x(g, g^{-1})
\theta_g(x, x) \chi(e(g) \otimes x^2).$$ (Undefined notation is explained below; $\gamma_x$ and $\theta_g$ are certain $2$-cochains determined by $\omega$.)
If $G$ is *abelian* then $D^{\omega}(G)$ is a Hopf algebra ([@MN01]), though perhaps with a non-trivial $\beta$ element, and for any $G$ it turns out that one can always *gauge* $\omega$, i.e. replace it by a cohomologous $3$-cocycle $\omega'$, in such a way that the antipode for $D^{\omega'}(G)$ is an involution. So (\[eq: kmmfsind\]) provides a preliminary solution to our problem, but it is unsatisfactory for the following reason: if we gauge $\omega$, the new $3$-cocycle $\omega'$ will give new values for the Frobenius-Schur indicators which in general are not the same as the original values. While this may not be an issue if one is interested in a fixed $D^{\omega}(G)$, there are both mathematical and physical reasons for insisting that the FS indicators for $D^{\omega}(G)$ be *robust*, that is they depend only on the cohomology class of $\omega$. From this standpoint, (\[eq: kmmfsind\]) is generally not what we are looking for. We need a more *functorial* approach.
One knows that if $\omega$ and $\omega'$ are cohomologous then $D^{\omega}(G)$ and $D^{\omega'}(G)$ are gauge-equivalent and that therefore the corresponding module categories are tensor equivalent (cf. [@Drin90], [@DPR90], [@Kassel]). Indeed, it follows from a result of Etinghof and Gelaki ([@EG02]) that the converse is also true, so that gauge-equivalence of the twisted doubles is the *same* as tensor equivalence of the module categories. So we are looking for invariants of such module categories with respect to tensor equivalence. Because Hopf algebras and twisted doubles are not closed with respect to gauge equivalence, this means that we have to work with the module categories of *arbitrary* semi-simple quasi-Hopf algebras.
Peter Bantay has introduced a notion of indicator into rational conformal field theory from a rather different point-of-view ([@Bantay97], [@Bantay00]). His point of departure is the Verlinde formula and the $S$ and $T$ matrices associated to a RCFT. To this modular data together with an irreducible character he associates a certain numerical expression and shows that it is equal once again to either $0, 1$ or $-1$. It is possible to evaluate Bantay’s indicator in case the matrices $S$ and $T$ are associated to a twisted double $D^{\omega}(G)$ ([@BantayConv]) and one obtains the expression $$\label{eq: bantayfsind}
|G|^{-1} \sum_{ x^{-1}gx=g^{-1}}\omega(g^{-1}, g, g^{-1})
\gamma_x(g, g^{-1}) \theta_g(x, x) \chi(e(g) \otimes x^2).$$ Compared to (\[eq: kmmfsind\]), (\[eq: bantayfsind\]) contains an extra term $\omega(g^{-1}, g, g^{-1})$. Furthermore, it is easy to see that (\[eq: bantayfsind\]) is robust in the previous sense.
Suppose that $H$ is any semi-simple quasi-Hopf algebra, and let $M$ be an irreducible $H$-module with character $\chi$. In the present paper we will construct a *canonical* central element $\nu_H$ of $H$ with the following properties:
> 1. $ \nu_H$ is invariant under *any* gauge transformation of $H$.
>
> 2. If $H$ is a Hopf algebra then $\nu_H$ coincides with (\[eq: centralhopfelt\]).
>
> 3. $
> \mbox{If} \ H = D^{\omega}(G) \ \mbox{then} \ \chi(\nu_H) \
> \mbox{coincides with Bantay's indicator (\ref{eq: bantayfsind})}.$
>
> 4. Asumme that $k$ is algebraically closed and char $k=0$.
>
> 1. $\chi(\nu_H) = 0, 1, \ \mbox{or} -1 $.
>
> 2. $ \chi(\nu_H) \neq 0$ if, and only if, ${^*\!M} \cong M$. In this case, $M$ admits a certain non-degenerate bilinear form $\<\cdot , \cdot\> $ such that $$\label{qhopfbilform}
> \<x, y\> = \<y, g^{-1} x\>$$ for all $x$, $y \in M$. Here, $g$ is a distinguished element of $H$, which we call the *trace element*, which is independent of $M$.
>
> 3. $\displaystyle \Tr(S) =
> \sum_{\chi \in \Irr(H)}\chi(\nu_H) \chi(\b^{-1})$.
>
Part (d) is the analog for general semi-simple quasi-Hopf algebras of the corresponding result in [@LM00] for Hopf algebras. The bilinear form $\<\cdot , \cdot\> $ has a certain adjointness property with respect to the antipode $S$ of $H$, and there are relations to an analog of a Theorem of Larson-Radford ($S$ is involutorial for semi-simple Hopf algebras). Namely, we show that for a semi-simple quasi-Hopf algebra the antipode is involutorial up to conjugation. The trace element $g$ plays an important role in our discussion of (d), in particular its properties lead to the fact that the category $H$-$\mod_{fin}$ of finite-dimensional H-modules is a pivotal category in the sense of Joyal and Street. For twisted doubles, $g$ coincides with $\beta$, while for Hopf algebras the Larson-Radford Theorem implies that $g = 1$.\
The proof that $\chi(\nu_H)$ takes only the values 0, 1 or -1 is somewhat elaborate. Indeed, in an earlier version of the present paper ([@MN03]) this had been left open. Subsequently, Pavel Etingof alerted us to the existence of his recent preprint with Nikshych and Ostrik [@ENO] on fusion categories, and suggested that some of the results obtained there could be used to help settle the issue of the values of our indicator. More precisely, Etingof pointed out that our trace element $g$ defines an isomorphism of tensor functors $Id \longrightarrow {^{**}?}$. This together with $S(g) =g^{-1}$ are the main ingredients in the proof.\
The paper is organized as follows: we cover some basic facts about quasi-Hopf algebras in Section 2, including several strategically important elements in $H \otimes H$ introduced by Hausser and Nill [@HNQA]. In Sections 3 and 4 we define the central element $\nu_H$ and establish that the family of Frobenius-Schur indicators $\chi(\nu_H)$ is a gauge invariant for semi-simple quasi-Hopf algebras. In section 5 we show that our indicators coincide with those of Bantay in the case of a twisted double. In Section 6 we introduce the trace element $g$ and establish the analog of the Larson-Radford Theorem, while Section 7 is devoted to further properties of $g$ as discussed above. Section 8 covers the relation of indicators to bilinear forms and completes the proof of (d)(i), and in Section 9 we return to the case of twisted doubles to complete the analysis in that case. For simplicity, we will only work on algebraically fields of characteristic zero in Section 7, 8 and 9.\
The authors are indebted to Pavel Etingof for his interest and extended correspondence, and also thank Peter Bantay and Susan Montgomery for helpful discussions.
Quasi-Hopf Algebras
===================
In this section we recall the definition of quasi-Hopf algebras and their properties described in [@Drin90] and [@Kassel]. Moreover, we recall some interesting results recently obtained in [@HNQA], [@HN991],[@HN992] and [@PV00]. In the sequel, we will use the notation introduced in this section. Throughout this paper, we will always assume that $k$ is a field and any algebras and vector spaces are over $k$. In section 7, 8 and 9, we will further assume $k$ to be an algebraically closed field of characteristic zero.\
A [*quasi-bialgebra*]{} over $k$ is a 4-tuple $(H, \Delta, \e,
\Phi)$, in which $H$ is an algebra over $k$, $\Delta: H \map H
\otimes H $ and $\e : H \map k$ are algebra maps, and $\Phi$ is an invertible element in $H \otimes H \otimes H$ satisfying the following conditions: $$\begin{gathered}
\label{1.1}
(\e \otimes id)\Delta(h) =h= (id \otimes \e)\Delta(h); \\
\label{1.2}
\Phi (\Delta \otimes id)\Delta(h)\Phi^{-1} = (id \otimes
\Delta)\Delta(h) \mbox{ for all } h \in H; \\
\label{1.3}
(id \otimes id \otimes \Delta)(\Phi)
(\Delta \otimes id \otimes id )(\Phi) =
(1 \otimes \Phi)( id \otimes\Delta \otimes id )(\Phi)(\Phi
\otimes 1);\\
\label{1.4}
(id \otimes \e \otimes id)(\Phi) = 1 \otimes 1\,.\end{gathered}$$ The maps $\Delta$, $\e$ and $\Phi$ are respectively called the diagonal map, counit, and associator of the quasi-bialgebra. If there is no ambiguity, we will simply write $H$ for the quasi-bialgebra $(H,
\Delta, \e, \Phi)$. Using (\[1.3\]), one can also easily see that $$\begin{gathered}
\label{1.5}
(\e \otimes id \otimes id)(\Phi) = 1 \otimes 1 =
(id \otimes id \otimes \e)(\Phi) \,.\end{gathered}$$ Moreover, the module category $H$-of the quasi-bialgebra $H$ is a tensor category (cf. [@Drin90] and [@Kassel] for the details).\
Following [@Kassel], a [*gauge transformation*]{} on a quasi-bialgebra $H=(H, \Delta, \e, \Phi)$ is an invertible element $F$ of $H \otimes H$ such that $$(\e \otimes id)(F)=1 =(id \otimes \e)(F)\,.$$ Using a gauge transformation on $H$, one can define an algebra map $\Delta_F: H \map H \otimes H$ by $$\label{2.6}
\Delta_F(h) = F \Delta(h) F^{-1}$$ for any $h \in H$, and an invertible element $\Phi_F$ of $H \otimes
H \otimes H$ by $$\label{2.7}
\Phi_F=(1 \otimes F) (id \otimes \Delta)(F) \Phi (\Delta \otimes
id)(F^{-1})(F^{-1} \otimes 1)\,.$$ Then $H_F=(H, \Delta_F, \e, \Phi_F)$ is also a quasi-bialgebra.\
Two quasi-bialgebras $A$ and $B$ are said to be [*gauge equivalent*]{} if there exists a gauge transformation $F$ on $B$ such that $A$ and $B_F$ are isomorphic as quasi-bialgebras. If $A$ and $B$ are gauge equivalent quasi-bialgebras, $A$-$\mod$, $B$-$\mod$ are equivalent tensor categories (cf. [@Kassel]). Conversely, if $A$, $B$ are finite-dimensional semi-simple quasi-bialgebra such that $A$-and $B$-are equivalent tensor categories, then $A$ and $B$ are gauge equivalent quasi-bialgebras (cf. [@EG02]).\
A quasi-bialgebra $( H, \Delta, \e,\Phi)$ is called a [*quasi-Hopf algebra*]{} if there exist an anti-algebra automorphism $S$ of $H$ and elements $\a, \b \in H$ such that for all element $h \in H$, we have $$\label{2.8}
\sum_{(h)} S(h_1)\a h_2 =\e(h)\a, \quad \sum_{(h)} h_1\b S(h_2)
=\e(h)\b\,$$ and $$\label{2.9}
\sum_i X_i \b S(Y_i)\a Z_i =1, \quad
\sum_i S(\overline{X}_i) \a \overline{Y}_i\b S(\overline{Z}_i) =1$$ where $\Phi=\sum_i X_i \otimes Y_i\otimes Z_i$, $\Phi^{-1}= \sum_i
\overline{X}_i \otimes \overline{Y}_i\otimes \overline{Z}_i$ and $\sum_{(h)} h_1\otimes h_2 = \Delta(h)$. We shall write $( H,
\Delta, \e,\Phi, \a, \b , S)$ for the complete data of the quasi-Hopf algebra and $S$ is called the *antipode* of $H$. When the context is clear, we will simply write $H$ for the quasi-Hopf algebra $(H, \Delta, \e,\Phi, \a, \b , S)$. One can easily see that a Hopf algebra is a quasi-Hopf algebra with $\Phi=1 \otimes 1
\otimes 1$ and $\a=\b=1$.\
Unlike a Hopf algebra, the antipode for a quasi-Hopf algebra is generally not unique.
[[@Drin90 Proposition 1.1]]{}\[prop2.1\] Let $H=(H, \Delta, \e,\Phi, \a, \b , S)$ be a quasi-Hopf algebra. If $u$ is a unit of $H$ then $H_u= (H, \Delta, \e,\Phi, u \a, \b u^{-1} , S_u)$ is also a quasi-Hopf algebra, where $S_u(h)=uS(h)u^{-1}$ for all $h \in H$. Conversely, for any $\a', \b' \in H$ and for any algebra anti-automorphism $S'$ of $H$ such that $H'=(H, \Delta, \e,\Phi, \a', \b' , S')$ is a quasi-Hopf algebra, then there exist a unique invertible element $u$ of $H$ such that $$H_u = H'\,.$$
If $F$ is a gauge transformation on the quasi-Hopf algebra $H=(H,
\Delta, \e,\Phi, \a, \b , S)$, we can define $\a_F$ and $\b_F$ by $$\a_F = \sum_i S(d_i) \a e_i
\quad\mbox{and}\quad \b_F = \sum_i f_i \b S(g_i)$$ where $F=\sum_i f_i \otimes g_i$ and $F^{-1} =\sum_i
d_i \otimes e_i$. Then, $H_F=(H, \Delta_F,
\e,\Phi_F, \a_F, \b_F , S)$ is also a quasi-Hopf algebra.\
The antipode of a Hopf algebra is known to be a anti-coalgebra map. For a quasi-Hopf algebra $H$, this is true up to conjugation. Following [@Drin90], we define $\g, \d \in H \otimes H$ by the formulae $$\begin{aligned}
\g & = & \sum_i S(U_i)\a V_i \otimes S(T_i) \a W_i\,,\\
\d & = & \sum_j K_j \b S(N_j) \otimes L_j \b S(M_j)\,,\end{aligned}$$ where $$\begin{aligned}
\sum_i T_i \otimes U_i \otimes V_i \otimes W_i &=& (1 \otimes \Phi^{-1})(id
\otimes id \otimes \Delta)(\Phi) \,,\\
\sum_j K_j \otimes L_j \otimes M_j \otimes N_j &=& (\Delta \otimes id
\otimes id )(\Phi)( \Phi^{-1}\otimes 1 ) \,.\end{aligned}$$ Then, $$\label{FH}
F_H=\sum_i (S \otimes S)(\Delta^{op}(\overline{X}_i)) \cdot \g \cdot
\Delta(\overline{Y}_i \b S(\overline{Z}_i))$$ is an invertible element of $H \otimes H$ where $\Phi^{-1}= \sum_i
\overline{X}_i \otimes \overline{Y}_i\otimes \overline{Z}_i$. Moreover, $$F_H \Delta(S(h))F_H^{-1} = (S \otimes S)(\Delta^{op}(h)\,.$$ for all $h \in H$.\
The category of finite-dimensional left $H$-module of a quasi-Hopf algebra $H$ with antipode $S$, denote by $H$-$\mod_{fin}$ is a rigid tensor category. Let $M$ be a finite-dimensional left $H$-module and $M'$ its $k$-linear dual. Then the $H$-action on $M'$, given by $$(h\cdot f)(m) = f(S(h)m)$$ for any $f \in M'$ and $m \in M$, defines a left $H$-module structure on $M'$. We shall denote by ${{^*\!M}}$ the left dual of the $M$ in $H$-$\mod_{fin}$. Similarly, the right dual of $M$, denote by $M^*$, is the $H$-module with the underlying $k$-linear space $M'$ with the $H$-action given by $$(h\cdot f)(m) = f(S^{-1}(h)m)$$ for any $f \in M'$ and $m \in M$ (cf. [@Drin90]).\
In [@HNQA], [@HN991] and [@HN992], Frank Hausser and Florian Nill introduced some interesting elements in $H \otimes H$ for any arbitrary quasi-Hopf algebra $H=(H, \Delta, \e,\Phi, \a,
\b , S)$ in the course of studying the corresponding theories of quantum double, integral and the fundamental theorem for quasi-Hopf algebras. These elements of $H \otimes H$ are given by $$\begin{gathered}
\label{R}
q_R = \sum X_i \otimes S^{-1} (\a Z_i)Y_i \,, \quad p_R= \sum
\overline{X}_i \otimes \overline{Y}_i \b
S(\overline{Z}_i)\,, \\
\label{L}
q_L = \sum S(\overline{X}_i) \a \overline{Y}_i \otimes
\overline{Z}_i\,, \quad p_L = \sum Y_iS^{-1}(X_i\b) \otimes Z_i
\,\end{gathered}$$ where $\Phi= \sum_i X_i \otimes Y_i \otimes Z_i$ and $\Phi^{-1}= \sum_i
\overline{X}_i \otimes \overline{Y}_i\otimes \overline{Z}_i$. One can show easily (cf. [@HNQA]) that they obey the relations (for all $a
\in H$) $$\begin{aligned}
\label{eq:qR}
(a\otimes 1)\, q_R &= & \sum (1 \otimes S^{-1}(a_2))\,q_R\, \Delta(a_1) , \\
\label{eq:qL}
(1 \otimes a)\, q_L &= & \sum (S(a_1)\otimes 1)\, q_L \, \Delta(a_2) , \\
\label{eq:pR}
p_R\, (a \otimes 1) &= & \sum \Delta(a_1)\, p_R \, (1 \otimes S(a_2)) , \\
\label{eq:pL}
p_L\, (1 \otimes a) &= & \sum\Delta(a_2)\,p_L\, (S^{-1}(a_1)\otimes
1)\,.\end{aligned}$$ where $\Delta(a)= \sum a_1 \otimes a_2$. Suppressing the summation symbol and indices, we write $q_R = q^1_R \otimes q^2_R$, etc. These elements also satisfy the identities (cf. [@HNQA]): $$\begin{aligned}
\label{2.16}
\Delta(q_R^1) \, p_R\, (1\otimes S(q_R^2)) &= & 1\otimes 1, \\
\label{2.17}
(1\otimes S^{-1}(p_R^2))\, q_R \, \Delta(p_R^1)& = & 1\otimes 1, \\
\label{2.18}
\Delta (q_L^2) \, p_L \, (S^{-1}(q_L^1) \otimes 1)&= &1 \otimes 1, \\
\label{2.19}
(S(p_L^1) \otimes 1)\, q_L \, \Delta(p_L^2) & = & 1\otimes 1.\end{aligned}$$ We will use these equations in the sequel.
Central Gauge Invariants for Semi-simple Quasi-Hopf Algebras
============================================================
Suppose that $H=(H, \Delta, \e, \Phi, \a, \b, S)$ is a finite-dimensional quasi-Hopf algebra. A left integral of $H$ is an element $l$ of $H$ such that $h l = \e(h) l$ for all $h \in H$. A right integral of $H$ can be defined similarly. It follows from [@HNQA] that the subspace of left (right) integrals of $H$ is of dimension 1. Moreover, if $H$ is semi-simple, the subspace of left integral is identical to the space of right integrals of $H$ and $\e(\Lambda) \ne 0$ for any non-zero left integral $\Lambda$ of $H$ (see also [@PV00]). We will call the two-sided integral $\Lambda$ of $H$ [*normalized*]{} if $\e(\Lambda)=1$.\
Let $\Lambda$ be a left integral of $H$. Then for any $a \in H$, $$\label{eq:left_integral}
\e(a)\Delta(\Lambda) =
\Delta(a) \Delta(\Lambda)\,.$$ Similarly, if $\Lambda'$ is a right integral of $H$, then we have $$\label{eq:right_integral}
\e(a)\Delta(\Lambda') =
\Delta(\Lambda') \Delta(a)\,.$$
We then have the following lemma.
\[pq\_1\] Let $H=(H, \Delta, \e, \Phi, \a, \b, S)$ be a finite-dimensional quasi-Hopf algebra.
1. If $\Lambda$ is a left integral of $H$, then for any $a \in H$, $$\begin{aligned}
\label{3.3}
(1 \otimes a)q_R \Delta(\Lambda) &=& (S(a) \otimes 1)q_R \Delta(\Lambda)\,,\\
\label{3.4}
(1 \otimes a)q_L \Delta(\Lambda) & =& (S(a) \otimes 1)q_L \Delta(\Lambda)\,, \\
\label{3.5}
\mbox{and}\quad (\b \otimes 1)q_L \Delta(\Lambda)&=& (\b \otimes 1)q_R \Delta(\Lambda)=\Delta(\Lambda)\,.\end{aligned}$$
2. If $\Lambda'$ is a right integral of $H$, then for any $a \in H$, $$\begin{aligned}
\label{3.6}
\Delta(\Lambda')p_R (a \otimes 1) & = & \Delta(\Lambda')p_R (1 \otimes S(a))
,\\
\label{3.7}
\Delta(\Lambda')p_L (a \otimes 1) &= & \Delta(\Lambda')p_L (1 \otimes S(a))\,,\\
\label{3.8}
\mbox{and}\quad\Delta(\Lambda') p_L (1 \otimes \a)&=&\Delta(\Lambda') p_R
(1 \otimes \a)=\Delta(\Lambda')\,.\end{aligned}$$
\(i) By the equations (\[eq:left\_integral\]) and (\[eq:qR\]), for any $a \in H$, $$\begin{aligned}
(a\otimes 1)q_R \Delta(\Lambda) & = & (1 \otimes S^{-1}(a_1)))q_R
\Delta(a_2)\Delta(\Lambda)\\
&=&(1 \otimes S^{-1}(a_1\e(a_2)))q_R
\Delta(\Lambda)\\
&=& ( 1 \otimes S^{-1}(a))q_R\Delta(\Lambda) .\end{aligned}$$ Hence, by substituting $a$ with $S(a)$, we prove equation (\[3.3\]). Now we have $$\begin{aligned}
\Delta(\Lambda) &=& (1\otimes S^{-1}(p_R^2))\, q_R \, \Delta(p_R^1)\Delta(\Lambda)\quad \mbox{ (by (\ref{2.17}))}\\
&=& (1\otimes S^{-1}(p_R^2\e(p_R^1))q_R \Delta(\Lambda) \quad \mbox{ (by (\ref{eq:left_integral}))}\\
& = & (1 \otimes S^{-1}(\b))q_R \Delta(\Lambda) \quad \mbox{ (by (\ref{1.5}))}\\
&=&(\b \otimes 1)q_R \Delta(\Lambda) \quad \mbox{ (by (\ref{3.3}))}\,.\end{aligned}$$ The remaining formulae in (i) and (ii) can be proved similarly using equations (\[1.5\]), (\[eq:qR\])-(\[2.19\]), (\[eq:left\_integral\]) and (\[eq:right\_integral\]).\
\[l3.2\]\[invariant\_1\] Let $H=(H, \Delta, \e, \Phi, \a, \b, S)$ be a finite-dimensional quasi-Hopf algebra and $F$ a gauge transformation on $H$. Suppose that $q_R^F, q_L^F, p_R^F, p_L^F$ are the corresponding $p$’s and $q$’s for $H_F$ defined in [(\[R\])]{} and [(\[L\])]{}.
1. If $\Lambda$ is a left integral of $H$, then $$q_R^F \Delta_F(\Lambda)=q_R\Delta(\Lambda)F^{-1} \,\quad
\mbox{and} \quad
q_L^F \Delta_F(\Lambda)=q_L\Delta(\Lambda)F^{-1}\,.$$
2. If $\Lambda'$ is a right integral of $H$, then $$\Delta_F(\Lambda')p_R^F = F \Delta(\Lambda')p_R
\,\quad
\mbox{and} \quad
\Delta_F(\Lambda') p_L^F = F \Delta(\Lambda')p_L\,.$$
\(i) Let $\Phi^{-1}=\sum_j \overline{X}_j \otimes \overline{Y}_j \otimes
\overline{Z}_j$, $F=\sum_i f_i \otimes g_i$ and $F^{-1}=\sum_l d_l \otimes e_l$. Then, we obtain $$\begin{aligned}
\Phi_F^{-1} & = &(F \otimes 1)(\Delta \otimes id)(F) \Phi^{-1}
(id \otimes \Delta)(F^{-1})(1 \otimes F^{-1}) \\
& = &(F \otimes 1)\left(\sum_{i,j,l} f_{i,1} \overline{X}_j d_l \otimes
f_{i,2} \overline{Y}_j e_{l,1} \otimes
g_i \overline{Z}_j e_{l,2}\right) (1 \otimes F^{-1})\end{aligned}$$ where $\Delta(f_i)=\sum f_{i,1} \otimes f_{i,2}$ and $\Delta(e_l)=\sum e_{l,1} \otimes e_{l,2}$. Thus, we have $$\begin{aligned}
q_L^F \Delta_F(\Lambda) & = & \left(\sum S(f_{i'}f_{i,1} \overline{X}_jd_l)\a_F g_{i'}f_{i,2} \overline{Y}_je_{l,1}
\otimes g_i \overline{Z}_j e_{l,2}\right)F^{-1} F\Delta(\Lambda) F^{-1} \\ \\
& = & \left(\sum S(f_{i,1} \overline{X}_jd_l)\a f_{i,2} \overline{Y}_je_{l,1}
\otimes g_i \overline{Z}_j e_{l,2}\right)\Delta(\Lambda) F^{-1}
\quad(\mbox{since } \sum S(f_{i'}) \a_F g_{i'} = \a)\\ \\
& = & \left(\sum S(f_{i,1} \overline{X}_jd_l \e(e_l))\a f_{i,2} \overline{Y}_j
\otimes g_i \overline{Z}_j \right)\Delta(\Lambda) F^{-1} \quad (\mbox{by }(\ref{eq:left_integral}))\\ \\
& = & \left(\sum S(f_{i,1} \overline{X}_j)\a f_{i,2} \overline{Y}_j
\otimes g_i \overline{Z}_j \right)\Delta(\Lambda) F^{-1}
\quad(\mbox{since } \sum d_l\e(e_l) = 1_H)\\
&=& \left(\sum S(\overline{X}_j)\a \e(f_i) \overline{Y}_j
\otimes g_i \overline{Z}_j \right)\Delta(\Lambda) F^{-1}
\quad(\mbox{since } \sum S(f_{i,1})\a f_{i,2} = \e(f_i)\a) \\ \\
&=& \left(\sum S(\overline{X}_j)\a \overline{Y}_j
\otimes \overline{Z}_j \right)\Delta(\Lambda) F^{-1}
\quad(\mbox{since } \sum \e(f_i)g_i=1_H )\\ \\
& =& q_L \Delta(\Lambda) F^{-1})\end{aligned}$$ The other three equations can be proved similarly.\
\[th3.3\]Let $H=(H, \Delta, \e, \Phi, \a, \b, S)$ be a finite-dimensional quasi-Hopf algebra. Suppose that $\Lambda$ is a two-sided integral of $H$. Then, the elements $$\label{eq:invariant_1}
q_R \Delta(\Lambda) p_R\,, \quad \quad q_R \Delta(\Lambda) p_L\,,\quad
q_L \Delta(\Lambda) p_R\,, \mbox{ and } \quad q_L \Delta(\Lambda) p_L$$ in $H \otimes H$ are invariant under gauge transformations. Moreover, $$m(q_R \Delta(\Lambda) p_R)=m(q_R \Delta(\Lambda) p_L)=m(q_L \Delta(\Lambda) p_R)=m(q_L \Delta(\Lambda) p_L)$$ where $m$ denote the multiplication of $H$. In addition, $m(q_R \Delta(\Lambda) p_R)$ is a central element of $H$.
It follows from Lemma \[invariant\_1\] that for any gauge transformation $F$ on $H$, $$q_*^F\Delta_F(\Lambda)=q_*\Delta(\Lambda)F^{-1},\quad\mbox{and} \quad
\Delta_F(\Lambda)p_*^F = F\Delta(\Lambda)p_*\,$$ where $q_*^F =q_L^F$ or $q_R^F$ and $p_*^F =p_L^F$ or $p_R^F$. Thus we have $$\begin{aligned}
q_*^F\Delta_F(\Lambda)p_*^F &= &q_*\Delta(\Lambda)F^{-1}p_*^F \\
& = & q_*F^{-1}\Delta_F(\Lambda)p_*^F \\
&= & q_* F^{-1}F\Delta(\Lambda)p_*\\
&= & q_*\Delta(\Lambda)p_*\,.\end{aligned}$$
Let $m$ denote the multiplication of $H$ and let $m_{RR}$, $m_{RL}$, $m_{LR}$ and $m_{LL}$ denote the elements $$m(q_R \Delta(\Lambda) p_R), \, m(q_R \Delta(\Lambda) p_L), \, m(q_L \Delta(\Lambda) p_R), \,\,\mbox{ and }
m(q_L \Delta(\Lambda) p_L)$$ respectively. Then for any $a \in H$, $$\begin{aligned}
S(a)m_{RR} & = &
m( (S(a) \otimes 1)q_R \Delta(\Lambda)p_R) \\
& =& m((1\otimes a)q_R \Delta(\Lambda)p_R) \quad \mbox{by Lemma \ref{pq_1}(i)}\\
& =& m( q_R \Delta(\Lambda)p_R(a\otimes 1))\\
& =& m(q_R \Delta(\Lambda)p_R (1\otimes S(a))) \quad \mbox{by Lemma \ref{pq_1}(ii)}\\
&=& m_{RR}S(a)\,.\end{aligned}$$ As $S$ is an automorphism, the above equation implies that $m_{RR}$ is in the center of $H$. Using the same kind of arguments, one can show that $m_{RL}, m_{LR}$ and $m_{LL}$ are each in the center of $H$.\
Let $Q_R$, $Q_L$, $P_R$ and $P_L$ denote the elements $$m(q_R \Delta(\Lambda)), \, m(q_L \Delta(\Lambda)),\,
m(\Delta(\Lambda) p_R)\,, \,\mbox{ and } m(\Delta(\Lambda) p_L)\,$$ respectively. Then, we have $$\label{eq:exchange_1}
\begin{aligned}
m_{RR} & = \sum q_R^1 \Lambda_1 p_R^1 q_R^2 \Lambda_2 p_R^2\\
&=\sum S(p_R^1) q_R^1 \Lambda_1 q_R^2 \Lambda_2 q_R^2 \quad \mbox{by Lemma \ref{pq_1}(i)}\\
& = S(p_R^1) Q_R p_R^2 \\
&= \sum_j S(\overline{X}_j) Q_R \b\overline{Y}_jS(\overline{Z}_j)
\end{aligned}$$ and $$\label{eq:exchange_1.2}
\begin{aligned}
m_{RR} & = \sum q_R^1 \Lambda_1 p_R^1 q_R^2 \Lambda_2 p_R^2\\
&=\sum q_R^1 \Lambda_1 p_R^1 \Lambda_2 p_R^2 S(q_R^2) \quad \mbox{by Lemma \ref{pq_1}(ii)}\\
& = q_R^1 P_R S(q_R^2)\\
&= \sum X_i P_R S(Y_i) \a Z_i
\end{aligned}$$ where $\Phi^{-1}=\sum_j \overline{X}_j \otimes \overline{Y}_j \otimes
\overline{Z}_j$ and $\Phi=\sum_i X_i \otimes Y_i \otimes
Z_i$. Similarly, $$\label{eq:exchange_2}
m_{LL} = S(\overline{X}_j) \a \overline{Y}_j P_L S(\overline{Z}_j) \\
= \sum X_i \b S(Y_i) Q_L Z_i\,.$$ By (\[3.5\]) and (\[3.8\]), we have $$\label{eq:PQ}
\begin{aligned}
Q_R = m_{RL}\a, &\quad Q_L =m_{LR}\a \,, \\
P_L = \b m_{RL}, &\quad P_R = \b m_{LR}\,.
\end{aligned}$$
Therefore, using equation (\[2.9\]), we have $$m_{RR} = S(\overline{X}_j) m_{RL}\a \overline{Y}_j \b S(\overline{Z}_j) =
m_{RL}S(\overline{X}_j) \a \overline{Y}_j \b S(\overline{Z}_j) = m_{RL}$$ Similarly, using equations (\[eq:exchange\_1.2\]), (\[eq:exchange\_2\]) and (\[eq:PQ\]) we can prove $$m_{RR}=m_{LR}=m_{LL}\,.$$\
In [@HNQA] and [@PV00], it is shown that a finite-dimensional quasi-Hopf algebra $H$ is semi-simple if, and only, if there exist a unique normalized two-sided integral. In this case, we have the following:
[Let $H=(H, \Delta, \e, \Phi, \a, \b, S)$ be a finite-dimensional semi-simple quasi-Hopf algebra and let $\Lambda$ be the unique normalized two-sided integral of $H$. We denote by $\nu_H$ the central element $$m(q_L \Delta(\Lambda)p_L)$$ discussed in Theorem \[th3.3\].]{}
\[cor3.5\] Let $H=(H, \Delta, \e, \Phi, \a, \b, S)$ be a finite-dimensional semi-simple quasi-Hopf algebra and $\Lambda$ the normalized two-sided integral of $H$. Then $\nu_H$ is invariant under gauge transformations, that is $$\nu_H = \nu_{H_F}\,$$ for any gauge transformation $F$ on $H$. Moreover, $$\b\a\nu_H = \nu_H\b\a=\sum(\Lambda_1\Lambda_2)$$ where $\sum \Lambda_1 \otimes \Lambda_2 = \Delta(\Lambda)$. In particular, if both $\a$ and $\b$ are units of $H$, then $$\nu_H=\sum(\Lambda_1\Lambda_2)(\b\a)^{-1} = (\b\a)^{-1}\sum(\Lambda_1\Lambda_2)\,.$$
The first statement follows immediately from Theorem \[3.3\]. By equation (\[3.5\]) and (\[3.8\]), we have $\b\nu_H\a= \sum(\Lambda_1\Lambda_2)$. Since $\nu_H$ is central, then the result follows.\
Let $H=(H, \Delta,\e, \Phi, \a, \b, S)$, $H'=(H', \Delta',\e', \Phi', \a', \b', S')$ be semisimple quasi-Hopf algebras. If $H$ and $H'$ are gauge equivalent quasi-bialgebras via the gauge transformation $F$ on $H$ and the quasi-bialgebra isomorphism $\sigma : H_F \map H'$, then $$\sigma(\nu_{H}) = \nu_{H'}\,.$$ In particular, if $u$ is a unit of $H$, then $\nu_{H_u} = \nu_{H}$.
Since $H_F$ and $H'$ are isomorphic quasi-bialgebras, $(H', \Delta', \e', \Phi', \sigma(\a_F),\sigma(\b_F), \sigma S \sigma^{-1})$ is a quasi-Hopf algebra. By Proposition \[prop2.1\], there exists a unit $u$ of $H'$ such that $$\label{eq:3.13}
\sigma S \sigma^{-1}(a) = u S'(a) u^{-1}, \quad \sigma(\b_F) = u \a'\, \mbox{ and }
\sigma(\b_F) = \b'u^{-1}\,.$$ for all $a \in H'$. Then, we have $$\label{eq:3.14}
\sigma S^{-1} \sigma^{-1}(a) = S'^{-1}(u) S'^{-1}(a) S'^{-1}(u^{-1})\,.$$ Let $\Lambda$ be the normalized two-sided integral of $H$. Since $\sigma$ is a quasi-bialgebra isomorphism, $\sigma(\Lambda)$ is then a two-sided integral of $H'$ and $$\e'(\sigma(\Lambda)) = \e(\Lambda) = 1\,.$$ Therefore, $\Lambda'=\sigma(\Lambda)$ is the unique normalized integral of $H'$. In particular, we have $$(\sigma\otimes \sigma)\Delta_F(\Lambda)= \sum \Lambda'_1 \otimes \Lambda'_2 \quad \mbox{ and }\quad
(\sigma \otimes \sigma \otimes \sigma )(\Phi_F)= \Phi'$$ where $\sum \Lambda'_1 \otimes \Lambda'_2 = \Delta'(\Lambda')$. Let $$\begin{aligned}
& \Phi_F = \sum X^F_i \otimes Y^F_i \otimes Z^F_i \,,\quad
\Phi_F^{-1} =
\sum \overline{X}^F_j \otimes \overline{Y}^F_j \otimes \overline{Z}^F_j\,,\\
& \Phi'= \sum X'_i \otimes Y'_i \otimes Z'_i\,,\quad
\Phi'^{-1} =
\sum \overline{X}'_j \otimes \overline{Y}'_j \otimes \overline{Z}'_j\,,\\
&\mbox{and} \quad(\sigma\otimes \sigma)\Delta_F(\Lambda) = \sum \Lambda_1^F \otimes \Lambda_2^F\,.
\end{aligned}$$ Then, $$\begin{aligned}
& \sigma(\nu_H) = \sigma(\nu_{H_F})\quad \quad \mbox{by Corollary (\ref{cor3.5})} \\
= \,\,& \sigma\left(\sum X^F_i \Lambda^F_1 \overline{X}^F_j S^{-1}(\a_F Z^F_i)
Y^F_i \Lambda^F_2 \overline{Y}^F_j \b_F S(\overline{Z}^F_j)\right)\\
= \,\,& \sum X'_i \Lambda'_1 \overline{X}'_j\, (\sigma S^{-1})(\a_F Z^F_i)
Y'_i \Lambda'_2 \overline{Y}'_j \sigma(\b_F)\, (\sigma S)(\overline{Z}^F_j)\\
= \,\,& \sum X'_i \Lambda'_1 \overline{X}'_j \,(\sigma S^{-1} \sigma^{-1})(Z'_i) \,
(\sigma S^{-1}\sigma^{-1})(\sigma(\a_F))
Y'_i \Lambda'_2 \overline{Y}'_j \sigma(\b_F) \,(\sigma S \sigma^{-1})(\overline{Z}'_j)\\
=\,\,& \sum X'_i \Lambda'_1 \overline{X}'_j \,S'^{-1}(u) S'^{-1} (Z'_i) S'^{-1}(\a')\,
Y'_i \Lambda'_2 \overline{Y}'_j \b'S'(\overline{Z}'_j)u^{-1} \quad \mbox{by (\ref{eq:3.13}) and (\ref{eq:3.14})}\\
= \,& \sum X'_i \Lambda'_1 \overline{X}'_j S'^{-1}(\a' Z'_i)
Y'_i \Lambda'_2 \overline{Y}'_j \b'S'(\overline{Z}'_j)u u^{-1} \quad \mbox{by Lemma \ref{pq_1}(ii)}\\
= \, & \nu_{H'}\,.
\end{aligned}$$
For any unit $u$ of $H$, $H$ and $H_u$ are obviously gauge equivalent as quasi-bialgebra under the gauge transformation $1 \otimes 1$ and the quasi-bialgebra isomorphism $id_H$. Hence, the second statement follows.
Frobenius-Schur Indicators
==========================
Let $(H, \Delta,\e, \Phi, \a, \b, S)$ be a semi-simple quasi-Hopf algebra over the field $k$. Let $M$ be an irreducible $H$-module with character $\chi$. We call $\chi(\nu_H)$ the Frobenius-Schur indicator of $\chi$ (or $M$). The family of Frobenius-Schur indicators $\{ \chi(\nu_H) \}$ is in fact an invariant of the tensor category $H$-for any semi-simple quasi-Hopf algebra $H$.
\[thm4.1\] Let $H=(H, \Delta,\e, \Phi, \a, \b, S)$ and $H'=(H', \Delta',\e',
\Phi', \a', \b', S')$ be finite-dimensional semi-simple quasi-Hopf algebras over an algebraically closed field $k$ of characteristic zero. If $H$-and $H'$-are equivalent as $k$-linear tensor categories, then the families of Frobenius-Schur indicators for $H$ and $H'$ are identical.
If $H$-and $H'$-are equivalent as $k$-linear tensor categories, then, by [@EG02 Theorem 6.1], $H$ and $H'$ are gauge equivalent quasi-bialgebras. Suppose that $F$ is a gauge transformation on $H$ and $\sigma : H_F \map H'$ is a quasi-bialgebra isomorphism. It follows from Corollary \[cor3.5\] that $$\sigma(\nu_H)=\nu_{H'}\,.$$ Let $\Irr(H)$, $\Irr(H')$ be the set of irreducible characters of $H$ and $H'$ respectively. Then, the map $\chi' \mapsto \chi'
\circ \sigma$ is a bijection from $\Irr(H')$ onto $\Irr(H)$. Moreover, for any irreducible character $\chi'$ of $H'$, $$\chi' \circ \sigma(\nu_H)= \chi'(\nu_{H'})\,.$$ Thus, $\{\chi'(\nu_{H'}) \}_{\chi' \in \Irr(H')}$ is identical of the family $\{\chi(\nu_{H}) \}_{\chi \in \Irr(H)}$.
[If $H$ is a semi-simple Hopf algebra, then $\Phi=1 \otimes
1\otimes 1$ and $\a=\b=1$. It follows from Corollary \[cor3.5\] that $$\nu_H= \sum \Lambda_1 \Lambda_2$$ where $\sum \Lambda_1 \otimes \Lambda_2=\Delta(\Lambda)$ and $\Lambda$ is the normalized two-sided integral of $H$. Thus, $\chi(\nu_H)$ coincides with the Frobenius-Schur indicator defined in [@LM00].]{}
As an application of Theorem, we give a simple alternative proof of the fact that $\BC [Q_8]$-and $\BC [D_8]$-are not equivalent as $\BC$-linear tensor categories where $Q_8$ and $D_8$ are the quaternion group and the dihedral group of order 8 respectively (cf. [@TaYa98]).
[[@TaYa98]]{} The $\BC$-linear categories $\BC[Q_8]$-and $\BC[D_8]$-are not equivalent as tensor categories.
Let $G=D_8$ or $Q_8$. Then, $G$ has four degree 1 characters and one degree 2 irreducible character $\chi_2$. Let $z$ be the non-trivial central element of $G$. Then $\chi_2(z)=-2$ and $\chi(z) = 1$ for any character $\chi$ of $G$ of degree 1. Since $\nu_G=\frac{1}{8}\sum_{g \in G} g^2$, one can easily obtain that $$\nu_{Q_8} = \frac{1}{8}(6z+2e), \quad \mbox{and}\quad \nu_{D_8}=
\frac{1}{8}(2z+6e)$$ where $e$ is the identity of the group. Thus, the family of Frobenius-Schur indicators for $Q_8$ is $\{1,1,1, 1, -1\}$ but the the family of Frobenius-Schur indicators for $D_8$ is $\{1,1,1,1,1\}$. By virtue of Theorem \[thm4.1\], $\BC[Q_8]$-and $\BC[D_8]$-are not equivalent as $\BC$-linear tensor categories.
Bantay’s Formula for Indicators of Twisted Quantum Doubles {#s:Bantay_Indicator}
==========================================================
In this section, we will show that if $H$ is a twisted quantum double of a finite group $G$ over the field $k$ such that $|G|^{-1}$ exists in $k$, then for any irreducible character $\chi$ of $H$, $\chi(\nu_H)$ is identical to Bantay’s formula (\[eq: bantayfsind\]). We begin with the definition of twisted quantum doubles of finite groups.\
Let $G$ be a finite group and $\omega: G \times G \times G \map k^\times$ be a normalized 3-cocycle; that is, a function such that $\omega(x,y,z)=1$ whenever one of $x,y$ or $z$ is equal to the identity element 1 of $G$ and which satisfies the functional equation $$\label{eq:3cocycle}
\omega(g,x,y)\omega(g, xy,
z)\omega(x,y,z)=\omega(gx,y,z)\omega(g,x,yz)\quad \mbox{for
any } g, x, y, z \in G\,.$$ For any $g \in G$, define the functions $\theta_g, \gamma_g :
G\times G \rightarrow k^\times$ as follows: $$\begin{aligned}
\theta_g(x,y) &=&\frac{\omega(g,x,y)\omega(x,y,(x y)^{-1}g x
y)}{\omega(x,x^{-1}g x,y)}\,,\label{eq0.001}\\
\gamma_g(x,y) &=& \frac{\omega(x,y,g)\omega(g, g^{-1}x g,
g^{-1}yg)}{\omega(x,g,
g^{-1}y g)}\,.\label{eq0.002}\end{aligned}$$ Let $\{e(g)|g \in G\}$ be the dual basis of the canonical basis of $k [G]$. The [*twisted quantum double*]{} $D^\w(G)$ of $G$ with respect to $\omega$ is the quasi-Hopf algebra with underlying vector space $k[G]' \otimes k[G]$. The multiplication, comultiplication and associator are given, respectively, by $$\label{multiplication}
(e(g) \otimes x)(e(h) \otimes y) =\theta_g(x,y)
\delta_{g, xhx^{-1}}e(g)\otimes x y\,,$$ $$\label{comultiplication}
\Delta(e(g)\otimes x) = \sum_{hk=g} \gamma_x(h,k) e(h)\otimes x
\otimes e(k) \otimes x\,,$$ $$\label{eq0.01}
\Phi = \sum_{g,h,k \in G} \omega(g,h,k)^{-1} e(g) \otimes 1 \otimes
e(h) \otimes 1 \otimes e(k) \otimes 1\,.$$ The counit and antipode are given by $$\label{eq:counit}
\varepsilon(e(g)\otimes x) = \delta_{g,1}$$ and
$$\label{eq:antipode}
S(e(g)\otimes x) =
\theta_{g^{-1}}(x,x^{-1})^{-1}\gamma_x(g,g^{-1})^{-1}e(x^{-1}g^{-1}x)\otimes
x^{-1}\,,$$
where $\delta_{g,1}$ is the Kronecker delta. The corresponding elements $\alpha$ and $\beta$ are $1_{D^{\w}(G)}$ and $\sum\limits_{g \in
G} \w(g,g^{-1},g)e(g)\otimes 1$ respectively (cf. [@DPR90]). Verification of the detail involves the following identities, which result from the 3-cocycle identity for $\w$: $$\label{eq:0.1}
\theta_z(a,b)\theta_z(a b,c) = \theta_{a^{-1}z a}(b,c)\theta_z(a,b
c)\,,$$ $$\label{eq:0.2}
\theta_y(a,b)\theta_z(a,b)\g_a(y,z)\g_b(a^{-1}y a,a^{-1}z
a)=\theta_{yz}(a,b)\g_{ab}(y,z)\,,$$ $$\label{eq:0.3}
\g_z(a,b)\g_z(a b,c)\w(z^{-1}a z,z^{-1}b z,z^{-1}c
z)=\g_z(b,c)\g_z(a,b c)\w(a,b,c)\,,$$ for all $a,b,c,y,z \in G$.
\[r0.1\] [The algebra $D^\w(G)$ is a semi-simple (cf. [@DPR90]). If $\omega=1$, then the twisted quantum double $D^{\omega}(G)$ identical to the Drinfeld double of the group algebra $k [G]$. However, $D^{\omega}(G)$ is not a Hopf algebra in general. Moreover, even if $\omega, \omega'$ differ by a coboundary, $D^\w(G)$ and $D^{\w'}(G)$ are not isomorphic as quasi-bialgebras. Nevertheless, they are [*gauge equivalent*]{}. In addition, if $G$ is abelian, $D^\w(G)$ also admits a Hopf algebra structure with the same underlying $\Delta$, $\e$ and $S$ (cf. [@MN01]). ]{}
Let $$\label{integral}
\Lambda = \frac{1}{|G|}\sum_{x \in G} e(1) \otimes x \in D^\omega(G)\,.$$ It is straightforward to show that $\Lambda$ is a left integral of $D^\omega(G)$. Moreover, $$\e(\Lambda)=1 \,.$$ After [@Pana98] and [@HNQA], this gives another proof of the semi-simplicity of $D^\omega(G)$. Note that $$\Delta(\Lambda) = \sum \Lambda_1 \otimes \Lambda_2 = \frac{1}{|G|}
\sum_{g,x \in G}
\gamma_x(g, g^{-1})e(g) \otimes x \otimes e(g^{-1}) \otimes x \,.$$ Since $\beta \alpha = \beta$ is invertible, it follows from Corollary \[cor3.5\] that $$\begin{aligned}
\nu_{D^\w(G)} &=&
\frac{1}{|G|}\left(\sum_{g \in G} \omega(g, g^{-1}, g)^{-1}(e(g)
\otimes 1)\right) \left(\sum_{g,x \in G}
\gamma_x(g, g^{-1})(e(g) \otimes x) (e(g^{-1}) \otimes x)\right)\\
&=& \frac{1}{|G|} \left(\sum_{g \in G} \omega(g^{-1}, g, g^{-1})(e(g)
\otimes 1)\right)
\left(\sum_{x^{-1}gx=g^{-1}}\gamma_x(g, g^{-1})\theta_g(x,x)(e(g)
\otimes x^2)\right)
\\
&=& |G|^{-1} \sum_{x^{-1}gx=g^{-1}}
\omega(g^{-1}, g, g^{-1})\gamma_x(g, g^{-1})\theta_g(x,x)(e(g) \otimes x^2).\end{aligned}$$ Here we have used the equality $$\w(g, g^{-1}, g)^{-1}= \w(g^{-1}, g, g^{-1})$$ which is readily derived from equation (\[eq:3cocycle\]). Thus for any irreducible character $\chi$ of $D^\w(G)$, the Frobenius-Schur indicator of $\chi$ is $$\chi(\nu_{D^\w(G)})=|G|^{-1} \sum_{x^{-1}gx=g^{-1}}
\omega(g^{-1}, g, g^{-1})\gamma_x(g, g^{-1})\theta_g(x,x)\chi(e(g) \otimes x^2)$$ as given by Bantay.
Trace Elements and Antipodes of Semi-simple Quasi-Hopf Algebras
===============================================================
It is proved by Larson and Radford [@LaRa87] [@LaRa88] that if char $k = 0$, the antipode of a semi-simple Hopf algebra over $k$ is an involution. However, the antipode of a semi-simple quasi-Hopf algebra $H$ could be of any order. Nevertheless, we prove an analog of the Larson-Radford theorem for a split semi-simple quasi-Hopf algebras $H$ over any field $k$: there exists a unit $u \in H$ such that the antipode of $H_u$ is an involution. To this end we introduce the *trace element* $g$ of a semi-simple quasi-Hopf algebra. This element will play a role throughout the remaining Sections of the paper.\
Let $H=(H, \Delta,\e, \Phi, \a, \b, S)$ be a finite-dimensional semi-simple quasi-Hopf algebra over $k$ and $\Lambda$ the normalized two-sided integral of $H$. By [@HNQA], there exists a functional $\lambda \in H'$, called the normalized left cointegral of $H$, given by the formula $$\label{eq:cointegral}
\lambda(x)=\sum_i b^i(x S^2(b_i)S(\b)\a)$$ for all $x \in H$, where $\{b_i\}$ is a basis of $H$ and $\{b^i\}$ is its dual basis (see [@HNQA] for the details of cointegral). The normalized left cointegral $\lambda$ admits the following properties :
1. $\lambda(\Lambda)=1$.
2. $\lambda(ab)$ ($a, b \in H$) defines a non-degenerate bilinear form on $H$.
3. For all $a, b \in H$, $$\label{eq:nakayama}
\lambda(ab)=\lambda(bS^2(a)) \,.$$
Let $\chi_{reg}$ denote the character of the left regular representation of $H$. The bilinear form on $H$ defined by $\<a,b\>_{reg}:=\chi_{reg}(ab)$ is then symmetric and non-degenerate. By the non-degeneracy of $\lambda$, there exists a unique element $g$ of $H$ such that $$\label{eq:trace_element}
\chi_{reg}(x) = \lambda(xg)$$ for all $x \in H$. We call $g$ the [*trace element*]{}.
[If char $k =0$, and $H$ is a finite-dimensional semi-simple Hopf algebra over $k$, then $S^2=id_H$. By (\[eq:cointegral\]), $$\lambda(x)=\sum_i b^i(xb_i)=\chi_{reg}(x)\,.$$ Thus, the trace element of $H$ is 1. ]{}
\[l:inner\] Let $H=(H, \Delta,\e, \Phi, \a, \b, S)$ be a finite-dimensional semi-simple quasi-Hopf algebra. Then the trace element $g$ of $H$ is invertible and $$S^2(a) =g^{-1}ag$$ for all $a \in H$. Moreover, $gS(g)$ is in the center of $H$ and $gS(g)=S(g)g$.
By (\[eq:trace\_element\]), the left annihilator of $g$ in $H$ is a subset of $\ker \chi_{reg}$. Since $H$ is semi-simple, $\ker \chi_{reg}$ does not contain any non-trivial left ideals of $H$. Therefore, the left annihilator of $g$ is trivial. Since the left regular representation of $H$ is faithful and finite-dimensional, $g$ is invertible. Thus, we have $$\lambda(ab)=\lambda(abg^{-1}g)= \chi_{reg}(abg^{-1})=\chi_{reg}(bg^{-1}a)=\lambda(bg^{-1}ag)$$ for all $x, y \in H$. By the non-degeneracy of $\lambda$ and (\[eq:nakayama\]), we obtain $$S^2(a)=g^{-1}ag$$ for all $a \in H$. In particular, $$S(g^{-1}ag)=S^3(a)=g^{-1}S(a)g\,.$$ Therefore, $$\label{eq:g}
gS(g)S(a)=S(a)gS(g)$$ for all $a \in H$ and hence $gS(g)$ is in the center of $H$. Taking $a=g^{-1}$ in (\[eq:g\]), the result in the last statement follows.\
\[l:u\] Let $A$ be a finite-dimensional split semi-simple algebra over $k$ and $S$ an algebra anti-automorphism on $A$ such that $S^2$ is inner. Then there exists a unit $u \in A$ such that $S_u^2 =id_A$ where $$S_u(x) = uS(x)u^{-1}$$ for all $x \in A$.
Without loss of generality, we can assume that $A$ is a direct sum of full matrix rings over $k$, say $A=\oplus_{i=1}^d
M_{n_i}(k)$. Let $\iota_i $ denote the natural embedding from $M_{n_i}(k)$ into $A$, $p_i$ the natural surjection from $A$ onto $M_{n_i}(k)$, and $A_i$ the image of $\iota_i$. Then, $A_1, \cdots
, A_d$ is the complete set of minimal ideals of $A$. Since $S$ is an algebra anti-automorphism, there exists a permutation $\sigma$ on $\{1,\dots, d\}$ such that $S(A_i) = A_{\sigma(i)}$ for all $i=1, \dots, d$. As $S^2$ is inner, $S^2(A_i)=A_i$ for all $i$ and so $\sigma^2=id$.\
Since $S(A_i)= A_{\sigma(i)}$, $M_{n_i}(k) =
M_{n_{\sigma(i)}}(k)$. Moreover, $p_j \circ S \circ \iota_i=0$ for $j \ne \sigma(i)$ and $p_{\sigma(i)} \circ S \circ \iota_i$ is an algebra anti-automorphism on $M_{n_i}(k)$. By the Skolem-Noether theorem, there exists an invertible matrix $u_{i} \in
M_{n_{\sigma(i)}}(k)$ such that $p_{\sigma(i)} \circ S\circ \iota
(x) =u_i^{-1}x^t u_i$ for any $x \in M_{n_i}(k)$ where $x^t$ is the transpose of $x$.\
Let $u=\sum_{i=1}^d \iota_{\sigma(i)}(u_i)$. Since $u_i$ is invertible in $M_{n_{\sigma(i)}}(k)$ for all $i$, $u$ is invertible in $A$. Since $S(A_i) = A_{\sigma(i)}$ is an ideal of $A$, $S_u(A_i) = A_{\sigma(i)}$. Then for any $x \in M_{n_i}(k)$, $$p_{\sigma(i)}(S_u(\iota_i(x))) = u_i(p_{\sigma(i)}\circ S \circ
\iota_i(x))u_i^{-1}=x^t\,.$$ Thus, $$\iota_{\sigma(i)}(x^t) = \iota_{\sigma(i)}\circ
p_{\sigma(i)}(S_u(\iota_i(x))) = S_u(\iota_i(x))\,,$$ and hence $$S^2_u(\iota_i(x))=S_u(\iota_{\sigma(i)}(x^t))=\iota_{\sigma^2(i)}((x^t)^t)
=\iota_i(x)$$ as $\sigma^2=id$. Therefore, $S^2_u(a)=a$ for all $a \in A_i$, $i=1, \dots, d$. Since $A=A_1 \oplus \cdots \oplus A_d$, $S^2_u=id_A$.
\[t:6.4\] Let $H=(H, \Delta,\e, \Phi, \a, \b, S)$ be a finite-dimensional split semi-simple quasi-Hopf algebra over $k$. Then there exists an invertible element $u$ of $H$ such that the antipode of $H_u$ is an involution.
It follows from Lemma \[l:inner\] or [@HNQA Proposition 5.6] that $S^2$ is inner. By Lemma \[l:u\], the result follows.\
\[r:equiv\_dual\] [Suppose $u$ is an invertible element of $H$ and $M$ a finite-dimensional left $H$-module. Let ${^+\!M}$, $^*\!M$ denote the left dual of $M$ in $H_u$-$\mod_{fin}$ and $H$-$\mod_{fin}$ respectively. Then, ${^+\!M}$ and $^*\!M$ are isomorphic left $H$-modules under the map $\phi_u : {^+\!M} \map {^*\!M}$ defined by $$\phi_u(f)(x)=f(ux)$$ for all $x \in M$ and $f \in M'$. In particular, $M \cong {^*\!M}$ if, and only if, $M \cong {^+\!M}$ as left $H$-modules (cf. [@Drin90 p1425]). ]{}
Pivotal Category Structure of [$H\mbox{-}\mod_{fin}$]{}
=======================================================
We begin (Theorem \[t:etingof\]) with Etingof’s observation that the trace element $g$ of a finite-dimensional semi-simple quasi-Hopf algebra $H=(H, \Delta,\e, \Phi, \a, \b, S)$ over an algebraically closed field of characteristic zero defines an isomorphism of tensor functors $$j: Id \map \,{^{**}?}\,.$$ Moreover, we prove that $S(g)=g^{-1}$, a fact that we will need in Section 8. A direct result of this is that $H$-$\mod_{fin}$ is a pivotal category in the sense of Joyal and Street (cf. [@FY]). For the remainder of this paper we will assume that $k$ is an algebraically closed field of characteristic zero.\
For simplicity, we write $\mathcal{C}$ for the semi-simple rigid tensor category $H$-$\mod_{fin}$ in this section. Obviously, $\mathcal{C}$ is a fusion category over $k$ (cf. [@ENO]). Recall from [@BaKi] that if $V \in \mathcal{C}$ and $f: V
\map\, {^*V}$ then the categorical trace of $f$ is the scalar $\tr_V(f)$ defined by $$\label{e:dim}
\ev_{^*\!V} \circ (f \otimes id) \circ \coev_V\,.$$ where $\ev_V: {^*V} \otimes V \map k$ and $\coev_V: k\map\, V \otimes {^*V}$ are evaluation and coevaluation maps.
Following [@Mu], for any simple object $V$ in $\mathcal{C}$ and an isomorphism $f: V \map\, {^*V}$, we define $$\label{e:squared_norm_1}
|V|^2=\tr_V(f)\,\tr_{^*\!V}({^*\!(f^{-1})})\,.$$ Clearly, $|V|^2$ is independent of the choice of $f$.\
By [@ENO], there exists an isomorphism of tensor functors $$j: Id \map \,{^{**}?}$$ such that for any simple object $V$ of $\mathcal{C}$, $$\label{e:dimensions}
\tr_V(j) = \FPdim(V)=\dim(V)\,$$ where $\FPdim(V)$ is the [*Frobenius-Perron dimension*]{} of $V$. Moreover, $$\label{e:squared_norm_2}
|V|^2 = \dim(V)^2\,.$$
Let $a$ be the unique invertible element of $H$ such that $$\label{e:a}
j_H(1)(f)=f(a)$$ for all $f \in {^*\!H}$. By the naturality of $j$, one can show that $$\label{e:inner_a}
S^2(x)=axa^{-1} \quad \mbox{ for all } x \in H\,,$$ and for any $V \in \mathcal{C}$, $j:V \map\, {^{**}V}$ is given by $$\label{e:j}
j_V(x)(f)= f(ax)$$ for all $x \in V$ and $f \in {^*V}$. Thus, by (\[e:dim\]) and (\[e:squared\_norm\_1\]), for any simple objective $V$ in $\mathcal{C}$ with character $\chi$, $$\label{e:dimV}
\dim(V)= \chi(a \b S(\a))\,,$$ and $$\label{e:squared norm}
|V|^2= \chi(a \b S(\a))\chi(a^{-1} S(\b) \a)\,.$$ Hence, by (\[e:dimV\]) and (\[e:squared\_norm\_2\]), we also have $$\label{e:dimV_2}
\dim(V)=\chi(a^{-1} S(\b) \a)$$ In fact, $a^{-1}$ is the trace element of $H$.\
\[t:etingof\] Let $H=(H, \Delta,\e, \Phi, \a, \b, S)$ be a finite-dimensional semi-simple quasi-Hopf algebra over $k$ and $g$ the trace element of $H$. Then the natural isomorphism $j_V: V \map\, {^{**}\!V}$ for any $V$ in $H$-$\mod_{fin}$, given by $$j_V(x)(f)=f(g^{-1}x)$$ for all $x \in V$ and $f \in {^{*}\!V}$, defines an isomorphism of the tensor functors $Id$ and ${^{**}?}$ such that $$\dim(V) = \chi(g^{-1} \b S(\a))$$ for any simple $H$-module $V$ with character $\chi$.
By the preceding discussion, it suffices to show that the element $a$ defined in (\[e:a\]) is identical to $g^{-1}$. By Lemma \[l:inner\] and (\[e:inner\_a\]), $ag$ is in the center of $H$. Therefore, it is enough to show that for any simple $H$-module $V$ with character $\chi$, $$\chi(a \b S(\a) ) =\chi( g^{-1} \b S(\a) )\,.$$ Let $e_V$ be the central idempotent of $H$ such that $$\chi(x)\dim(V)=\chi_{reg}(e_V x)$$ for all $x \in H$. Thus, we obtain $$\begin{aligned}
\chi(g^{-1}\b S(\a))\dim(V) & = & \chi_{reg}(e_V g^{-1}\b S(\a)) \\
&=& \chi_{reg}(e_V \b S(\a) g^{-1}) \\
&=& \lambda(e_V \b S(\a))\end{aligned}$$ where $\lambda$ is the normalized left cointegral of $H$. Let $\{b^i\}$ be the dual basis of the basis $\{b_i\}$ of $H$. Then, we have $$\begin{aligned}
\chi(g^{-1}\b S(\a))\dim(V) & = & \sum_i b^i(e_V \b S(\a)S^2(b_i) S(\b)\a) \\
&=& \chi_{reg}(e_V \b S(\a) a b_i a^{-1} S(\b)\a) \\ \\
&=& \chi(\b S(\a) a)\chi(a^{-1} S(\b)\a) \\ \\
&=& |V|^2 = \dim(V)^2 \quad \mbox{by (\ref{e:squared norm}) and
(\ref{e:squared_norm_2})}\,.\end{aligned}$$ Therefore, by (\[e:dimV\]), we obtain $$\chi(g^{-1}\b S(\a))=\dim(V)=\chi(a \b S(\a))\,.$$
\[t:7.2\] Let $H=(H, \Delta,\e, \Phi, \a, \b, S)$ be a finite-dimensional semi-simple quasi-Hopf algebra over $k$ and $g$ the trace element of $H$. Then $S(g)=g^{-1}$ and hence $$\label{e:pivotal}
{^*(j_V)}\circ j_{^*V} =id_{^*V}$$ for any $V \in H$-$\mod_{fin}$
Since $gS(g)$ is central, $gS(g)$ acts on any simple $H$-module $V$ as multiplication by a scalar $c_V\in k$. In order to show that $S(g)=g^{-1}$, it suffices to prove that $$c_V=1$$ for any simple $H$-module $V$.\
Let $V$ be a simple $H$-module with character $\chi$. Then the character of ${^*V}$ is ${^*\!\chi}$ given by $${^*\!\chi}=\chi\circ S\,.$$ By Theorem \[t:etingof\], (\[e:dimV\]) and (\[e:dimV\_2\]), we have $$\begin{aligned}
\dim({^*V})& = {^*\!\chi}(gS(\b)\a)=
\chi(S(\a)S^2(\b)S(g))\\
&= \chi(S(\a) g^{-1} \b gS(g)) = c_V \, \chi(S(\a) g^{-1}
\b)\\
& = c_V\, \dim(V)
\end{aligned}$$ Therefore, $c_V=1$. Equation (\[e:pivotal\]) follows easily from $S(g)=g^{-1}$.\
Theorem \[t:etingof\] and (\[e:pivotal\]) implies that $H$-$\mod_{fin}$ is indeed a pivotal category defined by Joyal-Street (cf. [@FY]). Nikshych also pointed out that (\[e:pivotal\]) can be proved using *weak Hopf algebras*.
Frobenius-Schur Indicators via Bilinear Forms with Adjoint $S$
==============================================================
Let $H=(H, \Delta,\e, \Phi, \a, \b, S)$ be a finite-dimensional semi-simple quasi-Hopf algebra over $k$, and $g$ the trace element of $H$. In this section, we will prove that for any simple left $H$-module $M$ with character $\chi$, the Frobenius-Schur indicator $\chi(\nu_H)$ of $\chi$ can only be 0, 1 or -1. It is non-zero if, and only if $M \cong {^*\!M}$. Moreover in this case, $M$ admits a non-degenerate bilinear form $\<\cdot, \cdot\>$ such that $\<h u, v\> = \< u, S(h)v\>$ for all $h \in H$, $u, v \in M$, and $$\<u, v\> =\chi(\nu_H) \<v, g^{-1}u\>\,.$$
Let $H=(H, \Delta,\e, \Phi, \a, \b, S)$ be a quasi-Hopf algebra over $k$, $M$ be a left $H$-module and $\<\cdot, \cdot\>$ a bilinear form on $M$.
1. The form is said to be $H$-invariant if $$\sum \<h_1 u , h_2 v\> = \e(h) \<u,v\>$$ for all $h \in H$ and $u,v \in V$ where $\sum h_1 \otimes h_2 = \Delta(h)$.
2. The antipode $S$ is said to be the adjoint of the form if $$\<h u , v\> = \<u, S(h) v\>$$ for all $h \in H$ and $u,v \in V$.
\[l:schur\] Let $H=(H, \Delta,\e, \Phi, \a, \b, S)$ be a quasi-Hopf algebra over $k$ and $M$ a simple left $H$-module. If $\<\cdot, \cdot\>_1$ and $\<\cdot, \cdot\>_2$ are non-degenerate bilinear forms on $M$ with the same adjoint $S$, then there exists a non-zero element $c \in
k$ such that $$\<u, v\>_1 = c \<u, v\>_2$$ for all $u, v \in M$.
Define $J_i : M \map {^*\!M}$ ( $i=1, 2$) by $$J_i(u)(v) = \< u, v\>_i$$ for $u, v \in M$. Since $\<\cdot, \cdot\>_1$ and $\<\cdot,
\cdot\>_2$ are non-degenerate bilinear forms on $M$ with the adjoint $S$, $J_1, J_2$ are isomorphisms of $H$-modules. In particular, $M$ and ${^*\!M}$ are isomorphic simple $H$-modules. By Schur’s lemma, $J_1=c J_2$ for some non-zero element $c \in k$ and so the result follows.\
\[l:dualbasis\] Let $H=(H, \Delta,\e, \Phi, \a, \b, S)$ be a finite-dimensional semi-simple quasi-Hopf algebra, $\Lambda$ the normalized two-sided integral of $H$ and $g$ the trace element of $H$. Suppose that $$q_R \Delta(\Lambda) p_R = \sum_{i=1}^n x_i \otimes y_i$$ where $\{x_i\}$ is basis of $H$. Then $\{S(x_i) g^{-1}, y_i\}$ is a pair dual bases with respect to $\<\cdot,\cdot\>_{reg}$.
Following [@HNQA], we define the elements $U, V \in H
\otimes H$ by $$\begin{aligned}
U &= &F_H^{-1}(S \otimes S)(q_R^{21})\,,\\
V& = &(S^{-1} \otimes S^{-1})(F_H^{21}p_R^{21})\end{aligned}$$ where $F_H$, $q_R, p_R \in H \otimes H$ are defined in (\[FH\]) and (\[R\]). By [@HNQA (7.3) and (7.4)], $$\begin{aligned}
q_R \Delta(\Lambda) p_R &=&
(q_L^2 \otimes 1)V\,\Delta(S^{-1}(q_L^1))\, \Delta(\Lambda)
\Delta(S(p_L^1))\,U(p_L^2 \otimes 1)\\
&=& (q_L^2 \e(S^{-1}(q_L^1)) \otimes 1)V\,\Delta(\Lambda)\,U (\e(S(p_L^1))p_L^2 \otimes 1)\\\end{aligned}$$ By [@Drin90 Remark 7], $\e \circ S = \e = \e \circ S^{-1}$. Therefore, $$q_L^2 \e(S^{-1}(q_L^1)) = \e(\a)1_H, \quad \mbox{and}\quad \e(S(p_L^1))p_L^2=\e( \b)1_H\,.$$ It follows from (\[2.9\]) that $\e(\a\b)=1$ and so $$q_R \Delta(\Lambda) p_R = \e(\a)\e(\b) V \Delta(\Lambda) U =V \Delta(\Lambda) U \,.$$ Let $\lambda$ be the normalized left cointegral of $H$. By [@HNQA Proposition 5.5], $$\sum_i S(x_i)\lambda(y_i a) = a$$ for all $a \in H$. In particular, $$a = (ag)g^{-1} = \sum_i S(x_i)g^{-1} \lambda(y_i ag) =
\sum_i S(x_i)g^{-1} \chi_{reg}(y_i a)\,.$$ Since $\{S(x_i)g^{-1}\}$ is also a basis of $H$, $\chi_{reg}(y_i S(x_j)g^{-1}) = \delta_{ij}$ and so $\{S(x_i) g^{-1}, y_i\}$ is a pair dual bases of $H$ with respect to $\<\cdot, \cdot\>_{reg}$.\
\[l:sum\_basis\] Let $A$ be a finite-dimensional semi-simple algebra over $k$ and $\{a_i, b_i\}$ a pair dual bases with respect to the form $\<\cdot, \cdot\>_{reg}$. Then $$\sum_i a_i b_i =1_A\,.$$
Without loss of generality, we may assume that $A=\oplus_{i=1}^d
M_{n_i}(k)$. Then $\chi_{reg}(x)= \sum_{i=1}^d n_i tr_i(x)$ where $tr_i(x)$ is the trace of the $i$th component matrix of $x$. Let $\{e^i_{lm}\}$ be the set of matrix units for the $i$th summand $M_{n_i}(k)$ of $A$. Following [@LM00], $\{n_i^{-1}e^i_{lm}, e^i_{ml}\}$ is a pair of dual basis with respect to $\<\cdot, \cdot\>_{reg}$. Thus, $$\sum_{i, l, m} n_i^{-1}e^i_{lm}e^i_{ml}= \sum_{i, l}e^i_{ll} =
1_A\,.$$ It follows from [@LM00 Lemma 2.6] that $$\sum_i a_i b_i = \sum_{i, l,m} n_i^{-1}e^i_{lm}e^i_{ml} =1_A \,.$$\
Let $H=(H, \Delta,\e, \Phi, \a, \b, S)$ be a finite-dimensional semi-simple quasi-Hopf algebra over $k$. Then trace element $g$ of $H$ is given by $$g = m\tau(S \otimes id)(q_R\Delta(\Lambda) p_R)$$ where $\Lambda$ is the normalized integral of $H$, $m$ is multiplication and $\tau$ the usual flip map.
Let $$q_R \Delta(\Lambda) p_R = \sum_i x_i \otimes y_i\,.$$ By Lemma \[l:dualbasis\] and Lemma \[l:sum\_basis\], we have $$\sum_i y_iS(x_i)g^{-1} =1$$ and so the result follows.\
Let $\{a_i, b_i\}$ be dual bases of the semi-simple quasi-Hopf algebra $H$ with respect to the form $\< \cdot, \cdot\>_{reg}$ discussed in Lemma \[l:dualbasis\]. For any $k$-involution $\mathcal{I}$ on $H$ and for any character $\chi$ of $H$, we define $$\mu_2(\chi, \mathcal{I})= \chi(\sum_i \mathcal{I}(a_i)b_i)\,.$$
\[r:mu2\] [Since $\sum_i a_i b_i = 1_H$ by Lemma \[l:sum\_basis\], the $\mu_2$ defined in [@LM00 Therorem 2.7] with respect to the $k$-involution $\mathcal{I}$ is given by $$\frac{\chi(1_H)}{\chi(\sum_i a_i b_i)}
\chi(\sum_i \mathcal{I}(a_i)b_i) = \chi(\sum_i \mathcal{I}(a_i)b_i)$$ which coincides with $\mu_2(\chi, \mathcal{I})$. ]{}
\[l:mu2\] Let $H=(H, \Delta,\e, \Phi, \a, \b, S)$ be a finite-dimensional semi-simple quasi-Hopf algebra over $k$, $g$ the trace element of $H$, and $M$ an irreducible $H$-module with character $\chi$. Then for any unit $u \in H$ such that $S_u$ is an involution, $$\mu_2(\chi, S_u)= c\,\chi(\nu_H)$$ where $c$ is the non-zero scalar given by $$c=\frac{\chi( u S(u^{-1})g^{-1})}{\dim M}\,.$$
If $u$ is a unit of $H$ such that $S_u$ is an involution, then for any $x \in H$, $$x= S_u^2(x)=uS(u^{-1}) S^2(x) S(u) u^{-1}$$ or equivalently $$S^2(x)=S(u)u^{-1} x u S(u^{-1})\,.$$ By Lemma \[l:inner\], $ uS(u^{-1})g^{-1}$ is in the center of $H$. Thus, $uS(u^{-1}) g^{-1}$ acts on $M$ as multiplication by the non-zero scalar $$c=\frac{\chi(uS(u^{-1})g^{-1}))}{\dim M}\,.$$ Suppose that $$q_R \Delta(\Lambda) p_R = \sum_i x_i \otimes y_i$$ as in Lemma \[l:dualbasis\] where $\Lambda$ is the normalized two-sided integral of $H$. Then we have $$\begin{aligned}
\mu_2(\chi, S_u) & = & \chi(\sum_i S_u(S(x_i)g^{-1}) y_i) \\
&=& \chi(\sum_i u S(g^{-1})S^2(x_i) u^{-1} y_i) \\
&=& \chi(\sum_i u S(g^{-1}) g^{-1} x_i g u^{-1} y_i) \\
&=&\chi(\sum_i u S(g^{-1})g^{-1}S(u^{-1})S(g) x_i y_i) \quad\mbox{(by Lemma \ref{pq_1})}\\
&=& \chi(\sum_i u S(u^{-1})g^{-1}x_i y_i) \quad
\mbox{(by Lemma \ref{l:inner})}\\\\
& =& c\, \chi(\nu_H)\,.\end{aligned}$$
\[t:main2\] Let $H=(H, \Delta,\e, \Phi, \a, \b, S)$ be a finite-dimensional semi-simple quasi-Hopf algebra over $k$, $g$ the trace element of $H$, and $M$ a simple $H$-module with character $\chi$. Then the Frobenius-Schur indicator $\chi(\nu_H)$ of $\chi$ satisfies the following properties:
1. $\chi(\nu_H) \ne 0$ if, and only if, $M \cong {^*\!M}$ as left $H$-modules.
2. For any non-zero $\kappa \in k$, $\chi(\nu_H) =\kappa$ if, and only if, $M$ admits a non-degenerate bilinear form $\< \cdot, \cdot\>$ with the adjoint $S$ such that $$\< x, y \> = \kappa\<y, g^{-1}x \>$$ for all $x, y \in M$.
3. The values of $\chi(\nu_H)$ can only be $0$, $1$ or $-1$.
Moreover, $$\Tr(S) = \sum_{\chi \in Irr(H)} \chi(\nu_H)\chi(g^{-1}).$$
By Theorem \[t:6.4\], there exists an unit $u \in H$ such that $S_u$ is an involution. As in the proof of Lemma \[l:mu2\], $uS(u^{-1}) g^{-1}$ is a central unit of $H$. Thus, $uS(u^{-1}) g^{-1}$ acts on $M$ as multiplication by the non-zero scalar $$c=\frac{\chi(uS(u^{-1})g^{-1})}{\dim M}\,.$$ Also, by [@LM00 Theorem 2.7] and Remark \[r:mu2\], the element $\mu_2(\chi, S_u) \ne 0$ if, and only if $M \cong {^+\!M}$ as left $H$-modules where ${^+\!M}$ is the left $H$-module with underlying space $M'$ and the $H$-action given by $$(hf)(x) = f(S_u(h)x)$$ for all $f \in M'$ and $h \in H$. Actually, ${^+\!M}$ is the left dual of $M$ in $H_u$-$\mod_{fin}$. It follows from Remark \[r:equiv\_dual\] that $\mu_2(\chi, S_u) \ne 0$ if, and only if $M \cong {^*\!M}$ as left $H$-modules. Hence, by Lemma \[l:mu2\], statement (i) follows.\
If $\chi(\nu_H) \ne 0$, then $\mu_2(\chi, S_u) \ne 0$ by Lemma \[l:mu2\]. By Remark \[r:mu2\] and [@LM00 Theorem 2.7(ii)], $M$ admits a non-degenerate bilinear form $(\cdot, \cdot)$ with adjoint $S_u$ such that $$(x, y) = \mu_2(\chi, S_u) (y,x)$$ for any $x, y \in M$. Define $$\<x, y\>= (x, uy)$$ for any $x, y \in M$. One can easily see that $\<\cdot, \cdot\>$ is a non-degenerate bilinear form on $M$ with adjoint $S$. Moreover, for any $x, y\in M$, $$\<x, y\> = (x, uy) = \mu_2(\chi, S_u) (uy,x)= \mu_2(\chi, S_u) (y,S_u(u) x)\,.$$ Thus, by Lemma \[l:mu2\], we obtain $$\<x, y\> =c\, \chi(\nu_H) (y, S_u(u)x)
= \chi(\nu_H) \<y, S(u)u^{-1}cx\> = \chi(\nu_H) \<y, g^{-1}x\>\,.$$
Conversely, suppose $M$ admits a non-degenerate bilinear form $\<\cdot, \cdot\>$ with adjoint $S$ and that there exists a non-zero element $\kappa$ of $k$ such that $$\< x, y \> = \kappa\<y, g^{-1} x \>$$ for all $x, y \in M$. Then the map $J: M \map {^*\!M}$, defined by $$J(x)(y)=\<x, y\>, \quad x, y \in M\,,$$ is an isomorphism of left $H$-modules. Thus, by (i), $\chi(\nu_H)\ne 0$. Hence, by above arguments, $M$ admits a non-degenerate bilinear form $\<\cdot, \cdot\>_0$ with adjoint $S$ such that $$\< x, y \>_0 = \chi(\nu_H) \<y, g^{-1} x \>_0$$ for all $x, y \in M$. By Lemma \[l:schur\], $\<\cdot, \cdot\>$ is a non-zero scalar multiple of $\<\cdot, \cdot\>_0$. Therefore, $$\kappa=\chi(\nu_H)$$ and this finishes the proof statement(ii).\
(iii) If $M$ is a simple $H$-module with character $\chi$ such that $\chi(\nu_H)\ne 0$, by (ii), $M$ admits a non-degenerate bilinear form $\<\cdot, \cdot\>$ with adjoint $S$ such that $$\<x,y\> = \chi(\nu_H) \<y, g^{-1}x \>$$ for all $x, y \in M$. Thus, we have $$\begin{aligned}
\<x, y\> & = \chi(\nu_H)^2 \< g^{-1}x, g^{-1}y\>= \chi(\nu_H)^2
\<x, S(g^{-1})g^{-1}y\>\\
& = \chi(\nu_H)^2 \<x, y\> \quad (\mbox{ by Theorem }\ref{t:7.2} )\,.
\end{aligned}$$ Therefore, $\chi(\nu_H)^2=1$ or equivalently $\chi(\nu_H)=\pm 1$.\
Let $\sum_i x_i \otimes y_i= q_R \Delta(\Lambda)p_R$ where $\Lambda$ is the normalized two-sided integral of $H$. By Lemma \[l:dualbasis\], $
\{S(x_i)g^{-1}, y_i\}
$ is a pair of dual bases of $H$ with respect to the form $\<\cdot , \cdot \>_{reg}$ on $H$. Therefore, we obtain $$\begin{aligned}
\Tr(S) & =& \sum_i \<S(S(x_i)g^{-1}), y_i\>_{reg} \\
&=& \sum_i \chi_{reg}(S(g^{-1})S^2(x_i)y_i)\\
&=& \sum_i\chi_{reg}(S(g^{-1})g^{-1} x_i g y_i)\\
&=& \sum_i \chi_{reg}(S(g^{-1})g^{-1} S(g) x_i y_i)
\quad\mbox{by Lemma \ref{pq_1}}\\
&=& \sum_i \chi_{reg}(g^{-1} x_i y_i)
\quad\mbox{by Lemma \ref{l:inner}}\\
&=& \chi_{reg}(g^{-1} \nu_H)\,.\end{aligned}$$ Since $\nu_H$ is in the center of $H$, for any irreducible $H$-module $M$ with character $\chi$, $\nu_H$ acts on $M$ as a multiplication by the scalar $$c_\chi = \chi(\nu_H)/\chi(1_H)\,.$$ Since $\displaystyle \chi_{reg}=\sum_{\chi \in Irr(H)} \chi(1_H)\chi$, we have $$\begin{aligned}
\Tr(S) & = &\sum_{\chi \in Irr(H)}\chi(1_H)
\chi(g^{-1} \nu_H) \\
&=& \sum_{\chi \in Irr(H)}\chi(1_H)
c_\chi \,\chi(g^{-1})\\
&=& \sum_{\chi \in Irr(H)}\chi(\nu_H)\chi(g^{-1})\,.\end{aligned}$$
[In [@FGSV99], Fuchs et al also define a notion of Frobenius-Schur indictor for simple objects in a sovereign $C^*$-category $\mathcal{C}$ such that $$id : {^*\!M} \map M^*$$ defines an isomorphism of the tensor functors ${^*?}$ and ${?^*}$. Let $k_M : M \map (^*\!M)^*={^{**}\!M}$ be the natural isomorphism of the underlying autonomous structure of $\mathcal{C}$. Then for any simple object $M$ in $\mathcal{C}$, the Frobenius-Schur indicator $c_M$ of $M$ is defined to be 0 if $M \not\cong {^*M}$ and $c$ if there exists a $H$-module isomorphism $J: M \longrightarrow {^*\!M}$ where $c$ given by the equation $$\label{e:Fuch_FS}
J^* \circ k_M = c \, J \,,$$ in which case the values of $c_M$ can only be 0, 1 or -1.\
The category $H$-$\mod_{fin}$ is not of this kind in general. Nevertheless, if one replaces $k_M$ in (\[e:Fuch\_FS\]) by $j_M : M \map\, {^{**}\!M}$, given by $$j_M(f)(x)=f(g^{-1}x) \quad \mbox{for all } x\in M\, \mbox{ and }f \in {^*\!M}\,,$$ one can still define [*Frobenius-Schur indicator*]{} $c_M$ for any simple $H$-module $M$ to be 0 if $M \not\cong {^*M}$ and $c$ if there exists a $H$-module isomorphism $J: M \longrightarrow {^*M}$ where $c$ given by the equation $$J^* \circ j_M = c\, J \,.$$ Theorem \[t:main2\] (i) and (ii) implies $c_M = \chi(\nu_H)$.\
]{}
Before closing this section, we will show that if $\a$ is a central unit, a bilinear form on a $H$-module $M$ is $H$-invariant if, and only if, $S$ is the adjoint of the form. Both semi-simple Hopf algebras over $k$ or twisted quantum doubles of finite groups are of this type.
\[p:adj\_inv\] Let $H=(H, \Delta,\e, \Phi, \a, \b, S)$ be a quasi-Hopf algebra over $k$ and $M$ a $H$-module. Then, the set $Inv(M)$ of $H$-invariant forms on $M$ and the set $Adj_S(M)$ of forms $M$ with adjoint $S$ are isomorphic as $k$-spaces. In addition, if $\a$ is a central unit of $H$, then $$Inv(M) = Adj_S(M)\,.$$
Note that both $Inv(M)$ and $Adj_S(M)$ are $k$-subspaces of $(M \otimes M)^*$. We define $\phi: Adj_S(M) \map (M \otimes M)^*$ and $\psi: Inv(M) \map (M \otimes M)^*$ by $$\begin{aligned}
\phi(\mathbf{b})(x \otimes y) &=& \mathbf{b}(x \otimes \a y)\\
\psi(\mathbf{b'})(x \otimes y) &=& \mathbf{b'}(p_L(x \otimes y))\end{aligned}$$ for any $x, y \in M$, $\mathbf{b} \in Adj_S(M)$ and $\mathbf{b'}\in Inv(M)$. Using (\[2.8\]), one can easily see that $$Im (\phi) \C Inv(M)\,.$$ By (\[eq:pL\]), $\psi(\mathbf{b'})$ has adjoint $S$ for any $H$-invariant form $\mathbf{b'}$ on $M$ and so $$Im (\psi) \C Adj_S(M)\,.$$
It follows easily from (\[2.18\]) that for any $\mathbf{b'} \in Inv(M)$ and $x, y \in M$, $$\begin{aligned}
\mathbf{b'}(x\otimes y)& =&
\mathbf{b'}(\Delta(q_L^2) p_L (S^{-1}(q_L^1) x \otimes y)) \\
& =& \mathbf{b'}(p_L (S^{-1}(q_L^1 \e(q_L^2)) x \otimes y) \\
&=& \psi(\mathbf{b'})((S^{-1}(\a) x \otimes y)\,.\end{aligned}$$ Since $\psi(\mathbf{b'}) \in Adj_S(M)$, $$\phi\circ\psi = id_{Inv(M)}\,.$$ On the other hand, by (\[2.9\]), for any $\mathbf{b} \in Adj_S(M)$ and $x, y \in M$, $$\psi\circ \phi(\mathbf{b})(x \otimes y) =
\mathbf{b}(p^1_L x \otimes \a p^2_L y) =
\mathbf{b}(x \otimes S(p_L^1)\a p^2_L y) = \mathbf{b}(x \otimes y)\,.$$ Therefore, $\phi: Adj_S(M) \map Inv(M)$ is a $k$-linear isomorphism.\
If $\a$ is a central unit, we consider the quasi-Hopf algebra $H_{\a^{-1}}$. Then, the corresponding $\phi$ is the identity map and so $$Adj_{S_{\a^{-1}}}(M)=Inv(M)\,.$$ Since $S_{\a^{-1}} =S$, the second statement follows.
Frobenius-Schur Indicators of Twisted Quantum Doubles of Finite Groups
======================================================================
We showed in section \[s:Bantay\_Indicator\] that for any simple module $M$ for $D^\w(G)$ with character $\chi$, Bantay’s formula of the indicator of $\chi$ is $\chi(\nu_{D^\w(G)})$. In this section, we will prove that the trace element of $D^\w(G)$ is $\b$ and the Frobenius-Schur indictor $\chi(\nu_{D^{\w}(G)})$ of $\chi$ is non-zero if, and only if, ${^*\!M} \cong M$. Moreover, the indicator of $\chi$ is $1$ (respectively $-1$) if and only if $M$ admits a $\b^{-1}$-symmetric (resp. $\b^{-1}$-skew symmetric) non-degenerate $D^{\w}(G)$-invariant bilinear form $\<\cdot, \cdot\>$, that is $$\<x,y\>= \<y, \b^{-1} x\>\quad (resp. \quad \<x,y\>= -\<y, \b^{-1} x\> )$$ for all $x, y \in M$.\
We first need the following formula (cf. [@AC92]) to compute the trace element of $D^\w(G)$.
\[l:8.1\] Let $\w:G \times G \times G \map k^\times $ be a normalized 3-cocycle of a finite group $G$ and let $S$ be the antipode of the quasi-Hopf algebra $D^\w(G)$ defined in Section \[s:Bantay\_Indicator\]. Then for any $g, x \in G$, $$\begin{aligned}
S^2(e(g)\otimes x) & = & \frac{\w(g^x, (g^{-1})^x, g^x)}{\w(g, g^{-1},g)}
\, e(g)\otimes x\,,\\
&=& \b^{-1} (e(g) \otimes x) \b\,.\end{aligned}$$
It follows from (\[eq:antipode\]) that $$\label{eq:S2}
S^2(e(g)\otimes x) = \left(
\theta_{g^{-1}}(x,x^{-1})\g_x(g,g^{-1})\theta_{g^x}(x^{-1},
x)\g_{x^{-1}}((g^{-1})^x,g^x) \right)^{-1} e(g)\otimes x\,,$$ where $g^x$ denotes the product $x^{-1}gx$. By the normality of $\w$ and (\[eq:0.1\]), $$\theta_g(x, x^{-1}) = \theta_{g^x}(x^{-1}, x)\,.$$ Thus, we have $$\begin{aligned}
&&\theta_{g^{-1}}(x,x^{-1})\g_x(g,g^{-1})\theta_{g^x}(x^{-1}, x)
\g_{x^{-1}}((g^{-1})^x,g^x) \\ \\
& = &\theta_{g^{-1}}(x,x^{-1})\theta_{g}(x, x^{-1})
\g_x(g,g^{-1})\g_{x^{-1}}((g^{-1})^x,g^x)\end{aligned}$$ By the normality of $\w$ and equation (\[eq:0.2\]), we have $$\begin{aligned}
&&\theta_{g^{-1}}(x,x^{-1})\g_x(g,g^{-1})\theta_{g^x}(x^{-1}, x)
\g_{x^{-1}}((g^{-1})^x,g^x)\\ \\
& = & \frac{\g_x(g,g^{-1})\g_{x^{-1}}((g^{-1})^x,g^x)}{\g_x(g, g^{-1})\g_{x^{-1}}(g^x, (g^{-1})^x)}\\ \\
&=& \frac{\g_{x^{-1}}((g^{-1})^x,g^x)}{\g_{x^{-1}}(g^x, (g^{-1})^x)}\,.\end{aligned}$$ By equation (\[eq:0.3\]), for any $z, a \in G$ we have $$\g_z(a,a^{-1}) \w(a^z, (a^{-1})^z, a^z)=\g_z(a^{-1}, a)\w(a,a^{-1},a)\,.$$ Hence we have $$\begin{aligned}
\nonumber
\frac{\g_{x^{-1}}((g^{-1})^x,g^x)}{\g_{x^{-1}}(g^x, (g^{-1})^x)} &=&
\frac{\w((g^x)^{x^{-1}}, ((g^{-1})^x)^{x^{-1}}, (g^x)^{x^{-1}})}{\w(g^x, (g^{-1})^x, g^x)}\\
\label{eq:0.4} \\\nonumber
&=&
\frac{\w(g, g^{-1}, g)}{\w(g^x, (g^{-1})^x, g^x)}\,.\\ \nonumber\end{aligned}$$ The second equation in the statement of the Lemma follows immediately from (\[multiplication\]).\
Let $\w:G \times G \times G \map k^\times $ be a normalized 3-cocycle of a finite group $G$. Then the trace element of the quasi-Hopf algebra $D^\w(G)$ is $\b$.
Using (\[eq:antipode\]), $S(\b)=\b^{-1}$. Suppose that $\{f_{g,x}\}_{g,x \in G}$ is the dual basis of $\{e(g) \otimes
x\}_{g,x \in G}$. Then, by (\[eq:cointegral\]), the normalized left cointegral of $D^\w(G)$ is given by $$\lambda(e(g) \otimes x) = \sum_{h,y \in G} f_{h,y}((e(g) \otimes x)
S^2(e(h) \otimes y)\b^{-1} )\,.$$ Using Lemma \[l:8.1\], we have $$\begin{aligned}
\lambda(e(g) \otimes x) & = & \sum_{h,y \in G} f_{h,y}((e(g) \otimes x)
\b^{-1} (e(h) \otimes y ))\\
& =& \chi_{reg}((e(g) \otimes x)\b^{-1})\\
& = & \lambda((e(g) \otimes x)\b^{-1} g) \,.\end{aligned}$$ By the non-degeneracy of $\lambda$, $\b^{-1} g =1$ and so $g= \b$.\
Let $\w:G \times G \times G \map k^\times $ be a normalized 3-cocycle of a finite group $G$. Suppose that $M$ is a simple $D^\w(G)$-module with character $\chi$. Then the Frobenius-Schur indicator $\chi(\nu_{D^\w(G)})$ of $\chi$ satisfies the following properties:
1. $\chi(\nu_{D^\w(G)}) = 0, 1$, or $-1$.
2. $ \chi(\nu_{D^\w(G)}) \neq 0$ if, and only if, ${^*\!M} \cong M$.
3. $\chi(\nu_{D^\w(G)})=1$ (respectively $-1$) if and only if $M$ admits a $\b^{-1}$-symmetric (resp. $\b^{-1}$-skew symmetric) non-degenerate $D^{\omega}(G)$-invariant bilinear form.
Moreover, $$\Tr(S) = \sum_{\chi \in Irr(D^\w(G))} \chi(\nu_{D^\w(G)}) \chi(\b^{-1})\,.$$
Statement (i), (ii) and the last statement are immediate consequences of Theorem \[t:main2\]. Since $\a=1$, by Proposition \[p:adj\_inv\], a bilinear form $\<\cdot, \cdot\>$ on $M$ is $D^\w(G)$-invariant if, and only if, $S$ is the adjoint of $\<\cdot, \cdot\>$. Thus, by Theorem \[t:main2\] (ii), the result in statement (iii) follows.
[FGSV99]{}
Daniel Altsch[ü]{}ler and Antoine Coste, *Quasi-quantum groups, knots, three-manifolds, and topological field theory*, Comm. Math. Phys. **150** (1992), no. 1, 83–107. [MR ]{}[94b:57006]{}
Peter Bantay, *The [F]{}robenius-[S]{}chur indicator in conformal field theory*, Phys. Lett. B **394** (1997), no. 1-2, 87–88. [MR ]{}[98c:81195]{}
[to3em]{}, *Frobenius-[S]{}chur indicators, the [K]{}lein-bottle amplitude, and the principle of orbifold covariance*, Phys. Lett. B **488** (2000), no. 2, 207–210. [MR ]{}[2001e:81094]{}
[to3em]{}, *Private communication, 2002*.
Bojko Bakalov and Alexander Kirillov, Jr., *Lectures on tensor categories and modular functors*, University Lecture Series, vol. 21, American Mathematical Society, Providence, RI, 2001. [MR ]{}[2002d:18003]{}
Charles W. Curtis and Irving Reiner, *Representation theory of finite groups and associative algebras*, Wiley Classics Library, John Wiley & Sons Inc., New York, 1988, Reprint of the 1962 original, A Wiley-Interscience Publication. [MR ]{}[90g:16001]{}
R. Dijkgraaf, V. Pasquier, and P. Roche, *Quasi-[H]{}opf algebras, group cohomology and orbifold models \[[M]{}[R]{} 92m:81238\]*, Integrable systems and quantum groups (Pavia, 1990), World Sci. Publishing, River Edge, NJ, 1992, pp. 75–98.
V. G. Drinfel’d, *Quasi-[H]{}opf algebras*, Leningrad Math. J. **1** (1990), 1419–1457.
Pavel Etingof and Shlomo Gelaki, *On families of triangular [H]{}opf algebras*, Int. Math. Res. Not. (2002), no. 14, 757–768. [MR ]{}[2002m:16036]{}
Pavel Etingof, Dmitri Nikshych, and Viktor Ostrik, *On fusion categories*, preprint **arXiv:math.QA/0203060**.
J. Fuchs, A. Ch. Ganchev, K. Szlach[á]{}nyi, and P. Vecserny[é]{}s, *[$S\sb 4$]{} symmetry of [$6j$]{} symbols and [F]{}robenius-[S]{}chur indicators in rigid monoidal [$C\sp *$]{} categories*, J. Math. Phys. **40** (1999), no. 1, 408–426. [MR ]{}[99k:81111]{}
Peter Freyd and David N. Yetter, *Coherence theorems via knot theory*, J. Pure Appl. Algebra **78** (1992), no. 1, 49–76. [MR ]{}[93d:18013]{}
Frank Hausser and Florian Nill, *Diagonal crossed products by duals of quasi-quantum groups*, Rev. Math. Phys. **11** (1999), no. 5, 553–629. [MR ]{}[2000d:81069]{}
[to3em]{}, *Doubles of quasi-quantum groups*, Comm. Math. Phys. **199** (1999), no. 3, 547–589. [MR ]{}[2000a:16075]{}
[to3em]{}, *Integral theory for quasi-hopf algebras*, preprint **arXiv.math.QA/9904164**.
Christian Kassel, *Quantum groups*, Springer-Verlag, New York, 1995.
Y. Kashina, G. Mason, and S. Montgomery, *Computing the [F]{}robenius-[S]{}chur indicator for abelian extensions of [H]{}opf algebras*, J. Algebra **251** (2002), no. 2, 888–913. [MR ]{}[1 919 158]{}
V. Linchenko and S. Montgomery, *A [F]{}robenius-[S]{}chur theorem for [H]{}opf algebras*, Algebr. Represent. Theory **3** (2000), no. 4, 347–355, Special issue dedicated to Klaus Roggenkamp on the occasion of his 60th birthday. [MR ]{}[2001k:16073]{}
Richard G. Larson and David E. Radford, *Semisimple cosemisimple [H]{}opf algebras*, Amer. J. Math. **109** (1987), no. 1, 187–195. [MR ]{}[89a:16011]{}
[to3em]{}, *Finite-dimensional cosemisimple [H]{}opf algebras in characteristic $0$ are semisimple*, J. Algebra **117** (1988), no. 2, 267–289. [MR ]{}[89k:16016]{}
Geoffrey Mason, *The quantum double of a finite group and its role in conformal field theory*, Groups ’93 Galway/St. Andrews, Vol. 2, Cambridge Univ. Press, Cambridge, 1995, pp. 405–417. [MR ]{}[97a:11067]{}
Geoffrey Mason and Siu-Hung Ng, *Group cohomology and gauge equivalence of some twisted quantum doubles*, Trans. Amer. Math. Soc. **353** (2001), no. 9, 3465–3509 (electronic). [MR ]{}[1 837 244]{}
[to3em]{}, *Central invariants and frobenius-schur indicators for semisimple quasi-hopf algebras*, preprint **arXiv.math.QA/0304156**.
Michael Mueger, *From subfactors to categories and topology i. frobenius algebras in and morita equivalence of tensor categories*, preprint **arXiv:math.CT/0111204**.
Florin Panaite, *A [M]{}aschke-type theorem for quasi-[H]{}opf algebras*, Rings, Hopf algebras, and Brauer groups (Antwerp/Brussels, 1996), Lecture Notes in Pure and Appl. Math., vol. 197, Dekker, New York, 1998, pp. 201–207. [MR ]{}[99k:16085]{}
Florin Panaite and Freddy Van Oystaeyen, *Existence of integrals for finite dimensional quasi-[H]{}opf algebras*, Bull. Belg. Math. Soc. Simon Stevin **7** (2000), no. 2, 261–264. [MR ]{}[2001f:16079]{}
Jean-Pierre Serre, *Linear representations of finite groups*, Springer-Verlag, New York, 1977, Translated from the second French edition by Leonard L. Scott, Graduate Texts in Mathematics, Vol. 42. [MR ]{}[56 \#8675]{}
Daisuke Tambara and Shigeru Yamagami, *Tensor categories with fusion rules of self-duality for finite abelian groups*, J. Algebra **209** (1998), no. 2, 692–707.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The relativistic interaction between the spin of an electron and its orbit in a solid is key to understanding novel phenomena in condensed matter physics, such as topologically-protected states, spin-orbit torques, and Majorana modes. Here we focus on the prototypical spin-orbit coupling process, the Rashba-Dresselhaus effect, and analyze how it is modified by lattice vibrations. We show that, in centrosymmetric non-magnetic crystals, dynamic spin splitting is symmetry-forbidden in thermodynamic equilibrium, but it is allowed in crystals with an out-of-equilibrium phonon population. Our findings resolve the current debate on the possibility of phonon-induced Rashba-Dresselhaus effects, reconcile conflicting experimental and theoretical evidence on the Rashba-Dresselhaus effect in hybrid perovskites, and establish the design principles for engineering spin texture in quantum matter via coherent phonons.'
author:
- Martin Schlipf
- Feliciano Giustino
bibliography:
- 'biblio.bib'
title: 'Dynamic Rashba-Dresselhaus Effect'
---
The interplay between crystal symmetry and spin-orbit coupling (SOC) is at the heart of many recent advances in condensed matter physics: SOC provides the coupling between ferromagnetism and ferroelectricity in proper multiferroic materials [@Kenzelmann2005; @cm07], is responsible for band inversion in topological insulators [@Hasan2010], and underpins chiral anomaly and Fermi arcs in Weyl semimetals [@Liu2014; @Xu2015]. In systems with broken time-reversal symmetry the SOC induces the anomalous Hall effect [@Nagaosa2010], and in systems with broken inversion symmetry it leads to [skyrmions]{} via the Dzyaloshinskii-Moriya interaction [@Fert2017].
These exotic states of quantum matter are typically investigated within a static framework, whereby atoms are considered to be immobile in their crystallographic sites. In this description the manifestation of SOC effects is dictated by the global symmetry of the crystal. However, [recently it has]{} been proposed that time-dependent lattice fluctuations and the resulting symmetry breaking might induce some form of dynamic Rashba-Dresselhaus (dRD) effect [@Monserrat2017a; @Niesner2018]. This notion generated considerable interest in the area of hybrid perovskites [@emd16; @Kepenekian2017; @Monserrat2017a; @Frohna2018; @Niesner2018; @McKechnie2018], since Rashba-Dresselhaus spin-splitting might induce an indirect optical gap [@ztlr15], and possibly explain the extraordinary lifetimes of charge carriers in these compounds [@Stranks2013; @Xing2013]. However, despite intense experimental and theoretical efforts in this area [@Monserrat2017a; @Niesner2018; @emd16; @Kepenekian2017; @Monserrat2017a; @Frohna2018; @Niesner2018; @McKechnie2018; @ztlr15], the possible existence of a dRD effect remains controversial; for example, recent measurements of the circular photogalvanic effect [@Niesner2018] and of second-harmonic-generation rotational anisotropy [@Frohna2018] reached conflicting conclusions. To shed light on this debate, here we develop a rigorous quantum many-body theory of the dynamic version of the Rashba-Dresselhaus [@Rashba1959; @brw15; @Manchon2015; @dres55] effect, and we examine how symmetry dictates the structure of spin-phonon interactions in non-magnetic centrosymmetric crystals.
The Rashba-Dresselhaus effect can be understood as a special case of the spin-orbit coupling Hamiltonian: $$\label{eq.1}
V_{\rm SOC} = -\frac{e \,\hbar}{4m^2 c^2}\, \bE\cdot \bs \times \bp~,$$ where $\hbar$, $e$, $m$, $c$, $\bs$, and $\bp$ denote the Planck constant, electron charge and mass, speed of light, Pauli vector, and electron momentum, respectively. $\bE$ is the electrostatic field experienced by the electrons. In the traditional Rashba effect one considers a uniform electric field along the Cartesian direction $\hat{z}$ (Fig. \[fig1\]A), and a parabolic electron band with minimum at the Brillouin-zone center, so that the coupling term $V_{\rm SOC}$ is proportional to $\sigma_x k_y \!-\!\sigma_y k_x$, with $\bk$ denoting the Bloch wavevector of the electron. The resulting band structure and spin texture for the Rashba and Dresselhaus effects are shown in Fig. \[fig1\]B and Fig. \[fig1\]C, respectively. This situation is also representative of polar crystals, which can sustain a finite macroscopic electric field [@Rashba1959].
We now focus on centrosymmetric non-magnetic crystals, for which the Rashba-Dresselhaus effect is symmetry-forbidden. In this case the electric field $\bE$ of Eq. is replaced by the gradient of the single-particle potential energy, $ \bE = (1/e)\nabla V$. In density-functional theory calculations $V$ corresponds to the Kohn-Sham potential, and depends parametrically on the atomic coordinates. [Intuitively one]{} might think that, as the atoms fluctuate around their equilibrium sites, at any given time the crystal is in a broken-symmetry state, therefore it should exhibit a non-vanishing Rashba-Dresselhaus spin splitting induced by the [ fluctuations of the lattice]{}. This is precisely what is observed in Born-Oppenheimer molecular dynamics simulations [@emd16]. Albeit intuitive, this picture is incomplete since the quantum nature of lattice vibrations is not taken into account. In the following we overcome this limitation by [developing a quantum many-body theory of this effect]{}.
In second-quantized notation, the phonon-induced spin-orbit coupling Hamiltonian reads: $$\label{eq.2}
\Vtot = N^{-\frac{1}{2}} \sum_{\,\bk, \bq,s,s'}\,\, g_{s' s}
(\bk,\bq)\, \hcd_{\bk+\bq s'}\hc_{\bk s}(\ha_{\bq}+\had_{-\bq})~,$$ where $\bk$ and $\bq$ are electron and phonon wavevectors belonging to a uniform Brillouin zone grid with $N$ points, $\hcd_{\bk s}$/$\hc_{\bk s}$ and $\had_{\bq}$/$\ha_{\bq}$ are creation/annihilation operators for electrons and phonons, respectively, and $g_{s' s}(\bk,\bq)$ is the spin-phonon coupling matrix element. In this notation the subscript $s=1,2$ denotes the Kramers-degenerate spinors for the same $\bk$. For notational simplicity we omit electron band and phonon branch indices; complete expressions and detailed derivations are provided in the . The matrix elements $g_{s' s}(\bk,\bq)$ are obtained by taking the first-order variation of $V_{\rm SOC}$ in Eq. with respect to [collective displacements]{} of the atoms along each vibrational eigenmode (see ).
In order to establish whether $\Vtot$ can lift the spin degeneracy of the band structure, we consider the electronic state $|\bk s\> = \hcd_{\bk s}|0\>$, obtained by creating one electron in the spinor with wavevector $\bk$ and spin label $s$ from the many-body Fermi vacuum and phonon vacuum $|0\>$. We start by considering zero temperature for simplicity. Since the [spinors]{} $\kone$ and $\ktwo$ are degenerate as a consequence of parity and time-reversal symmetry, in order to determine the perturbed energies we must diagonalize the perturbation matrix $\<\bk s|\Vtot|\bk s'\>$. However, this matrix vanishes identically, for each term of $\Vtot$ either creates or annihilates one phonon, so that $\Vtot|\bk s'\>$ and $|\bk s\>$ differ in their phonon occupations [and must therefore]{} be orthogonal. We deduce that the dynamic Rashba-Dresselhaus effect is forbidden to first order in perturbation theory at zero temperature.
The next step is to check whether the spin degeneracy is lifted at the second order of perturbation theory. In this case the matrix to be diagonalized is (see Sec. C): $$\label{eq.3}
V_{s's} = {\sum_{i}}^\prime \, \< \bk s'| \Vtot| i \> \<i|\Vtot|\bk s\>/(\ve_{\bk}-\ve_i)~,$$ where $\ve_{\bk}$ is the unperturbed energy of the states $\kone$ and $\ktwo$, and $|i\>$ represents a coupled electron-phonon state with energy $\ve_i$. The prime indicates that the states $\kone$ and $\ktwo$ are omitted from the sum. As we show in the Sec. B, the transformation laws of phonons under time-reversal symmetry impose the following constraint on the spin-phonon matrix elements: $$\label{eq.4}
g_{s's}(\bk,\bq) = g_{ss'}^*(\bk+\bq,-\bq)~.$$ Furthermore, the transformation laws of both electrons and phonons under inversion symmetry require: $$\label{eq.5}
g_{s' s}(\bk, \bq) = {-(-1)^{s'-s}}\,
e^{-i\varphi_{\bq}} g_{\bar s' \bar s}^*(\bk, \bq)~,$$ and $$\label{eq.6}
g_{s' s}(\bk,\bq) = - e^{i(\varphi_{-\bq}+\chi_{-\bk s}-\chi_{-\bk-\bq s'})}
g_{s' s}(-\bk,-\bq)~,$$ respectively. Here $\bar s$ is the spin label other than $s$, and $\chi_{\bk s}$ and $\varphi_\bq$ are the phases between a spinor or phonon eigenmode and their inversion partner, respectively. By substituting Eqs. , , and in Eq. we find $V_{11}=V_{22}$ and $V_{12} = V_{21} = 0$. Therefore at zero temperature the dynamic Rashba-Dresselhaus effect is also forbidden to second order in perturbation theory (Fig. \[fig2\]A).
We can generalize these considerations to the situation of a crystal in thermodynamic equilibrium at the temperature $T>0$ as follows. If we replace the phonon vacuum by an excited state with a single phonon, for example $\hat{a}^\dagger_{\bq}|0\rangle$, then the first-order perturbation matrix elements $\<\bk s|\hat{a}_{\bq} \,\Vtot \,\hat{a}^\dagger_{\bq}|\bk s'\>$ vanish as they contain the product of three ladder operators. More generally, for any many-body eigenstate of the unperturbed Hamiltonian, the matrix elements will contain an odd number of ladder operators, hence the first-order correction will also vanish identically. At second order in perturbation theory, the analysis can be performed along similar lines as for the zero-temperature case. For each eigenstate of the unperturbed Hamiltonian we determine the spin splitting using Eq. , and then we perform a canonical average of the corrections to obtain the finite-temperature result. After carrying out the algebra in Fock space (see Sec. D and Fig. S1) we obtain: $$\begin{gathered}
\label{eq.7}
\Delta \ve_{\bk s} = \frac{1}{N}\sum_{\bq\ne 0}
\sum_{s'=1,2} |g_{s' s}(\bk,\bq)|^2
\\ \times
\biggl(\frac{1+n_{\bq}}
{\ve_{\bk}-\ve_{\bk+\bq}-\hbar\w_{\bq}} +\frac{n_{\bq}}
{\ve_{\bk}-\ve_{\bk+\bq}+\hbar\w_{\bq}}\biggr),
\end{gathered}$$ where $\w_\bq$ are the phonon frequencies, and the temperature dependence enters via the Bose-Einstein occupation factors $n_\bq$. For simplicity in this expression we omit vacuum polarization and spin-flip contributions, which are discussed in the Sec. D. Equation constitutes the generalization of the Fan-Migdal self-energy to non-collinear spin systems with SOC [@Giustino2017]. Now the transformation laws in Eq. dictate that the sum $\sum_{s'=1,2} |g_{s' s}(\bk,\bq)|^2$ appearing in Eq. be independent of the spinor label $s$. Therefore the energy shift is the same for both spinors, $\Delta \ve_{\bk\,1} = \Delta \ve_{\bk\,2}$. We deduce that, for a system in equilibrium at finite temperature, dynamic Rashba-Dresselhaus spin-splitting is forbidden also to second order in perturbation theory. Using similar arguments it is possible to further verify that the so-called Debye-Waller self-energy, which corresponds to the second-order electron-phonon Hamiltonian evaluated to first order in perturbation theory [@Giustino2017], does not lead to spin splitting.
One could go on to consider perturbation theory beyond the second order, or generalize the results to non-linear spin-phonon coupling beyond Eq. , however the above conclusions would not change. In fact, it is not difficult to [show]{} that dynamic Rashba-Dresselhaus spin-splitting is forbidden for any many-body [*eigenstate*]{} of the system, irrespective of the formal structure of the electron-phonon Hamiltonian. This is a direct consequence of symmetry: the invariance of the interacting many-body Hamiltonian with respect to parity and time-reversal operations implies Kramers’ degeneracy on all many-body eigenstates (Fig. \[fig2\]B). Since thermal equilibrium is obtained from a canonical average over Kramers-degenerate many-body eigenstates, and this average is a linear operation, we obtain that a system in [*thermal equilibrium*]{} cannot exhibit dynamic Rashba-Dresselhaus spin-splitting, to any order of perturbation. This result is not surprising if we consider that the quantum probability distribution of the atoms in a centrosymmetric system is parity-even in thermal equilibrium, as shown in Fig. \[fig2\]C. In this case the quantum thermal average of the displacement of every atom $\kappa$ vanishes identically, $\< \Delta {\bm \tau}_\kappa \>_T$ = 0. Therefore thermal motion does not break inversion symmetry, and the Rashba-Dresselhaus effect is forbidden. The quantum nature of atomic vibrations plays a central role in this argument.
One exception to the above rule occurs for those [*anharmonic*]{} systems where soft phonons lead to ferroelectric distortions in thermodynamic equilibrium (Fig. \[fig2\]D). In the ferroelectric phase the average displacements are nonzero with respect to the parent centrosymmetric phase, $\< \Delta {\bm \tau}_\kappa \>_T \ne 0$ therefore Eq. and cease to hold. Under these conditions, a dRD splitting from the phonon-assisted SOC in Eq. is allowed to second order in perturbation theory.
These considerations provide us with a clue for reconciling the experimental observations of Ref. and on the hybrid organic-inorganic perovskite methylammonium lead iodide (CH$_3$NH$_3$PbI$_3$, “MAPI”). Below 160 K MAPI crystallizes in a orthorhombic structure with space group $Pnma$ ( Fig. S2A), while between 160 K and 330 K the system is tetragonal with space group $I4/mcm$ ( Fig. S2B). Ref. reports measurements of the circular photogalvanic effect (CPE) in MAPI, whereby spin-polarized carriers can selectively be excited via photons with opposite helicity. These authors observe spin-polarized photocurrents below the absorption edge in the tetragonal phase, but not in the orthorhombic phase. Ref. reports measurements of second-harmonic generation (SHG) for MAPI in the tetragonal structure. The SHG signal is sensitive to symmetry breaking, and the authors find no such breaking in measurements taken over several minutes with a beam spot size of $\sim$10 microns. For the orthorhombic phase, the absence of a bulk CPE is in agreement with our finding that dRD splitting is forbidden in the orthorhombic phase of MAPI, because the space group $Pnma$ has inversion symmetry. For the tetragonal phase, the observation of a bulk CPE signal can be explained by noting that, while the $I4/mcm$ space group is also centrosymmetric, the $C_{3v}$ symmetry of methylammonium does not match the $D_{2d}$ point group symmetry of its Wyckoff site. This mismatch is responsible for the eight-fold orientational disorder of the molecular cation observed in neutron scattering experiments [@Whitfield2016]. As a consequence of orientational disorder, parity is lifted over length scales comparable with the crystal unit cell, therefore spin-splitting is [*allowed*]{}. On the other hand, the structure appears as centrosymmetric to measurements that probe [[*many*]{}]{} crystal unit cells, therefore a SHG signal is not expected, in agreement with Ref. [@Frohna2018]. This picture is consistent with the recent proposal by Ref. that MAPI constitutes a ‘polymorphous’ structure, in the sense that the average crystal structure and space-group symmetry differ from the local symmetry at the scale of the crystal unit cell. This phenomenology is also reminiscent of relaxor ferroelectrics, which consist of ferroelectric nanodomains, but exhibit no macroscopic ferroelectric [polarization [@Li2018]. In relaxor ferroelectrics, even though the macroscopic]{} polarization vanishes, the dielectric response exhibits a broad frequency-dependent maximum which does not follow a Curie-Weiss law, a [ signature]{} of the relaxor structure [@Ahn2016] [and the local symmetry breaking]{}.
It has been estimated that in the $I4/mcm$ structure the MA cations change their orientation over a timescale ranging from 0.5 ps in nuclear magnetic resonance experiments [@Wasylishen1985] to 14 ps in quasielastic neutron scattering measurements [@Leguy2015]. Similar re-orientation times are obtained by first-principles molecular dynamics simulations [@Bokdam2018]. This timescale is much longer than the carrier lifetimes near the band egdes (6-30 fs, [@Schlipf2018]), therefore in the CPE measurements of Ref. the electrons effectively experience a quasi-static crystal potential. This indicates that, rather than a dynamic Rashba-Dresselhaus spin-splitting, the CPE measurements are probing a quasi-static, disorder-induced spin-splitting. [In this situation Eq. and do not hold due to symmetry-breaking, therefore a temperature-dependent dRD splitting from the phonon-assisted SOC is [*allowed*]{}, in line with the experimental observations. However, since the]{} orientational disorder also breaks the translational invariance of the lattice, the very notion of band structure becomes questionable as the Bloch theorem ceases to hold. Hence the observed spin-polarized photocurrent cannot be taken as evidence for the Rashba-like band splitting as in the example of Fig. \[fig1\]B.
We now consider the question on whether the dRD effect is possible in non-polymorphous centrosymmetric crystals, such as for example MAPI in the $Pnma$ phase. Since we established that the dRD is forbidden at thermodynamic equilibrium, we are left to examine the case of a non-thermal, out-of-equilibrium phonon population. To this end we consider a quantum state which is [*not*]{} an eigenstate of the unperturbed Hamiltonian. Being this state [*non-stationary*]{}, we focus on a snapshot at a given time $t$. As a representative example of non-equilibrium state we consider a coherent phonon. Coherent phonons are minimum-uncertainty states whose time evolution resembles classical dynamics. In the textbook example of the quantum harmonic oscillator, a coherent state has a Gaussian probability distribution as in the ground state, rigidly translated away from the equilibrium coordinate, see Fig. \[fig2\]E. More general out-of-equilibrium states can be built from superpositions of coherent phonons and do not modify the following discussion.
A coherent state associated with the phonon of frequency $\w_\bq$ can be written as ( Sec. E): $$\label{eq.8}
|\bk s, u(t)\> = \hcd_{\bk s} \,{\rm exp}(-N u/2)
\,{\rm exp}\left( N^{1/2} u^{1/2} \,\had_\bq \right) |0\>~,$$ where $u(t)$ is proportional to the instantaneous fluctuation amplitude, and $u^2$ quantifies the number of excited phonons in this mode inside each unit cell of the crystal. The expectation value of the atomic displacements in this quantum state is $$\label{eq.10}
\< \Delta {\bm \tau}_\kappa \>_t = 2 u(t)\,
[\hbar/2M_\kappa \w_0]^{1/2}\,\be_{\kappa}~,$$ where $M_\kappa$ is the mass of the $\kappa$-th atom, and $\be_{\kappa}$ is the phonon polarization. Equation shows that, unlike in the case of thermodynamic equilibrium, where $\< \Delta {\bm \tau}_\kappa \>_T$ = 0 and parity is preserved, this coherent state may induce average atomic displacements $\< \Delta {\bm \tau}_\kappa \>_t \ne 0$ that break inversion symmetry at the time $t$. Accordingly we expect that dRD may be allowed for this state. A closer inspection reveals indeed that the expectation values of $\Vtot$ taken between two Kramers-degenerate states $|\bk s,u\>$ and $|\bk \bar s,u\>$ is not bound to vanish, since coherent states are eigenvectors of the annihilation operator $\had_\bq$. In the Sec. E and Sec. G we derive the conditions that this coherent state must satisfy in order to observe a nonzero dRD spin splitting. The analysis can be summarized as follows: (i) A nonzero spin splitting is allowed only for zone-center phonons, $\bq=0$; (ii) Among zone-center phonons, only odd-parity modes can induce spin splitting; (iii) While allowed by symmetry, the spin-splitting induced by polar longitudinal-optical phonons is negligibly small. Taken together, these findings indicate that only zone-center, odd-parity transverse-optical phonons can induce an observable dRD effect in the framework of coherent oscillations. These modes are precisely the [*infrared active*]{} (IR) phonons in polar crystals.
To quantify the magnitude of the dRD effect for out-of-equilibrium phonons, we diagonalize the many-body Hamiltonian in the presence of the coupling from Eq. (see ). The main result is that, as a consequence of Eq. , there exists a spin splitting which increases linearly with $\bk$ near the zone center, and which vanishes identically at $\bk=0$. This is precisely the hallmark of the Rashba effect encoded in Eq. . The resulting electronic bands are shown in Fig. \[fig1\]D. The Rashba energy, defined as the separation between the crossing point and the band minima (Fig. \[fig1\]D), takes the form (details in the Sec. F): $$\label{eq.9}
E_{\rm R}(t) = \frac{m^* u^2(t)}{\hbar^2}
\sum_{s,s'}\left| \frac{\D }{\D \hat{k}}g_{s' s}(\bk, 0)\right|^2~,$$ where $m^*$ is the band mass and $\D/\D \hat{k}$ is the directional derivative $\bk/|\bk|\cdot\nabla_\bk$ evaluated at $\bk=0$. In the case of driven oscillations of the type $u(t) =A \cos(\w_0 t)$, the Rashba energy will fluctuate with a a period $\pi/\w_0$.
{width="80.00000%"}
In order to illustrate the concept out-of-equilibrium dRD, we consider MAPI in the orthorhombic phase. This system admits twenty IR-active optical phonons associated with the deformation of the PbI$_6$ octahedra. In Fig. \[fig3\][A and C]{} we show the Rashba energy from Eq. for these modes, by plotting $E_{\rm R}/u^2$ [for the valence and conduction band extrema, respectively.]{} The largest spin splitting for the conduction bands is found for the $B_{2{\rm u}}$ mode at 3.2 meV, which corresponds to the Pb-I-Pb rocking vibration (Fig. \[fig3\][C]{}). The instantaneous band structure and the associated spin texture are shown in Fig. \[fig3\][B]{}. It is evident that the Rashba-Dresselhaus phenomenology of Fig. \[fig1\]B is fully reproduced. In this case the spin-splitting is found in the reciprocal-space plane that is perpendicular to the polarization vector of the IR phonon. For the valence bands the largest splitting is found with the $B_{3\rm u}$ mode at 7.7 meV (Fig. \[fig3\]C), which corresponds to the Pb-I stretching vibration.
Is this spin-splitting measurable in ultrafast experiments? In Ref. it was shown that the electron lifetimes near the conduction band edges of MAPI [are]{} $\sim$10 fs at 160 K. Given the period of the coherent oscillation [along the $B_{2{\rm u}}$ mode]{} is $1.3$ ps, [during this oscillation the]{} electrons effectively experience a quasi-static potential, [hence]{} spin-split bands should be measurable. The maximum splitting is limited by how much energy the crystal can [absorb]{} upon thermalization of the coherent phonon. If the energy of the coherent state [ were]{} converted entirely into thermal energy, then the system would reach a temperature $T$ given by $(3/2) N_{\rm at} k_{\rm B} T = N_{\rm ph}\hbar\w_0$, where $N_{\rm at}$ is the number of atoms per unit cell, $k_{\rm B}$ the Boltzmann constant, and $N_{\rm ph}=u^2$ is the number of phonons excited per unit cell. For MAPI to remain in the orthorhombic phase ($T<160$ K), we find that the upper bound on the phonon number is $N_{\rm ph} = 100$, having assumed that only the inorganic scaffold thermalizes ($N_{\rm at} = 16$). In Fig. \[fig3\]D we show that, even for considerably smaller vibrational amplitudes, Rashba energies in excess of $30$ meV should be within reach. Energy splittings of this magnitude are well within the resolution of modern time- and angle-resolved photoelectron spectroscopy [@Zhou2018].
Coherent states also offer a key to rationalize the dRD observed in [*ab initio*]{} molecular dynamics simulations of MAPI [@emd16]. In these simulations the atomic nuclei are described as classical particles, therefore each instantaneous configuration is very similar to a superposition of quasiclassical coherent states. Being broken-symmetry states as shown in Fig. \[fig2\]F, these configurations exhibits a RD effect when taken individually. However, to obtain thermodynamic averages it is necessary to perform time-averages over long simulation trajectories, as dictated by the ergodic theorem. For well-converged simulations these averages cannot exhibit spin-splitting due to parity and time-reversal symmetry, unless the system exhibit polymorphous behavior [@Zunger2019]; [but]{} in the latter case the simulations are probing quasi-static spin-splitting as opposed to the dynamic Rashba-Dresselhaus effect.
The use of light to induce dynamic symmetry breaking is rapidly emerging as a new frontier in research on quantum matter. For example light-matter coupling has been used to induce metal-insulator transitions [@Cavalleri2004], superconductivity [@Fausti2011], and symmetry-breaking in charge-transfer crystals [@Collet2003] and perovskites [@Nozawa2005]. [Furthermore]{} it has been proposed [@Radaelli2018] that metastable crystal structures with long-range ferroelectric or ferromagnetic order can be stabilized using nonlinear optical techniques [@Siders2003; @Foerst2011]. Based on our findings, we propose to revisit the structures for which a dynamic ferroelectic order has been induced [@Collet2003; @Nozawa2005], for such an order must be accompanied by a non-equilibrium dRD effect. More generally, we propose that it should be possible to look for phonon-induced Rashba-Dresselhaus physics among materials with strong IR activity. In this context, our dynamic Rashba energy in Eq. could serve as a figure of merit to systematically screen such a broad [library of materials]{}.
In conclusion, using [an [*ab initio*]{}]{} quantum many-body formalism we showed that (i) dynamic, phonon-induced Rashba-Dresselhaus spin-splitting is forbidden in centrosymmetric and non-magnetic crystals at equilibrium; (ii) this effect is allowed in the case of polymorphous crystals but it relates to quasi-static structural fluctuations and not to phonons; (iii) a spin-splitting can be induced and controlled by driving a crystal out-of-equilibrium via coherently-excited infrared-active optical phonons. These findings provide a unified explanation for seemingly inconsistent theoretical and experimental studies on hybrid perovskites, and open the way to the rational design of photo-induced dynamic Rashba-Dresselhaus [spin]{} physics.
Work by M.S. was supported by the Leverhulme Trust under award RL-2012-001. Work by F.G. was supported by the U.S. Department of Energy (DOE), Office of Science, Basic Energy Sciences (BES) under Award DE-SC0020129.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Using face algebras (i.e. algebras of L-operators of IRF models), we construct modular tensor categories with positive definite inner product, whose fusion rules and S-matrices are the same as (or slightly different from) those obtained by $U_q (\frak{sl}_{N})$ at roots of unity. Also we obtain state-sums of ABF models on framed links which give quantum $SU(2)$-invariants of corresponding 3-manifolds.'
author:
- Takahiro Hayashi
date: |
$\qquad$\
Graduate School of Mathematics, Nagoya University,\
$\quad\;$ Furo-cho, Chikusa-ku, Nagoya 464, Japan
title: 'Face algebras and unitarity of $\text{SU(N)}_{\text{L}}$-TQFT'
---
=0truecm =0truecm \[section\] \[thm\][Corollary]{} \[thm\][Proposition]{} \[thm\][Lemma]{} \[section\] \[section\]
introduction
============
As is well known, quantum groups have their origin in the theory of quantum inverse scattering method. More specifically, they first appeared as so-called algebras of L-operators of lattice models (of vertex type). For example, the simplest quantum group $U_q(\frak{sl}(2))$ can be viewed as the algebra of L-operators of 6-vertex model without spectral parameter. It seems that it is worth trying to study algebras of L-operators independently from the framework of Drinfeld-Jimbo algebra.
By investigating algebraic structure of lattice models of face type, we found a new class of quantum groups, which is called the class of [*face algebras*]{} (cf. [@subf]-[@cpcq] and also [@BS; @JurcoSchupp; @Schauenburgface]). It contains all bialgebras as a subclass. Moreover, as well as bialgebras, face algebras produce monoidal categories as their (co-)module categories.
In this paper, we give a detailed study of face algebras ${{\frak S}(A_{N-1};t)_{\epsilon}}$, which are obtained as algebras of L-operators of RSOS models of type $A_{N-1}$ $(N \geq 2)$ (cf. [@JMO]), where $\epsilon = \pm 1$ and $t$ denotes a primitive $2(N + L)$-th root of unity with $L \geq 1$. We also give two applications of ${{\frak S}(A_{N-1};t)_{\epsilon}}$ to 3-dimensional topological quantum field theory (TQFT) and corresponding quantum invariants of 3-manifolds.
We show that the algebra ${{\frak S}(A_{N-1};t)_{\epsilon}}$ is finite-dimensional cosemisimple and that its dual is a $C^*$-algebra for a suitable $t$. Also, we classify irreducible comodules of ${{\frak S}(A_{N-1};t)_{\epsilon}}$ and determine their dimensions. Moreover, we construct various structures on ${{\frak S}(A_{N-1};t)_{\epsilon}}$, such as the antipode, the braiding and the ribbon functional. The algebra ${{\frak S}(A_{N-1};t)_{\epsilon}}$ is constructed as a quotient of the face version ${{\frak A}(w_{N,t,\epsilon})}$ of FRT construction modulo one additional relation “$\det = 1$”, where $w_{N,t,\epsilon}$ is the Boltzmann weight of RSOS models of type $A_{N-1}$ without the spectral parameter and $\det$ denotes the “(quantum) determinant” of ${{\frak A}(w_{N,t,\epsilon})}$. Since the representation theory of ${{\frak A}(w_{N,t,\epsilon})}$ is relatively easily established using a result on Iwahori-Hecke algebras due to H. Wenzl, the core of our work is to study the properties of the element $\det$ or, corresponding “exterior” algebra.
Next, we explain the unitarity of 3-dimensional TQFT briefly. Roughly speaking, a 3-dimensional TQFT is a map which assigns to each 3-cobordism $(M, \partial_- M,$ $\partial_+ M)$, a linear map $\tau (M)\!: \EuScript{T}(\partial_- M) \to \EuScript{T}(\partial_+ M)$. Here, by a 3-dimensional cobordism, we mean a compact 3-dimensional manifold $M$ whose boundary is a disjoint union of two closed surfaces $\partial_- M, \partial_+ M$. A 3-dimensional TQFT is called unitary if $\EuScript{T}(\partial_{\pm} M)$ are (finite-dimensional) Hilbert spaces and $\tau (-M) = \tau (M)^*$ for each 3-cobordism $(M, \partial_- M, \partial_+ M)$. It is established in [@Turaev] that to obtain a (unitary) 3-dimensional MTC, it suffices to construct a (unitary) modular tensor category (MTC) (i.e. a braided category which satisfies certain additional properties). The most important examples of MTC are constructed as semisimple quotients $\cal{C} (\frak{g}, q)$ of some module categories of Drinfeld-Jimbo algebra $U_q (\frak{g})$ at roots $q$ of unity (cf. [@ReshetikhinTuraev; @Andersen; @GelfandKazhdan; @Kirillov], and see also [@Finkelberg; @TuraevWenzl; @TuraevWenzl2] for other construction of MTC’s). For a suitable $q$, it also is expected that $\cal{C} (\frak{g}, q)$ is a unitary MTC. However, it seems that it is not easy to verify it directly, since $U_q (\frak{g})$ is non-semisimple and cannot have a $C^*$-algebra structure (cf. A. Kirillov, Jr, [@Kirillov] and V. Tuaev and H. Wenzl [@TuraevWenzl2]).
The first application of ${{\frak S}(A_{N-1};t)_{\epsilon}}$ is to show that the category $\cal{C}_{\frak{S}}(A_{N-1},t)_{\epsilon}$ of all finite-dimensional right ${{\frak S}(A_{N-1};t)_{\epsilon}}$-comodules is a MTC, and that (for a suitable $t$), the category $\cal{C}^u_{\frak{S}}(A_{N-1},t)_{\epsilon}$ of all finite-dimensional unitary ${{\frak S}(A_{N-1};t)_{\epsilon}}$-comodules is a unitary MTC, whose fusion rules agree with those of $\cal{C} (\frak{sl}_N, q)$. Here “unitary” comodule means a comodule with inner product which satisfies some conditions. The quantum dimensions and $S$-matrices of $\cal{C}_{\frak{S}}(A_{N-1},t)_{\epsilon}$ are the same as (or slightly different from) those of $\cal{C} (\frak{sl}_N, q)$ (according to the choice of $\epsilon$ and another sign parameter $\iota$ when $N$ is even). Although we use $U_q (\frak{sl}_N)$ to obtain some combinatorial formulas, the essential part of our theory is independent from $U_q (\frak{sl}_N)$. Hence the equivalence of $\cal{C} (\frak{sl}_N, q)$ and $\cal{C}_{\frak{S}}(A_{N-1},t)_{\epsilon}$ is left as an open problem. However, by the result of Kazhdan and Wenzl [@KazhdanWenzl], these two categories are equivalent up to a “twist.”
Unlike the module category of $U_q(\frak{g})$, the category $\cal{C}_{\frak{S}}(A_{N-1},t)_{\epsilon}$ itself is semisimple. Moreover, it has a apparent similarity to the spaces of the conformal blocks of Wess-Zumino-Witten (WZW) models. We hope that there exists a direct connection between ${{\frak S}(A_{N-1};t)_{\epsilon}}$ and WZW models, similarly to Drinfeld [@DrinfeldGal].
The second application is to give an explicit description of the quantum $SU(2)$-invariant $\tau (M)$ of closed 3-manifolds $M$, which is associated with $\cal{C}_{\frak{S}}(A_1,t)_{\epsilon}$. More precisely, we express $\tau (M)$ as a state sum on each generic link diagram $D$ which represents $M$. It gives a direct connection between the invariant and ABF model.
The paper is organized as follows. We start in Sect. 2-4, by recalling elementary properties of face algebras ${\frak H}$, various structures on ${\frak H}$ and their relations to the comodule category of ${\frak H}$. In Sect. 5, we recall the notion of star-triangular (Yang-Baxter) face models and flat face models. The later is introduced in [@gal], and is a variant of Ocneanu’s notion of flat biunitary connection in operator algebra. Flat face models play an crucial role to the determination of the representation theory of ${{\frak S}(A_{N-1};t)_{\epsilon}}$. In Sect. 6-7, we define the algebras ${{\frak S}(A_{N-1};t)_{\epsilon}}$ and state the main result on the representation theory of them. In Sect. 9-10, we construct several structures on ${{\frak S}(A_{N-1};t)_{\epsilon}}$, which we call the transpose, the costar structure, the antipode and the ribbon functional. Consequently, we see that $\cal{C}_{\frak{S}}(A_{N-1},t)_{\epsilon}$ (resp. $\cal{C}^u_{\frak{S}}(A_{N-1},t)_{\epsilon}$) is a (unitary) ribbon category. In Sect. 11, we prove that $\cal{C}_{\frak{S}}(A_{N-1},t)_{\epsilon}$ is a modular tensor category, by computing its $S$-matrix. Section 8 and Section 12 are devoted to some technical calculations on face analogues of the exterior algebras. In Sect. 13-14, we give an explicit description of $\cal{C}_{\frak{S}}(A_1,t)_{\epsilon}$ and give a state sum expression of the quantum $SU(2)$-invariant stated above.
After submitting the manuscript, one of the referee informed me the existence of the following two papers;
H. Wenzl, $C^*$-tensor categories from quantum groups, [*J. Amer. Math. Soc.*]{} [**11**]{} (1988), pp. 261-282.
C. Blanchet, Heck algebras, modular categories and 3-manifolds quantum invariants, preprint.
The former gives a proof of the unitarity of $\cal{C} (\frak{g}, q)$ for each $\frak{g}$. The latter gives a construction of modular tensor categories via Iwahori-Hecke algebras at root of unity.
Throughout this paper, we use Sweedler’s sigma notation for coalgebras, such as $(\Delta \otimes \mathrm{id}) (\Delta (a)) =
\sum_{(a)} a_{(1)} \otimes a_{(2)} \otimes a_{(3)}$ (cf. [@Sweedler]).
face algebras
=============
In this section, we give the definition of the face algebra and various structures on it. Let ${\frak H}$ be an algebra over a field ${\Bbb K}$ equipped with a coalgebra structure $({\frak H},\Delta,{\varepsilon})$. Let ${\EuScript V}$ be a finite non-empty set and let ${e_{\lambda}}$ and ${{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}$ $(\lambda \in {{\EuScript V}})$ be elements of ${{\frak H}}$. We say that $({{\frak H}}, {{e_{\lambda}},{{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}})$ is a [*${{\EuScript V}}$-face algebra*]{} if the following relations are satisfied: $$\Delta (ab) = \Delta(a) \Delta(b),
\label{D(ab)}$$ $${e_{\lambda}}{e_{\mu}}= \delta_{\lambda \mu} {e_{\lambda}},
\quad {{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{{\stackrel{\scriptscriptstyle\circ}{e}_{\mu}}}= \delta_{\lambda \mu} {{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}},
\quad {e_{\lambda}}{{\stackrel{\scriptscriptstyle\circ}{e}_{\mu}}}= {{\stackrel{\scriptscriptstyle\circ}{e}_{\mu}}}{e_{\lambda}},
\label{ee}$$ $$\sum_{\nu \in {{\EuScript V}}} \, {e_{\nu}}= 1 = \sum_{\nu \in {\EuScript V}} {{{\stackrel{\scriptscriptstyle\circ}{e}_{\nu}}}},
\label{sume}$$ $$\Delta({{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}}) = \sum_{\nu \in {\EuScript V}} {{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\nu}}\otimes {{\stackrel{\scriptscriptstyle\circ}{e}_{\nu}}}{e_{\mu}},
\quad
{\varepsilon}({{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}}) = \delta_{\lambda \mu} ,
\label{D(ee)}$$ $${\varepsilon}(ab) = \sum_{\nu \in {\EuScript V}} {\varepsilon}(a{e_{\nu}}) {\varepsilon}({{\stackrel{\scriptscriptstyle\circ}{e}_{\nu}}}b)
\label{e(ab)}$$ for each $a, b \in {\frak H}$ and $\lambda, \mu \in {\EuScript V}$. We call elements ${e_{\lambda}}$ and ${{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}$ [*face idempotents*]{} of ${\frak H}$. It is known that bialgebra is an equivalent notion of ${\EuScript V}$-face algebra with $\mathrm{card}({\EuScript V}) = 1$. For a ${\EuScript V}$-face algebra ${\frak H}$, we have the following formulas: $${\varepsilon}({{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}a) = {\varepsilon}({e_{\lambda}}a),
\quad {\varepsilon}(a {{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}) = {\varepsilon}(a {e_{\lambda}}),
\label{e(eae)}$$ $${\sum_{(a)}}a_{(1)} {\varepsilon}({e_{\lambda}}a_{(2)} {e_{\mu}}) = {e_{\lambda}}a {e_{\mu}},
\label{ae(eae)}$$ $${\sum_{(a)}}{\varepsilon}({e_{\lambda}}a_{(1)} {e_{\mu}})a_{(2)} = {{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}a {{\stackrel{\scriptscriptstyle\circ}{e}_{\mu}}},
\label{e(eae)a}$$ $$\Delta (a) = \sum_{\nu, \xi} {\sum_{(a)}}{e_{\nu}}a_{(1)} {e_{\xi}}\otimes {{\stackrel{\scriptscriptstyle\circ}{e}_{\nu}}}a_{(2)} {{\stackrel{\scriptscriptstyle\circ}{e}_{\xi}}},
\label{D(a)}$$ $${\sum_{(a)}}{e_{\lambda}}a_{(1)} {e_{\mu}}\otimes a_{(2)} =
{\sum_{(a)}}a_{(1)} \otimes {{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}a_{(2)} {{\stackrel{\scriptscriptstyle\circ}{e}_{\mu}}},
\label{eae*a}$$ $$\Delta ({{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}}a {{\stackrel{\scriptscriptstyle\circ}{e}}}_{\lambda'} e_{\mu'}) =
{\sum_{(a)}}{{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}a_{(1)} {{\stackrel{\scriptscriptstyle\circ}{e}}}_{\lambda'} \otimes {e_{\mu}}a_{(2)} e_{\mu'}
\label{D(eeaee)}$$ for each $a \in {\frak H}$ and $\lambda, \mu, \lambda', \mu' \in {\EuScript V}$.
Let ${\EuScript G}$ be a finite oriented graph with set of vertices ${{\EuScript V}}$ = ${{{\EuScript G}}^0}$. For an edge ${{\bold p}}$, we denote by ${{\frak {s}}({\bold p})}$ and ${{\frak {r}}({\bold p})}$ its [*source*]{} ([*start*]{}) and its [*range*]{} ([*end*]{}) respectively. For each $m \geq 1$, we denote by ${{{\EuScript G}}^m = {\coprod}_{\lambda, \mu \in {\EuScript V}}{{\EuScript G}}_{\lambda \mu}^m}$ the set of [*paths*]{} of ${\EuScript G}$ of [*length*]{} $m$, that is, ${\bold p}\in {{\EuScript G}}_{\lambda \mu}^m$ if ${\bold p}$ is a sequence $({\bold p}_1, \ldots, {\bold p}_m)$ of edges of ${\EuScript G}$ such that ${\frak {s}}({\bold p}):= {\frak {s}}({\bold p}_1) = \lambda$, ${\frak {r}}({\bold p}_n) = {\frak {s}}({\bold p}_{n+1})\,$ $(1 \leq n < m)$ and ${\frak {r}}({\bold p}):= {\frak {r}}({\bold p}_m) = \mu$. Let ${{\frak H}({\EuScript G})}$ be the linear span of the symbols ${e{{\bold p}\choose {\bold q}}}$ $ ( {\bold p},{\bold q}\in {{\EuScript G}}^m, m \geq 0) $. Then ${{\frak H}({\EuScript G})}$ becomes a ${\EuScript V}$-face algebra by setting $${{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}= \sum_{\mu \in {\EuScript V}}e{\lambda \choose \mu},
\quad {e_{\mu}}= \sum_{\lambda \in {\EuScript V}}e{\lambda \choose \mu},
\label{eHGi}$$ $$e{{\bold p}\choose {\bold q}} e{{\bold r}\choose {\bold s}} =
{\delta}_{{\frak {r}}({\bold p}) {\frak {s}}({\bold r})} \, {\delta}_{{\frak {r}}({\bold q}) {\frak {s}}({\bold s})}
\: e{{\bold p}\cdot{\bold r}\choose {\bold q}\cdot{\bold s}},
\label{epqers}$$ $$\Delta \left( e{{\bold p}\choose {\bold q}} \right) =
\sum_{{\bold t} \in {\EuScript G}^m}
e{{\bold p}\choose {\bold t}} \otimes e{{\bold t} \choose {\bold q}},
\quad
{\varepsilon}\left( e{{\bold p}\choose {\bold q}} \right) = {\delta}_{{\bold p}{\bold q}}
\label{D(epq)}$$ for each ${\bold p},{\bold q}\in {{\EuScript G}}^m$ and $ {\bold r},{\bold s}\in {{\EuScript G}}^n$ $( m,n \geq 0)$. Here for paths ${\bold p}= ({\bold p}_1,\ldots,{\bold p}_m)$ and ${\bold r}= ({\bold r}_1,\ldots,{\bold r}_n)$, we set ${\bold p}\cdot {\bold r}= ({\bold p}_1,\ldots,{\bold p}_m,{\bold r}_1,\ldots,{\bold r}_n)$ if ${\frak {r}}({\bold p}) = {\frak {s}}({\bold r})$. We note that ${\Bbb K}{\EuScript G}^m = \bigoplus_{{\bold p}\in {\EuScript G}^m} {\Bbb K}{\bold p}$ $(m \geq 0)$ becomes a right ${\frak H}({\EuScript G})$-comodule via $${\bold q}\mapsto
\sum_{{\bold p}\in {\EuScript G}^m}
{\bold p}\otimes e{{\bold p}\choose {\bold q}}.$$
Every finitely generated ${\EuScript V}$-face algebra is isomorphic to a quotient of ${{\frak H}({\EuScript G})}$ for some ${\EuScript G}$.
We say that a linear map $S\!: {{\frak H}} \to {{\frak H}}$ is an [*antipode*]{} of ${{\frak H}}$, or $({{\frak H}}, S)$ is a [*Hopf*]{} ${{\EuScript V}}$-[*face algebra*]{} if $${\sum_{(a)}}S(a_{(1)})a_{(2)} = \sum_{\nu \in {\EuScript V}} {\varepsilon}(a {e_{\nu}}) {e_{\nu}},
\quad
{\sum_{(a)}}a_{(1)}S(a_{(2)}) = \sum_{\nu \in {\EuScript V}} {\varepsilon}({e_{\nu}}a){{{\stackrel{\scriptscriptstyle\circ}{e}_{\nu}}}},
\label{S(a)a}$$ $${\sum_{(a)}}S(a_{(1)})a_{(2)}S(a_{(3)}) = S (a)
\label{S(a)aS(a)}$$ for each $a \in {\frak H}$. An antipode of a ${\EuScript V}$-face algebra is an antialgebra-anticoalgebra map, which satisfies $$S({{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}}) = {{\stackrel{\scriptscriptstyle\circ}{e}_{\mu}}}{e_{\lambda}}\quad (\lambda, \mu \in {{\EuScript V}}).
\label{S(ee)}$$ The antipode of a ${\EuScript V}$-face algebra is unique if it exists.
Let ${{\frak H}}$ be a ${{\EuScript V}}$-face algebra with product $m$ and let ${{\cal R}^+}$ and ${{\cal R}^-}$ be elements of $({{\frak H}} \otimes{{\frak H}})^*$. We say that $({{\frak H}},{{\cal R}^{\pm}})$ is a [*coquasitriangular*]{} (or [*CQT*]{}) ${{\EuScript V}}$-face algebra if the following relations are satisfied $${{\cal R}^+}m^*(1) = {{\cal R}^+},
\quad
m^*(1) {{\cal R}^-}= {{\cal R}^-},
\label{Rpm(1)}$$ $${{\cal R}^-}{{\cal R}^+}= m^*(1),
\quad
{{\cal R}^+}{{\cal R}^-}= (m^{\text{op}})^*(1),
\label{RmRp}$$ $$\quad
{{\cal R}^+}m^*(X) {{\cal R}^-}= (m^{\text{op}})^*(X)
\quad (X \in {\frak H}^*),
\label{RmXR}$$ $$\label{mid(R)}
(m {\otimes}\text{id})^* ({{\cal R}^+}) = {{\cal R}^+}_{13} {{\cal R}^+}_{23},
\quad
(\text{id}{\otimes}m)^* ({{\cal R}^+}) = {{\cal R}^+}_{13} {{\cal R}^+}_{12}.$$ Here, as usual, for each $Z \in ({\frak H}\otimes {\frak H})^*$ and $\{ i,j,k \} = \{1,2,3 \}$, we define $Z_{ij} \in ({\frak H}^{\otimes 3})^*$ by $Z_{ij} (a_1, a_2, a_3) = Z(a_i, a_j) {\varepsilon}(a_k)$ $(a_1, a_2, a_3 \in {\frak H})$. We note, for example, that the second formula of is equivalent to $$\label{Rp(a,bc)}
{{\cal R}^+}(a,\, bc) = {\sum_{(a)}}{{\cal R}^+}(a_{(1)},\, c) {{\cal R}^+}(a_{(2)},\, b)
\quad (a, b, c \in {\frak H}).$$ It is known that ${{\cal R}^{\pm}}$ satisfies $$\label{mid(Rm)}
(m {\otimes}\text{id})^* ({{\cal R}^-}) = {{\cal R}^-}_{23} {{\cal R}^-}_{13},
\quad
(\text{id}{\otimes}m)^* ({{\cal R}^-}) = {{\cal R}^-}_{12} {{\cal R}^-}_{13},$$ $$\label{Rpeeee}
{{\cal R}^+}({{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}}a {{\stackrel{\scriptscriptstyle\circ}{e}_{\nu}}}{e_{\xi}},\, b) = {{\cal R}^+}(a,\, {{\stackrel{\scriptscriptstyle\circ}{e}_{\mu}}}{e_{\xi}}b {{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\nu}}),$$ $$\label{Rmeeee}
{{\cal R}^-}({{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}}a {{\stackrel{\scriptscriptstyle\circ}{e}_{\nu}}}{e_{\xi}},\, b) = {{\cal R}^-}(a,\, {{\stackrel{\scriptscriptstyle\circ}{e}_{\nu}}}{e_{\lambda}}b {{\stackrel{\scriptscriptstyle\circ}{e}_{\xi}}}{e_{\mu}}),$$ $$\label{Rpeea}
{{\cal R}^+}({{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}},\, a) = {\varepsilon}({e_{\mu}}a {e_{\lambda}}),
\quad
{{\cal R}^+}(a,\, {{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}}) = {\varepsilon}({e_{\lambda}}a {e_{\mu}}),$$ $$\label{Rmeea}
{{\cal R}^-}({{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}},\, a) = {\varepsilon}({e_{\lambda}}a {e_{\mu}}),
\quad
{{\cal R}^-}(a,\, {{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}}) = {\varepsilon}({e_{\mu}}a {e_{\lambda}})$$ for each $\lambda, \mu, \nu, \xi \in {\EuScript V}$, $a,b \in {\frak H}$. If, in addition, ${\frak H}$ has an antipode $S$, then $$\label{Sid(R)}
(S \otimes \text{id})^* ({{\cal R}^+}) = {{\cal R}^-},
\quad
(\text{id} \otimes S)^* ({{\cal R}^-}) = {{\cal R}^+}.$$ For a CQT Hopf ${{\EuScript V}}$-face algebra ${{\frak H}}$, we define its [*Drinfeld functionals*]{} ${\cal U}_{i} \in {\frak H}^*$ $(i = 1,2)$ via $$\label{Udef}
{\cal U}_1 (a) = {\sum_{(a)}}{{\cal R}^+}(a_{(2)}, S(a_{(1)})),
\quad
{\cal U}_2 (a) = {\sum_{(a)}}{{\cal R}^-}(S(a_{(1)}), a_{(2)})
\quad (a \in {\frak H}).$$ The Drinfeld functionals are invertible as elements of the dual algebra ${{\frak H}}^*$ and satisfy the following relations [@gsg]: $$\label{Um}
{\cal U}_1^{-1}(a) = {\sum_{(a)}}{{\cal R}^-}(S(a_{(2)}),a_{(1)}),
\quad
{\cal U}_2^{-1}(a) = {\sum_{(a)}}{{\cal R}^+}(a_{(1)},S(a_{(2)})),$$ $$\label{U(ee)}
{\cal U}_{i}^{\pm 1}({{{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}} {e_{\mu}}) = {\delta}_{\lambda \mu},
\quad
{\cal U}_{i}^{\pm}({{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}a {{\stackrel{\scriptscriptstyle\circ}{e}_{\mu}}}) = {\cal U}_{i}^{\pm}({e_{\lambda}}a {e_{\mu}}),$$ $$\begin{gathered}
\label{m*(U)}
m^*({\cal U}_1) = {{\cal R}^-}{{\cal R}^-}_{21} ({\cal U}_1 \otimes {\cal U}_1)
= ({\cal U}_1 \otimes {\cal U}_1) {{\cal R}^-}{{\cal R}^-}_{21}, \\
m^*({\cal U}_2)
= {{\cal R}^+}_{21} {{\cal R}^+}({\cal U}_2 \otimes {\cal U}_2)
= ({\cal U}_2 \otimes {\cal U}_2 ) {{\cal R}^+}_{21} {{\cal R}^+},
\nonumber\end{gathered}$$ $$\label{UXU-}
{\cal U}_i X {\cal U}_i^{-1} =(S^2)^*(X),$$ $$\label{S(U)}
S^*({\cal U}_1^{\pm 1}) = {\cal U}_2^{\mp 1},
\quad
S^*({\cal U}_2^{\pm 1}) = {\cal U}_1^{\mp 1}$$ for each $i = 1,2$, $a \in {\frak H}$, $X \in {\frak H}^*$ and $\lambda, \mu \in {\EuScript V}$. In particular, $S$ is bijective and ${\cal U}_1{\cal U}_2^{-1}$ is a central element of ${{\frak H}}^*$. Let $({{\frak H}},{{\cal R}^{\pm}})$ be a CQT Hopf ${\EuScript V}$-face algebra and $\cal{V}$ an invertible central element of ${{\frak H}}^*$. We say that $\cal{V}$ is a [*ribbon functional*]{} of ${\frak H}$, or $({\frak H},\cal{V})$ is a [*coribbon Hopf*]{} ${\EuScript V}$-[*face algebra*]{} if $$m^*({\cal V}) = {{\cal R}^-}{{\cal R}^-}_{21} ({\cal V} \otimes {\cal V}),
\label{m(V)}$$ $$S^* (\cal{V}) = \cal{V}.
\label{S(V)}$$ Let $g$ (resp. ${\cal G}$) be an element of (resp. a linear functional on) a ${\EuScript V}$-face algebra ${{\frak H}}$. We say that $g$ (resp. ${\cal G}$) is [*group-like*]{} if the following relations - (resp. -) are satisfied: $$\label{D(g)}
\Delta (g) = \sum_{\nu \in {{\EuScript V}}} g {e_{\nu}}\otimes g {{\stackrel{\scriptscriptstyle\circ}{e}_{\nu}}},$$ $$\label{gee}
g {{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}}= {{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}}g,
\quad
{\varepsilon}(g {{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}}) = \delta_{\lambda \mu},$$ $$\label{G(ab)}
{\cal G}(ab) = \sum_{\nu \in {{\EuScript V}}}{\cal G} (a {e_{\nu}}) {\cal G}({{{\stackrel{\scriptscriptstyle\circ}{e}_{\nu}}}} b),$$ $$\label{G(eae)}
{\cal G}({{{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}}a{{{\stackrel{\scriptscriptstyle\circ}{e}_{\mu}}}}) = {\cal G}({e_{\lambda}}a {e_{\mu}}),
\quad
{\cal G}({{{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}}{e_{\mu}}) = \delta_{\lambda \mu}$$ for each $a,b \in {\frak H}$ and $\lambda, \mu \in {{\EuScript V}}$. If ${\frak H}$ has an antipode $S$, then every group-like element $g$ and group-like functional ${\cal G}$ are invertible and satisfy $$S(g) = g^{-1},
\quad
S^*({\cal G}) = {\cal G}^{-1}.$$ We denote by $\mathrm{GLE} ({\frak H})$ the set of all group-like elements of ${\frak H}$.
Let ${{\frak H}}$ be a CQT Hopf ${\EuScript V}$-face algebra and ${\cal V}$ an invertible element of ${\frak H}^*$. Then $({\frak H},\cal{V})$ is a coribbon Hopf ${\EuScript V}$-face algebra if and only if $\cal{M} = \cal{U}_1 \cal{V}^{-1}$ is group-like and satisfies the following relations: $$\label{MXM-}
\qquad
{\cal M} X {\cal M}^{-1} =
(S^2)^* (X)
\quad (X \in {\frak H}^*),$$ $$\label{M2}
\cal{M}^2 = \cal{U}_1 \cal{U}_2. \qquad$$
For a coribbon Hopf ${\EuScript V}$-face algebra $({\frak H}, \cal{V})$, we call $\cal{M} = \cal{U}_1 \cal{V}^{-1}$ the [*modified ribbon functional*]{} of ${\frak H}$ corresponding to $\cal{V}$.
\[L=1\] When $L =1$, the algebra ${{\frak S}(A_{N-1};t)_{\epsilon}}$ is rather degenerate. Hence, we treat this case here separately from the case $L \geq 2$. (cf. [@wzw]). Let $N \geq 2$ be an integer and ${\EuScript V}$ the cyclic group ${\Bbb Z}/ N {\Bbb Z}$. Let $t \in {\Bbb C}$ be a primitive $2(N + 1)$-th root of unity and $\epsilon$ either $1$ or $-1$. We define the ${\EuScript V}$-face algebra ${{\frak S}(A_{N-1};t)_{\epsilon}}$ to be the ${\Bbb C}$-linear span of the symbols $e^i_j (m)$ $(i,j,m \in {\EuScript V})$ equipped with the structure of ${\EuScript V}$-face algebra given by $$\label{eijpeklq}
e^i_j [p] e^k_l [q] =
\delta_{i+p, k} \delta_{j+q, l} e^i_j [p+q]
\quad (p, q \in {\Bbb Z}_{\geq 0}),$$ $$ \Delta( e^i_k (m) ) =
\sum_j e^i_j (m) \otimes e^j_k (m),
\quad
{\varepsilon}(e^i_j (m)) = \delta_{ij},$$ $$ {{\stackrel{\scriptscriptstyle\circ}{e}}}_i = \sum_j e^i_j (0),
\quad
e_j = \sum_i e^i_j (0).$$ Here, in , we set $$e^i_j [qN + r] = (- \epsilon)^{(i-j)q(N-1)}
e^i_j (r + N {\Bbb Z})$$ for each $q \in {\Bbb Z}$ and $0 \leq r < N$. The algebra ${{\frak S}(A_{N-1};t)_{\epsilon}}$ becomes a coribbon Hopf ${\EuScript V}$-face algebra via $$ S (e^i_j [p]) = e^{j+p}_{i+p} [-p],$$ $$ {{\cal R}^+}(e^i_j [p], e^k_l [q]) =
\delta_{i, k+q} \delta_{jk}
\delta_{i+p, l+q} \delta_{j+p, l}
(- \zeta t)^{- pq},$$ $$ \cal{V}_{\iota} (e^i_j [p]) =
\delta_{ij} \iota^p (- \zeta t)^{p^2},$$ where $i,j,k,l \in {\EuScript V}$, $p, q \in {\Bbb Z}_{\geq 0}$, $\zeta$ denotes a solution of $\zeta^N = \epsilon^{N-1} t$, $\iota = \pm 1$ if $N \in 2 {\Bbb Z}$ and $\iota = 1$ if $N \in 1 + 2 {\Bbb Z}$.
comodules of face algebras
==========================
In this section, we recall categorical properties of comodules of face algebras (cf. [@fa]). We refer the readers to [@Kassel] for the terminologies on monoidal (or tensor) categories. Let $M$ be a right comodule of a ${\EuScript V}$-face algebra ${\frak H}$. We define its [*face space decomposition*]{} $M = \bigoplus_{\lambda, \mu \in {{\EuScript V}}}M(\lambda, \mu)$ by $$M(\lambda, \mu) = \Bigl\{ \sum_{(u)}u_{(0)}
\varepsilon ({e_{\lambda}}u_{(1)} {e_{\mu}})
\Bigm| u \in M \Bigr\}.$$ Here we denote the coaction $M \to M \otimes {\frak H}$ by $u \mapsto \sum\nolimits_{(u)}u_{(0)} \otimes u_{(1)}$ $(u \in M)$. Let $N$ be another ${\frak H}$-comodule. We define the [*truncated tensor product*]{} $M \bar{\otimes} N$ to be the ${\frak H}$-comodule given by $$\label{MNlm}
(M \bar{\otimes} N) (\lambda, \mu) =
{\bigoplus}_{\nu \in {{\EuScript V}}}
M(\lambda, \nu) \otimes N(\nu, \mu),$$ $$\begin{gathered}
\qquad\qquad\qquad
u \otimes v \mapsto
\sum_{(u),(v)} u_{(0)} \otimes v_{(0)} \otimes u_{(1)}v_{(1)} \\
(u \in M(\lambda, \nu), v \in N(\nu, \mu),
\lambda, \mu, \nu \in {{\EuScript V}}).
\qquad\end{gathered}$$ With this operation, the category $\bold{Com}^f_{{\frak H}}$ of all finite-dimensional right ${\frak H}$-comodules becomes a monoidal (or tensor) category whose unit object ${\Bbb K}{\EuScript V}$ is given by $$\label{unitcomdef}
{\Bbb K}{\EuScript V}= \bigoplus_{\mu \in {\EuScript V}} {\Bbb K}\mu;
\quad
\mu \mapsto
\sum_{\lambda \in {\EuScript V}} \lambda \otimes {{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}}\quad (\mu \in {\EuScript V}).$$ If, in addition, ${\frak H}$ is CQT, then $\bold{Com}^f_{{\frak H}}$ becomes a braided monoidal category with braiding $c_{MN}\!: M \bar{\otimes} N
\cong N \bar{\otimes} M$ given by $$\begin{gathered}
\label{cdef}
\qquad\qquad
c_{MN} (u \otimes v)
= \sum_{(u),(v)} v_{(0)} \otimes u_{(0)} {\cal R}^+(u_{(1)},v_{(1)})
\\ (u \in M(\lambda, \nu), v \in N(\nu, \mu),
\lambda, \mu, \nu \in {{\EuScript V}}).
\qquad \end{gathered}$$ Next, suppose that ${\frak H}$ has a bijective antipode and $M$ is finite-dimensional. Then there exists a unique right ${\frak H}$-comodule $M \spcheck$ such that whose underlying vector space is the dual of $M$ and that the coaction satisfies $$\label{leftdualdef}
\sum_{(u)} \langle v, u_{(0)} \rangle S(u_{(1)}) =
\sum_{(v)} \langle v_{(0)}, u \rangle v_{(1)}
\quad (u \in M, v \in M \spcheck).$$ As vector spaces, we have $$\label{Mcheckij}
M \spcheck (\lambda, \mu) \cong M(\mu, \lambda)^*
\quad (\lambda, \mu \in {\EuScript V}).$$ The comodule $M \spcheck$ becomes a left dual object of $M$ via maps $$\label{bMdef}
b_M\!:\,
{\Bbb K}{\EuScript V}\to M \bar{\otimes} M \spcheck ;
\quad
\lambda \mapsto
\sum_{\mu \in {\EuScript V}}
\sum_{{\bold p}\in \EuScript{M}_{\lambda \mu}}
{\bold p}\otimes {\bold p}\spcheck
\quad (\lambda \in {\EuScript V}),$$ $$\label{dMdef}
d_M\!:\,
M \spcheck \bar{\otimes} M \to {\Bbb K}{\EuScript V}\, ;
\quad
{\bold p}\spcheck \otimes {\bold q}\mapsto
\delta_{{\bold p}{\bold q}}\, \mu
\quad
({\bold p}\in \EuScript{M}_{\nu \lambda},
{\bold q}\in \EuScript{M}_{\nu \mu}),$$ where $\EuScript{M}_{\lambda \mu}$ denotes a basis of $M (\lambda, \mu)$ and $\{ {\bold p}\spcheck\, |\, {\bold p}\in
\EuScript{M}_{\lambda \mu} \}$ its dual basis. By replacing $S$ in with $S^{-1}$, we obtain another ${\frak H}$-comodule structure on $M^*$, which gives the right dual $M^{\land}$ of $M$. We note that the canonical linear isomorphism $I_{M^{\land}}\!: M^{\land} \to M^{\lor}$ satisfies $$\label{IXu}
I_{M^{\land}} (X u) =
(S^{-2})^* (X) I_{M^{\land}} (u)$$ for each $X \in {\frak H}^*$ and $u \in M^{\land}$. Here, as usual, we regard $M$ as a left ${\frak H}^*$-module via $$\label{Xu}
X u =
\sum_{(u)} u_{(0)} \langle X, u_{(1)} \rangle
\quad (u \in M, X \in {\frak H}^*).$$ Finally, suppose that ${\frak H}$ is a coribbon Hopf ${\EuScript V}$-face algebra. Then $\bold{Com}^f_{{\frak H}}$ becomes a ribbon category with twist $\theta_M\!: M \cong M$ given by $$\theta_M (u) =
\cal{V}^{-1} u
\quad (u \in M).$$
\[Trq\] Let ${\frak H}$ be a coribbon Hopf ${\EuScript V}$-face algebra with modified ribbon functional $\cal{M}$ such that its unit comodule ${\Bbb K}{\EuScript V}$ is absolutely irreducible. Then, the quantum trace of the ribbon category $\bold{Com}^f_{{\frak H}}$ is given by $$\begin{aligned}
\mathrm{Tr}_q (f) & =
\frac{1}{\mathrm{card} ({\EuScript V})} \mathrm{Tr} (\cal{M} f) \\
& =
\label{Trqf}
\mathrm{Tr} \left( (\cal{M} f) |_{M(\lambda, -)} \right)
\end{aligned}$$ for each $M \in \mathrm{ob} ( \bold{Com}^f_{{\frak H}} )$ and $f \in {\mathrm{End}}_{{\frak H}} (M)$, where $\lambda$ is an arbitrary element of ${\EuScript V}$ and $M ( \lambda, -) = \bigoplus_{\mu} M (\lambda, \mu)$.
We define $u_M\!: M \to M^{\lor \lor}$ by the composition $$\begin{gathered}
\qquad
M \cong M \bar{\otimes} {\Bbb K}{\EuScript V}@>{{\mathrm{id}}\bar{\otimes} b}>>
M \bar{\otimes} M^{\lor} \bar{\otimes} M^{\lor \lor}
@>{c \bar{\otimes} {\mathrm{id}}}>> \\
M^{\lor} \bar{\otimes} M \bar{\otimes} M^{\lor \lor}
@>{d \bar{\otimes} {\mathrm{id}}}>>
{\Bbb K}{\EuScript V}\bar{\otimes} M^{\lor \lor}
\cong M^{\lor \lor}.
\qquad
\end{gathered}$$ Then, as a consequence of the fact that $\bold{Com}_{{\frak H}}^f$ is a rigid braided monoidal category, we have $$\label{dMcMM}
d_M \circ c_{M M^{\lor}} =
d_{M^{\lor}} \circ ( u_M \bar{\otimes} {\mathrm{id}}_{M^{\lor}} ).$$ On the other hand, since $u_M (v) = I_M (\cal{U}_1 v)$ $(v \in M)$, we have $$\label{utheta}
(u_M \circ \theta_M) (v) = I_M (\cal{M} v)
\quad (v \in M),$$ where $I_M$ is as in . Hence, by the definition of ${\mathrm{Tr}}_q$, we obtain $$\begin{aligned}
{\mathrm{Tr}}_q (f) (\lambda) & =
d_{M^{\lor}} \circ ( u_M \theta_M f \bar{\otimes} {\mathrm{id}}_{M^{\lor}} )
\circ b_M (\lambda) \\
= & \sum_{\mu \in {\EuScript V}}
\sum_{{\bold p}\in \EuScript{M}_{\lambda \mu}}
\langle I_M ( \cal{M} f ({\bold p}) ), {\bold p}^{\lor}
\rangle\, \lambda \\
& =
\mathrm{Tr} \left( (\cal{M} f) |_{M(\lambda, -)} \right)
\lambda
\end{aligned}$$ as required, where $\EuScript{M}_{\lambda \mu}$ and ${\bold p}^{\lor}$ are as in .
We say that a right comodule $M$ of a ${\EuScript V}$-face algebra ${\frak H}$ is [*group-like*]{} if $\dim (M(\lambda, \mu)) = \delta_{\lambda \mu}$ for each $\lambda, \mu \in {\EuScript V}$. For each $g \in \mathrm{GLE} ({\frak H})$, we can define a group-like comodule ${\Bbb K}{\EuScript V}g = \bigoplus_{\mu \in {\EuScript V}} {\Bbb K}\mu g$ by the coaction $\mu g \mapsto \sum_{\lambda} \lambda g
\otimes {{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}}g$. Conversely, we have the following.
\[glcom\] For each group-like comodule $M$ and its basis $g_{\mu} \in M(\mu, \mu)$ $(\mu \in {\EuScript V})$, we obtain a group-like element $g$ by ${{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}{e_{\mu}}g = g \binom{\lambda}{\mu}$ and $g_{\mu} \mapsto \sum_{\lambda} g_{\lambda}
\otimes g \binom{\lambda}{\mu}$. Moreover, every group-like element is obtained in this manner.
Let ${\frak H}$ be a ${\EuScript V}$-face algebra and $\{ L_{\psi}\, |\, \psi \in {\boldsymbol\Lambda}\}$ its finite-dimensional comodules such that ${\frak H}\cong \bigoplus_{\psi \in {\boldsymbol\Lambda}}
{\mathrm{End}}( L_{\psi})^*$ as coalgebras. Let $g$ be a central group-like element of ${\frak H}$. We say that $g$ is [*simply reducible*]{} if there exists a set $\overline{{\boldsymbol\Lambda}}$ and a bijection $\varphi\!: \overline{{\boldsymbol\Lambda}} \times {\Bbb Z}_{\geq 0}$ $\cong$ ${\boldsymbol\Lambda}$ such that $L_{\varphi (\lambda, n)}$ $\cong$ ${\Bbb K}{\EuScript V}g^n \bar{\otimes} L_{\varphi (\lambda, 0)}$ for each $\lambda \in \overline{{\boldsymbol\Lambda}}$ and $n \geq 0$.
\[sympredgle\] Let ${\frak H}$, $g$ etc. be as above. Then, the element $g$ is not a zero-divisor of ${\frak H}$. The quotient $\overline{{\frak H}}= {\frak H}/ (g-1)$ is isomorphic to $\bigoplus_{\lambda \in \overline{{\boldsymbol\Lambda}}} {\mathrm{End}}(L_{\lambda})^*$ as coalgebras, where $L_{\lambda}$ is $L_{\varphi (\lambda, 0)}$ viewed as an $\overline{{\frak H}}$-comodule. As $\overline{{\frak H}}$-comodules, we have $L_{\varphi (\lambda, n)} \cong$ $L_{\lambda}$ for each $\lambda \in \overline{{\boldsymbol\Lambda}}$ and $n$.
compact face algebras
=====================
Let ${\frak H}$ be a ${\EuScript V}$-face algebra over the complex number field $\Bbb C$ and $\times\!: {\frak H}\to {\frak H};$ $a \mapsto a^{\times}$ an antilinear map. We say that $({\frak H}, \times)$ is a [*costar*]{} ${\EuScript V}$-[*face algebra*]{} [@cpt] if $$(a^{\times})^{\times} = a,
\quad
(ab)^{\times} = a^{\times} b^{\times},$$ $$\Delta(a^{\times})
= \sum_{(a)} a_{(2)}^{\times}
\otimes a_{(1)}^{\times},$$ $${e_{\lambda}}^{\times} = {{\stackrel{\scriptscriptstyle\circ}{e}_{\lambda}}}$$ for each $\lambda \in {\EuScript V}$ and $a, b \in {\frak H}$. Let $[t_{pq}]_{p,q \in I}$ be a finite-size matrix whose entries are elements of ${\frak H}$. Then $[t_{pq}]$ is called a [*unitary matrix corepresentation*]{} if $\Delta(t_{pq}) = \sum_{r} t_{pr} \otimes t_{rq}$, ${\varepsilon}(t_{pq}) = \delta_{pq}$ and $t_{pq}^{\times} = t_{qp}$. A costar ${\EuScript V}$-face algebra ${\frak H}$ is called [*compact*]{} if ${\frak H}$ is spanned by entries of unitary matrix corepresentations (cf. [@Koornwinder; @cpt]). For each costar ${\EuScript V}$-face algebra ${\frak H}$, its dual ${\frak H}^*$ becomes a $*$-algebra via $$\langle X^*, a \rangle =
\overline{ \langle X, a^{\times} \rangle}
\quad (X \in {\frak H}^*, a \in {\frak H}).$$ When ${\frak H}$ is a Hopf ${\EuScript V}$-face algebra, we also set $a^* = S( a^{\times} )$ and $X^{\times} = S^* (X)^*$ for each $a \in {\frak H}$ and $X \in {\frak H}^*$. These operations satisfy $$\label{a**}
(a^*)^* = a,
\quad
(X^{\times})^{\times} = X
\quad (a \in {\frak H}, X \in {\frak H}^*).$$
For each compact Hopf ${\EuScript V}$-face algebra ${\frak H}$, its [*Woronowicz functional*]{} [@Koornwinder; @cpt] $\cal{Q} \in {\frak H}^*$ is a unique group-like functional such that $$\label{QXQ-}
\qquad
\cal{Q} X \cal{Q}^{-1} =
(S^2)^* (X)
\quad (X \in {\frak H}^*),$$ $$\label{TrQ}
\sum_p \cal{Q} (t_{pp}) =
\sum_p \cal{Q}^{-1} (t_{pp}),$$ and that $[\cal{Q} (t_{pq})]$ is a positive matrix, where $[t_{pq}]$ denotes an arbitrary unitary matrix corepresentation. The functional $\cal{Q}$ also satisfies $$\label{Q*}
\cal{Q}^* = \cal{Q},
\quad
\cal{Q}^{\times} = \cal{Q}^{-1}.$$ Let $({\frak H},\times)$ be a costar ${\EuScript V}$-face algebra and $M$ a finite-dimensional ${\frak H}$-comodule equipped with a Hilbert space structure $(\,|\,)$. We say that $(M,(\,|\,))$ is [*unitary*]{} if $$\qquad
\sum_{(u)} (u_{(0)}\, |\, v)\, u_{(1)} =
\sum_{(v)} (u\, |\, v_{(0)})\, v_{(1)}^{\times}
\quad (u, v\in M).$$ We note that $(M,(\,|\,))$ is unitary if and only if gives a $*$-representation of ${\frak H}^*$. A costar ${\EuScript V}$-face algebra $({\frak H},\times)$ is compact if and only if every finite-dimensional ${\frak H}$-comodule is unitary for some $(\,|\,)$.
Let ${\frak H}$ be a compact ${\EuScript V}$-face algebra. The unit comodule ${\Bbb C}{\EuScript V}$ becomes a unitary comodule via $( \lambda\, |\, \mu ) = \delta_{\lambda \mu}$. For each unitary comodules $M$ and $N$, $M \bar{\otimes} N$ becomes a unitary comodule via $(u \otimes v\, |\, u^{\prime} \otimes v^{\prime} )$ $=$ $(u\, |\, u^{\prime}) (v\, |\, v^{\prime} )$ $(u \in M(\lambda, \nu), v \in N(\nu, \mu),
u^{\prime} \in M(\lambda^{\prime}, \nu^{\prime}),
v^{\prime} \in N(\nu^{\prime}, \mu^{\prime}) )$. If ${\frak H}$ has an antipode, then the left dual $M \spcheck$ becomes a unitary comodule via $$(u\, |\, v) =
( \cal{Q} \Upsilon^{-1} (v)\, |\, \Upsilon^{-1} (u) ),$$ where $\Upsilon\!: M \to M \spcheck$ denotes the antilinear isomorphism defined by $$\langle \Upsilon (u),\, v \rangle
= (v\, |\, u)
\quad (u, v \in M).$$
The proof of Part (1) and Part (2) is straightforward. Using the second equality of , we obtain $(S^2)^* (X^{\times}) =$ $S^* (X^*)$. Using this together with and $$\label{XUu}
X \Upsilon (u) = \Upsilon (X^{\times} u)
\quad (u \in M, X \in {\frak H}^*),$$ we obtain $$\begin{aligned}
(u\, |\, X v)
& =
( S^* (X^*) \cal{Q} \Upsilon^{-1} (v)\, |\, \Upsilon^{-1} (u) ) \\
& =
( \cal{Q} \Upsilon^{-1} (v)\, |\, (X^*)^{\times} \Upsilon^{-1} (u) ) \\
& = (X^* u\, |\, v)
\end{aligned}$$ for each $u, v \in M \spcheck$, as required.
By the proposition above, the category $\bold{Com}_{{\frak H}}^{fu}$ of all finite-dimensional unitary comodules of a compact (Hopf) ${\EuScript V}$-face algebra ${\frak H}$ becomes a (rigid) monoidal category. Let $M$ and $N$ be unitary ${\frak H}$-comodules and $f\!: M \to N$ an ${\frak H}$-comodule map. We define a linear map $\bar{f}\!: N \to M$ by $(\overline{f} (n)\, |\, m) = (n\, |\, f(m))$. We have $$\overline{\overline{f}} = f,
\quad
\overline{f \circ g} =
\overline{g} \circ \overline{f},
\quad
\overline{f \bar{\otimes} g} =
\overline{f} \bar{\otimes} \overline{g},$$ $$\overline{f + g} =
\overline{f} + \overline{g},
\quad
\overline{c f} = \overline{c} \overline{f}
\quad (c \in {\Bbb C}).$$
For each compact Hopf ${\EuScript V}$-face algebra ${\frak H}$ and its comodule $M$, we have $$\label{dMbar}
\overline{d_M} = \left(
(I_{M^{\land}} \circ \cal{Q}) \bar{\otimes} {\mathrm{id}}_M \right)
\circ b_{M^{\land}},$$ $$\label{bMbar}
\overline{b_M} = d_{M^{\land}} \circ
( {\mathrm{id}}_M \bar{\otimes} ( (I_{M^{\land}} \circ \cal{Q})^{-1} ).$$
Using and , we obtain $$\label{q|Qp}
( {\bold q}^{\lor}\, |\, \cal{Q} {\bold p}^{\lor})
=
\overline{ \langle {\bold p}^{\lor},\,
\Upsilon^{-1} ({\bold q}^{\lor}) \rangle},
\quad
( \Upsilon ({\bold q})\, |\, {\bold r}^{\lor})
=
\overline{ \langle {\bold r}^{\lor},\,
\cal{Q} {\bold q}\rangle},$$ where $\{ {\bold p}\}$ and $\{ {\bold p}^{\lor} \}$ is as in . Using the first equality of , we obtain $$\begin{aligned}
\left(
{\bold q}^{\lor} \otimes {\bold r}\, |\,
( (I_{M^{\land}} \circ \cal{Q}) \bar{\otimes} {\mathrm{id}}_M )
\circ b_{M^{\land}} (\lambda) \right)
& =
\sum_{\mu \in {\EuScript V}} \sum_{{\bold p}\in \EuScript{M}_{\mu \lambda}}
\left(
{\bold q}^{\lor} \otimes {\bold r}\, |\,
\cal{Q} {\bold p}^{\lor} \otimes {\bold p}\right)
\\ =
\sum_{\mu \in {\EuScript V}}
\sum_{{\bold p}\in \EuScript{M}_{\mu \lambda}}
\left( {\bold r}\, |\,
\langle {\bold p}^{\lor}, \Upsilon^{-1} ({\bold q}^{\lor})
\rangle {\bold p}\right)
& =
\, \delta_{\lambda {\frak {r}}({\bold q})}\, \delta_{{\bold q}{\bold r}},
\end{aligned}$$ where the first equality follows from . This proves . Similarly, follows from the second equality of .
Let ${\frak H}$ be a costar CQT ${\EuScript V}$-face algebra. We say that ${\frak H}$ is of [*unitary type*]{} if $$\label{R(atbt)}
{{\cal R}^+}(a^{\times}, b^{\times}) =
\overline{{{\cal R}^-}(a, b)}
\quad (a, b \in {\frak H}).$$ In this case, the Drinfeld functionals of ${\frak H}$ satisfy $$\label{U*}
\cal{U}_1^* = \cal{U}_2,
\quad
\cal{U}_2^* = \cal{U}_1.$$
When $q > 0$, the function algebras of the usual quantum groups (such as $\mathrm{Fun}(SL_q (N))$, $\mathrm{Fun}(Sp_q (2N))$) are both CQT and compact. However, they are not of unitary type but rather of “Hermitian type.”
\[Q=M\] For each compact CQT Hopf ${\EuScript V}$-face algebra ${\frak H}$ of unitary type, its Woronowicz functional is a modified ribbon functional of ${\frak H}$. Moreover, the corresponding ribbon functional $\cal{V}_{\cal{Q}}$ satisfies $$\label{V*}
\cal{V}_{\cal{Q}}^{\;*} = \cal{V}_{\cal{Q}}^{-1}.$$
We call $\cal{V}_{\cal{Q}}$ the [*canonical ribbon functional*]{} of $({\frak H}, \times)$. We note that the expression $\cal{U}_1 = \cal{V}_{\cal{Q}} \cal{Q}$ gives the “polar decomposition” of $\cal{U}_1$.
Let ${\frak H}$ be a compact CQT Hopf ${\EuScript V}$-face algebra of unitary type, equipped with the canonical ribbon functional. By and , we have $$\overline{c_{M N}} = (c_{M N})^{-1},
\quad
\overline{\theta_M} = \theta_M^{-1}.$$ Moreover, by , , , we have $$ \overline{b_M} = d_M \circ c_{M M^{\lor}}
\circ
( \theta_M \bar{\otimes} {\mathrm{id}}_{M^{\lor}} ).$$ Furthermore, by and the defining properties of $\cal{Q}$, we have $$ {\mathrm{Tr}}_q (f \overline{f}) =
\frac{1}{{\mathrm{card}}({\EuScript V})}
{\mathrm{Tr}}( \overline{f} \cal{Q} f ) > 0$$ for every $f \ne 0$. Thus, the category $\bold{Com}_{{\frak H}}^{fu}$ is a unitary ribbon category (cf. [@Turaev]).
Let $L=1$ and ${{\frak S}(A_{N-1};t)_{\epsilon}}$ as in Example \[L=1\]. Then ${{\frak S}(A_{N-1};t)_{\epsilon}}$ is a compact CQT Hopf ${\EuScript V}$-face algebra of unitary type with costar structure $e^i_j (m)^{\times} = e^j_i (m)$. Its Woronowicz functional agrees with the counit ${\varepsilon}$.
flat face models
================
Let ${\EuScript G}$ be a finite oriented graph with set of vertices ${\EuScript V}$. We say that a quadruple $\left( {\bold r}\frac[0pt]{{\bold p}}{{\bold q}} {\bold s}\right)$ or a diagram $$\begin{CD}
\lambda @>{{\bold p}}>> \mu \\
@V{{\bold r}}VV @VV{{\bold s}}V \\
\nu @>{{\bold q}}>> \xi \\
\end{CD}
\label{face}$$ is a [*face*]{} if ${\bold p},{\bold q}, {\bold r}, {\bold s}\in {{\EuScript G}}^1$ and $${\frak {s}}({\bold p}) = \lambda = {\frak {s}}({\bold r}),
\quad {\frak {r}}({\bold p}) = \mu = {\frak {s}}({\bold s}),
\quad {\frak {r}}({\bold r}) = \nu = {\frak {s}}({\bold q}),
\quad {\frak {r}}({\bold q}) = \xi = {\frak {r}}({\bold s}).
\label{facecond}$$ When ${\EuScript G}$ has no multiple edge, we also write ${\left( \scriptstyle{{\bold r}}
\textstyle{\frac[0pt]{{\bold p}}{{\bold q}}} \scriptstyle{{\bold s}} \right)} =$ $\left( {\lambda \, \mu}
\atop {\nu \, \xi} \right)$. We say that $({\EuScript G}, w)$ is a (${\EuScript V}$-)[*face model*]{} over a field ${\Bbb K}$ if $w$ is a map which assigns a number $w \!\! \left[ {\bold r}\frac[0pt]{{\bold p}}{{\bold q}} {\bold s}\right]$ $\in {\Bbb K}$ to each face $\left( {\bold r}\frac[0pt]{{\bold p}}{{\bold q}} {\bold s}\right)$ of ${\EuScript G}$. We call $w$ [*Boltzmann weight*]{} of $({\EuScript G}, w)$. For convenience, we set $w \!\! \left[ {\bold r}\frac[0pt]{{\bold p}}{{\bold q}} {\bold s}\right] = 0$ unless $\left( {\bold r}\frac[0pt]{{\bold p}}{{\bold q}} {\bold s}\right)$ is a face. For a face model $({\EuScript G},w)$, we identify $w$ with the linear operator on ${\Bbb K}{\EuScript G}^2$ $=$ $\bigoplus_{{\bold p}\in {\EuScript G}^2} {\Bbb K}{\bold p}$ given by $$w({\bold p}\cdot {\bold q}) =
\sum_{{\bold r}\cdot {\bold s}\in {{\EuScript G}}^2}
w \!\! \left[ {\bold r}\frac[0pt]{{\bold p}}{{\bold s}} {\bold q}\right]
{\bold r}\cdot {\bold s}\quad ({\bold p}\cdot {\bold q}\in {\EuScript G}^2).$$ For $m \geq 2$ and $1 \leq i < m$, we define an operator $w_i = w_{i/m}$ on ${\Bbb K}{\EuScript G}^m$ by $w_{i/m} ( {\bold p}\cdot {\bold q}\cdot {\bold r})$ $=$ ${\bold p}\otimes w ({\bold q}) \otimes {\bold r}$ $({\bold p}\in {\EuScript G}^{i-1}, {\bold q}\in {\EuScript G}^{2}, {\bold r}\in {\EuScript G}^{m-i-1})$, where we identify $({\bold p}_1, \ldots, {\bold p}_m) \in {\EuScript G}^m$ with ${\bold p}_1 \otimes \ldots \otimes {\bold p}_m \in
({\Bbb K}{\EuScript G}^1)^{\otimes m}$.
A face model is called [*invertible*]{} if $w$ is invertible as an operator on ${\Bbb K}{\EuScript G}^2$. An invertible face model is called [*star-triangular*]{} (or [*Yang-Baxter*]{}) if $w$ satisfies the braid relation $w_1 w_2 w_1$ $=$ $w_2 w_1 w_2$ in ${\mathrm{End}}({\Bbb K}{\EuScript G}^3)$.
For a star-triangular face model $({\EuScript G}, w)$, the operators $w_{i/m}$ $(1 \leq i < m)$ define an action of the $m$-string braid group $\frak{B}_m$ on ${\Bbb K}{\EuScript G}^m_{\lambda \mu}$ for each $m \geq 2$ and $\lambda, \mu \in {\EuScript V}$. The following proposition gives a face version of the FRT construction.
\[FRT\] Let $({\EuScript G},w)$ be a ${\EuScript V}$-face model and ${\frak H}({\EuScript G})$ as in $\S$ 1. Let ${\frak I}$ be an ideal of ${\frak H}({\EuScript G})$ generated by the following elements: $$\sum_{{\bold r}\cdot {\bold s}\in {{\EuScript G}}^2}
w \!\! \left[ {\bold r}\frac[0pt]{{\bold p}}{{\bold s}} {\bold q}\right]
e{{\bold a}\cdot {\bold b}\choose {\bold r}\cdot {\bold s}} -
\sum_{{\bold c}\cdot {\bold d}\in {{\EuScript G}}^2}
w \!\! \left[ {\bold a}\frac[0pt]{{\bold c}}{{\bold b}} {\bold d}\right]
e{{\bold c}\cdot {\bold d}\choose {\bold p}\cdot {\bold q}}
\quad ({\bold p}\cdot {\bold q},\, {\bold a}\cdot {\bold b}\in {\EuScript G}^2).$$ Then ${\frak I}$ is a coideal of ${\frak H}({\EuScript G})$ and the quotient ${{\frak A}(w)}:= {\frak H}({\EuScript G}) / {\frak I}$ becomes a ${\EuScript V}$-face algebra. If $({\EuScript G},w)$ is star-triangular, then there exist unique bilinear pairings ${{\cal R}^{\pm}}$ on ${{\frak A}(w)}$ such that $({{\frak A}(w)}, {{\cal R}^{\pm}})$ is a CQT ${\EuScript V}$-face algebra and that $$\label{UnivR=w}
{{\cal R}^+}\left( e{{\bold p}\choose {\bold q}}, e{{\bold r}\choose {\bold s}} \right)
= w \!\! \left[ {\bold r}\frac[0pt]{{\bold q}}{{\bold p}} {\bold s}\right]$$ for each ${\bold p},{\bold q}, {\bold r}, {\bold s}\in {{\EuScript G}}^1$.
We denote the image of $e \binom{{\bold p}}{{\bold q}}$ by the projection ${\frak H}({\EuScript G}) \to {{\frak A}(w)}$ again by $e \binom{{\bold p}}{{\bold q}}$. Then $\frak{A}^m (w)$ $:=$ $\sum_{{\bold p}, {\bold q}\in {\EuScript G}^m} {\Bbb K}e \binom{{\bold p}}{{\bold q}}$ becomes a subcoalgebra of ${{\frak A}(w)}$ for each $m \geq 0$. As the usual FRT construction (cf. [@qcg Proposition 2.1]), we have the following.
\[Aw\*\] For each star-triangular face model $({\EuScript G}, w)$, we have $\frak{A}^m (w)^* \cong {\mathrm{Hom}}_{\frak{B}_m} ({\Bbb K}{\EuScript G}^m)$ $(m \geq 2)$ as ${\Bbb K}$-algebras.
We say that $\left( {\bold r}\frac[0pt]{{\bold p}}{{\bold q}} {\bold s}\right)$ or is a [*boundary condition*]{} of size $m \times n$ if ${\bold p},{\bold q}\in {{\EuScript G}}^n, {\bold r}, {\bold s}\in {{\EuScript G}}^m$ and the relation is satisfied for some $\lambda, \mu, \nu, \xi$. For a face model $({\EuScript G},w)$, we define its [*partition function*]{} to be an extension $w: \{ \text{ boundary conditions of size}
\; m \times n ; \; m,n \geq 1 \} \to {\Bbb K}$ of the map $w$ which is determined by the following two recursion relations:
$$\begin{gathered}
w \!\! \left[ {\bold r}\frac[0pt]{{\bold p}\cdot {{\bold p}}^{\prime}}{{\bold q}\cdot {{\bold q}}^{\prime}} {\bold s}\right]
=
\sum_{{\bold a}\in {{\EuScript G}}^m}
w \!\! \left[ {\bold r}\frac[0pt]{{\bold p}}{{\bold q}} {\bold a}\right]
w \!\! \left[ {\bold a}\frac[0pt]{{{\bold p}}^{\prime}}{{{\bold q}}^{\prime}} {\bold s}\right], \\
w \!\! \left[ {\bold r}\cdot {{\bold r}}^{\prime} \; \frac[0pt]{{\bold p}}{{\bold q}}
\; {\bold s}\cdot {{\bold s}}^{\prime} \right]
=
\sum_{{\bold a}\in {{\EuScript G}}^n}
w \!\! \left[ {\bold r}\frac[0pt]{{\bold p}}{{\bold a}} {\bold s}\right]
w \!\! \left[ {{\bold r}}^{\prime} \frac[0pt]{{\bold a}}{{\bold q}} {{\bold s}}^{\prime} \right] \\
\qquad\qquad\qquad\qquad\qquad\qquad\quad
\left( {\bold p},{\bold q}\in {{\EuScript G}}^n, {{\bold p}}^{\prime},{{\bold q}}^{\prime}
\in {{\EuScript G}}^{n^{\prime}},
{\bold r},{\bold s}\in {{\EuScript G}}^m, {{\bold r}}^{\prime},{{\bold s}}^{\prime}
\in {{\EuScript G}}^{m^{\prime}} \right). \nonumber\end{gathered}$$
Also, we set $w \!\! \left[ {\bold r}\frac[0pt]{{\bold p}}{{\bold q}} {\bold s}\right] = {\delta}_{{\bold p}{\bold q}}$ (respectively $w \!\! \left[ {\bold r}\frac[0pt]{{\bold p}}{{\bold q}} {\bold s}\right] = {\delta}_{{\bold r}{\bold s}}$) if ${\bold r},{\bold s}\in {{\EuScript G}}^0$ (respectively ${\bold p},{\bold q}\in {{\EuScript G}}^0$). With this notation, the relation holds for every star-triangular face model $({\EuScript G},w)$ and ${\bold p},{\bold q}\in {{\EuScript G}}^n, {\bold r}, {\bold s}\in {{\EuScript G}}^m (m,n \geq 0)$. Next, we recall the notion of flat face model [@gal], which is a variant of A. Ocneanu’s notion of flat biunitary connection (cf. [@Ocneanu]). Let $({\EuScript G}, w)$ be an invertible face model with a fixed vertex $* \in {\EuScript V}= {{\EuScript G}}^0$. We assume that ${\boldsymbol\Lambda}_{{\EuScript G}}^m \ne \emptyset$ for each $m \geq 0$ and that ${\EuScript V}= \bigcup_{m \geq 0} {\EuScript V}(m)$, where ${\boldsymbol\Lambda}_{{\EuScript G}}^m$ $=$ ${\boldsymbol\Lambda}_{{\EuScript G}*}^m$ and ${\EuScript V}(m)$ $=$ ${\EuScript V}(m)_{{\EuScript G}*}$ are defined by $$\begin{aligned}
\label{LambdaGdef}
& {\boldsymbol\Lambda}_{{\EuScript G}} =
\bigl\{ (\lambda, m) \in {\EuScript V}\times {\Bbb Z}_{\geq 0} \bigm|
{\EuScript G}^m_{* \lambda} \ne \emptyset \bigr\}, \\
& {\boldsymbol\Lambda}_{{\EuScript G}}^m =
{\boldsymbol\Lambda}_{{\EuScript G}} \cap ( {\EuScript V}\times \{ m \} ), \\
& {\EuScript V}(m) =
\bigl\{ \lambda \in {\EuScript V}\bigm|
(\lambda, m) \in {\boldsymbol\Lambda}_{{\EuScript G}} \bigr\}.\end{aligned}$$ For each $m \geq 0$, we define the algebra ${\mathrm{Str}^{m} ({\EuScript G}, *)}$ by $${\mathrm{Str}^{m} ({\EuScript G}, *)} =
\prod_{\lambda \in {\EuScript V}(m)}
{\mathrm{End}}({\Bbb K}{\EuScript G}^m_{* \lambda})$$ and call it [*string algebra*]{} of $({\EuScript G}, w, *)$. For each $m, n \geq 0$, we define the algebra map $\iota_{mn}\!: {\mathrm{Str}^{m} ({\EuScript G}, *)} \to {\mathrm{Str}^{m+n} ({\EuScript G}, *)}$ by $\iota_{mn} (x) ({\bold p}\cdot {\bold q}) = x{\bold p}\otimes {\bold q}$ $({\bold p}\in {\EuScript G}^m_{* \lambda}, {\bold q}\in {\EuScript G}^n_{\lambda \mu})$. For each $1 \leq i < m$, we define the element ${{\stackrel{\scriptscriptstyle *}{w}}}_i = {{\stackrel{\scriptscriptstyle *}{w}}}_{i/m}$ of ${\mathrm{Str}^{m} ({\EuScript G}, *)}$ to be the restriction of $w_{i/m}$ on ${\Bbb K}{\EuScript G}^m_{*-}$, where ${\EuScript G}^m_{*-}$ $=$ $\coprod_{\lambda \in {\EuScript V}} {\EuScript G}^m_{* \lambda}$. We say that $({\EuScript G}, w, *)$ is a [*flat face model*]{} if $$\label{flatdef}
\iota_{mn} (x) {{\stackrel{\scriptscriptstyle *}{w}}}_{nm} \iota_{nm} (y) {{\stackrel{\scriptscriptstyle *}{w}}}_{nm}^{-1} =
{{\stackrel{\scriptscriptstyle *}{w}}}_{nm} \iota_{nm} (y) {{\stackrel{\scriptscriptstyle *}{w}}}_{nm}^{-1} \iota_{mn} (x)$$ for each $x \in {\mathrm{Str}^{m} ({\EuScript G}, *)}$ and $y \in {\mathrm{Str}^{n} ({\EuScript G}, *)}$ $(m, n \geq 0)$, where ${{\stackrel{\scriptscriptstyle *}{w}}}_{mn} \in {\mathrm{Str}^{m+n} ({\EuScript G}, *)}$ is defined by $$\label{wstamndef}
{{\stackrel{\scriptscriptstyle *}{w}}}_{mn} =
({{\stackrel{\scriptscriptstyle *}{w}}}_n {{\stackrel{\scriptscriptstyle *}{w}}}_{n+1} \cdots {{\stackrel{\scriptscriptstyle *}{w}}}_{m+n-1})
({{\stackrel{\scriptscriptstyle *}{w}}}_{n-1} {{\stackrel{\scriptscriptstyle *}{w}}}_{n} \cdots {{\stackrel{\scriptscriptstyle *}{w}}}_{m+n-2})
\cdots
({{\stackrel{\scriptscriptstyle *}{w}}}_{1} {{\stackrel{\scriptscriptstyle *}{w}}}_{2} \cdots {{\stackrel{\scriptscriptstyle *}{w}}}_{m}).$$ For each flat ${\EuScript V}$-face model $({\EuScript G}, w, *)$, $n \geq 0$ and $\lambda, \mu \in {\EuScript V}$, there exists a unique left action $\Gamma$ of ${\mathrm{Str}^{n} ({\EuScript G}, *)}$ on ${\Bbb K}{\EuScript G}^n_{\lambda \mu}$ such that $$\label{Gammadef}
{\bold p}\otimes (\Gamma (x) {\bold q}) =
{{\stackrel{\scriptscriptstyle *}{w}}}_{nm} \iota_{nm} (x) {{\stackrel{\scriptscriptstyle *}{w}}}_{nm}^{-1}
({\bold p}\cdot {\bold q})$$ for each $m \geq 0$, ${\bold p}\in {\EuScript G}^m_{* \lambda}, {\bold q}\in {\EuScript G}^n_{\lambda \mu}$ and $x \in {\mathrm{Str}^{n} ({\EuScript G}, *)}$. Using this action, we define [*costring algebra*]{} $$\mathrm{Cost} (w, *) =
\bigoplus_{m \geq 0} \mathrm{Cost}^m (w, *)$$ to be the quotient ${\EuScript V}$-face algebra of $\bigoplus_{m \geq 0}
{\mathrm{End}}_{{\Bbb K}} ({\Bbb K}{\EuScript G}^m)^* \cong {\frak H}({\EuScript G})$ given by $$\mathrm{Cost}^m (w, *) =
{\mathrm{End}}_{{\mathrm{Str}^{m} ({\EuScript G}, *)}} ({\Bbb K}{\EuScript G}^m)^* .$$ For each $\lambda, \mu \in {\EuScript V}$ and $(\nu, m) \in {\boldsymbol\Lambda}_{{\EuScript G}}$, we define the non-negative integer $N^{\mu}_{\lambda \nu} (m)$ by the irreducible decomposition of ${\Bbb K}{\EuScript G}^m_{\lambda \mu}$: $$\label{frdef}
[{\Bbb K}{\EuScript G}^m_{\lambda \mu}] =
\sum_{\nu \in {\EuScript V}(m)} N^{\mu}_{\lambda \nu} (m)
[{\Bbb K}{\EuScript G}^m_{* \nu}],$$ where, for each ${\mathrm{Str}^{m} ({\EuScript G}, *)}$-module $V$, $[V]$ denotes the element of the Grothendieck group $K_0 ({\mathrm{Str}^{m} ({\EuScript G}, *)})$ corresponding to $V$ (see e.g. [@ChariPressley §5.1]). We call $N^{\mu}_{\lambda \nu} (m)$ [*fusion rules*]{} of $({\EuScript G}, w, *)$,
\[Coststrthm\] Let $({\EuScript G}, w, *)$ be a flat ${\EuScript V}$-face model with fusion rule $N_{\lambda \nu}^{\mu} (m)$. For each $(\lambda, m) \in {\boldsymbol\Lambda}_{{\EuScript G}}$, up to isomorphism there exists a unique right $\mathrm{Cost}^m (w, *)$-comodule $L_{(\lambda, m)}$ such that $$\label{dimLlm*m}
\dim L_{(\lambda, m)}(*, \mu) = \delta_{\lambda \mu}$$ for each $\mu \in {\EuScript V}$. As coalgebras, we have $$\mathrm{Cost}^m (w, *) \cong
\bigoplus_{\lambda \in {\EuScript V}(m)}
\mathrm{End} (L_{(\lambda, m)})^*.$$ In the corepresentation ring $K_0 (\bold{Com}^f_{\mathrm{Cost} (w, *)})$, we have $$[L_{(*,0)}] = 1,$$ $$[L_{(\lambda,m)}] [L_{(\mu,n)}]
= \sum_{\nu \in {\EuScript V}(m+n)} N^{\nu}_{\lambda \mu} (n) [L_{(\nu, m+n)}].$$ Moreover, for each $\mathrm{Cost}^m (w, *)$-comodule $M$, we have $$\label{decompform}
[M] = \sum_{\lambda \in {\EuScript V}(m)}
\dim \left( M (*, \lambda) \right) [L_{(\lambda, m)}].$$ We have $$\label{dimLnmlm}
\dim \left( L_{(\nu, m)} (\lambda, \mu) \right) =
N^{\mu}_{\lambda \nu} (m).$$
For each flat star-triangular face model, we have $$\Gamma ({{\stackrel{\scriptscriptstyle *}{w}}}_{i/n}) = w_{i/n}
\quad (n \geq 2,\, 1 \leq i < n).$$
By the braid relation, we have $$\label{wnmwiwnm}
{{\stackrel{\scriptscriptstyle *}{w}}}_{nm} {{\stackrel{\scriptscriptstyle *}{w}}}_{i/n+m} {{\stackrel{\scriptscriptstyle *}{w}}}_{nm}^{-1} =
{{\stackrel{\scriptscriptstyle *}{w}}}_{i+m / n+m}$$ for each $1 \leq i < n$. Using this together with $\iota_{nm} ({{\stackrel{\scriptscriptstyle *}{w}}}_{i/n})$ $=$ ${{\stackrel{\scriptscriptstyle *}{w}}}_{i/n+m}$, we obtain $$\label{pGwq}
{\bold p}\otimes (\Gamma ({{\stackrel{\scriptscriptstyle *}{w}}}_{i/n}) {\bold q}) =
{\bold p}\otimes w_{i/n} {\bold q}$$ for each $m \geq 0$, ${\bold p}\in {\EuScript G}^m_{* \lambda}$ and ${\bold q}\in {\EuScript G}^n_{\lambda \mu}$ as required.
\[flatcond\] Let $({\EuScript G}, w)$ be a star-triangular ${\EuScript V}$-face model with a fixed vertex $* \in {\EuScript V}$. Then $({\EuScript G}, w, *)$ is flat if ${\Bbb K}{\EuScript G}^m_{* \lambda}$ is an absolutely irreducible $\frak{B}_m$-module for each $(\lambda, m) \in {\boldsymbol\Lambda}_{{\EuScript G}}$. In this case, we have $\mathrm{Cost} (w, *)$ $=$ ${{\frak A}(w)}$ as quotients of ${\frak H}({\EuScript G})$.
Using , we see that holds for every $x \in {\mathrm{Str}^{m} ({\EuScript G}, *)}$ and $y = {{\stackrel{\scriptscriptstyle *}{w}}}_i$ $(1 \leq i < n)$. Hence $({\EuScript G}, w, *)$ is flat if ${\mathrm{Str}^{m} ({\EuScript G}, *)} = \langle {{\stackrel{\scriptscriptstyle *}{w}}}_1, \ldots,
{{\stackrel{\scriptscriptstyle *}{w}}}_{m-1} \rangle$ for each $m > 1$. The second assertion follows from Proposition \[Aw\*\] and the lemma above.
$SU(N)_L$-SOS models
====================
In order to construct the algebras ${{\frak S}(A_{N-1};t)_{\epsilon}}$, we first recall ${SU(N)}_L$-SOS models (without spectral parameter) [@JMO], which are equivalent to H. Wenzl’s representations of Iwahori-Hecke algebras (cf. [@Wenzl]) and also, the monodromy representations of the braid group arising from conformal field theory (cf. A. Tsuchiya and Y. Kanie [@TK]). Let $N \geq 2$ and $L \geq 2$ be integers. For each $1 \leq i \leq N$, we define the vector $\hat{i} \in \Bbb{R}^N$ by $\hat{1} = (1 - 1/N, -1/N, \ldots, -1/N)$, $\ldots$, $\hat{N} = (- 1/N, \ldots, -1/N, 1 -1/N)$. Let ${\EuScript V}= {\EuScript V}_{NL}$ be the subset of $\Bbb{R}^N$ given by
$$\label{Vdef}
{{\EuScript V}}_{NL} =
\bigl\{ \lambda_1 \hat{1} + \cdots \lambda_N \hat{N}
\bigm| \lambda_1, \ldots, \lambda_N \in {\Bbb Z},
L \geq {\lambda}_1 \geq \dots \geq {\lambda}_{N} = 0
\bigr\}.$$
For $\lambda \in {\EuScript V}$, we define integers $\lambda_1, \ldots, \lambda_N$ and $| \lambda |$ by $\lambda = \sum_i \lambda_i \hat{i}$, $\lambda_N = 0$ and $| \lambda | = \sum_i \lambda_i$. For $m \geq 0$, we define a subset ${{\EuScript G}}^m$ of ${{\EuScript V}}^{m+1}$ by $${{\EuScript G}}^m =
{{\EuScript V}}^{m+1} \cap
\bigl\{ {\bold p}= (\lambda\, |\, i_1, \ldots, i_m)
\bigm| \lambda\in {\EuScript V},\,
1 \leq i_1, \ldots, i_m \leq N \bigr\},$$ where for $\lambda \in \Bbb{R}^N$ and $1 \leq i_1, \ldots, i_m \leq N$, we set $$(\lambda\, |\, i_1, \ldots, i_m) =
(\lambda, \lambda + \hat{i}_1, \ldots,
\lambda + \hat{i}_1 + \cdots + \hat{i}_m).$$ Then $({\EuScript V},{{\EuScript G}}^1)$ defines an oriented graph ${\EuScript G}= {{\EuScript G}}_{N,L}$ and ${{\EuScript G}}^m$ is identified with the set of paths of ${\EuScript G}$ of length $m$. For ${\bold p}= (\lambda\, |\, i,j)$, we set ${{\bold p}}^{\dag} = (\lambda\, |\, j,i)$ and $$d({\bold p})
= {\lambda}_i - {\lambda}_j + j - i.$$ We define subsets ${{\EuScript G}}^2[\to]$, ${{\EuScript G}}^2[\;\downarrow\;]$ and ${{\EuScript G}}^2[\searrow]$ of ${{\EuScript G}}^2$ by $$\begin{aligned}
{{\EuScript G}}^2[\to] & =
\bigl\{ {\bold p}\in {{\EuScript G}}^2
\bigm| {\bold p}^{\dag} = {\bold p}\bigr\},
\\
{{\EuScript G}}^2[\;\downarrow\;] & =
\bigl\{ {\bold p}\in {{\EuScript G}}^2
\bigm|
{\bold p}^{\dag} \not\in {\EuScript G}^2 \bigr\},
\\
{{\EuScript G}}^2[\searrow] & =
\bigl\{ {\bold p}\in {{\EuScript G}}^2
\bigm| {\bold p}\not= {\bold p}^{\dag} \in {\EuScript G}^2
\bigr\}. \end{aligned}$$ Let $t \in{\Bbb C}$ be a primitive $2(N+L)$-th root of $1$. Let $\epsilon$ be either $1$ or $-1$ and $\zeta$ a nonzero parameter. We define a face model $({\EuScript G},w_{N,t,\epsilon}) = ({{\EuScript G}}_{N,L},w_{N,t, \epsilon, \zeta})$ by setting $$w_{N,t,\epsilon} \!
\begin{bmatrix}
\lambda & \lambda + \hat{i} \\
\lambda + \hat{i} & \lambda + \hat{i} + \hat{j}
\end{bmatrix}
= - \zeta^{-1} t^{-d({\bold p})} \frac{1}{[d({\bold p})]},$$ $$w_{N,t,\epsilon} \!
\begin{bmatrix}
\lambda & \lambda + \hat{i} \\
\lambda + \hat{j} & \lambda + \hat{i} + \hat{j}
\end{bmatrix}
= \zeta^{-1} \epsilon \, \frac{[d({\bold p})-1]}{[d({\bold p})]},$$ $$w_{N,t,\epsilon} \!
\begin{bmatrix}
\lambda & \lambda + \hat{k} \\
\lambda + \hat{k} & \lambda + 2 \hat{k}
\end{bmatrix}
= \zeta^{-1} t$$ for each ${\bold p}= (\lambda\, |\, i,j) \in
{\EuScript G}^2[\searrow] \amalg {\EuScript G}^2 [\,\downarrow\,]$ and $(\lambda\, |\, k,k) \in {\EuScript G}^2[\to]$, where $[n] = (t^n-t^{-n})/(t-t^{-1})$ for each $n \in {\Bbb Z}$. We call $({\EuScript G},w_{N,t,\epsilon})$ [*$SU(N)_L$-SOS model*]{} (without spectral parameter) [@JMO]. It is known that $({\EuScript G},w_{N,t,\epsilon})$ is star-triangular. Moreover, H. Wenzl [@Wenzl] showed that ${\Bbb C}{\EuScript G}^m_{0 \lambda}$ is an irreducible $\frak{B}_m$-module for each $m \geq 0$ and $\lambda \in \Lambda^m_{{\EuScript G}0}$. Therefore $({\EuScript G},w_{N,t,\epsilon},0)$ is flat by Proposition \[flatcond\]. In [@GoodmanWenzl], F. Goodman and H. Wenzl showed that the fusion rule of $({\EuScript G},w_{N,t,\epsilon},0)$ agrees with that of $SU(N)_L$-WZW model. We give another proof of their result in the next section.
\[rembase\] (1) Strictly speaking, Wenzl deals with $w_{N,t,\epsilon}$ only when $\epsilon = - 1$. However, it is clear that his arguments are applicable to the case $\epsilon = 1$. The results for $\frak{A}(w_{N,t,1})$ also follows from those of $\frak{A}(w_{N,t,-1})$, since the former is a 2-cocycle deformation of the latter (cf. [@DT]). Hence, $\epsilon$ may be viewed as a gauge parameter.\
(2) In order to avoid using square roots of complex numbers, we use a different normalization of $w_{N,t, -1}$ from Wenzl [@Wenzl]. For each ${\bold p}\in {\EuScript G}^m$ $(m \geq 1)$, we define $\kappa ({\bold p}) \in {\Bbb C}$ by $$\begin{gathered}
\label{kappadef}
\kappa ({\bold p}\cdot {\bold q}) = \kappa ({\bold p}) \kappa ({\bold q})
\quad ({\bold p}\in {\EuScript G}^m, {\bold q}\in {\EuScript G}^n, m, n > 0), \\
\kappa (\lambda\, |\, i) =
\prod_{k = i+1}^{N} A_{d(0\, |\, i,k)+1}
A_{d(0\, |\, i,k)+2} \cdots A_{d(\lambda\, |\, i,k)}
\quad ((\lambda\, |\, i) \in {\EuScript G}^1),
\nonumber \end{gathered}$$ where $A_d = a_d / \sqrt{a_d a_{-d}}$ and $a_d = [d+1] / [2] [d]$. Note that $\kappa ({\bold p})$ satisfies $$\frac{\kappa ({\bold p})}{\kappa ({\bold p}^{\dag})} =
A_{d ({\bold p})}$$ for each ${\bold p}\in {\EuScript G}^2 [\searrow]$. By replacing the basis $\{ {\bold p}\}$ of ${\Bbb C}{\EuScript G}^m$ with $\{ \kappa ({\bold p}) {\bold p}\}$, we obtain Wenzl’s original expression of the Hecke algebra representation. It is also useful to use $\{ \kappa ({\bold p})^2 {\bold p}\}$ instead of $\{ {\bold p}\}$ (see §12). The corresponding Boltzmann weight $w^{\Sigma}_{N,t,\epsilon}$ satisfies $$w^{\Sigma}_{N,t,\epsilon} \!
\left[ {\bold r}\frac[0pt]{{\bold p}}{{\bold s}} {\bold q}\right] :=
\left(
\frac{\kappa ({\bold p}\cdot {\bold q})}{\kappa ({\bold r}\cdot {\bold s})}
\right)^2
w_{N,t,\epsilon} \!
\left[ {\bold r}\frac[0pt]{{\bold p}}{{\bold s}} {\bold q}\right] =
w_{N,t,\epsilon} \!
\left[ {\bold p}\frac[0pt]{{\bold r}}{{\bold q}} {\bold s}\right].$$ We call $\{ {\bold p}\}$ and $\{ \kappa ({\bold p})^2 {\bold p}\}$ [*rational basis of type*]{} $\Omega$ and [*type*]{} $\Sigma$ respectively.
The algebra ${{\frak S}(A_{N-1};t)_{\epsilon}}$
===============================================
Applying Proposition \[FRT\] to $({\EuScript G},w_{N,t,\epsilon, \zeta})$, we obtain a CQT ${\EuScript V}$-face algebra ${{\frak A}(w_{N,t,\epsilon,\zeta})}$ $=$ $\mathrm{Cost} (w_{N,t,\epsilon}, 0)$. In order to define the “(quantum) determinant” of ${{\frak A}(w_{N,t,\epsilon})}$, we introduce an algebra $\Omega = \Omega_{N,L, \epsilon}$, which is a face-analogue of the exterior algebra. It is defined by generators $\omega ({\bold p})$ $({\bold p}\in {\EuScript G}^m; m \geq 0)$ with defining relations: $${\sum_{k \in {\EuScript V}}}\omega (k) \, = \, 1,$$ $$\omega ({\bold p}) \omega ({\bold q}) \, = \,
\delta_{{\frak {r}}({\bold p}) {\frak {s}}({\bold q})} \, \omega ({\bold p}\cdot {\bold q}),$$ $$\omega ({\bold p}) = - \epsilon\, \omega ({\bold p}^{\dag})
\quad ({\bold p}\in {\EuScript G}^2 [\searrow]),$$ $$\omega ({\bold p}) = 0
\quad ({\bold p}\in {\EuScript G}^2[\to]).$$ It is easy to verify that $\Omega^m:= \sum_{{\bold p}\in {\EuScript G}^m} {\Bbb C}\omega ({\bold p}) $ becomes an ${{\frak A}(w_{N,t,\epsilon})}$-comodule via $$\label{comOdef}
\omega ({\bold q}) \mapsto
\sum_{{\bold p}\in {\EuScript G}^m} \omega ({\bold p}) \otimes e \binom{{\bold p}}{{\bold q}}
\quad ({\bold q}\in {\EuScript G}^m)$$ for each $m \geq 0$.
For each $m \geq 0$, we set $$B \Omega^m =
\bigl\{ (\lambda,\,
\lambda + \sum_{k \in I} \, \hat{i}\, )
\in {\EuScript V}^2 \bigm|
I \subset \{ 1, \ldots, N \},\,
\mathrm{card} (I) = m \bigr\}.$$ Also we define ${\EuScript{L}}\!: {\EuScript G}^m \to {\Bbb Z}_{\geq 0}$ by $$\EuScript{L}
(\lambda\, |\, i_1, \ldots, i_m)
=
\mathrm{Card} \{ (k,l) |
1 \leq k < l \leq N, i_k < i_l \}$$
For each $(\lambda,\mu) \in B \Omega^m$, ${\EuScript G}^m_{\lambda \mu} \not= \emptyset$ and $\omega_m (\lambda, \mu)$ $:=$ ${(- \epsilon)^{\EuScript{L} ( {\bold p})}} \omega ({\bold p})$ does not depend on the choice of ${\bold p}\in {\EuScript G}^m_{\lambda \mu}$. Moreover $\left\{ \omega_m (\lambda, \mu) \vert
(\lambda, \mu) \in B \Omega^m \right\}$ is a basis of $\Omega^m$. In particular, $\Omega^m = 0$ if $m > N$.
We will prove this lemma by means of Bergman’s diamond lemma [@Bergman], or rather its obvious generalization to the quotient algebras of ${\Bbb C}\langle {\EuScript G}\rangle$, where $\langle {\EuScript G}\rangle = \coprod_m {\EuScript G}^m$. We define a “reduction system” $S = S_1 \coprod S_2 \subset
\langle {\EuScript G}\rangle \times {\Bbb C}\langle {\EuScript G}\rangle$ by setting $$\begin{gathered}
S_1 = \bigl\{ ({\bold p}, - \epsilon\, {\bold p}^{\dag}) \bigm|
{\bold p}= (\lambda\, |\, i,j) \in {\EuScript G}^2 [\searrow],\, i < j \bigr\},
\\
S_2 = \bigl\{ ({\bold p}, 0) \bigm|
{\bold p}= (\lambda\, |\, i_1, \ldots, i_m) \in {\EuScript G}^m,\, m \geq 2,\,
\mathrm{card} \{ i_1, \ldots, i_m \} < m \bigr\}.
\end{gathered}$$ It is straightforward to verify that the quotient ${\Bbb C}\langle {\EuScript G}\rangle /
\langle W - f \, \vert \, (W,f) \in S \rangle$ is isomorphic to $\Omega$ and that all ambiguities of $S$ are resolvable. Next, we introduce a semigroup partial order $\leq$ on $\langle {\EuScript G}\rangle$ by setting $(\lambda\, |\, i_1, \ldots, i_m) <
(\lambda\, |\, j_1, \ldots, j_n) $ if either $m < n$, or $m = n$ and $i_1 = j_1$, $\ldots$, $i_{k-1} = j_{k-1}$, $i_k > j_k$ for some $1 \leq k \leq m$. Then $\leq$ is compatible with $S$ and satisfies the descending chain condition. This completes the proof of the proposition.
For each $0 \leq m \leq N$, we set $\Lambda_m = \hat{1}+ \cdots + \hat{m}$. As an immediate consequence of and the the proposition above, we obtain the following result.
\[O=L\] For each $0 \leq m \leq N$, we have $\Omega^m \cong L_{(\Lambda_m, m)}$ as ${{\frak A}(w_{N,t,\epsilon})}$-comodules.
Now we define the “determinant” ${\det} = \sum_{\lambda, \mu \in {\EuScript V}} \det \binom{\lambda}{\mu}$ of ${{\frak A}(w_{N,t,\epsilon})}$ to be the group-like element which corresponds to the group-like comodule $\Omega^N$ and its basis $\bigl\{ \bar{\omega} (\lambda)\bigr\}$ via Lemma \[glcom\], where $\bar{\omega} (\lambda) =
D (\lambda) \omega_{N} (\lambda, \lambda)$ and $$D(\lambda) =
\prod_{1 \leq i < j \leq N}
\frac{[d(\lambda\, |\, i,j)]}{[d(0\, |\, i,j)]}
\quad (\lambda \in {\EuScript V}).$$ Explicitly, we have $$\label{detformula}
\det \binom{\lambda}{\mu} =
\frac{D(\mu)}{D(\lambda)}
\sum_{{\bold p}\in {\EuScript G}^N_{\lambda \lambda}}
(- \epsilon)^{{\EuScript{L}}({\bold p}) + {\EuScript{L}}({\bold q})}
e \binom{{\bold p}}{{\bold q}},$$ where ${\bold q}$ denotes an arbitrary element of ${\EuScript G}^N_{\mu \mu}$. By and for $g = \det$, the quotient $${{\frak S}(A_{N-1};t)_{\epsilon}}:= {{\frak A}(w_{N,t,\epsilon})}/ (\det - 1)$$ naturally becomes a ${\EuScript V}$-face algebra, which we call $SU(N)_{L}$-[*SOS algebra*]{}.\
The proof of the following lemma will be given in §8 and §13.
\[cformula\] For each ${\bold p}\in {\EuScript G}^1$, we have $$\begin{aligned}
\label{c_OmegaNOmega1}
c_{\Omega^N \Omega^1}
\left( \bar{\omega} ({\frak {s}}({\bold p})) \otimes
\omega ({\bold p}) \right) =
\epsilon^{N-1} \zeta^{- N} t
\omega ({\bold p}) \otimes
\bar{\omega} ({\frak {r}}({\bold p})), \\
c_{\Omega^1 \Omega^N}
\left( \omega ({\bold p}) \otimes
\bar{\omega} ({\frak {r}}({\bold p}))
\right) =
\epsilon^{N-1} \zeta^{- N} t
\bar{\omega} ({\frak {s}}({\bold p}))
\otimes \omega ({\bold p}).
\end{aligned}$$
\[SSisCQT\] The element $\det$ belongs to the center of ${{\frak A}(w_{N,t,\epsilon})}$. Moreover, if $\zeta$ satisfies $$\label{etaeq}
\zeta^N =
\epsilon^{N-1} t,$$ then $$\label{R(det-1,a)}
{{\cal R}^{\pm}}(\det -1, a) = 0 = {{\cal R}^{\pm}}(a, \det -1)
\quad (a \in {{\frak A}(w_{N,t,\epsilon,\zeta})}).$$ Hence, ${{\frak S}(A_{N-1};t)_{\epsilon}}$ naturally becomes a quotient CQT ${\EuScript V}$-face algebra of ${{\frak A}(w_{N,t,\epsilon,\zeta})}$.
(cf. [@gd]) By computing the coaction of ${{\frak A}(w_{N,t,\epsilon})}$ on $\omega ({\bold p}) \otimes \bar{\omega} ({\frak {r}}({\bold p}))$, in two ways via , we obtain the first assertion. We show the first equality of for $\pm = +$ and $a = e \binom{{\bold p}}{{\bold q}}$ $({\bold p}, {\bold q}\in {\EuScript G}^m, m \geq 0)$. By and , it suffices to show $$\label{R(det,e)}
{{\cal R}^+}\left( \det \binom{{\frak {r}}({\bold p})}{{\frak {s}}({\bold p})},\, e \binom{{\bold p}}{{\bold q}} \right)
= \delta_{{\bold p}{\bold q}}
\quad ({\bold p}, {\bold q}\in {\EuScript G}^m).$$ For $m = 0$, this follows from and . By computing the left-hand side of via , we obtain for $m = 1$. For $m \geq 2$, follows from and for $g = \det$ by induction on $m$.
Since the braiding of ${{\frak S}(A_{N-1};t)_{\epsilon}}$ depends on the choice of the discrete parameter $\zeta$ satisfying , we sometimes write ${{\frak S}(A_{N-1};t)_{\epsilon,\zeta}}$ instead of ${{\frak S}(A_{N-1};t)_{\epsilon}}$.\
To state our first main result, we recall the [*fusion rule of*]{} $SU(N)_L$-$WZW model$ in conformal field theory. By [@GoodmanNakanishi], it is characterized as the structure constant $N_{\lambda \mu}^{\nu}$ of a commutative ${\Bbb Z}$-algebra $\EuScript{F}$ (called the [*fusion algebra of*]{} $SU(N)_L$-$WZW model$) with free basis $\{ \chi_{\lambda} \}_{\lambda \in {\EuScript V}}$ (i.e., $\chi_{\lambda} \chi_{\mu}$ $=$ $\sum_{\nu \in {\EuScript V}} N_{\lambda \mu}^{\nu} \chi_{\nu}$) such that $$\label{NlLmm}
N^{\mu}_{\lambda \Lambda_m} =
\begin{cases}
1 & (\lambda, \mu) \in B \Omega^m \\
0 & \text{otherwise}
\end{cases}$$ for each $\lambda, \mu \in {\EuScript V}$ and $0 \leq m < N$. See Kac [@Kac] or Walton [@Walton] for an explicit formula of $N_{\lambda \mu}^{\nu}$.
\[SSrep\] For each $\lambda \in {\EuScript V}$, up to isomorphism there exists a unique right ${{\frak S}(A_{N-1};t)_{\epsilon}}$-comodule $L_{\lambda}$ such that $$\label{dimLl0m}
\dim L_{\lambda}(0, \mu) = \delta_{\lambda \mu}
\quad (\mu \in {\EuScript V}).$$ Moreover, we have $${{\frak S}(A_{N-1};t)_{\epsilon}}\cong
\bigoplus_{\lambda \in {\EuScript V}} \mathrm{End} (L_{\lambda})^*$$ as coalgebras. In particular, $L_{\lambda}$ is irreducible for each $\lambda \in {\EuScript V}$\
The corepresentation ring $K_0 ( \bold{Com}_{{{\frak S}(A_{N-1};t)_{\epsilon}}} )$ is identified with the fusion algebra $\EuScript{F}$ of $SU(N)_L$-WZW model via $\chi_{\lambda} = [ L_{\lambda} ]$. That is, we have $$\label{L0=1}
[L_{0}] = 1,$$ $$[L_{\lambda}] [L_{\mu}] =
\sum_{\nu \in {\EuScript V}} N^{\nu}_{\lambda \mu} [L_{\nu}].$$ We have $$\dim \left( L_{\nu} (\lambda, \mu) \right) =
N^{\mu}_{\lambda \nu}.$$
It is easy to verify that $(\lambda, m) \in \boldsymbol\Lambda_{{\EuScript G}_{N,L}}$ if and only if $m \in | \lambda | + N {\Bbb Z}_{\geq 0}$. Since ${\Bbb C}{\EuScript V}\det^n \bar{\otimes} L_{(\lambda,m)}$ $\cong$ $L_{(\lambda,m + Nn)}$ by and , we see that $\det$ satisfies all conditions of Lemma \[sympredgle\], where $\bar{\boldsymbol\Lambda} = {\EuScript V}$ and $\varphi (\lambda, n) = (\lambda, | \lambda | + Nn)$. Therefore, we have Part (1), and $$[L_{\lambda}] [L_{\mu}] =
\sum_{\nu \in {\EuScript V}} N^{\nu}_{\lambda \mu}
(| \mu |)[L_{\nu}],
\quad
\dim \left( L_{\nu} (\lambda, \mu) \right) =
N^{\mu}_{\lambda \nu} (m)
\quad ( (\nu, m) \in \boldsymbol\Lambda_{{\EuScript G}_{N,L}}).$$ In particular, $\tilde{N}^{\nu}_{\lambda \mu}$ $:=$ $N^{\nu}_{\lambda \mu} (m)$ does not depend on the choice of $m$. Using Proposition \[O=L\], and , we obtain $$[L_{\lambda}] [L_{\Lambda_m}] =
\sum_{\mu \in {\EuScript V}}
\dim ( L_{\lambda} \bar{\otimes} \Omega^m) (0, \mu)
[L_{\mu}]
=
\sum_{\mu \in {\EuScript V}} \dim
\left( \Omega^m (\lambda, \mu) \right)
[L_{\mu}].$$ Thus the numbers $\tilde{N}^{\nu}_{\lambda \mu}$ satisfies the condition of $N^{\nu}_{\lambda \mu}$ stated above.
The element $\det$ is not a zero-divisor of ${{\frak A}(w_{N,t,\epsilon})}$. In particular, we have $\det \binom{\lambda}{\mu} \not= 0$ for each $\lambda, \mu \in {\EuScript V}$. Moreover, we have $$\label{GLEAS}
\mathrm{GLE} \left( {{\frak A}(w_{N,t,\epsilon})}\right) =
\bigl\{ \sum_{\lambda, \mu \in {\EuScript V}}
\frac{c (\lambda)}{c (\mu)} \, {{\stackrel{\scriptscriptstyle\circ}{e}}}_{\lambda} e_{\mu}
{\det}^m \bigm| m \in {\Bbb Z}_{\geq 0},\,
c(\lambda) \in {\Bbb C}^{\times} \,\, (\lambda \in {\EuScript V}) \bigr\}.$$
The first assertion follows from the fact that $\det$ is simply reducible (see the proof of the theorem above). By Theorem \[Coststrthm\] (1), every group-like comodule of ${{\frak A}(w_{N,t,\epsilon})}$ is isomorphic to $(\Omega^N)^{\overline{\otimes} m}$ for some $m \geq 0$. Hence the second assertion follows from Lemma \[glcom\].
Module structure of $\Omega$
============================
Let ${\frak H}$ be a CQT ${\EuScript V}$-face algebra over $\Bbb{K}$. As in case ${\frak H}$ is a CQT bialgebra, the correspondence $a \mapsto {{\cal R}^+}(\,,a)$ defines an antialgebra-coalgebra map from ${\frak H}$ into the dual face algebra ${\frak H}^{\circ}$ (cf. [@fa]). Let $W$ be a right ${\frak H}$-comodule. Combining the above map with the left action of ${\frak H}^*$ on $W$, we obtain a right action of ${\frak H}$ on $W$ given by $$\label{wa}
wa = \sum_{(w)} w_{(0)} {{\cal R}^+}(w_{(1)}, a)
\quad (w \in W, a \in {\frak H}).$$ Let $V$ be another ${\frak H}$-comodule. Then we have $$\label{(vw)a}
\left( v \otimes w \right) a =
\sum_{(a)} v a_{(1)} \otimes w a_{(2)}
\quad (v \in V(\lambda,\nu), w \in W(\nu,\mu), a \in {\frak H}).$$ If ${\frak H}$ has an antipode and $W$ is finite-dimensional, then we have $$\label{<va,w>}
\langle v a, w \rangle =
\langle v, w S^{-1} (a) \rangle
\quad (v \in W \spcheck, w \in W),$$ by and .
In this section, we give an explicit description of the right ${{\frak A}(w_{N,t,\epsilon})}$-module structure of $\Omega$. By and , we obtain the following.
For each ${\bold s}\in {\EuScript G}^m$ and ${\bold p}, {\bold q}\in {\EuScript G}^n$ $(n \geq 0)$, we have $$\label{osepq}
\omega ({\bold s}) \, e \binom{{\bold p}}{{\bold q}} =
\sum_{{\bold r}\in {\EuScript G}^m}
w_{N,t,\epsilon} \!
\left[ {\bold p}\frac[0pt]{{\bold s}}{{\bold r}} {\bold q}\right]\,
\omega ({\bold r})
\quad ({\bold s}\in {\EuScript G}^m, {\bold p}, {\bold q}\in {\EuScript G}^n, n \geq 0).$$ In particular, we have $$\label{Olmepq}
\Omega (\lambda, \mu) e \binom{{\bold p}}{{\bold q}} \in
\delta_{\lambda, {\frak {s}}({\bold p})} \delta_{\mu, {\frak {s}}({\bold q})}\,
\Omega ({\frak {r}}({\bold p}), {\frak {r}}({\bold q})) .$$
\[actOmega\] Let $(\lambda, \mu)$ be an element of $B \Omega^m$ $(m > 0)$ and ${\bold p}$ $=$ $(\lambda\, |\, i_1, \ldots, i_m)$ an element of ${\EuScript G}^m_{\lambda \mu}$. Define the set $I$ and $C(\lambda\, |\, k,l) \in {\Bbb C}$ $(k \not= l)$ by $I = \{ i_1, \ldots, i_m \}$ and $$C(\lambda\, |\, k, l) =
\frac{[d(\lambda\, |\, k, l) + 1]}{[d(\lambda\, |\, k, l)]}$$ respectively. Then for each $(\lambda\, |\, i), (\mu\, |\, j) \in {\EuScript G}^1$, we have: $$\begin{gathered}
\label{iinjnin}
\omega ({\bold p})
e\! \binom{\lambda\, |\, i}{\mu\, |\, j}
\, = \,
(- \zeta)^{- m} t^{-d(\lambda |\, i,j)}
\frac{1}{[d(\lambda |\, i,j)]}
\prod_{k \in I \setminus \{ i \} }
C(\lambda\, |\, i,k)\,
\omega (\lambda + \hat{i}\, |\, i_2, \ldots, i_m, j)
\\
(i = i_1, j \not\in I),
\end{gathered}$$ $$\begin{gathered}
\label{iinjin}
\omega ({\bold p})
e \binom{\lambda\, |\, i}{\mu\, |\, j}
\; = \;
- \delta_{ij}
(- \zeta)^{- m} t
\prod_{k \in I \setminus \{ i \}}
C(\lambda\, |\, i,k)
\, \omega_m (\lambda + \hat{i}\, |\, i_2, \ldots, i_m, i) \\
(i = i_1, j \in I), \qquad
\end{gathered}$$ $$\begin{gathered}
\label{ininjnin}
\omega_m ({\bold p})
e \binom{\lambda\, |\, i}{\mu\, |\, j}
\; = \;
\delta_{ij}
(\epsilon \zeta^{- 1})^m
\prod_{k \in I} C(\lambda\, |\, i,k)
\, \omega_m (\lambda + \hat{i}\, |\, i_1, \ldots, i_m) \\
(i, j \not\in I), \qquad\qquad
\end{gathered}$$ $$\label{ininjin}
\qquad
\omega_m ({\bold p})
e \binom{\lambda\, |\, i}{\mu\, |\, j}
\; = \; 0
\qquad\quad
(i \not\in I, j \in I).$$
These formulas are proved by induction on $m$ in a similar manner. Here we give the proof of . Since $\Omega^m$ is a quotient module of ${\Bbb C}{\EuScript G}^{m-1} \bar{\otimes} {\Bbb C}{\EuScript G}^1$, the left-hand side of is rewritten as $\sum_q A_q B_q$ with $$ A_q = \omega_{m-1} (\lambda, \nu)
e\! \binom{\lambda\, |\, i}{\nu\, |\, q},
\quad
B_q = \omega_1 (\nu, \mu)
e\! \binom{\nu\, |\, q}{\mu\, |\, j}$$ by , where $\nu = \lambda + \hat{i}_1 + \cdots + \hat{i}_{m-1}$ and the summation is taken over for all $1 \leq q \leq N$ such that $(\nu\, |\, q) \in {\EuScript G}^1$. Since $(\nu + \hat{q}, \mu + \hat{j}) \in {\EuScript G}^1$ only if $q = i_m$ or $j$, we have $$\label{oe=AB+AB}
\omega_m (\lambda, \mu)
e\! \binom{\lambda\, |\,i}{\mu\, |\,j}
=
\begin{cases}
A_{i_m} B_{i_m} + A_j B_j
& (\nu\, |\, j) \in {\EuScript G}^1\\
A_{i_m} B_{i_m}
& \mathrm{otherwise.}
\end{cases}$$ Using inductive assumption, we see that the right-hand side of equals $$\begin{gathered}
\label{iinjninpre}
(- \zeta)^{- m} t^{-d(i,j)}
\left(
\frac{1}{[d(i, i_m)] [d(i_m, j)]} +
\frac{[d(i_m, j) - 1]}{[d(i, j)] [d(i_m, j)]}
\right) \\
\prod_{n=2}^{m-1}
C(\lambda\, |\, i, i_n)\,
\omega_m (\lambda + \hat{i}, \mu + \hat{j})
\qquad\qquad
\end{gathered}$$ if $(\nu\, |\, j) \in {\EuScript G}^1$, where $d(k, l) = d (\lambda\, |\, k, l)$. Applying $[a + b + 1] + [a][b] =$ $[a + 1][b + 1]$ to $a = d(i, i_m)$ and $b = d(i_m, j) -1$, we see that equals the right-hand side of (even if $(\nu\, |\, j) \not\in {\EuScript G}^1$). Next suppose that $(\nu\, |\, j) \not\in {\EuScript G}^1$. It suffices to verify that the second term in the parentheses of is zero. In case $j = 1$, we obtain $\nu_1 = L$. Using this together with $1 \not\in I$, we see that $\lambda_1 = L$. On the other hand, since $(\mu\, |\, 1) \in {\EuScript G}^1$, we have $L - 1 \geq \mu_1 = L - \delta_{i_m N}$. Hence, $i_m = N$ and $d(i_m, j) -1 = - L - N$. In case $j > 1$, we obtain $d(i_m, j) -1 = 0$ in a similar manner. Thus we complete the proof of .
The following lemma is frequently used in the sequel.
\[irrG1\] As a right ${{\frak A}(w_{N,t,\epsilon})}$-module, ${\Bbb C}{\EuScript G}^1 \cong \Omega^1$ is irreducible. Hence ${\Bbb C}{\EuScript G}^1$ is also irreducible as a left ${{\frak A}(w_{N,t,\epsilon})}^*$-module.
Let $W$ be a non-zero submodule of ${\Bbb C}{\EuScript G}^1$. Since $W = \bigoplus_{\lambda \mu} W {{\stackrel{\scriptscriptstyle\circ}{e}}}_{\lambda} e_{\mu}$ and $W {{\stackrel{\scriptscriptstyle\circ}{e}}}_{\lambda} e_{\mu} \subset {\Bbb C}{\EuScript G}^1_{\lambda \mu}$, we have ${\bold s}_0 \in W$ for some ${\bold s}_0 = (\lambda\, |\, i) \in {\EuScript G}^1$. To show ${\Bbb C}{\EuScript G}^1 = {\bold s}_0\, {{\frak A}(w_{N,t,\epsilon})}$, we introduce the oriented graph $\EuScript{H}$ determined by $\EuScript{H}^0 = {\EuScript G}^1$ and $$\mathrm{card} \left( \EuScript{H}^1_{{\bold s}{\bold r}} \right) =
\begin{cases}
1 & {\bold s}\, e\! \binom{{\bold p}}{{\bold q}} \in {\Bbb C}^{\times} {\bold r}\quad (\exists {\bold p}, {\bold q}\in {\EuScript G}^1) \\
0 & \mathrm{otherwise}.
\end{cases}$$ It suffices to show that $\langle \EuScript{H} \rangle_{{\bold s}_0\, {\bold r}} \not=$ $\emptyset$ for every ${\bold r}\in {\EuScript G}^1$, where $\langle \EuScript{H} \rangle_{{\bold s}\, {\bold r}} =$ $\cup_m \EuScript{H}^m_{{\bold s}\, {\bold r}}$. We note that $$\label{H1necond3}
\EuScript{H}^1_{(\mu\, |\, j)\, (\mu + \hat{j}\, |\, k)}
\not= \emptyset
\quad \mathrm{if}\;\,
(\mu\, |\, j), (\mu + \hat{j}\, |\, k)
\in {\EuScript G}^1\; \mathrm{and}\; j \not= k,$$ $$\label{H1necond1}
\EuScript{H}^1_{(\mu\, |\, j)\, (\mu + \hat{k}\, |\, j)}
\not= \emptyset
\quad \mathrm{if} \;\,
(\mu\, |\, j), (\mu + \hat{k}\, |\, j)
\in {\EuScript G}^1\; \mathrm{and}\; j \not= k,$$ $$\label{H1necond2}
\EuScript{H}^1_{(\mu\, |\, j)\, (\mu + \hat{j}\, |\, j)}
\not= \emptyset
\quad \mathrm{if}\;\,
(\mu\, |\, j, j) \in {\EuScript G}^2.$$ Using , we obtain $\langle \EuScript{H} \rangle_{{\bold s}_0\, {\bold s}_1}$ $\not=$ $\emptyset$, where ${\bold s}_1 =$ $(L \Lambda_{i-1} +
\sum_{k=i}^{N-1} \lambda_k \hat{k}\,|\, i)$. Suppose $i \not= N$. Using , and then using , we obtain $\langle \EuScript{H} \rangle_{{\bold s}_1\, {\bold s}_2}$ $\not=$ $\emptyset$, where ${\bold s}_2 =$ $((L-1) \Lambda_{N-1} + \Lambda_{i-1}\, |\,i)$. Using and respectively, we obtain $\langle \EuScript{H} \rangle_{{\bold s}_2\, {\bold s}_3},$ $\langle \EuScript{H} \rangle_{{\bold s}_4\, (0 | 1)}$ $\not=$ $\emptyset$ and $\langle \EuScript{H} \rangle_{{\bold s}_3\, {\bold s}_4}$ $\not=$ $\emptyset$ respectively, where ${\bold s}_3 =$ $((L-1) \Lambda_{N-1} + \Lambda_{N-2}\, |\,N-1)$. and ${\bold s}_4 =$ $(\Lambda_{N-2}\, |\, N-1)$. Therefore we obtain $\langle \EuScript{H} \rangle_{{\bold s}_0 (0|1)}$ $\not=$ $\emptyset$ if $i \not= N$. By similar consideration, we also obtain $\langle \EuScript{H} \rangle_{{\bold s}_0 (0|1)}$ $\not=$ $\emptyset$ in case $i = N$, and also, $\langle \EuScript{H} \rangle_{(0|1) {\bold r}}$ $\not=$ $\emptyset$ for every ${\bold r}\in {\EuScript G}^1$. Thus, we have verified the first assertion. The second assertion is obvious since the image of $a \mapsto {{\cal R}^+}(\, , a)$ is a subalgebra of ${{\frak A}(w_{N,t,\epsilon})}^*$.
Now we begin to prove Lemma \[cformula\]. By and Lemma \[actOmega\], we have $$\bar{\omega} (\nu)
e \binom{\lambda\, |\, i}{\mu\, |\, j}
\; = \;
\epsilon^{N-1} \zeta^{- N} t
\delta_{ij} \delta_{\nu \lambda} \delta_{\nu \mu}
\bar{\omega} (\nu + \hat{i}).$$ Using and this equality, we see that both $\Omega^N \bar{\otimes} {\Bbb C}{\EuScript G}^1 \to {\Bbb C}{\EuScript G}^1$; $\bar{\omega} ({\frak {s}}({\bold p})) \otimes {\bold p}$ $\mapsto$ ${\bold p}\otimes \bar{\omega} ({\frak {r}}({\bold p}))$ and ${\Bbb C}{\EuScript G}^1 \bar{\otimes} \Omega^N \to {\Bbb C}{\EuScript G}^1$; ${\bold p}\otimes \bar{\omega} ({\frak {s}}({\bold p}))$ $\mapsto$ ${\bold p}\otimes \bar{\omega} ({\frak {r}}({\bold p}))$ are isomorphisms of right ${{\frak A}(w_{N,t,\epsilon})}$-modules. Hence, by Lemma \[irrG1\] and Schur’s Lemma, we have $$\label{cONO1=theta}
c_{\Omega^N \Omega^1}
(\bar{\omega} ({\frak {s}}({\bold p})) \otimes {\bold p})
= \vartheta\,
{\bold p}\otimes \bar{\omega} ({\frak {r}}({\bold p}))
\quad ({\bold p}\in {\EuScript G}^1)$$ for some constant $\vartheta$. We will prove $\vartheta = \epsilon^{N-1} \zeta^{- N} t$ in §12.
transposes and complex conjugates
=================================
The following proposition is an immediate consequence of the following reflection symmetry: $$w_{N,t,\epsilon} \!
\left[ {\bold r}\frac[0pt]{{\bold p}}{{\bold s}} {\bold q}\right] =
\left(
\frac{\kappa ({\bold r}\cdot {\bold s})}{\kappa ({\bold p}\cdot {\bold q})}
\right)^2
w_{N,t,\epsilon} \!
\left[ {\bold p}\frac[0pt]{{\bold r}}{{\bold q}} {\bold s}\right],$$ where $\kappa$ is as in .
\[trA\] There exists an algebra-anticoalgebra map ${{\frak A}(w_{N,t,\epsilon})}\to {{\frak A}(w_{N,t,\epsilon})};$ $a \mapsto a^{{{\intercal}}}$ given by $$\label{tencdef}
e \binom{{\bold p}}{{\bold q}}^{{{\intercal}}} =
\left( \frac{\kappa ({\bold p})}{\kappa ({\bold q})} \right)^2
e \binom{{\bold q}}{{\bold p}}
\quad ({\bold p}, {\bold q}\in {\EuScript G}^m, m \geq 0).$$ Moreover it satisfies $(a^{{{\intercal}}})^{{{\intercal}}}$ $=$ $a$ and $$\label{R&tenc}
{{\cal R}^{\pm}}\left(a^{{{\intercal}}},\, b^{{{\intercal}}} \right) =
{{\cal R}^{\pm}}\left(b,\, a \right)$$ for each $a, b \in {{\frak A}(w_{N,t,\epsilon})}$ and $\zeta \in {\Bbb C}^{\times}$.
The following proposition is needed to construct the “cofactor matrix.”
\[trS\] The element $\det$ satisfies ${\det}^{{{\intercal}}} = \det$. Hence, ${{\intercal}}$ induces an algebra-anticoalgebra involution of ${{\frak S}(A_{N-1};t)_{\epsilon}}$, which satisfies .
Since ${\det}^{{{\intercal}}}$ is a group-like element of ${{\frak A}(w_{N,t,\epsilon})}$, we have $${\det}^{{{\intercal}}} = g \det;
\quad
g = \sum_{\lambda, \mu \in {\EuScript V}}
\frac{c(\lambda)}{c(\mu)} \, {{\stackrel{\scriptscriptstyle\circ}{e}}}_{\lambda} e_{\mu}$$ by , where $c(\lambda)$ $(\lambda \in {\EuScript V})$ denotes some nonzero constant. Since both $\det$ and ${\det}^{{{\intercal}}}$ are central and $\det$ is not a zero divisor, $g$ is central. Hence, by Lemma \[irrG1\] and Schur’s lemma, we have ${\bold p}\, g = c \, {\bold p}$ $({\bold p}\in {\EuScript G}^1)$ for some $c \in {\Bbb C}$. Hence we have $c(\lambda) = c^{| \lambda |} c(0)$ for each $\lambda \in {\EuScript V}$. In order to prove $c = 1$, we compute ${\det \binom{0}{\hat{1}}}^{{{\intercal}}}$ in two ways. Using $$[d({\bold p}) - 1] \, e \! \binom{{\bold p}}{{\bold q}} =
- \epsilon [d({\bold p}) + 1] \, e \! \binom{{\bold p}^{\dag}}{{\bold q}}
\quad ({\bold p}\in \ {{\EuScript G}}^2[\;\searrow\;], \,
{\bold q}\in {{\EuScript G}}^2[\;\downarrow\;]),$$ we obtain $$\sum_{i=1}^{k} {(- \epsilon)^{\EuScript{L} ( {\bold p}_i )}}
e \binom{{\bold p}_i}{0\, |\, 1, \ldots, N} =
{(- \epsilon)^{\EuScript{L} ( {\bold p}_k )}} [k]^2 \,
e \binom{{\bold p}_k}{0\, |\, 1, \ldots, N}$$ by induction on $k$, where ${\bold p}_i = (\hat{1}\, |\, 2, \ldots, i, 1, i+1, \ldots, N)$. Substituting $k = N$ in this equality, we get $$\begin{aligned}
\det {\binom{0}{\hat{1}}}^{{{\intercal}}} & = \,\,
c^{|\hat{1}| - |0|}
\det \binom{\hat{1}}{0}
\\
& =
(- \epsilon)^{N(N-1)/2 + {\EuScript{L}}{({\bold p}_N)}}
c [N]\, e \!
\begin{pmatrix}
\hat{1}\,|\!\!\! & 2,\!\!\! & 3,\!\!\! & \ldots,\!\!\! & N,\!\!\! & 1 \\
0\,|\!\!\! & 1,\!\!\! & 2,\!\!\! & \ldots,\!\!\! & N - 1,\!\!\! & N \\
\end{pmatrix}.
\end{aligned}$$ On the other hand, using and , we see that the right-hand side of the above equality agrees with $c \,{\det \binom{0}{\hat{1}}}^{{{\intercal}}}$. This completes the proof of the proposition.
Next suppose $t = \exp (\pm \frac{\pi i}{N + L})$, or $- \exp (\pm \frac{\pi i}{N + L})$ with $N + L \in 2 {\Bbb Z}$. Then, we have $\kappa ({\bold p}) > 0$ for each ${\bold p}\in {\EuScript G}^m$ $(m > 0)$. Moreover, for each $\zeta$ with $| \zeta | = 1$, the Boltzmann weight $w_{N,t}$ satisfies $$\label{wNt*}
w_{N,t,\epsilon}^{-1} \!
\left[ {\bold r}\frac[0pt]{{\bold p}}{{\bold s}} {\bold q}\right] =
\left(
\frac{\kappa ({\bold r}\cdot {\bold s})}{\kappa ({\bold p}\cdot {\bold q})}
\right)^2 \,
\overline{w_{N,t,\epsilon} \!
\left[ {\bold p}\frac[0pt]{{\bold r}}{{\bold q}} {\bold s}\right]}.$$ Similarly to Proposition \[trA\] and Proposition \[trS\], we obtain the following.
\[SSiscpt\] Set $t = \pm \exp (\pm \frac{\pi i}{N + L})$ if $N + L \in 2 {\Bbb Z}$, and $t = \exp (\pm \frac{\pi i}{N + L})$ if $N + L \in 1 + 2 {\Bbb Z}$. Then for each solution $\zeta$ of , both ${{\frak S}(A_{N-1};t)_{\epsilon,\zeta}}$ and ${{\frak A}(w_{N,t,\epsilon,\zeta})}$ are compact CQT ${\EuScript V}$-face algebras of unitary type with costar structure $$\label{costardef}
e \binom{{\bold p}}{{\bold q}}^{\times} =
\left( \frac{\kappa ({\bold p})}{\kappa ({\bold q})} \right)^2
e \binom{{\bold q}}{{\bold p}}
\quad ({\bold p}, {\bold q}\in {\EuScript G}^m, m \geq 0).$$
In fact, both ${{\frak S}(A_{N-1};t)_{\epsilon}}$ and ${{\frak A}(w_{N,t,\epsilon})}$ are spanned by unitary matrix corepresentations $[e_u \binom{{\bold p}}{{\bold q}}]_{{\bold p}, {\bold q}\in {\EuScript G}^m}$ $(m \geq 0)$ given by $$\label{esymdef}
e_u \binom{{\bold p}}{{\bold q}} =
\frac{\kappa ({\bold q})}{\kappa ({\bold p})} \,
e \binom{{\bold p}}{{\bold q}}
\quad ({\bold p}, {\bold q}\in {\EuScript G}^m, m \geq 0).$$
Antipodes and ribbon functionals
================================
The ${\EuScript V}$-face algebra ${{\frak S}(A_{N-1};t)_{\epsilon}}$ has an antipode given by $$\label{Santipode}
S \left( e \binom{\lambda\,|\,i}{\mu\,|\,j} \right) =
(- \epsilon)^{i+j}
\frac{D(\lambda + \hat{i})}{D(\mu + \hat{j})}
\sum_{{\bold a}} (- \epsilon)^{{\EuScript{L}}({\bold a}) + {\EuScript{L}}({\bold b})} e \binom{{\bold a}}{{\bold b}},
$$ where ${\bold b}$ denotes an arbitrary element of ${\EuScript G}^{N-1}_{\lambda + \hat{i}\,\lambda}$ and the summation is taken over all ${\bold a}\in {\EuScript G}^{N-1}_{\mu + \hat{j}\, \mu}$. Moreover, we have $$\label{S2formula}
S^2 \left( e \binom{{\bold p}}{{\bold q}} \right) =
\frac{D({\frak {r}}({\bold p}))D({\frak {s}}({\bold q}))}{D({\frak {s}}({\bold p}))D({\frak {r}}({\bold q}))} \,
e \binom{{\bold p}}{{\bold q}}
\quad ({\bold p}, {\bold q}\in {\EuScript G}^m, m \geq 0).$$
(cf. [@TakeuchiMat], [@qcg]) Let $Y_{(\lambda\, |\,i)\,(\mu\, |\,j)}$ denote the right-hand side of viewed as an element of ${{\frak A}(w_{N,t,\epsilon})}$. By [@cpt §7], it suffices to verify that $$\label{cofactor}
\sum_{{\bold r}\in {\EuScript G}^1} Y_{{\bold p}{\bold r}} X_{{\bold r}{\bold q}} =
\delta_{{\bold p}{\bold q}} e_{{\frak {r}}({\bold p})} \det,
\quad
\sum_{{\bold r}\in {\EuScript G}^1} X_{{\bold p}{\bold r}} Y_{{\bold r}{\bold q}} =
\delta_{{\bold p}{\bold q}} {{\stackrel{\scriptscriptstyle\circ}{e}}}_{{\frak {s}}({\bold p})} \det,$$ where $X_{{\bold p}{\bold q}}$ $=$ $e \binom{{\bold p}}{{\bold q}}$ $({\bold p}, {\bold q}\in {\EuScript G}^1)$. We define a basis $\bigl\{ \bar{\omega} (\tilde{{\bold p}})
\bigm| {\bold p}\in {\EuScript G}^1 \bigr\}$ of $\Omega^{N-1}$ by $\bar{\omega} (\lambda + \hat{i}, \lambda) =
(- \epsilon)^{i-1} D(\lambda + \hat{i})
\omega_{N-1} (\lambda + \hat{i}, \lambda)$, where $\tilde{{\bold p}} = (\mu, \lambda) \in B \Omega^{N-1}$ for ${\bold p}= (\lambda, \mu) \in {\EuScript G}^1$. Then, the coaction of ${{\frak A}(w_{N,t,\epsilon})}$ on $\Omega^{N-1}$ is given by $\bar{\omega} (\tilde{{\bold q}}) \mapsto
\sum_{{\bold p}\in {\EuScript G}^1} \bar{\omega} (\tilde{{\bold p}}) \otimes Y_{{\bold q}{\bold p}}$ and the multiplication of $\Omega$ gives maps $$\begin{gathered}
\Omega^{N-1} \bar{\otimes} \Omega^1 \to \Omega^N;
\quad
\bar{\omega} (\tilde{{\bold q}}) \otimes \omega({\bold p}) \mapsto
\delta_{{\bold p}{\bold q}} \bar{\omega} ({\frak {r}}({\bold p})), \\
\Omega^1 \bar{\otimes} \Omega^{N-1} \to \Omega^N;
\quad
\omega({\bold p}) \otimes \bar{\omega} (\tilde{{\bold q}}) \mapsto
\delta_{{\bold p}{\bold q}} (- \epsilon)^{N-1}
\frac{D({\frak {r}}({\bold p}))}{D({\frak {s}}({\bold p}))}
\bar{\omega} ({\frak {s}}({\bold p})).
\end{gathered}$$ Since these maps are compatible with the coaction of ${{\frak A}(w_{N,t,\epsilon})}$, we have the first formula of and $$\sum_{{\bold r}} W_{{\bold r}{\bold p}} Y_{{\bold q}{\bold r}} =
\delta_{{\bold p}{\bold q}} e_{{\frak {s}}({\bold p})} \det,$$ where $W_{{\bold p}{\bold q}} \in {{\frak A}(w_{N,t,\epsilon})}$ denotes the right-hand side of . Applying ${{\intercal}}$ to this equality, we obtain $$\sum_{{\bold r}} X_{{\bold p}{\bold r}} Z_{{\bold r}{\bold q}} =
\delta_{{\bold p}{\bold q}} {{\stackrel{\scriptscriptstyle\circ}{e}}}_{{\frak {s}}({\bold p})} \det;
\quad
Z_{{\bold p}{\bold q}} =
\frac{D({\frak {r}}({\bold p})) D({\frak {s}}({\bold q})) \kappa ({\bold p})^2}
{D({\frak {s}}({\bold p})) D({\frak {r}}({\bold q})) \kappa ({\bold q})^2} \,
Y_{{\bold q}{\bold p}}^{{{\intercal}}}.$$ Computing $\sum_{{\bold r}{\bold s}} Y_{{\bold p}{\bold r}} X_{{\bold r}{\bold s}} Z_{{\bold s}{\bold q}}$ in two ways, we obtain $Y_{{\bold p}{\bold q}} = Z_{{\bold p}{\bold q}}$. This proves the second equality of . Finally, Computing $\sum_{{\bold r}{\bold s}} W_{{\bold r}{\bold q}} S ( X_{{\bold s}{\bold r}} ) S^2 (X_{{\bold p}{\bold s}})$ in the algebra ${{\frak S}(A_{N-1};t)_{\epsilon}}$ in two ways, we obtain .
\[SSrib\] For each $t$ and $\zeta$, ${{\frak S}(A_{N-1};t)_{\epsilon,\zeta}}$ becomes a coribbon Hopf ${\EuScript V}$-face algebra, whose braiding, antipode $S$ and modified ribbon functional $\cal{M} = \cal{M}_{1}$ are given by , and the following formulas: $$\label{Mformula}
\cal{M} \left( e \binom{{\bold p}}{{\bold q}} \right) =
\delta_{{\bold p}{\bold q}} \,
\frac{D({\frak {r}}({\bold p}))}{D({\frak {s}}({\bold p}))}
\quad ({\bold p}, {\bold q}\in {\EuScript G}^m, m \geq 0).$$ Moreover, we have $$\label{d=D}
\dim_q (L_{\lambda}) = D(\lambda)$$ for each $\lambda \in {\EuScript V}$. When $N$ is even, there exists another ribbon functional $\cal{M}_{-1}$ given by $$\label{M-formula}
\cal{M}_{-1} \left( e \binom{{\bold p}}{{\bold q}} \right) =
\delta_{{\bold p}{\bold q}} \,
(-1)^{ | {\frak {r}}({\bold p}) | - | {\frak {s}}({\bold p}) | }
\frac{D({\frak {r}}({\bold p}))}{D({\frak {s}}({\bold p}))}
\quad ({\bold p}, {\bold q}\in {\EuScript G}^m, m \geq 0).$$ The quantum dimension of the corresponding ribbon category is given by $$\label{d=-D}
\dim_q^{-1} (L_{\lambda}) = (-1)^{| \lambda |} D(\lambda).$$
Let $\cal{M} \in {\frak H}^*$ be as in . Using , we see that $\cal{M}$ satisfies . Hence, it suffices to verify that $$\label{U1U2=M2}
(\cal{U}_1 \cal{U}_2) \left( e \binom{{\bold p}}{{\bold q}} \right)
= \cal{M}^2 \left( e \binom{{\bold p}}{{\bold q}} \right)
\quad ({\bold p}, {\bold q}\in {\EuScript G}^m).$$ for each $m \geq 0$. By and , $\cal{U}_1 \cal{M}^{-1}$ is a central element of ${{\frak A}(w_{N,t,\epsilon})}^*$. Hence by Lemma \[irrG1\] and Schur’s lemma, we have $\cal{U}_1 {\bold p}= \vartheta \cal{M} {\bold p}$ $({\bold p}\in {\EuScript G}^1)$ for some constant $\vartheta$. Using , and , we compute $$\begin{aligned}
\label{Uformula}
\cal{U}_1^{-1} \left( e \binom{0\, |\, 1}{0\, |\, 1}
\right) = &
\sum_{{\bold r}\in {\EuScript G}^1}
{{\cal R}^+}\left( S^2 \left( e \binom{{\bold r}}{0\, |\, 1} \right),\,
e \binom{0\, |\, 1}{{\bold r}} \right) \\
= &
\sum_{\nu = 2 \hat{1}, \hat{1} + \hat{2}}
\frac{D(0)D(\nu)}{D(\hat{1})^2}
w_{N,t}
\begin{bmatrix}
0 & \hat{1} \\
\hat{1} & \nu
\end{bmatrix} \\
= & \;
\zeta^{-1} t^{N} \frac{1}{[N]}.
\end{aligned}$$ This shows that $\vartheta = \zeta t^{-N}$, and similarly, we obtain $\cal{U}_2 {\bold p}= \vartheta^{-1} \cal{M} {\bold p}$ $({\bold p}\in {\EuScript G}^1)$. This proves for $m = 1$. For $m \geq 2$, follows from for $m = 1$ by induction on $m$, using the fact that both $\cal{M}^2$ and $\cal{U}_1 \cal{U}_2$ are group-like (cf. , ). The second assertion follows from and Lemma \[Trq\].
We denote by ${{\frak S}(A_{N-1};t)_{\epsilon,\zeta}^{\iota}}$ the coribbon Hopf ${\EuScript V}$-face algebra $({{\frak S}(A_{N-1};t)_{\epsilon,\zeta}}, \cal{M}_{\iota})$, where $\iota = \pm 1$ if $N$ is even and $\iota = 1$ if $N$ is odd.
If $t = \exp (\pm \frac{\pi i}{N + L})$, or $t = - \exp (\pm \frac{\pi i}{N + L})$ and $N$ is odd, then we have $D (\lambda ) > 0$ for every $\lambda \in {\EuScript V}$. If $t = - \exp (\pm \frac{\pi i}{N + L})$ and $N$ is even, then we have $(-1)^{| \lambda |} D (\lambda ) > 0$.
Straightforward.
When $t = \exp (\pm \frac{\pi i}{N + L})$, or $t = - \exp (\pm \frac{\pi i}{N + L})$ and $N, L \in 1 + 2 {\Bbb Z}$, the Woronowicz functional $\cal{Q}$ of ${{\frak S}(A_{N-1};t)_{\epsilon}}$ is given by . While when $t = - \exp (\pm \frac{\pi i}{N + L})$ and $N, L \in 2 {\Bbb Z}$ , $\cal{Q}$ is given by .
We will prove the first assertion. We set $$M =
\left[ \cal{M} \left(
e_u \binom{{\bold p}}{{\bold q}}
\right) \right]_{{\bold p}, {\bold q}\in {\EuScript G}^1}
\quad
Q =
\left[ \cal{Q} \left(
e_u \binom{{\bold p}}{{\bold q}} \right) \right]_{{\bold p}, {\bold q}\in {\EuScript G}^1}$$ where $\cal{Q}$ denotes the Woronowicz functional of ${{\frak S}(A_{N-1};t)_{\epsilon}}$ and $e_u \binom{{\bold p}}{{\bold q}}$ is as in . By , and Lemma \[irrG1\], we have $M = \vartheta Q$ for some $\vartheta$. Since $M$ is positive by and the lemma above, we have $\vartheta > 0$. Since the quantum dimension satisfies $\dim_q L = \dim_q L\spcheck$ for every $L$, we have $\mathrm{Tr} (M) = \mathrm{Tr} (M^{-1})$. By , this proves $M = Q$. Now the assertion $\cal{M} \bigl( e_u \binom{{\bold p}}{{\bold q}} \bigr)$ $=$ $\cal{Q} \bigl( e_u \binom{{\bold p}}{{\bold q}} \bigr)$ $({\bold p}, {\bold q}\in {\EuScript G}^m)$ easily follows from the fact that both $\cal{M}$ and $\cal{Q}$ are group-like, by induction on $m$.
The modular tensor category
===========================
Let $\cal{C}$ be a ribbon category, which is additive over a field $\Bbb{K}$. We say that $\cal{C}$ is [*semisimple*]{} if there exist a set ${\EuScript V}_{\cal{C}}$, an involution ${}^{\lor}\!: {\EuScript V}_{\cal{C}} \to {\EuScript V}_{\cal{C}}$, an element $0 \in {\EuScript V}_{\cal{C}}$ and simple objects $L_{\lambda}^{\cal{C}}$ $(\lambda \in {\EuScript V}_{\cal{C}})$ such that every object of $\cal{C}$ is isomorphic to a finite direct sum of $L_{\lambda}^{\cal{C}}$’s, and that $$L_0^{\cal{C}} \cong \bold{1},
\quad
(L_{\lambda}^{\cal{C}})^{\lor}
\cong
L_{\lambda^{\lor}}^{\cal{C}},$$ $$\cal{C} (L_{\lambda}^{\cal{C}}, L_{\mu}^{\cal{C}}) =
\begin{cases}
\Bbb{K} & (\lambda = \mu) \\
0 & (\lambda \ne \mu) \\
\end{cases}$$ for each $\lambda, \mu \in {\EuScript V}^{\cal{C}}$, where $\bold{1}$ denotes the unit object of $\cal{C}$. For a semisimple ribbon category $\cal{C}$, we define its [*fusion rule*]{} $N^{\nu}_{\lambda \mu}$ $(\lambda, \mu, \nu \in {\EuScript V}_{\cal{C}})$ and [*S-matrix*]{} $S^{\cal{C}}$ $=$ $[S^{\cal{C}}_{\lambda \mu}]_{\lambda, \mu \in {\EuScript V}_{\cal{C}}}$ by $$\begin{gathered}
[L_{\lambda}^{\cal{C}}][L_{\mu}^{\cal{C}}] =
\sum_{\nu \in {\EuScript V}_{\cal{C}}}
N^{\nu}_{\lambda, \mu} [L_{\nu}^{\cal{C}}], \\
S^{\cal{C}}_{\lambda \mu}
=
\mathrm{Tr}_q (c_{L_{\mu}^{\cal{C}} L_{\lambda}^{\cal{C}}}
\circ
c_{L_{\lambda}^{\cal{C}} L_{\mu}^{\cal{C}}}).\end{gathered}$$ By definition, we have $$\label{Sl0C}
S^{\cal{C}}_{\lambda 0} =
\dim_q L_{\lambda}^{\cal{C}}.$$ Since the twist $\theta$ satisfies $$\label{thetaVW}
c_{W V} \circ c_{V W} =
\theta_{V \otimes W} \circ
(\theta_V \otimes \theta_W)^{-1},$$ $S^{\cal{C}}$ satisfies $$\label{SlmC}
S_{\lambda \mu}^{\cal{C}} =
\sum_{\nu \in {\EuScript V}_{\cal{C}}}
\frac{\theta_{\nu}}{\theta_{\lambda} \theta_{\mu}}
N_{\lambda \mu}^{\nu}
\dim_q (L_{\nu}^{\cal{C}}),$$ where $\theta_{\lambda} \in \Bbb{K}$ is defined by $\theta_{L_{\lambda}^{\cal{C}}} =
\theta_{\lambda}
\mathrm{id}_{L_{\lambda}^{\cal{C}}}$. Moreover, $S = S^{\cal{C}}$ satisfies the following [*Verlinde’s formula*]{} (cf. [@Verlinde; @MooreSeiberg; @Turaev]): $$\label{Verlinde}
{S_{\nu,0}}
\sum_{\xi \in {\EuScript V}} N_{\lambda \mu}^{\xi}
S_{\xi \nu}
= S_{\lambda \nu} S_{\mu \nu},$$ where $\lambda$, $\mu$ and $\nu$ denote arbitrary elements of ${\EuScript V}= {\EuScript V}_{\cal{C}}$. Let $\cal{C}$ be a semisimple ribbon category. We say that $\cal{C}$ is a [*modular tensor category*]{} (or [*MTC*]{}) if ${\EuScript V}_{\cal{C}}$ is finite and the matrix $S^{\cal{C}}$ is invertible. If, in addition, $\cal{C}$ is unitary as a ribbon category, then it is called a [*unitary MTC*]{}. It is known that each (unitary) MTC gives rise to a (unitary) 3-dimensional topological quantum field theory (TQFT), hence, in particular, an invariant of 3-manifolds of Witten-Reshetikhin-Turaev type (cf. V. Turaev [@Turaev]).
The most well-known example of MTC is obtained as a certain semisimple quotient $\cal{C} (\frak{g}, \kappa)$ of a category of representations of the quantized enveloping algebra $U_q (\frak{g})$ of finite type in the case when $q$ is a root of unity [@Andersen; @GelfandKazhdan; @Kirillov; @TuraevWenzl]. When $\frak{g} = \frak{sl}_N$, the simple objects $L_{\lambda}^U$ of $\cal{C} (\frak{sl}_N, N + L)$ $(L \geq 1)$ are also indexed by the set ${\EuScript V}= {\EuScript V}_{N L}$ given by and the fusion rules agree with those of $SU(N)_L$-WZW model. The quantum dimension and the constant $\theta_{\lambda}$ for $\cal{C} (\frak{sl}_N, N + L)$ are given by
$$\label{dimqU}
\dim_q ( L_{\lambda}^U ) = D (\lambda)_{t_0},
\quad
\theta_{\lambda} = \zeta_0^{\, (\lambda | \lambda + 2 \rho )^{\sim}}$$
respectively, where $\zeta_0 = \exp (\frac{\pi i}{N (N + L)})$, $t_0 = \zeta_0^N$, $\rho = \Lambda_1 + \cdots + \Lambda_{N-1}$, $(\,|\,)^{\sim} = N (\, |\, )$ and $(\,|\,)$ denotes the usual inner product of ${\Bbb R}^N$. Moreover, the S-matrix of $\cal{C} (\frak{sl}_N, N + L)$ is given by $S^U = S^{1} (\zeta_0)$. Here, for each primitive $2 N (N + L)$-th root $\zeta$ of unity, we define the matrix $S^{\iota} (\zeta)$ by the following Kac-Peterson formula (cf. [@Kirillov]): $$\label{SlmU}
S^{\iota} (\zeta)_{\lambda \mu} =
\iota^{| \lambda | + |\mu | }
\frac{\sum_{w \in \frak{S}_N} (-1)^{l (w)}
\zeta^{-2 (w (\lambda + \rho)\, |\, \mu + \rho )^{\sim}}}
{\sum_{w \in \frak{S}_N} (-1)^{l (w)}
\zeta^{-2 (w (\rho)\, |\, \rho)^{\sim}}},$$ where $\iota = \pm 1$ if $N \in 2 {\Bbb Z}$, $\iota = 1$ if $N \in 1 +2 {\Bbb Z}$ and the action of the symmetric group $\frak{S}_N$ on $\sum_i {\Bbb C}\hat{i}$ is given by $w \hat{i} = \widehat{w(i)}$. Note that $(\lambda\, |\, \mu)^{\sim} \in {\Bbb Z}$ for every $\lambda, \mu \in \bigoplus_i {\Bbb Z}\Lambda_i$.
Let $\zeta$ be a primitive $2 N(N + L)$-th root of unity and $t = \zeta^N$. Then the matrix $S = S^{\iota}(\zeta)$ is both symmetric and invertible, and satisfies Verlinde’s formula . Moreover, we have: $$\label{Szl0}
S^{\iota}(\zeta)_{\lambda 0} =
\iota^{| \lambda |} D(\lambda)_t,$$ $$\label{Szqr}
S^{\iota}(\zeta)_{\Lambda_q \Lambda_r} =
\iota^{q + r} \sum_s
(\zeta t)^{-2qr} t^{2s (q + r -s + 1)}
D ( \Lambda_{q+r-s} + \Lambda_{s} )_t.$$ for each $\lambda \in {\EuScript V}_{NL}$ and $0 < r \leq q < N$, where the summation in is taken over $\max \{0, q + r -N \} \leq s \leq r$.
Suppose $\zeta = \zeta_0$. Then the formula follows from for $\cal{C} (\frak{sl}_N, N + L)$, and . In this case, the other assertions also follow from the results for $\cal{C} (\frak{sl}_N, N + L)$. For other $\zeta$, the assertions follow from Galois theory for ${\Bbb Q}(\zeta_0)/ {\Bbb Q}$.
Let $N_{\lambda \mu}^{\nu}$ be the fusion rules of $SU(N)_L$-WZW models and let $S$ and $S^{\prime}$ be symmetric matrices whose entries are indexed by ${\EuScript V}= {\EuScript V}_{NL}$. If these satisfy Verlinde’s formula , $ S_{\lambda 0}$ $=$ $S^{\prime}_{\lambda 0} \ne 0$ $(\lambda \in {\EuScript V})$ and $$\label{Sqr=Sqr}
\quad
S_{\Lambda_q \Lambda_r} =
S^{\prime}_{\Lambda_q \Lambda_r}
\quad (0 < r \leq q < N),$$ then we have $S = S^{\prime}$.
We recall that there exists an algebra surjection from ${\Bbb Z}[x_1, \ldots, x_N]^{\frak{S}_N}$ onto $\EuScript{F}$ (cf. Theorem \[SSrep\] (2)), which sends the Schur function $s_{(\lambda_1, \ldots, \lambda_N)}$ (see e.g. [@Macdonald]) to $[ L_{\lambda} ]$ for each $\lambda \in {\EuScript V}$ (see e.g. [@GoodmanNakanishi]). For each $\xi = (\xi_1, \ldots, \xi_m)$ such that $N > \xi_1 \geq \ldots \geq \xi_m > 0$, we define $E_{\xi} \in \EuScript{F}$ to be the image of the elementary symmetric function $e_{\xi}$ via this map, that is $E_{\xi} = [\Omega^{\xi_1}] \cdots [\Omega^{\xi_m}]$. Since $\{ e_{\xi} \}$ is a basis of ${\Bbb Z}[x_1, \ldots, x_N]^{\frak{S}_N}$, $ \{ E_{\xi} \}$ spans $\EuScript{F}$. We define the symmetric bilinear forms $\cal{S}$ and $\cal{S}^{\prime}$ on $\EuScript{F}$ by setting $$\cal{S}\!
\left( [L_{\lambda}],\, [L_{\mu}] \right)
=
S_{\lambda \mu},
\quad
\cal{S}^{\prime}\!
\left( [L_{\lambda}],\, [L_{\mu}] \right)
=
S^{\prime}_{\lambda \mu}.$$ Then, Verlinde’s formula for $S$ is rewritten as $$\label{Verlinde2}
\cal{S}\!
\left(
ab,\, \frac{[L_{\nu}]}{S_{\nu 0}}
\right) =
\cal{S}\! \left(
a,\, \frac{[L_{\nu}]}{S_{\nu 0}}
\right)
\cal{S}\! \left(
b,\, \frac{[L_{\nu}]}{S_{\nu 0}}
\right),$$ where $\nu \in {\EuScript V}$ and $a,b \in \EuScript{F}$. By and this formula, we obtain $$\cal{S}
(E_{\xi},\, [\Omega^r]) =
\cal{S}^{\prime}
(E_{\xi},\, [\Omega^r])$$ for each $\xi$ and $0 \leq r < N$, or equivalently, we obtain $$\cal{S}
([L_{\lambda}],\, [\Omega^r]) =
\cal{S}^{\prime}
([L_{\lambda}],\, [\Omega^r])$$ for each $\lambda \in {\EuScript V}$ and $0 \leq r < N$. Repeating similar consideration, we conclude that $S_{\lambda \mu}$ $=$ $S^{\prime}_{\lambda \mu}$ holds for every $\lambda, \mu \in {\EuScript V}$.
For $\frak{S} = {{\frak S}(A_{N-1};t)_{\epsilon,\zeta}^{\iota}}$, we denote the semisimple ribbon category $\bold{Com}^f_{\frak{S}}$ (resp. unitary ribbon category $\bold{Com}^{fu}_{\frak{S}}$) by $\cal{C}_{\frak{S}}(A_{N-1}, t)_{\epsilon, \zeta}^{\iota}$ (resp. $\cal{C}^u_{\frak{S}}(A_{N-1}, t)_{\epsilon, \zeta}$).
\[ComSSisMTC\] Let $N \geq 2$ and $L \geq 1$ be integers, $\iota = \pm 1$ if $N \in 2 {\Bbb Z}$, $\iota = 1$ if $N \in 1 + 2 {\Bbb Z}$ and $\epsilon = \pm 1$. Let $\zeta$ be a primitive $2 N (N + L)$-th root of unity.\
Suppose $N$ is odd or $\epsilon = 1$ and set $t = \zeta^N$. Then the category $\cal{C}_{\frak{S}}(A_{N-1}, t)_{\epsilon, \zeta}^{\iota}$ is a modular tensor category with S-matrix $S^{\iota}(\zeta)$.\
Suppose $N$ is even, $\epsilon = - 1$ and $t\,:= - \zeta^N$ is a primitive $2 (N + L)$-th root of unity. Note that this implies $L \in 2 {\Bbb Z}$ . Then the category $\cal{C}_{\frak{S}}(A_{N-1}, t)_{-1, \zeta}^{\iota}$ is a modular tensor category with S-matrix $S^{- \iota}(\zeta)$.
We will prove Part (2). By , , , and $D(\lambda)_{-t} = (-1)^{| \lambda |} D(\lambda)_t$, we have $S_{\lambda 0}^{\frak{S}}$ $=$ $S^{- \iota} (\zeta)_{\lambda 0}$. Hence by the lamma above, it suffices to show that $S_{\Lambda_q \Lambda_r}^{\frak{S}}$ $=$ $S^{- \iota} (\zeta)_{\Lambda_q \Lambda_r}$ for each $0 < r \leq q < N$. As we will see in the next section, the action of $c_{\Omega^r \Omega^q} \circ c_{\Omega^q \Omega^r}$ on the one-dimensional space $( \Omega^q \bar{\otimes} \Omega^r)
(0,\, \Lambda_{q+r-s} + \Lambda_{s})$ $=$ ${\Bbb C}\omega_q (0,\, \Lambda_q) \otimes
\omega_r (\Lambda_q,\, \Lambda_{q+r-s} + \Lambda_{s})$ is given by the scalar $(\zeta t)^{-2qr} t^{2s (q + r -s + 1)}$. Hence, by , we obtain $$\begin{gathered}
\label{SLLS=Tr}
S_{\Lambda_q \Lambda_r}^{\frak{S}} =
\sum_s
{\mathrm{Tr}}\bigl( \cal{M}_{\iota} \circ c_{\Omega^r \Omega^q}
\circ c_{\Omega^q \Omega^r}
\!\!
\bigm|_{( \Omega^q \bar{\otimes} \Omega^r)
(0,\, \Lambda_{q+r-s} + \Lambda_{s})}
\bigr) \\
= (- \iota)^{p + q} \sum_s
(\zeta t^{\prime})^{-2qr} {t^{\prime}}^{2s (q + r -s + 1)}
D(\Lambda_{q+r-s} + \Lambda_{s})_{t^{\prime}}, \\\end{gathered}$$ where $t^{\prime} = \zeta^N$ and the summation is taken over $\max \{0, q + r -N \} \leq s \leq r$. Since the right-hand side of equals $S^{- \iota} (\zeta)_{\Lambda_q \Lambda_r}$ by , this completes the proof of Part (2).
Let $N \geq 2$ and $L \geq 1$ be integers, $\epsilon = \pm 1$ and $t = \exp (\pm \frac{\pi i}{N + L})$. If $N + L \in 1 + 2{\Bbb Z}$, $\cal{C}^u_{\frak{S}}(A_{N-1}, t)_{\epsilon, \zeta}$ is a unitary MTC provided that $\epsilon = 1$ or $N$ is odd, where $\zeta$ denotes an arbitrary primitive $2 N (N + L)$-th root of unity such that $\zeta^N = t$. If $N + L \in 2{\Bbb Z}$, $\cal{C}^u_{\frak{S}}(A_{N-1}, \pm t)_{\epsilon, \zeta}$ is a unitary MTC for each primitive $2 N (N + L)$-th root $\zeta$ of unity such that $\zeta^N = \pm \epsilon^{N-1} t$.
When $N \in 1 + 2 {\Bbb Z}$, $\frak{S}(A_{N-1},t)_{-1, \zeta}^{\iota}$ is isomorphic to a 2-cocycle deformation of $\frak{S}(A_{N-1},t)_{1, \zeta}^{\iota} $. Hence $\cal{C}_{\frak{S}}(A_{N-1}, t)_{1, \zeta}^{\iota}$ and $\cal{C}_{\frak{S}}(A_{N-1}, t)_{-1, \zeta}^{\iota}$ are equivalent.
For $\frak{g}$ $=$ $\frak{so}_N$ and $\frak{sp}_N$, a category-theoretic construction of unitary MTC’s related to $\cal{C}(\frak{g}, \kappa)$ is given by Turaev and Wenzl [@TuraevWenzl2].
Braidings on $\Omega$
=====================
In this section, we give some explicit calculation of the braiding $c_{q, r} := c_{\Omega^q \Omega^r}$ in order to complete the proof of Lemma \[cformula\] and Theorem \[ComSSisMTC\]. Since the braiding is a natural transformation and the multiplication of $\Omega$ gives a ${{\frak S}(A_{N-1};t)_{\epsilon}}$-comodule map $m_{q, r}\!:
\Omega^q \bar{\otimes} \Omega^r \to \Omega^{q+r}$, $c_{q, r}$ satisfies $$\label{cOq+q'Or}
c_{q+q^{\prime}, r} \circ
(m_{q, q^{\prime}} \bar{\otimes} \mathrm{id}_{\Omega^r}) =
(\mathrm{id}_{\Omega^r}
\bar{\otimes} m_{q, q^{\prime}}) \circ
(c_{q, r} \bar{\otimes}
\mathrm{id}_{\Omega^{q^{\prime}}}) \circ
(\mathrm{id}_{\Omega^q} \bar{\otimes}
c_{q^{\prime}, r}),$$ $$\label{cOqOr+r'}
c_{q, r+r^{\prime}} \circ
(\mathrm{id}_{\Omega^q} \bar{\otimes} m_{r, r^{\prime}})
=
(m_{r, r^{\prime}} \bar{\otimes}
\mathrm{id}_{\Omega^q}) \circ
(\mathrm{id}_{\Omega^r}
\bar{\otimes} c_{q, r^{\prime}}) \circ
(c_{q, r}
\bar{\otimes} \mathrm{id}_{\Omega^{r^{\prime}}}).$$
For each $1 \leq p < N$ and $1 \leq q \leq N-p$, we have $$\begin{gathered}
\label{cOqO1}
c_{q, 1}
\bigl(
\omega \left( \Lambda_p\, |\, p+1,
\ldots, p+q \right)
\otimes \omega \left( \Lambda_{p+q}\, |\, 1 \right)
\bigr) \\
= - (- \zeta)^{-q} t^{p+1}
\frac{[q]}{[p+1]} \,
\omega \left( \Lambda_p\, |\, p+1 \right)
\otimes
\omega \left( \Lambda_{p+1}\, |\,
p+2, \ldots, p+q, 1 \right) \\
+ (- \zeta)^{-q} (- \epsilon)^q
\frac{[p+q+1]}{[p+1]} \,
\omega \left( \Lambda_p\, |\, 1 \right)
\otimes
\omega \left( \Lambda_{p} + \hat{1}\, |\,
p+1, \ldots, p+q \right). \end{gathered}$$
Suppose is valid for each $p$ and for some $q < N-p$. Then, by , we obtain $$\label{cOq+1O1}
c_{1+q, 1}
\bigl(
\omega \left( \Lambda_p\, |\, p+1,
\ldots, p+q+1 \right)
\otimes \omega \left( \Lambda_{p+q+1}\, |\, 1 \right)
\bigr) \qquad\qquad\qquad$$ $$\begin{gathered}
= \;
(\mathrm{id}_{\Omega_1} \bar{\otimes} m_{1, q})
\circ
(c_{1, 1} \bar{\otimes} \mathrm{id}_{\Omega^q})
\Bigl[ \Bigr.
\omega (\Lambda_p | p+1) \otimes \\
\bigl\{ \bigr.
- (- \zeta)^{-q} t^{p+2}
\frac{[q]}{[p+2]} \,
\omega \left( \Lambda_{p+1}\, |\, p+2 \right)
\otimes
\omega \left( \Lambda_{p+2}\, |\,
p+3, \ldots, p+q+1, 1 \right) \\
+ (- \zeta)^{-q} (- \epsilon)^q
\frac{[p+q+2]}{[p+2]} \,
\omega \left( \Lambda_{p+1}\, |\, 1 \right)
\otimes
\omega \left( \Lambda_{p+1} + \hat{1}\, |\,
p+2, \ldots, p+q+1 \right)
\bigl. \bigr\}
\Bigl. \Bigr]
\end{gathered}$$ $$\begin{gathered}
= \;
\biggl\{ \biggr.
-
t^{p+2}
\frac{[q]}{[p+2]} \,
w_{N,t}\!
\begin{bmatrix}
\Lambda_p & \Lambda_{p+1} \\
\Lambda_{p+1} & \Lambda_{p+2}
\end{bmatrix}
+
\frac{[p+q+2]}{[p+2]}
w_{N,t}\!
\begin{bmatrix}
\Lambda_p & \Lambda_{p+1} \\
\Lambda_{p+1} & \Lambda_{p+1} + \hat{1}
\end{bmatrix}
\biggl. \biggr\} \\
(- \zeta)^{-q}
\omega (\Lambda_p | p+1) \otimes
\omega \left( \Lambda_{p+1}\, |\, p+2,
\ldots, p+q+1, 1 \right) \\
+ (- \zeta)^{-q} (- \epsilon)^q
\frac{[p+q+2]}{[p+2]} \,
w_{N,t}\!
\begin{bmatrix}
\Lambda_p & \Lambda_{p+1} \\
\Lambda_p + \hat{1} & \Lambda_{p+1} + \hat{1}
\end{bmatrix} \\
\omega \left( \Lambda_p\, |\, 1 \right)
\otimes
\omega \left( \Lambda_p + \hat{1}\, |\,
p+1, \ldots, p+q+1 \right) .
\end{gathered}$$ Computing the right-hand side of the above equality, we obtain for $q+1$.
Using for $p = 1$, $q = N-1$ together with , we obtain $$ c_{N, 1}
\bigl(
\omega \left( 0 |\, 1,
\ldots, N \right)
\otimes \omega \left(0\, |\, 1 \right)
\bigr)
=
- (- \zeta)^{- N} t\,
[N]\, \omega \left( 0\, |\, 1 \right)
\otimes
\omega \left( 1\, |\, 2, \ldots, N, 1 \right).$$ This shows that the constant $\vartheta$ in equals $\epsilon^{N-1} \zeta^{- N} t$ and completes the proof of Lemma \[cformula\].
We have the following relations: $$\begin{gathered}
\label{cOqO1*}
c_{q, 1}
\Bigl(
\omega \left( \Lambda_p\, |\, p+1,
\ldots, p+s, 1, \ldots, q-s \right)
\otimes \omega \left(
\Lambda_{p+s} + \Lambda_{q-s}\, |\, q-s+1 \right)
\Bigr) \\
\in - (- \zeta)^{-q} t^{p-q+s+1}
\frac{[s]}{[p+1]}\,
\omega \left( \Lambda_p\, |\, p+1 \right)
\otimes
\omega \left( \Lambda_{p+1}\, |\,
p+2 \ldots, p+s, 1, \ldots, q-s+1 \right) \\
+
\Omega^1 \left( \Lambda_p\, |\, 1 \right)
\otimes
\Omega^q \left( \Lambda_{p} + \hat{1},
\Lambda_{p+s} + \Lambda_{q-s+1} \right) \\
(1 \leq p < N, 1 \leq s \leq N-p, s < q \leq p+2s-1),
\end{gathered}$$ $$\begin{gathered}
\label{cOqOr}
c_{q, r}
\Bigl(
\omega \left( \Lambda_p\, |\, p+1,
\ldots, p+q \right)
\otimes \omega \left(
\Lambda_{p+q}\, |\, 1, \ldots, r \right)
\Bigr) \\
\in (- 1)^r (- \zeta)^{- qr} t^{pr+r}
\frac{[q]! \, [p]!}{[p+r]! \, [q-r]!} \,
\omega \left( \Lambda_p\, |\, p + 1, \ldots, p + r \right)
\qquad\qquad \\
\qquad\qquad \otimes
\omega \left( \Lambda_{p+r}\, |\,
p+r+1, \ldots, p+q, 1, \ldots, r \right) \\
\qquad\qquad\qquad\qquad +
\sum_{\lambda \ne \Lambda_{p+r}}
\Omega^r
\left( \Lambda_p, \lambda \right)
\otimes
\Omega^q \left( \lambda,
\Lambda_{p+q} + \Lambda_{r} \right) \\
(0 \leq p \leq N-1, 0 \leq q \leq N-p, 0 \leq r \leq q),
\end{gathered}$$ where $[n]! = [n] \cdots [2][1]$ and $[0]! = 1$.
The relation follows from , and $$\begin{gathered}
c_{q-s, 1}
\bigl(
\omega \left( \Lambda_{p+s}\, |\, 1,
\ldots, q-s \right)
\otimes \omega \left( \Lambda_{p+s} + \Lambda_{q-s}\,
|\, q-s+1 \right)
\bigr) \\
= (- \zeta t)^{s - q}
\omega \left( \Lambda_{p+s}\, |\, 1 \right)
\otimes
\omega \left( \Lambda_{p+s} + \hat{1}\, |\,
2, \ldots, q-s+1 \right).
\end{gathered}$$ The relation is easily proved by induction on $r$, using and .
Using , and $$\begin{gathered}
c_{q, r-s} \bigl(
\omega \left( 0\, |\, 1, \ldots, q \right)
\otimes \omega \left( \Lambda_q\, |\, q+1,
\ldots, q+r-s \right) \bigr) \\
= (- \zeta t )^{q s - q r}
\omega \left( 0\, |\, 1, \ldots, r - s \right)
\otimes \omega \left( \Lambda_{r-s}\, |\, r-s+1,
\ldots, q+r-s \right), \end{gathered}$$ we obtain the following.
For each $0 \leq q, r < N$ and $\max \{ 0, q+r-N \} \leq s \leq \min \{q,r\}$, we have $$\begin{gathered}
c_{q, r}
\bigl(
\omega \left( 0\, |\, 1, \ldots, q \right)
\otimes \omega \left( \Lambda_q\, |\, q+1,
\ldots, q+r-s, 1, \ldots s
\right)
\bigr) \\
= \; (-1)^s (- \zeta t)^{- qr} t^{s(q+r-s+1)}
\frac{[q]!\,[r-s]!}{[r]!\,[q-s]!} \,
\qquad \qquad\qquad \\ \qquad\qquad
\omega \left( 0\, |\, 1, \ldots, r \right)
\otimes \omega \left( \Lambda_r\, |\, r+1,
\ldots, q+r-s, 1, \ldots, s
\right). \\
\end{gathered}$$
As an immediate consequence of the lemma above, we see that $c_{r,q} \circ c_{q,r}$ acts on $( \Omega^q \bar{\otimes} \Omega^r)
(0,\, \Lambda_{q+r-s} + \Lambda_{s})$ as the scalar $(\zeta t)^{- 2 qr} t^{2 s(q+r-s+1)}$. Thus we complete the proof of Theorem \[ComSSisMTC\].
ABF models and $SU(2)_L$-SOS algebras
=====================================
In this section, we give an explicit description of the representation theory of ${{\frak S}(A_{1};t)_{\epsilon}}$. We identify ${\EuScript G}= {\EuScript G}_{2,L}$ with the Dynkin diagram of type $A_{L+1}$: $$\begin{matrix}
& & & & & & & \\
\scriptstyle{0} & & \scriptstyle{1} & & &
\scriptstyle{L - 1} & & \scriptstyle{L} \quad\,\, \\
\circ & \overrightarrow{ \longleftarrow \!\!\!-\!\!\!-\!\!\!-\!\!\!-} &
\circ & \overrightarrow{ \longleftarrow \!\!\!-\!\!\!-\!\!\!-\!\!\!-} &
\cdot \quad \cdot \quad \cdot &
\circ & \overrightarrow{ \longleftarrow \!\!\!-\!\!\!-\!\!\!-\!\!\!-} &
\circ \quad .
\end{matrix}$$ Also, we identify ${\EuScript V}$ and ${\EuScript G}_{ij}^k$ with $\left\{ 0, 1, \cdots, L \right\}$ and $$\left\{ \left. (i_0, i_1, \cdots, i_k) \right| \,
0 \leq i_0, \cdots, i_k \leq L, \,\,
\left| i_{\nu} - i_{\nu - 1} \right| = 1 \,
(1 \leq \nu \leq k) \right\}$$ respectively. We define the set $\cal{B}$ by $$\cal{B} = \left\{ \left. \binom{k}{ij} \right| \,
{{i, j, k \in {\EuScript V}, \, |i - j | \leq k \leq i + j,}
\atop {i + j + k \in 2 \Bbb{Z}, \, i + j + k \leq 2L}}
\right\}.$$ Then, we have $$N_{ij}^k =
\begin{cases}
1 & \left( \binom{k}{ij} \in \cal{B} \right) \\
0 & \left( \text{otherwise} \right)
.
\end{cases}
$$ In order to simplify the formula for quantum invariants stated in the introduction, we use the rational basis of type $\Sigma$ instead of type $\Omega$ (cf. Remark \[rembase\]). The corresponding Boltzmann weight $w = w_{2,t,\epsilon}^{\Sigma}$ is given by $$w
\begin{bmatrix}
i & i \pm 1 \\
i \pm 1 & i
\end{bmatrix}
= - \zeta^{-1}\, \frac{\pm t^{\mp (i + 1)}}{[i + 1]},
\quad
w
\begin{bmatrix}
i & i \pm 1 \\
i \mp 1 & i
\end{bmatrix}
= \zeta^{-1} \epsilon\, \frac{[i + 1 \pm 1]}{[i + 1]},$$ $$w
\begin{bmatrix}
i & i \pm 1 \\
i \pm 1 & i \pm 2
\end{bmatrix}
= \zeta^{-1} t,
\quad
w \Bigl[ \text{otherwise} \Bigr] = 0.$$ Next, we recall a realization of $L_k$ introduced in [@subf]. Let $\Sigma$ be an algebra generated by the symbols $\sigma ({\bold p})$ $({\bold p}\in {\EuScript G}^k, k \geq 0)$ with defining relations: $${\sum_{k \in {\EuScript V}}}\sigma (k) \, = \, 1,$$ $$\sigma ({\bold p}) \sigma ({\bold q}) \, = \,
\delta_{{\frak {r}}({\bold p}) {\frak {s}}({\bold q})} \, \sigma ({\bold p}\cdot {\bold q}),$$ $$\qquad
\sigma (i, i+ 1, i) = \epsilon \sigma (i, i - 1, i)
\quad (0 < i < L),$$ $$\sigma (0, 1, 0) = \sigma (L, L - 1, L) = 0.$$ We define the grading $\Sigma = \bigoplus_{k \geq 0} \Sigma^k$ via $\Sigma^k =
\mathrm{span} \{ \sigma({\bold p}) \left. \right| {\bold p}\in {\EuScript G}^k \}$. Then each component $\Sigma^k$ becomes a right ${{\frak S}(A_{1};t)_{\epsilon}}$-comodule via $$\sigma ({\bold q}) \mapsto
\sum_{{\bold p}\in {\EuScript G}^k} \sigma ({\bold p}) \otimes e \binom{{\bold p}}{{\bold q}}
\quad ({\bold q}\in {\EuScript G}^k).$$ For each $\binom{k}{ij} \in \cal{B}$, the element $\sigma_k (i,j)\!:= \epsilon^{{\EuScript{L}}({\bold q})}\sigma ({\bold q})$ does not depend on the choice of ${\bold q}\in {\EuScript G}^k_{ij}$, where ${\EuScript{L}}$ is as in §7, that is, $${\EuScript{L}}(i, i - 1, \cdots, (i + j - k)/2, \cdots, j - 1, j)
= 0,$$ $${\EuScript{L}}\left( ( \cdots, n, n + 1, n, \cdots ) \right)
= {\EuScript{L}}\left( (\cdots, n, n - 1, n, \cdots) \right) + 1.
\label{sgndef}$$ It is easy to see that $\Sigma^k (i,j) = \Bbb{C} \sigma_k (i,j)$ for each $\binom{k}{ij} \in \cal{B}$ and that $\{ \left. \sigma_k (i,j) \right|
\binom{k}{ij} \in \cal{B} \}$ is a linear basis of $\Sigma$. Since $\dim \Sigma^k (0,l) = \dim \Sigma^k (l,0) = \delta_{kl}$, we have $\Sigma^k \cong L_{k} \cong (\Sigma^k) \spcheck$ by Theorem \[SSrep\] and . More explicitly, we have the following.
The map $\Sigma^k \to (\Sigma^k)^{\lor};$ $\sigma_k (i,j) \mapsto c \binom{k}{ij}
\sigma^{\lor}_k (i,j)$ gives an identification of ${{\frak S}(A_{1};t)_{\epsilon}}$-comodules, where $\{ \sigma^{\lor}_k (j,i) \}$ denotes the dual basis of $\{ \sigma_k (i,j) \}$ and the constant $c \binom{k}{ij}$ is given by $$\label{ckijformula}
c \binom{k}{ij}
=
(- \epsilon)^{(i-j)/2}
\frac{[(i + j + k)/2 + 1]! \, [(i - j + k)/2]! \, [(- i + j + k)/2]!}
{[i + 1] \, [(i + j - k)/2]!}.$$ Under this identification, the maps $d_{\Sigma^k}$ and $b_{\Sigma^k}$ in - are given by $$\label{dSigma}
d_{\Sigma^k} (i) =
\sum_j c \binom{k}{ji}^{-1}
\sigma_k (i,j) \otimes \sigma_k (j,i),$$ $$b_{\Sigma^k}
\left( \sigma_k (i,j) \otimes \sigma_k (j,l) \right)
= \delta_{il}\, c \binom{k}{ij}\, i$$ respectively, where the summation in is taken over all $j \in {\EuScript V}$ such that $\binom{k}{ji} \in \cal{B}$.
It suffices to show the first assertion. Since $\Sigma^k (i,j) = {\Bbb C}\sigma_k (i,j)$ and $(\Sigma^k) \spcheck (i,j)$ $ =$ ${\Bbb C}\sigma^{\lor}_k (i,j)$, there exists an isomorphism $\Sigma^k \cong (\Sigma^k) \spcheck$ of the form $\sigma_k (i,j) \mapsto c \binom{k}{ij} \sigma^{\lor}_k (i,j)$ for some nonzero constant $c \binom{k}{ij}$ $(\binom{k}{ij} \in \cal{B})$. To compute $c \binom{k}{ij}$, we consider the ${{\frak S}(A_{1};t)_{\epsilon}}$-right module structure on $\Sigma$ given by . Similarly to Lemma \[actOmega\], we obtain $$\label{sijeij}
\sigma_k (i, j)\, e\! \binom{i,\, i \pm 1}{j,\, j \pm 1} =
\zeta^{- k} {\epsilon}^{(\pm i \mp j + k)/2}
t^{(\mp i \pm j + k)/2}
\frac{[ \frac{i + j \mp k}{2} + 1]}{[i + 1]}
\sigma_k (i \pm 1, j \pm 1).$$ On the other hand, by , we obtain $$\label{Sformulasu2}
S \left(
e\! \binom{i,\, i \pm 1}{j,\, j \pm 1} \right) =
\frac{[j + 1]}{[i + 1]}\,
e\! \binom{j \pm 1,\, j}{i \pm 1,\, i},
\quad
S \left(
e\! \binom{i,\, i \pm 1}{j,\, j \mp 1} \right) =
- \epsilon\, \frac{[j + 1]}{[i + 1]}\,
e\! \binom{j \mp 1,\, j}{i \pm 1,\, i}.$$ Using this together with and ${}_-$, we obtain $$\label{svijei+j+}
\sigma^{\lor}_k (i,j)\,
e\! \binom{i, i+1}{j, j+1} =
\zeta^{- k} {\epsilon}^{(i - j + k)/2}
t^{(- i + j + k)/2}
\frac{[\frac{i + j + k}{2} + 2]}{[i + 2]}
\sigma^{\lor}_k (i + 1, j + 1).$$ By ${}_+$ and , we obtain $$\frac{c \binom{k}{i + 1, j + 1}}{c \binom{k}{i, j}}
=
\frac{[i + 1]\, [(i + j + k)/2 + 2]}{[i + 2]\, [(i + j - k)/2 + 1]}.$$ Similarly, by computing $\sigma_k (i, j)\, e\! \binom{i, i \pm 1}{j, j \mp 1}$, we obtain $$\frac{c \binom{k}{i \pm 1, j \mp 1}}{c \binom{k}{i, j}}
= - \epsilon\,
\frac{[i + 1]\, [(\pm i \mp j + k)/2 + 1]}{[i + 1 \pm 1]\, [(\mp i \pm j + k)/2]}.$$ By solving these recursion relations under some initial condition, we get .
The braiding $c\!: \Sigma^{m} \bar{\otimes} \Sigma^{n} \tilde{\to}
\Sigma^{n} \bar{\otimes} \Sigma^{m}$ and its inverse are given by $$c^{\pm 1} \left( \sigma_m (h,i) \otimes \sigma_n (i,k) \right)
=
\sum_j w^{\pm}_{mn} { \!\! \left[ {h \,\, i} \atop {j \,\, k} \right]}
\sigma_n (h,j) \otimes \sigma_m (j,k),
\label{wmndef}$$ $$\label{wmnformula}
w^{\pm}_{mn} { \!\! \left[ {h \,\, i} \atop {j \,\, k} \right]}
:= \sum_{\bold{u} \in {\EuScript G}_{hj}^n} \sum_{\bold{v} \in {\EuScript G}_{jk}^m}
\epsilon^{{\EuScript{L}}(\bold{u}) + {\EuScript{L}}(\bold{v}) + {\EuScript{L}}({\bold q}) + {\EuScript{L}}({\bold r})}
w^{\pm} \!\!
\left[ \bold{u} \,\, \frac[0pt]{{\bold q}}{\bold{v}} \,\, {\bold r}\right],$$ where ${\bold q}$ and ${\bold r}$ denote arbitrary elements of ${\EuScript G}^m_{hi}$ and ${\EuScript G}^n_{ik}$ respectively, and the summation in is taken over all $j \in {\EuScript V}$ such that $\binom{n}{hj}, \binom{m}{jk} \in \cal{B}$.\
The ribbon functional of $\frak{S}(A_1, t)_{\epsilon, \zeta}^{\iota}$ acts on $L_i$ as the scalar $\theta_i^{-1}$ given by $$\label{thetasu2}
\theta_i = \iota^i \zeta^{i (i + 2)}.$$
Since the braiding is a natural transformation and the map ${\Bbb C}{\EuScript G}^m \to \Sigma^m$; ${\bold p}\mapsto \sigma ({\bold p})$ is a ${{\frak S}(A_{1};t)_{\epsilon}}$-comodule map, we have $$c^{\pm 1} \left( \sigma_m (h,i) \otimes \sigma_n (i,k) \right)
=
\sum_j
\sum_{\bold{u} \in {\EuScript G}_{hj}^n} \sum_{\bold{v} \in {\EuScript G}_{jk}^m}
\epsilon^{{\EuScript{L}}({\bold q}) + {\EuScript{L}}({\bold r})}
w^{\pm} \!\!
\left[ \bold{u} \,\, \frac[0pt]{{\bold q}}{\bold{v}} \,\, {\bold r}\right]
\sigma (\bold{u}) \otimes \sigma (\bold{v}).$$ Since $\sigma ({\bold p}) = \epsilon^{{\EuScript{L}}(\bold{{\bold p}})}
\sigma_m (i, j)$ for each ${\bold p}\in {\EuScript G}^m_{ij}$, this proves Part (1). When $i = 1$, follows from the proof of Proposition \[SSrib\]. For $i > 1$, follows from by induction on $i$.
The state sum invariants
========================
Let $L$ be a positive integer and let $\epsilon, \iota$ be elements of $\{ \pm 1\}$ such that $\epsilon = 1$ if $L$ is a odd integer. Let $t$ be a primitive $2 (L + 2)$-th root of unity and $\zeta$ a solution of $\zeta^2 = \epsilon\, t$. Applying the general theory of TQFT to the MTC $\cal{C}_{\frak{S}}(A_1, t)_{\epsilon, \zeta}^{\iota}$, we obtain an invariant $\tau_{\epsilon, \zeta}^{\iota}$ of oriented 3-manifolds. To give an explicit description of $\tau_{\epsilon, \zeta}^{\iota}$, we prepare some terminologies on link diagrams. Let $D$ be a generic link diagram in $\Bbb{R} \times (0,\,1)$ (viewed as a union of line segments $\overline{AB}$), which presents a framed link $L$ with components $K_1, \ldots, K_p$ (see e.g. [@Kassel]). A point of $D$ is called [*extremal*]{} if the height function on $D$ attains its local maximum or local minimum in this point, where the height function $\mathrm{ht}$ is the restriction of the projection $\Bbb{R} \times (0,\,1) \to (0,\,1)$ on $D$. A point of $D$ is called [*singular*]{} if it is either an extremal point or a crossing point. We denote by $\sharp D$ the set of all singular points of $D$. Let $\cal{E}$ be the set of all connected components of $D \setminus \sharp D$. We say that $E \in \cal{E}$ [*belongs*]{} to $K_q$ ($1 \leq q \leq p$) if $E$ is a subset of the image of $K_q$ via the projection $L \to D$. Let c be a map from $\{K_1, \ldots, K_p \}$ to ${\EuScript V}$. We say that a map $\lambda\!: \cal{E} \to \cal{B};$ $E \mapsto \binom{\lambda_1 (E)}{\lambda_2 (E) \lambda_3 (E)}$ is a [*state*]{} on $D$ of [*color*]{} $c$ if $\lambda_1 (E) = c(K_q)$ for each component $K_q$ and $E \in \cal{E}$ belonging to $K_q$. We denote by $\bold{c}_{\lambda}$ the color of a state $\lambda$, and by $\cal{S} (D)$ the set of all states on $D$. Figure (A) shows a state $\lambda$ on a diagram of the Hopf link with $6$ singular points, such that $\bold{c}_{\lambda}(K_1) =1$, $\bold{c}_{\lambda}(K_2) =2$.
(400,380)(-70,-80) (120,0) (120,240) (180,60) (180,180) (240,0) (240,240) (90,110)[(30,20)\[l\][$\binom{2}{31}$]{}]{} (245,110)[(30,20)\[l\][$\binom{1}{23}$]{}]{} (125,30)[(30,20)\[l\][$\binom{1}{12}$]{}]{} (212,30)[(30,20)\[l\][$\binom{2}{02}$]{}]{} (125,192)[(30,20)\[l\][$\binom{1}{01}$]{}]{} (212,192)[(30,20)\[l\][$\binom{2}{11}$]{}]{} (30,174)[(30,20)\[l\][$\binom{1}{32}$]{}]{} (308,174)[(30,20)\[l\][$\binom{2}{L\, L-2}$]{}]{} (34,40)[(30,20)\[l\][$K_1$]{}]{} (312,40)[(30,20)\[l\][$K_2$]{}]{} (150,-60)[(30,20)\[l\][Figure (A)]{}]{} (0,120)[(1,1)[117]{}]{} (0,120)[(1,-1)[117]{}]{} (183,63)[(1,1)[57]{}]{} (186,54)[(1,-1)[51]{}]{} (174,66)[(-1,1)[54]{}]{} (177,57)[(-1,-1)[54]{}]{} (183,183)[(1,1)[54]{}]{} (186,174)[(1,-1)[54]{}]{} (174,186)[(-1,1)[51]{}]{} (177,177)[(-1,-1)[57]{}]{} (360,120)[(-1,1)[117]{}]{} (360,120)[(-1,-1)[117]{}]{}
\
Next, we assign a complex number $\langle \lambda | A \rangle$ for each state $\lambda \in \cal{S} (D)$ and singular point $A \in \sharp D$ as follows: When $(\lambda, A)$ is as in Figure (B) or Figure (C), then we set $$\langle \lambda | A \rangle
= c \binom{l}{hi} \delta_{hk} \delta_{ij}
\quad \mathrm{or} \quad
\langle \lambda | A \rangle
= c \binom{l}{ih}^{-1} \delta_{hk} \delta_{ij}$$ respectively, where $c \binom{l}{hi}$ is as in .
(200,95) (51.5,67.0)[(-2,-3)[20]{}]{} (53.7,67.0)[(2,-3)[20]{}]{} (51.4,65)[(5,5)\[l\][$\circ$]{}]{} (51,70)[(5,5)\[l\][$A$]{}]{} (27,19)[(30,20)\[l\][$\binom{l}{hi}$]{}]{} (70,19)[(30,20)\[l\][$\binom{l}{jk}$]{}]{} (46,0)[(30,20)\[l\][Figure (B)]{}]{} (123.8,39.9)[(2,3)[20]{}]{} (121.8,39.9)[(-2,3)[20]{}]{} (121.4,36.4)[(5,5)\[l\][$\circ$]{}]{} (121,31.4)[(5,5)\[l\][$A$]{}]{} (97,67)[(30,20)\[l\][$\binom{l}{hi}$]{}]{} (140,67)[(30,20)\[l\][$\binom{l}{jk}$]{}]{} (116,0)[(30,20)\[l\][Figure (C)]{}]{}
\
(200,95) (53.4,55.1)[(2,3)[16]{}]{} (51.8,53)[(-2,-3)[16]{}]{} (50,57.9)[(-2,3)[14.3]{}]{} (55,50.5)[(2,-3)[14.3]{}]{} (51.4,51.4)[(5,5)\[l\][$\circ$]{}]{} (44.9,51.4)[(5,5)\[l\][$A$]{}]{} (31,77)[(30,20)\[l\][$\binom{m}{hi}$]{}]{} (66,77)[(30,20)\[l\][$\binom{n}{jk}$]{}]{} (31,12)[(30,20)\[l\][$\binom{n}{cd}$]{}]{} (66,12)[(30,20)\[l\][$\binom{m}{ef}$]{}]{} (46,0)[(30,20)\[l\][Figure (D${}_{+}$)]{}]{} (125.2,57.8)[(2,3)[14.3]{}]{} (119.8,50.8)[(-2,-3)[14.3]{}]{} (121.8,54.9)[(-2,3)[15.7]{}]{} (123.6,52.9)[(2,-3)[15.7]{}]{} (121.4,51.4)[(5,5)\[l\][$\circ$]{}]{} (114.9,51.4)[(5,5)\[l\][$A$]{}]{} (101,77)[(30,20)\[l\][$\binom{m}{hi}$]{}]{} (136,77)[(30,20)\[l\][$\binom{n}{jk}$]{}]{} (101,12)[(30,20)\[l\][$\binom{n}{cd}$]{}]{} (136,12)[(30,20)\[l\][$\binom{m}{ef}$]{}]{} (116,0)[(30,20)\[l\][Figure (D${}_{-}$)]{}]{}
\
When $(\lambda, A)$ is as in Figure (D${}_{\pm}$), we set $$\langle \lambda | A \rangle =
\delta_{ij} \delta_{hc} \delta_{de} \delta_{fk}
w_{mn}^{\pm} { \!\! \left[ {h \,\, i} \atop {e \,\, f} \right]},$$ where $w_{mn}^{\pm}$ is as in .
The following result follows from §13 by a method quite similar to “vertex models on link invariants” (see e.g. [@Turaev] Appendix II), hence we omit the proof.
Let $\tau_{\epsilon, \zeta}^{\iota}$ be the invariant of closed oriented $3$-manifolds associated with the modular tensor category $\cal{C}_{\frak{S}}(A_1, t)_{\epsilon, \zeta}^{\iota}$ cf. [@Turaev]. Let $M$ be a $3$-manifold obtained by surgery on $S^3$ along a framed link $L$ with $p$ components $K_1, \ldots, K_p$. Let $D$ be a generic diagram in $\Bbb{R} \times (0,\,1)$ which presents $L$. Then $\tau_{\epsilon, \zeta}^{\iota}$ is given by $$\label{tauMformula}
\tau_{\epsilon, \zeta}^{\iota} (M) =
\Delta^{\sigma (L)}
\EuScript{D}^{- \sigma (L)- p -1}
\sum_{\lambda \in \cal{S} (D)} \,
\prod_{q =1}^{p} \iota^{\bold{c}_{\lambda}(K_q)}\,
[\bold{c}_{\lambda}(K_q) + 1] \,
\prod_{A \in \sharp D} \langle \lambda | A \rangle.$$ Here $\Delta$ denotes a fixed square root of $\sum_{i \in {\EuScript V}} [i + 1]^2$, $\EuScript{D} = \sum_{i \in {\EuScript V}} \iota^i \zeta^{- i (i + 2)} [i + 1]^2$ and $\sigma (L)$ denotes the signature of the linking matrix of $L$.
[99]{} H. Andersen, Tensor products of quantized tilting modules, [*Commun. Math. Phys.*]{} [**149**]{} (1992), 149-159. G. Bergman, The diamond lemma for ring theory, [*Adv. Math.*]{} [**29**]{} (1978), 178-218. G. B$\Ddot{\mathrm{o}}$hm and K. Szlach$\Acute{\mathrm{a}}$nyi, A coassociative $\mathrm{C}^*$-quantum group with non-integral dimensions, [*Lett. Math. Phys.*]{} [**35**]{} (1996), 437-456. V. Chari and A. Pressley, “A guide to quantum groups,” Cambridge University Press, Cambridge, 1994. Y. Doi and M.Takeuchi, Multiplication alternation by two-cocycles — The quantum version —, [*Commun. Alg.*]{} [**22**]{} (1994), 5715-5732. V. G. Drinfeld, On quasitriangular Quasi-Hopf algebras and a group closely connected with $Gal (\bar{\Bbb{Q}}/\Bbb{Q})$ [*Leningrad Math. J.*]{} [**2**]{} (1991), 829-860. M. Finkelberg, An equivalence of fusion categories, [*Geom. Funct. Analysis*]{} [**6**]{} (1996), 249-267. S. Gelfand and D. Kazhdan, Examples of tensor categories, [*Invent. Math.*]{} [**109**]{} (1992), 595-617. F. Goodman and T. Nakanishi, Fusion algebras in integrable systems in two dimensions, [*Phys. Lett. B*]{} [**262**]{} (1991), 259-264. F. Goodman and H. Wenzl, Littlewood-Richardson coefficients for Hecke algebras at roots of unity, [*Adv. Math.*]{} [**82**]{} (1990), 244-265. T. Hayashi, An algebra related to the fusion rules of Wess-Zumino-Witten models, [*Lett. Math Phys.*]{} [**22**]{} (1991), 291-296. T. Hayashi, Quantum deformation of classical groups, [*Publ. RIMS, Kyoto Univ.*]{} [**28**]{} (1992), 57 - 81. T. Hayashi, Quantum groups and quantum determinants, [*J. Algebra*]{} [**152**]{} (1992), 146-165. T. Hayashi, Quantum group symmetry of partition functions of IRF models and its application to Jones’ index theory, [*Commun. Math. Phys.*]{} [**157**]{} (1993), 331-345. T. Hayashi, Face algebras and their Drinfeld doubles, [*in*]{} “Proceedings of Symposia in Pure Mathematics,” Vol 56, Part 2, American Mathematical Society, 1994. T. Hayashi, Face algebras I — A generalization of quantum group theory, to appear in [*J. Math. Soc. Japan.*]{} T. Hayashi, Compact quantum groups of face type, [*Publ. RIMS, Kyoto Univ.*]{} [**32**]{} (1996), 351 - 369. T. Hayashi, Galois quantum groups of II$_1$-subfactors, preprint. T. Hayashi, Face algebras II — Standard generator theorems, in preparation. T. Hayashi, Quantum groups and quantum semigroups, to appear in [*J. Algebra*]{}. T. Hayashi, in preparation. T. Hayashi, in preparation. M. Jimbo, T. Miwa and M. Okado, Solvable lattice models related to the vector representation of classical simple Lie algebras, [*Commun. Math. Phys.*]{} [**116**]{} (1988), 507-525. B. Jurčo and P. Schupp, AKS scheme for face and Calgero-Moser-Sutherland type models, preprint. V. Kac, “Infinite dimensional Lie algebras,” 3rd ed., Cambridge Univ. press, 1990. C. Kassel, “Quantum groups,” Springer-Verlag, New York, 1995. D. Kazhdan and H. Wenzl, Reconstructing monoidal categories, [*Adv. in Soviet Math.*]{} [**16**]{} (1993),111-136. A. Kirillov, Jr, On an inner product in modular tensor categories, [*J. of AMS*]{} [**9**]{} (1996),1135-1169. T. Koornwinder, Compact quantum groups and $q$-special functions, preprint. R. Larson and J. Towber, Two dual classes of bialgebras related to the concepts of “quantum group” and “quantum Lie algebra,” [*Commun. Alg.*]{} [**19**]{} (1991), 3295-3345. I. Macdonald, “Symmetric functions and Hall polynomials,” 2nd ed., Oxford Univ. press, Oxford, 1995. G. Moore and N. Seiberg, Classical and quantum conformal field theory, [*Commun. Math. Phys.*]{} [**123**]{} (1989), 177-254. A. Ocneanu, Quantized group, string algebras and Galois theory for algebras, In “Operator algebras and applications, Vol2”. London Math. Soc. Lecture note series [**136**]{} (1989), 119-172. N. Reshetikhin and V. Turaev, Invariants of 3-manifolds via link polynomials and quantum groups, [*Invent. Math.*]{} [**103**]{} (1991), 547-598. N. Reshetikhin, L. Takhtadzhyan and L. Faddeev, Quantization of Lie groups and Lie algebras, [*Leningrad Math. J.*]{} [**1**]{} (1990), 193-225. P. Schauenburg, Face algebras are ${\times}_R$-bialgebras, preprint. M. Sweedler, “Hopf algebras,” Benjamin Inc., New York, 1969. M. Takeuchi, Matric bialgebras and quantum groups, [*Israel J. Math.*]{} [**72**]{} (1990), 232-251. A. Tsuchiya and Y. Kanie, Vertex operators in conformal field theory on ${\Bbb P}^1$ and monodromy representations of braid groups, In “Adv. Stud. Pure Math. Vo16”. V. Turaev, “Quantum invariants of knots and 3-manifolds,” Walter de Gruyter, Berlin, New York, 1994. V. Turaev and H. Wenzl, Quantum invariants of 3-manifolds associated with classical simple Lie algebras, [*Int. J. of Modern Math.*]{} [**4**]{} (1993), 323-358. V. Turaev and H. Wenzl, Semisimple and modular categories from link invariants, [*Math. Ann*]{} [**309**]{} (1997), 411-461. E. Verlinde, Fusion rules and modular transformations in 2D conformal field theory, [*Nucl. Phys.*]{} [**B300**]{} (1988), 360-376. M. Walton, Algorithm for WZW fusion rules: A proof, [*Phys. Lett. B*]{} [**241**]{} (1990), 365-368. H. Wenzl, Hecke algebras of type $A_n$ and subfactors, [*Invent. Math.*]{} [**92**]{} (1988), 349-383. H. Wenzl, $C^*$ tensor categories from quantum groups, [*J. AMS*]{} [**11**]{} (1998), 261-282. E. Witten, Quantum field theory and the Jones polynomial, [*Comm. Math. Phys.*]{} [**121**]{} (1989), 351-399.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Detecting temporal extents of human actions in videos is a challenging computer vision problem that requires detailed manual supervision including frame-level labels. This expensive annotation process limits deploying action detectors to a limited number of categories. We propose a novel method, called WSGN, that learns to detect actions from *weak supervision*, using only video-level labels. WSGN learns to exploit both video-specific and dataset-wide statistics to predict relevance of each frame to an action category. This strategy leads to significant gains in action detection for two standard benchmarks THUMOS14 and Charades. Our method obtains excellent results compared to state-of-the-art methods that uses similar features and loss functions on THUMOS14 dataset. Similarly, our weakly supervised method is only 0.3% mAP behind a state-of-the-art supervised method on challenging Charades dataset for action localization.'
author:
- Basura Fernando
- Cheston Tan Yin Chet
- Hakan Bilen
bibliography:
- 'wacv.bib'
title: Weakly Supervised Gaussian Networks for Action Detection
---
Introduction
============
Related Work {#s:related}
============
Problem, approach and model {#s:model}
===========================
Experiments {#s:experiments}
===========
Conclusion {#s:conclusion}
==========
\
**Acknowledgment**: This research was supported by the National Research Foundation, Singapore (Award Number: NRF2015-NRF-ISF001-2451), the National Research Foundation Singapore under its AI Singapore Programme (Award Number: AISG-RP-2019-010), and the Agency for Science, Technology and Research (A\*STAR) under its AME Programmatic Funding Scheme (Project \#A18A2b0046).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Collective dynamics result from interactions among noisy dynamical components. Examples include heartbeats, circadian rhythms, and various pattern formations. Because of noise in each component, collective dynamics inevitably involve fluctuations, which may crucially affect functioning of the system. However, the relation between the fluctuations in isolated individual components and those in collective dynamics is unclear. Here we study a linear dynamical system of networked components subjected to independent Gaussian noise and analytically show that the connectivity of networks determines the intensity of fluctuations in the collective dynamics. Remarkably, in general directed networks including scale-free networks, the fluctuations decrease more slowly with the system size than the standard law stated by the central limit theorem. They even remain finite for a large system size when global directionality of the network exists. Moreover, such nontrivial behavior appears even in undirected networks when nonlinear dynamical systems are considered. We demonstrate it with a coupled oscillator system.'
author:
- |
Naoki Masuda${}^{1,2}$, Yoji Kawamura${}^{3}$, and Hiroshi Kori${}^{4,2*}$\
\
\
${}^{1}$ Graduate School of Information Science and Technology,\
The University of Tokyo,\
7-3-1 Hongo, Bunkyo, Tokyo 113-8656, Japan \
${}^2$ PRESTO, Japan Science and Technology Agency,\
4-1-8 Honcho, Kawaguchi, Saitama 332-0012, Japan \
${}^3$ Institute for Research on Earth Evolution,\
Japan Agency for Marine-Earth Science and Technology,\
3173-25 Showa-machi, Kanazawa-ku, Yokohama, Kanagawa 236-0001, Japan\
${}^4$ Division of Advanced Sciences, Ochadai Academic Production,\
Ochanomizu University,\
2-1-1, Ohtsuka, Bunkyo-ku, Tokyo 112-8610, Japan\
\
$^*$ Author for correspondence ([email protected])
title: Collective fluctuations in networks of noisy components
---
Introduction {#sec:introduction}
============
Understanding fluctuations in dynamically ordered states and physical objects, which consist of networks of interacting components, is an important issue in many disciplines ranging from biology to engineering. When each constituent component of a system is noisy due to, [*e.g.*]{}, thermal fluctuations, it generally occurs that the entire system collectively fluctuates in time. Such collective fluctuations may be advantageous or disadvantageous in functioning of the systems depending on situations. For example, reduction in noise is likely to improve information processing in retinal neural networks [@Lamb76; @Smith95; @Devries02; @Bloomfield04]. Precision of biological circadian clocks [@Wilders93; @Enright80; @Garcia04pnas; @Herzog04] may be improved by reduction in collective fluctuations (*i.e.*, fluctuations in collective activities). On the other hand, maintaining a certain amount of fluctuations in an ordered state is advantageous for stochastic resonance [@Gammaitoni98] and Brownian motors [@Reimann02].
Despite the relevance of collective fluctuations in a variety of systems, theoretical frameworks that formulate collective fluctuations are missing. The central limit theorem states that, if the dynamical order is simply the averaged activity of noisy components, the standard deviation of the collective fluctuation would decrease with the number $N$ of noisy components as $N^{-1/2}$. However, scaling is unclear in systems of interacting components. Clarifying the property of collective fluctuations in such systems will give us insights into the mechanisms and design principles underlying the regulation of noise in, for example, living organisms and chemical reactions, and also into possible controls of fluctuations in collective dynamics.
In this study, we analyze an ensemble of components subjected to independent Gaussian noise that interact on general networks, including complex networks and regular lattices. We first consider a linear dynamical system, which can be regarded as linearization of various systems, such as networks of periodic or chaotic oscillators [@kuramoto84; @pikovsky01], the overdamped limit of elastic networks [@togashi07], a consensus problem treated in control theory [@olfatisaber07]. We show that collective fluctuations are determined by the connectivity of networks. It turns out that the scaling $N^{-1/2}$ is the tight lower bound, which is obtained for undirected networks. General directed networks yield a slower or nonvanishing decay of collective fluctuations with an increase in $N$. We then argue such nontrivial behavior appears even in undirected networks when nonlinear systems are considered. In particular, we show that linearization of coupled nonlinear oscillator systems on undirected media yields linear dynamics on asymmetric networks, such that the slow decay of the collective fluctuation is relevant.
Model and analysis
==================
Consider a network of $N$ components obeying $$\dot x_i= \sum_{j=1}^N w_{ij}(x_j-x_i)+\sqrt{D_i} \xi_i(t), \quad (1\le i\le N),
\label{eq:linear}$$ where $x_i$ is the state (or the position) of the $i$th component, $\sqrt{D_i}$ is the intensity of noise, $\xi_i$ is the independent Gaussian (generally colored) noise, and $w_{ij}$ is the intensity of coupling and can be also regarded as originating from the Jacobian matrix of underlying nonlinear dynamical systems such as coupled oscillator systems that we consider later. We allow negative weights and asymmetric coupling; $w_{ij}$ can be negative or different from $w_{ji}$. [Equation ]{} is a multivariate Ornstein-Uhlenbeck process [@Riskenbook; @Vankampenbook].
For convenience, we represent [Eq. ]{} as $$\dot {\bm x}=-L\bm x + \bm p,$$ where ${\bm x} \equiv (x_1\; \ldots \;
x_N)^{\top}$ ($\top$ denotes the transpose), $\bm p \equiv (\sqrt{D_1}
\xi_1\; \ldots \; \sqrt{D_N} \xi_N)^{\top}$, and $L=(L_{ij})$ is the asymmetric Laplacian defined by $L_{ij}=\delta_{ij}\sum_{i^{\prime}\neq i}w_{ii^{\prime}}
-(1-\delta_{ij})w_{ij}$ [@Arenas08; @mkk09njp]. $L$ always has a zero eigenvalue with the right eigenvector $\bm u \equiv (1\; \ldots\; 1)^{\top}$, *i.e.*, $L\bm u=0$. This eigenvector is associated with a global translational shift in state $\bm x$ and corresponds to the fact that such a shift keeps [Eq. ]{} invariant. We assume the stability of the ordered state represented by $x_1=\ldots=x_N$ in the absence of the noise (*i.e.*, $D_i=0$ for all $i$); the system relaxes to the ordered state from any initial condition. This is equivalent to assuming that the real parts of all the eigenvalues of $L$ are positive except for one zero eigenvalue, *i.e.*, $0\equiv
\lambda_1 < {\rm Re} \lambda_2 \le \ldots \le {\rm Re} \lambda_N$. This is a nontrivial condition for general networks with negative weights. However, for networks with only non-negative weights, *i.e.*, $w_{ij}\ge 0$ ($1\le i, j\le N$), this property holds true when the network is strongly connected or all the nodes are reachable by a directed path from a single node [@Ermentrout92; @Agaev00; @Arenas08].
We are concerned with collective fluctuations in dynamics given by [Eq. ]{}. To quantify their intensity, we decompose $\bm x$ as $$\bm x(t) = y(t)\bm u + \bm \rho(t).
\label{eq:mode decomposition}$$ $y(t)$ describes the one-dimensional component along $\bm u$, and $\bm
\rho(t)$ is the ($N-1$)-dimensional remainder mode. Note that $y(t) =
\bm v \bm x(t)$, where the row vector $\bm v\equiv (v_1 \ldots v_N)$ is the left eigenvector of $L$ corresponding to the zero eigenvalue, *i.e.*, $\bm v L=0$, and is normalized as $\bm v \bm u =1$, *i.e.*, $\sum_{i=1}^N v_i=1$ (see Appendix A for detailed descriptions). We call $y(t)\bm u$ the collective mode. In the absence of noise, the dynamical equation for $y(t)$ is given by $$\dot y = \bm v \dot {\bm x} = -
\bm v L \bm x =0.$$ Therefore, $y(t)$ is a conserved quantity of the dynamics. The remainder mode $\bm \rho(t)$ is associated with relative motions among the components. Because of the stability assumption, $\bm \rho(t)$ asymptotically vanishes with characteristic time $({\rm Re}~\lambda_2)^{-1}$. Therefore, all the values of $x_i$ ($1\le i \le N$) eventually go to the same value $y$ that is determined by the initial condition, *i.e.*, $y=\bm v
\bm x(0)$.
In the presence of noise, we obtain $$\dot y = \bm v \dot {\bm x} = \bm v (-L \bm x + \bm
p) = \bm v \bm p = \sum_{i=1}^N v_i \sqrt{D_i}
\xi_i(t).$$ Because $\xi_i$ is the independent Gaussian noise, this equation reduces to $$\dot y(t) = \sqrt{\sum_{i=1}^N v_i^2 D_i} \xi (t) \equiv \sigma \xi (t),
\label{eq:dynamics of collective mode}$$ where $\xi(t)$ is the Gaussian noise having the same statistical property as that of each $\xi_i(t)$. Thus, $y(t)$ performs the Brownian motion with effective noise strength $\sigma$ and is unbounded. The remainder mode $\bm \rho (t)$ fluctuates around zero because of its decaying nature. Therefore, the long-time behavior of $x_i(t) = y(t) + \rho_i(t)$ is approximately described by a single variable $y(t)$ for any $i$. We denote as $\sigma$, which depends on the structure of the network, the intensity of collective fluctuations. $\sigma$ can be calculated for given network.
In practice, the average activity of the population, $\bar x\equiv \sum_{i=1}^N
x_i/N$, but not the activity at individual nodes, may be observed. Because $\bar x = y + \sum_{i=1}^N \rho_i/N$ and $\sum_{i=1}^N \rho_i/N$ can be neglected in a long run, $\sigma$ also characterizes the fluctuations of $\bar x$.
Collective fluctuations in various networks
===========================================
General properties
------------------
We assume for simplicity that $D_i=1$ ($1\le i\le N$) so that $\sigma=\sqrt{\sum_{i=1}^N v_i^2}$. It is straightforward to extend the following results to the case of heterogeneous $D_i$. The vector $\bm v$ is uniform, *i.e.*, $v_i=1/N$ ($1\le i\le N$) if and only if $k_i^{\rm in}=k_i^{\rm out}$ ($1\le i\le N$), where $k_i^{\rm in}\equiv \sum_{j=1}^N w_{ij}$ and $k_i^{\rm
out}\equiv \sum_{j=1}^N w_{ji}$ are indegree and outdegree, respectively [@mkk09njp]. Undirected networks satisfy this condition. In this case, we obtain $\sigma=N^{-1/2}$, which agrees with the central limit theorem. The normalization condition $\sum_{i=1}^N v_i=1$ guarantees that $\sigma\ge
N^{-1/2}$ for any $\bm v$. Therefore, undirected networks are the best for reducing collective fluctuations. In the case of directed or asymmetrically weighted networks, $v_i$ is generally heterogeneous, and $\sigma > N^{-1/2}$. We will show later that this is also the case for nonlinear systems on undirected networks. When the weight $w_{ij}$ is nonnegative for any $i$ and $j$, the Perron–Frobenius theorem guarantees that $v_i$ is nonnegative for all $i$ [@Hornbook]. In this case, we obtain $$\frac{1}{\sqrt{N}}\le \sigma \le 1.$$ The case $\sigma=1$ is realized by a feedforward network, in which a certain component $i_0$ has no inward connection (*i.e.*, $k_{i_0}^{\rm in}=0$). Then, $v_{i_0}=1$ and $v_i= 0$ for $i \neq i_0$, which yields $\sigma=1$ irrespective of $N$; the collective fluctuations are not reduced at all with an increase in $N$. When negative weights are allowed, some elements of $\bm v$ may assume negative values. Then, $\sigma$ may be larger than $1$, in which case collective fluctuations are larger than individual noise.
We note that $\sigma^{2}$ is the so-called inverse participation ratio [@Derrida87]. $\sigma^{-2}$ can be interpreted as the effective number of components that participate in collective activities; the remaining components are slaved.
Directed scale-free networks
----------------------------
We demonstrate our theory by using some example networks. First, we consider directed scale-free networks, schematically shown in [Fig. ]{}\[fig:schem nets\](a) in which $k_i^{\rm in}$ and $k_i^{\rm out}$ independently follow the distributions $p(k^{\rm in})\propto k^{-\gamma_{\rm in}}$ and $p(k^{\rm out})\propto k^{-\gamma_{\rm out}}$, respectively. By assuming that the values of $v_i$ of adjacent nodes are independent of each other, we obtain $$\sum_{j=1}^N w_{ji} v_j \approx
\sum_{j=1}^N w_{ji} \bar{v} = k_i^{\rm out} \bar{v},$$ where $$\bar{v}\equiv \frac{\sum^N_{i=1}v_i}{N}=\frac{1}{N}.$$ Therefore, $$v_i
=\frac{\sum_{j=1}^N w_{ji} v_j}{\sum_{j=1}^N w_{ij}} \approx
\frac{k_i^{\rm out}/k_i^{\rm in}}{\sum_{j=1}^N\left(k_j^{\rm out}/k_j^{\rm
in}\right)}.$$ This approximation is sufficiently accurate for uncorrelated networks [@mkk09njp]. For $p(k^{\rm in})\propto k^{-\gamma_{\rm in}}$ and $p(k^{\rm out})\propto k^{-\gamma_{\rm out}}$, we obtain $$v_i\approx \frac{k_i^{\rm out}/k_i^{\rm
in}}{N\left<k^{\rm out}\right>\left<(k^{\rm in})^{-1}\right>}$$ and $$\sigma\approx
\sqrt{\frac{\left<\left(k^{\rm out}\right)^2\right>
\left<\left(k^{\rm in}\right)^{-2}\right>}
{N\left<k^{\rm out}\right>^2\left<\left(k^{\rm in}\right)
^{-1}\right>^2}},$$ where $\left< \cdot \right>$ is the ensemble average. When $\gamma_{\rm out}<2$, a winner-take-all network is generated [@Albert02rmp; @Dorogovtsev08rmp], and there exists a node $i$ such that $k_i^{\rm out}=O(N)$ and $v_i=O(1)$. When $\gamma_{\rm out}\ge 2$, the extremal criterion results in the maximum degree increasing with $N$ as $N^{1/(\gamma_{\rm out}-1)}$ ($\gamma_{\rm out}\ge 2$) in many networks [@Newman05comp; @Dorogovtsev08rmp]. Then, we obtain [@Sood08pre] $$\left<k^{\rm out}\right>\propto
\left\{\begin{array}{ll}
N^{2-\gamma_{\rm out}},& (\gamma_{\rm out}< 2),\\
\ln N,& (\gamma_{\rm out}=2),\\
O(1),& (\gamma_{\rm out}>2),
\end{array}\right.$$ $$\left<\left(k^{\rm out}\right)^2\right>\propto
\left\{\begin{array}{ll}
N^{-\gamma_{\rm out}+3},& (\gamma_{\rm out}< 2),\\
N^{(-\gamma_{\rm out}+3)/(\gamma_{\rm out}-1)},& (2\le
\gamma_{\rm out}< 3),\\
\ln N,& (\gamma_{\rm out}=3),\\
O(1),& (\gamma_{\rm out}>3),
\end{array}\right.$$ and $$\left<\left(k^{\rm in}\right)^{-1}\right>,
\left<\left(k^{\rm in}\right)^{-2}\right>
=O(1).$$ Therefore, we obtain $$\sigma\propto \left\{\begin{array}{ll}
1,& (\gamma_{\rm out}<2),\\
1/\ln N,& (\gamma_{\rm out}=2),\\
N^{-1+(\gamma_{\rm out}-1)^{-1}},& (2\le \gamma_{\rm out}<3),\\
N^{-1/2}(\ln N)^{1/2},& (\gamma_{\rm out}= 3),\\
N^{-1/2},& (\gamma_{\rm out}> 3).
\end{array}\right.
\label{eq:sf theory}$$ The fairly heterogeneous case $\gamma_{\rm out}<2$, in which the average outdegree diverges as $N\to\infty$, effectively yields a feedforward network. The case $\gamma_{\rm out}\ge 3$, where the second moment of the outdegree converges for $N\to\infty$, reproduces the central limit theorem. The latter result is shared by the directed version of the conventional random graph. The case $2\le\gamma_{\rm out}<3$ yields a nontrivial dependence of $\sigma$ on $N$. In [Fig. ]{}\[fig:sigma sf\](a), we compare the scaling exponent $\beta$, where $\sigma\propto N^{-\beta}$ obtained from the theory (solid line; [Eq. ]{}) and numerical simulations of the configuration model [@Boccaletti06; @Albert02rmp] with the power-law degree distribution with minimum degree 3 (open circles). The fitting procedure is explained in [Fig. ]{}\[fig:sigma sf\](b). Equation roughly explains numerically obtained values of $\beta$.
Directed lattices
-----------------
The second example is the directed one-dimensional chain of $N$ nodes depicted in [Fig. ]{}\[fig:schem nets\](b). We set $w_{i+1,i}=1$ ($1\le
i\le N-1$), $w_{i-1,i}=\epsilon$ ($2\le i\le N$), and $w_{j,i}=0$ ($j\neq i-1, i+1$). For this network, by solving $\bm v L=0$, we analytically obtain $$v_i = \frac{(1 -
\epsilon)\epsilon^{i-1}}{1 - \epsilon^N},\quad (1\le i\le N)$$ and $$\sigma = \sqrt{\frac{1 - \epsilon}{1 + \epsilon} \, \frac{1 + \epsilon^N}{1 - \epsilon^N}}.$$ Values of $\sigma$ for various $\epsilon$ and $N$ are plotted by solid lines in [Fig. ]{}\[fig:sigma linear\](a). Interestingly, for $\epsilon\neq 1$, $\lim_{N \to \infty} \sigma = \sqrt{(1 - \epsilon)/(1
+ \epsilon)}$; $\sigma$ is nonvanishing. We have also analytically derived $\sigma$ for directed $d$-dimensional lattices (see Appendix B). The results for the two-dimensional lattice depicted in [Fig. ]{}\[fig:schem nets\](c) are plotted by solid lines in [Fig. ]{}\[fig:sigma linear\](b). To confirm our theory, we also carried out direct numerical simulations of [Eq. ]{} with Gaussian white noise for these directed lattices. The results indicated by circles in [Fig. ]{}\[fig:sigma linear\] indicate an excellent agreement with our theory.
A similar result is obtained for the Cayley tree (see Appendix C).
Oscillator dynamics
===================
As an application of our theory to nonlinear systems, we examine noisy and rhythmic components. As a general, tractable, yet realistic model, we consider a network of phase oscillators [@winfree67; @kuramoto84; @kiss07], whose dynamical equation is given by $$\dot \phi_i = \omega_i +
\sum_{j=1}^N A_{ij} f(\phi_j-\phi_i)+ \sqrt{D_i} \xi_i(t),
\quad (1\le i\le N),
\label{eq:phase}$$ where $\phi_i\in [0,2\pi)$ and $\omega_i$ are the phase and the intrinsic frequency of the $i$th oscillator, respectively, $A_{ij}$ is the intensity of coupling, and $f(\cdot)$ is a $2\pi$–periodic function. We assume that, in the absence of noise, all the oscillators are in a fully phase-locked state, *i.e.*, $\phi_i(t) = \Omega t + \psi_i$, where $\Omega$ and $\psi_i$ are the constants derived from $\dot \phi_i = \Omega$ ($1\le i \le N$). Under sufficiently weak noise, we can linearize [Eq. ]{} around the phase-locked state. Letting $x_i=\phi_i-(\psi_i+\Omega t)$, we obtain [Eq. ]{}, where $w_{ij} = A_{ij}
f^{\prime}(\psi_j-\psi_i)$ is the effective weight. The validity of linearizing [Eq. ]{} for small noise intensity is tested by carrying out direct numerical simulations of [Eq. ]{} with $\omega_i=\omega$ ($1\le i\le N$) and $f(\phi)=\sin \phi$. The relationship $\sigma\approx \sqrt{\sum_{i=1}^N
D_i v_i^2}$ is satisfied in the directed one- and two-dimensional lattices, as shown in [Fig. ]{}\[fig:sigma osc\](a) and (b), respectively.
When there is some dispersion in $\psi_i$ in a phase-locked state, the relation $\sigma\approx N^{-1/2}$ may be violated even in undirected networks. This is because the effective weight is generally asymmetric (*i.e.*, $w_{ij}\neq w_{ji}$) unless $f(\cdot)$ is an exact odd function. In reality, $f(\cdot)$ is usually not an odd function [@kuramoto84; @brown04; @galan05; @kiss07]. As an example, we consider target patterns (*i.e.*, concentric traveling waves), which naturally appear in spatially extended oscillator systems [@kuramoto84; @mikhailov06]. We carry out direct numerical simulations of [Eq. ]{} on the two-dimensional [*undirected*]{} lattice with linear length $\sqrt{N}=50$, $f(\phi)=\sin(\phi-\alpha)+\sin \alpha$, and $\alpha=\pi/4$. Such a function may be analytically derived from a general class of coupled oscillators [@kuramoto84], and it approximates a variety of real systems [@galan05; @tsubo07; @kiss07]. We set $\omega_i=\omega_0 +
\Delta\omega$ $(\Delta\omega \ge 0)$ for $4 \times 4$ pacemaker oscillators in the center and $\omega_i=\omega_0$ for the other oscillators, where $\omega_0$ is arbitrary and set to $1$. A target pattern is formed when there is sufficient heterogeneity in the intrinsic frequency [@kuramoto84]. A region with high intrinsic frequency acts as a pacemaker. A snapshot for $\Delta \omega=0.3$ is shown in [Fig. ]{}\[fig:phase eigenvector\](a). As observed, the radial phase gradient is approximately constant, which makes the effective network similar to the directed two-dimensional lattice depicted in [Fig. ]{}\[fig:schem nets\](c). Therefore, as shown in [Fig. ]{}\[fig:phase eigenvector\](b), $v_i$ calculated numerically decreases almost exponentially with the distance from the center. We find that the dependence of $\sigma$ on $N$, shown in [Fig. ]{}\[fig:phase eigenvector\](c), is similar to that for directed lattices.
We emphasize that the network is undirected (*i.e.*, $A_{ij}=A_{ji}$). We have also theoretically confirmed that our results are valid for the continuous oscillatory media under spatial block noise, which models chemical reaction–diffusion systems (see Appendix D).
Conclusions
===========
In summary, we have obtained the analytical relationship between collective fluctuations and the structure of networks. In undirected networks, the fluctuations decrease with the system size $N$ as $N^{-1/2}$; this result agrees with the central limit theorem. In general directed networks, the collective fluctuations decay more slowly. For example, in directed scale-free networks, we obtain $N^{-\beta}$ with $0 < \beta <
1/2$. In networks with global directionality, the fluctuations do not vanish for a large system size. We have also demonstrated that such nontrivial dependence appears even in undirected networks when nonlinear systems are considered. We have focused on systems of nonleaky components. Results for coupled leaky components will be reported elsewhere.
Our results are distinct from earlier results demonstrating the breach of the central limit theorem due to heavy-tailed noise [@Bouchaud90] or the correlation between the noise in different elements [@Kaneko90; @Zohary94].
Finally, because our theory is based on a general linear model, it can be tested in a variety of experimental systems. An ideal experimental protocol is provided by photo-sensitive Belousov-Zhabotinsky reaction systems, in which the heterogeneity, noise intensity, and system size can be precisely controlled by light stimuli [@mikhailov06]. Experiments with coupled oscillatory cells, such as cardiac cells and neurons under an appropriate condition, would be also interesting.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank Istvan Z. Kiss, Norio Konno, Yoshiki Kuramoto and Ralf Tönjes for their valuable discussions. N.M. acknowledges the support through the Grants-in-Aid for Scientific Research (Nos. 20760258 and 20540382) from MEXT, Japan.
Appendix A: Derivation of the collective mode {#appendix-a-derivation-of-the-collective-mode .unnumbered}
=============================================
To derive the collective mode $y(t)\bm u$, we note that there exists a nonsingular matrix $P$ such that $\tilde{L} \equiv P^{-1}LP$ is its Jordan canonical form [@Hornbook; @Arenas08]. We assume that $\tilde{L}_{11}=\lambda_1=0$ and $\tilde{L}_{1i}= \tilde{L}_{i1}=0$ ($2\le i\le N$) without loss of generality. The submatrix $(\tilde{L}_{ij})$ ($2\le i,j\le N$) corresponds to the $N-1$ modes with the eigenvalues $\lambda_2,\ldots, \lambda_N$. Because the first column of $LP=P\tilde{L}$ is equal to $(0\;\ldots\; 0)^{\top}$, the first column of $P$ is equal to the right eigenvector of $L$ corresponding to $\lambda_1=0$, *i.e.*, $\bm u = (1\;\ldots\; 1)^{\top}$. Because the first row of $P^{-1}L=\tilde{L}P^{-1}$ is equal to $(0\;\ldots\; 0)$, the first row of $P^{-1}$ is equal to the left eigenvector of $L$ corresponding to $\lambda_1=0$, *i.e.*, $\bm v = (v_1\; v_2\;\ldots\; v_N)$. The normalization is given by $\sum_{i=1}^N v_i=1$. Under the variable change $(y\; \bm y_{\rm r})
\equiv P^{-1}\bm x
\in \mathbb{R}^N$, where $y\in \mathbb{R}$ and $\bm y_{\rm r}\in
\mathbb{R}^{N-1}$, the coupling term is transformed into $-\tilde{L} (y\; \bm y_{\rm r})$. Then, in the absence of the dynamical noise, $y=\sum_{i=1}^N v_i x_i$ is a conserved quantity, which is the collective mode. $\bm \rho(t)$ in [Eq. ]{} is given by $P_{\rm r}\bm y_{\rm r}$, where $P_{\rm r}$ is the $N$ by $N-1$ matrix satisfying $P=(\bm u\; P_{\rm r})$.
Appendix B: Collective fluctuations in regular lattices with arbitrary dimensions {#appendix-b-collective-fluctuations-in-regular-lattices-with-arbitrary-dimensions .unnumbered}
=================================================================================
Consider a directed two-dimensional square lattice with a root node. As depicted in [Fig. ]{}\[fig:schem nets\](c), the edges descending from the root node and those approaching the root node in terms of the graph-theoretic distance are given weight 1 and $\epsilon$ ($0\le\epsilon\le 1$), respectively. We define layers such that the layer $\ell$ ($\le \ell_{\max}$) is occupied by the nodes whose distance from the root node is equal to $\ell$. Layer 0 contains the root node only, and layer $\ell$ ($\ge 1$) contains $4\ell$ nodes. We consider the lattice within a finite range specified by $\ell\le \ell_{\max}$. Note the difference from the case of the one-dimensional chain examined in the main text ([Fig. ]{}\[fig:schem nets\](b)), where the root node is located at the periphery of the chain. However, the scaling of $\sigma$ is not essentially affected by this difference.
The symmetry guarantees that the four nodes in layer 1 have the same value of $v_i$. Consider a node in [Fig. ]{}\[fig:schem nets\](c) that is labeled 2 and adjacent to two nodes labeled 1. There are four such nodes. The equation in $\bm v
L=0$ corresponding to this node is given by $(2+2\epsilon)v_2=2\epsilon v_1+2v_3$. The other four nodes labeled 2 in [Fig. ]{}\[fig:schem nets\](c) yield a different equation $(1+3\epsilon)v_2=\epsilon v_1+3v_3$. Similarly, we obtain $(2+2\epsilon)v_\ell=2\epsilon v_{\ell-1}+
2v_{\ell+1}$ for all but four nodes in layer $\ell$. The other four nodes satisfy $(1+3\epsilon)v_\ell=\epsilon v_{\ell-1}+
3v_{\ell+1}$. Despite this inhomogeneity, $v_{\ell}\propto \epsilon^{\ell}$ satisfies all these equations. By counting the number of nodes in each layer, the properly normalized solution is given by $$v_\ell = \left[ T_{\ell_{\max}}^{(2)}(\epsilon) \right]^{-1} \, \epsilon^\ell, \qquad
(0\le \ell\le \ell_{\max}),$$ and $$\sigma
= \frac{\sqrt{T_{\ell_{\max}}^{(2)}(\epsilon^2)}}
{T_{\ell_{\max}}^{(2)}(\epsilon)},
\label{eq:sigma z2}$$ where $$\begin{aligned}
T_{\ell_{\max}}^{(2)}(z) &=& 1+ 4\sum_{\ell=1}^{\ell_{\max}} \ell z^\ell
\nonumber\\
&=& \frac{(1+z)^2 - 4[ 1 + {\ell_{\max}} (1 - z) ] z^{\ell_{\max}+1}}
{(1 - z)^2}.\end{aligned}$$ The difference between the one- and two-dimensional cases lies in the number of nodes in each layer, which affects the normalization of $v_\ell$ and hence the value of $\sigma$. In the limit of a purely feedforward network, $\sigma$ is independent of the system size, *i.e.*, $\lim_{\epsilon \to 0} \sigma = 1$. In the case of undirected networks, the central limit theorem is recovered, *i.e.*, $\lim_{\epsilon \to 1} \sigma = N^{-1/2}$. In the limit of infinite space, we obtain $$\lim_{{\ell_{\max}}\to\infty}\sigma=
\frac{(1-\epsilon)(1+\epsilon^2)}{(1+\epsilon)^3}.$$
For a general dimension $d$, layer 0 has a single root node, and layer $\ell$ ($1\le \ell\le \ell_{\max}$) has $$N_{\ell}^{(d)} \equiv
\sum_{d^{\prime}=1}^{\ell} \frac{d!}{d^{\prime}!(d-d^{\prime})!}
\frac{(\ell-1)!}{(d^{\prime}-1)!(\ell-d^{\prime})!}
2^{d^{\prime}}$$ nodes. $d^{\prime}$ is the number of coordinates among the $d$ coordinates to which nonzero values are assigned, and the factor $2^{d^{\prime}}$ takes care of the fact that reversing the sign of any coordinate does not change the layer of the node. Similar to the case of the two-dimensional lattice, the value of $v_i$ for any node in layer $\ell$ in a $d$-dimensional lattice, denoted by $v_\ell^{(d)}$, is given by $$v_\ell^{(d)} = \left[ T_{\ell_{\max}}^{(d)}(\epsilon) \right]^{-1} \, \epsilon^\ell, \qquad
(0\le \ell\le \ell_{\max}),
\label{eq:vl spatial}$$ where $$T_{\ell_{\max}}^{(d)}(z)
= 1 + \sum_{\ell=1}^{\ell_{\max}} N_{\ell}^{(d)}z^{\ell}
\label{eq:T_lmax^{(d)}(z)}$$ From [Eq. ]{}, we obtain $$\sigma = \frac{\sqrt{T_{\ell_{\max}}^{(d)}(\epsilon^2)}}
{T_{\ell_{\max}}^{(d)}(\epsilon)}.
\label{eq:sigma vs T_lmax^{(d)}}$$
Note that $\lim_{\epsilon \to 0} \sigma = 1$ and $\lim_{\epsilon \to 1} \sigma = N^{-1/2}$. In the limit $\ell_{\max}\to\infty$, [Eq. ]{} becomes $$\begin{aligned}
\lim_{\ell_{\max}\to\infty}T_{\ell_{\max}}^{(d)}(z)
&=& 1 + \sum_{d^{\prime}=1}^{\infty}
\frac{d!}{d^{\prime}!(d-d^{\prime})!} 2^{d^{\prime}}
\sum_{\ell=d^{\prime}}^{\infty}\frac{(\ell-1)!}
{(d^{\prime}-1)!(\ell-d^{\prime})!}z^{\ell}\nonumber\\
&=& 1+\sum_{d^{\prime}=1}^{\infty}
\frac{d!}{d^{\prime}!(d-d^{\prime})!} 2^{d^{\prime}}
\left(\frac{z}{1-z}\right)^{d^{\prime}}\nonumber\\
&=& \left(\frac{1+z}{1-z}\right)^d.
\label{eq:T_lmax^{(d)}(z) lmax=infty}\end{aligned}$$ Substituting [Eq. ]{} into [Eq. ]{} yields $$\lim_{\ell_{\max}\to\infty}\sigma=\left[
\frac{\left(1-\epsilon\right)
\left(1+\epsilon^2\right)}{\left(1+\epsilon\right)^3} \right]^{d/2}.$$
Appendix C: Collective fluctuations in the Cayley tree {#appendix-c-collective-fluctuations-in-the-cayley-tree .unnumbered}
======================================================
Consider a Cayley tree with degree $k$ and a specific root node. We assume that the maximum distance from the root node is equal to $\ell_{\max}$. The edges descending from the root node and those approaching the root node are assigned weight 1 and $\epsilon$, respectively. The exact value of $v_i$ in layer $\ell$, denoted by $v_{\ell}$ without confusion, is obtained via $$\left[1+\left(k-1\right)\epsilon\right]
v_{\ell}=\epsilon v_{\ell-1}+\left(k-1\right)v_{\ell+1},\quad (\ell\ge 1).
\label{eq:linear system tree}$$ By solving [Eq. ]{}, we obtain $$v_\ell = \frac{1 - (\epsilon k)}{1 - (\epsilon k)^{\ell_{\max}+1}}
\, \epsilon^{\ell}, \qquad
(0\le \ell\le \ell_{\max}).
\label{eq:vl tree}$$ From [Eq. ]{}, we obtain $$\sigma = \frac{1 - (\epsilon k)}{1 - (\epsilon k)^{\ell_{\max}+1}} \,
\sqrt{
\frac{1 - \left( \epsilon^2 k \right)^{\ell_{\max}+1}}
{1 - \left( \epsilon^2 k \right)}
}.$$ Note that $\lim_{\epsilon \to 0} \sigma = 1$ and $\lim_{\epsilon \to 1} \sigma
= \sqrt{(1 - k)/(1 - k^{\ell_{\max}+1})} = N^{-1/2}$. The infinite-size limit exists only when $\epsilon k < 1$, and it is equal to $$\lim_{\ell_{\max}\to\infty} \sigma =
\frac{1-\epsilon k}{\sqrt{1-\epsilon^2 k}}.$$
Appendix D: Target patterns in continuous media under spatial block noise {#appendix-d-target-patterns-in-continuous-media-under-spatial-block-noise .unnumbered}
=========================================================================
We show that our results for the coupled oscillator system in the $d$-dimensional lattice are also valid for that in the continuous Euclidean space. We assume that Gaussian spatial block noise is applied. This type of noise has been used in experiments [@mikhailov06].
We consider the $d$-dimensional nonlinear phase diffusion equation given by $$\partial_t \phi(\bm{r}, t)
= \omega + \nu \nabla^2 \phi + \mu \left( \nabla \phi \right)^2 + s(\bm{r}),
\label{eq:target 1}$$ where $\bm {r}\in \mathbb{R}^d$ is the spatial coordinate, $\omega > 0$ is the intrinsic frequency, $\nu > 0$ is the diffusion constant, and $\mu > 0$ is the coefficient of the nonlinear term [@kuramoto84]. The term $s(\bm{r})$ represents the localized heterogeneity, which is positive near the origin and vanishing otherwise.
The synchronous solution corresponding to the target pattern is written as $\phi(\bm{r}, t) = \Omega t + \psi(\bm{r})$, where $\Omega$ and $\psi(\bm{r})$ satisfy $$\Omega
= \omega + \nu \nabla^2 \psi + \mu \left( \nabla \psi \right)^2 + s(\bm{r}).
\label{eq:target 2}$$ Let $x(\bm r, t)$ be a small deviation from the target pattern defined by $x \equiv
\phi - (\Omega t + \psi)$. Linearizing [Eq. ]{} using $x(\bm r,t)$, we obtain $\partial _t x(\bm{r}, t) = \mathcal{L} x$, where the linear operator $\mathcal{L}$ is given by $$\mathcal{L} x
= \nu \nabla^2 x + 2 \mu \left( \nabla \psi \right) \cdot \left( \nabla x \right).$$ We define the inner product as $$\left[ x_1(\bm{r}), x_2(\bm{r}) \right] = \int d\bm{r} \, x_1(\bm r)
x_2(\bm r).$$ We define the adjoint operator $\mathcal{L}^\dag$ as $[x_1,
\mathcal{L}x_2] = [\mathcal{L}^\dag x_1, x_2]$, *i.e.*, $$\mathcal{L}^\dag x
= \nu \nabla^2 x - 2 \mu \nabla \cdot \left( x \nabla \psi \right).$$ Note that $\mathcal{L}$ is self-adjoint when $\nabla \psi = 0$.
Because of the translational symmetry in [Eq. ]{} with respect to $\phi$, $\mathcal{L}$ has one zero eigenvalue. Let the right and left eigenfunctions of $\mathcal{L}$ corresponding to the zero eigenvalue be $u(\bm{r})$ and $v(\bm{r})$, respectively, *i.e.*, $\mathcal{L}
u=0$ and $\mathcal{L}^\dag v =0$. Trivially, $u(\bm{r}) = 1$. The normalization condition $[v(\bm{r}), u(\bm{r})]=1$ then implies that $\int d\bm{r} \, v(\bm{r}) = 1$.
Now, we introduce the perturbation to [Eq. ]{} as follows: $$\partial_t \phi(\bm{r}, t)
= \omega + \nu \nabla^2 \phi + \mu \left( \nabla \phi \right)^2 + s(\bm{r})
+ \sqrt{D} \, \xi(\bm{r}, t),$$ where $\xi(\bm{r}, t)$ represents a weak perturbation to the target pattern. Similarly to [Eq. ]{}, we decompose $x$ into $$x(\bm r,t) = y(t) u(\bm r) + \rho(\bm r,t),$$ where $y(t) u(\bm r)$ is the collective mode. The dynamical equation for $y$ is then obtained as $$\dot y = \sqrt{D} \int d\bm{r} \, v(\bm{r}) \xi(\bm{r}, t).
\label{eq:dTheta/dt target}$$ Let us assume that $\xi(\bm{r},t)$ is the Gaussian spatial block noise characterized by $$\xi\left( \bm{r}, t \right) = \xi_{\bm{\ell}}\left( t \right), \qquad
\bm{r} \in \mathbb{R}^d(\bm{\ell}),
\label{eq:block noise target}$$ $$\left\langle \xi_{\bm{\ell}}\left( t \right) \xi_{\bm{\ell}'}\left( t' \right) \right\rangle
= \delta_{\bm{\ell}, \bm{\ell}'} C\left( \left| t - t' \right| \right),$$ where $\bm{\ell}$ is the vector index for the block $\mathbb{R}^d(\bm{\ell})$. Using [Eq. ]{}, [Eq. ]{} is transformed into $$\dot y = \sqrt{D} \sum_{\bm{\ell}} \, v_{\bm{\ell}} \, \xi_{\bm{\ell}}(t),
\label{eq:dTheta/dt block target}$$ where $$v_{\bm{\ell}} = \int_{\mathbb{R}^d(\bm{\ell})} d\bm{r} \, v(\bm{r}).$$ From [Eq. ]{}, we find that the intensity of the collective fluctuation is given by $$\sigma = \sqrt{D\sum_{\bm{\ell}} v_{\bm{\ell}}^2}.$$ Note that $v_{\bm{\ell}}$ satisfies the normalization condition as follows: $$\sum_{\bm{\ell}} v_{\bm{\ell}}
= \sum_{\bm{\ell}} \int_{\mathbb{R}^d(\bm{\ell})} d\bm{r} \, v(\bm{r})
= \int d\bm{r} \, v(\bm{r}) = 1.$$
[10]{}
Lamb T D and Simon E J 1976 [*J. Physiol. (London)*]{} [**263**]{} 257
Smith R G and Vardi N 1995 [*Vis. Neurosci.*]{} [**12**]{} 851
DeVries S H, Qi X, Smith R, Makous W and Sterling P 2002 [*Curr. Biol.*]{} [**12**]{} 1900
Bloomfield S A and Völgyi B 2004 [*Vis. Res.*]{} [**44**]{} 3297
Wilders R and Jongsma H J 1993 [*Biophys. J.*]{} [**65**]{} 2601
Enright J T 1980 [*Science*]{} [**209**]{} 1542
Garcia-Ojalvo J, Elowitz M B and Strogatz S H 2004 [*Proc. Natl. Acad. Sci. USA*]{} [**101**]{} 10955
Herzog E D, Aton S J, Numano R, Sakaki Y and Tei H 2004 [*J. Biol. Rhythms*]{} [**19**]{} 35
Gammaitoni L, Hänggi P, Jung P and Marchesoni F 1998 [*Rev. Mod. Phys.*]{} [**70**]{} 223
Reimann P 2002 [*Phys. Rep.*]{} [**361**]{} 57
Kuramoto Y 1984 (New York: Springer)
Pikovsky A, Rosenblum M and Kurths J 2001 (Cambridge: Cambridge University Press)
Togashi Y and Mikhailov A 2007 [*Proc. Natl. Acad. Sci. USA*]{} [**104**]{} 8697
Olfati-Saber R, Fax J A and Murray R M 2007 [*Proc. IEEE*]{} [**95**]{} 215
Risken H 1989 2nd edn (Berlin: Springer)
Van Kampen N G 2007 3rd edn (Amsterdam: Elsevier)
Arenas A, Díaz-Guilera A, Kurths J, Moreno Y and Zhou C 2008 [*Phys. Rep.*]{} [**469**]{} 93
Masuda N, Kawamura Y and Kori H 2009 [*New J. Phys.*]{} [**11**]{} 113002
Ermentrout G B 1992 [*SIAM J. Appl. Math.*]{} [**52**]{} 1665
Agaev R P and Chebotarev P Y 2000 [*Autom. Remote Control*]{} [**61**]{} 1424
Horn R A and Johnson C R 1985 (Cambridge: Cambridge University Press)
Derrida B and Flyvbjerg H 1987 [*J. Phys. A: Math. Gen.*]{} [**20**]{} 5273
Albert R and Barabási A-L 2002 [*Rev. Mod. Phys.*]{} [**74**]{} 47
Dorogovtsev S N, Goltsev A V and Mendes J F F 2008 [*Rev. Mod. Phys.*]{} [**80**]{} 1275
Newman M E J 2005 [*Contem. Phys.*]{} [**46**]{} 323
Sood V, Antal T and Redner S 2008 [*Phys. Rev. E*]{} [**77**]{} 041121
Boccaletti S, Latora V, Moreno Y, Chavez M and Hwang D-U 2006 [*Phys. Rep.*]{} [**424**]{} 175
Winfree A T 1967 [*J. Theor. Biol.*]{} [**16**]{} 15
Kiss I Z, Rusin C G, Kori H and Hudson J L 2007 [*Science*]{} [**316**]{} 1886
Brown E, Moehlis J and Holmes P 2004 [*Neural Comput.*]{} [**16**]{} 673
Galán R F, Ermentrout G B and Urban N N 2005 [*Phys. Rev. Lett.*]{} [**94**]{} 158101
Mikhailov A S and Showalter K 2006 [*Phys. Rep.*]{} [**425**]{} 79
Tsubo Y, Takada M, Reyes A D and Fukai T 2007 [*Eur. J. Neurosci.*]{} [ **25**]{} 3429
Bouchaud J P and Georges A 1990 [*Phys. Rep.*]{} [**195**]{} 127
Kaneko K 1990 [*Phys. Rev. Lett.*]{} [**65**]{} 1391
Zohary E, Shadlen M N and Newsome W T 1994 [*Nature*]{} [**370**]{} 140
![Schematic of (a) directed scale-free network, (b) directed chain, and (c) directed two-dimensional lattice. The numbers in (b) indicate the indices of the nodes, while those in (c) indicate the layer index.[]{data-label="fig:schem nets"}](schematic){width="15cm"}
![(a) Scaling exponent $\beta$ for $\sigma\propto N^{-\beta}$ in scale-free networks with $\gamma_{\rm in}=\gamma_{\rm out}$. The solid line is the theoretical prediction given by [Eq. ]{}. The open circles are obtained numerically as follows. For each network, we calculate the eigenvector $\bm v$ to obtain $\sigma$. Then, $\beta$ is determined from the best linear fit between $\beta$ and $N$ on the logarithmic scale, as described in (b). (b) Determination of $\beta$. Data points are generated as an average value of $\beta$ for each of $N=100$, 200, 400, $\ldots$, 12800. The results obtained from the direct numerical simulations are shown by circles. For demonstration, the results for $\gamma_{\rm in}=\gamma_{\rm out}=2.05$, 2.5, 3, 3.5, and 4 are shown. By assuming $\sigma\propto N^{-\beta}$, we regress $\log\sigma$ against $\log N$ by the best linear fit (solid lines). The slope gives an estimate of $-\beta$. The Pearson correlation coefficient is large ($>$ 0.99) for each value of $\gamma_{\rm out}$ analyzed in (a). []{data-label="fig:sigma sf"}](sigma-sf "fig:"){width="8cm"} ![(a) Scaling exponent $\beta$ for $\sigma\propto N^{-\beta}$ in scale-free networks with $\gamma_{\rm in}=\gamma_{\rm out}$. The solid line is the theoretical prediction given by [Eq. ]{}. The open circles are obtained numerically as follows. For each network, we calculate the eigenvector $\bm v$ to obtain $\sigma$. Then, $\beta$ is determined from the best linear fit between $\beta$ and $N$ on the logarithmic scale, as described in (b). (b) Determination of $\beta$. Data points are generated as an average value of $\beta$ for each of $N=100$, 200, 400, $\ldots$, 12800. The results obtained from the direct numerical simulations are shown by circles. For demonstration, the results for $\gamma_{\rm in}=\gamma_{\rm out}=2.05$, 2.5, 3, 3.5, and 4 are shown. By assuming $\sigma\propto N^{-\beta}$, we regress $\log\sigma$ against $\log N$ by the best linear fit (solid lines). The slope gives an estimate of $-\beta$. The Pearson correlation coefficient is large ($>$ 0.99) for each value of $\gamma_{\rm out}$ analyzed in (a). []{data-label="fig:sigma sf"}](showfit-sigma-sf "fig:"){width="8cm"}
![Collective fluctuations for (a) the directed one-dimensional chain and (b) the directed two-dimensional lattice for various $N$ and $\epsilon$. The solid lines and the circles represent the theoretical and the numerical results, respectively. In (b), $\ell_{\max}$ is the maximum distance from the center of the lattice. For both networks, we set $D_i=1$ ($1\le i\le N$) and simulate [Eq. ]{} with the initial condition $x_i(t=0)=0$ ($1\le i\le N$). We measured $\sigma$ as the standard deviation of $\bar{x}(t=10200)-\bar{x}(t=200)$ obtained by conducting 2000 trials, which is then normalized by $\sqrt{10000}$. We disregard the first 200 time units as transient.[]{data-label="fig:sigma linear"}](sigma-l-1dim "fig:"){width="8cm"} ![Collective fluctuations for (a) the directed one-dimensional chain and (b) the directed two-dimensional lattice for various $N$ and $\epsilon$. The solid lines and the circles represent the theoretical and the numerical results, respectively. In (b), $\ell_{\max}$ is the maximum distance from the center of the lattice. For both networks, we set $D_i=1$ ($1\le i\le N$) and simulate [Eq. ]{} with the initial condition $x_i(t=0)=0$ ($1\le i\le N$). We measured $\sigma$ as the standard deviation of $\bar{x}(t=10200)-\bar{x}(t=200)$ obtained by conducting 2000 trials, which is then normalized by $\sqrt{10000}$. We disregard the first 200 time units as transient.[]{data-label="fig:sigma linear"}](sigma-l-2dim "fig:"){width="8cm"}
![Collective fluctuation for coupled phase oscillators in (a) directed one-dimensional chain and (b) directed two-dimensional lattice. The solid lines represent the theoretical results, and the circles represent the numerical results obtained by the direct numerical simulations of [Eq. ]{} with $f(\phi)=\sin \phi$. We set $D_i=0.01$ ($1\le i\le N$) and start with $x_i=0$ ($1\le i\le N$). We measure $\sigma$ as the standard deviation of $\bar{x}(t=10200) - \bar{x}(t=200)$ obtained by conducting 2000 trials, which is then normalized by $\sqrt{10000}$.[]{data-label="fig:sigma osc"}](sigma-o-1dim "fig:"){width="8cm"} ![Collective fluctuation for coupled phase oscillators in (a) directed one-dimensional chain and (b) directed two-dimensional lattice. The solid lines represent the theoretical results, and the circles represent the numerical results obtained by the direct numerical simulations of [Eq. ]{} with $f(\phi)=\sin \phi$. We set $D_i=0.01$ ($1\le i\le N$) and start with $x_i=0$ ($1\le i\le N$). We measure $\sigma$ as the standard deviation of $\bar{x}(t=10200) - \bar{x}(t=200)$ obtained by conducting 2000 trials, which is then normalized by $\sqrt{10000}$.[]{data-label="fig:sigma osc"}](sigma-o-2dim "fig:"){width="8cm"}
![(color online) Numerical results of the coupled oscillators on the two-dimensional undirected lattice. (a) Snapshot of $\sin \phi_i$ and (b) eigenvector $\bm v$ (log scale) for $\Delta\omega=0.3$, where $r_1$ and $r_2$ denote the spatial coordinates. (c) The dependence of $\sigma$ on system size $N$. []{data-label="fig:phase eigenvector"}](2dim-wave){width="8cm"}
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
We have calculated the $E1$ and $E2$ contributions to the low-energy ${}^8{\rm B}+ {}^{208}{\rm Pb} \rightarrow p+{}^7{\rm Be}
+{}^{208}{\rm Pb}$ Coulomb dissociation cross sections using the kinematics of a recent experiment at RIKEN. Using a potential model description of the ${}^7{\rm Be} (p,\gamma) {}^8{\rm B}$ reaction, we find that the $E2$ contributions cannot [*a priori*]{} be ignored in the analysis of the data. Its inclusion reduces the extracted ${}^7{\rm Be} (p,\gamma) {}^8{\rm B}$ $S$-factor at solar energies by about 25%.
address: |
W. K. Kellogg Radiation Laboratory, 106-38\
California Institute of Technology, Pasadena, California 91125 USA
author:
- 'K. Langanke and T. D. Shoppa'
title: The contribution to the Coulomb dissociation cross section
---
The ${}^7{\rm Be} (p,\gamma) {}^8{\rm B}$ reaction plays a crucial role in the solar neutrino puzzle, as its rate is directly proportional to the flux of those high-energy neutrinos to which the ${}^{37}$Cl and Kamiokande detectors are particularly sensitive [@Bahcall]. While the energy dependence of the low-energy ${}^7{\rm Be} (p,\gamma) {}^8{\rm B}$ cross section is believed to be sufficiently well known [@Johnson], the absolute cross section at solar energies ($E\approx20$ keV) is rather uncertain as the two measurements of the cross section that extend lowest in energy disagree by about 25% in magnitude [@Kavanagh; @Filippone]. The recent availability of radioactive beam facilities offers the possibility of resolving this discrepancy indirectly by measuring the Coulomb dissociation of a ${}^8$B nucleus in the field of a heavy-target nucleus like ${}^{208}$Pb. Performing such an experiment at carefully chosen kinematics to minimize nuclear-interaction effects and assuming the break-up as a one-step process in which a single virtual photon is absorbed, the Coulomb dissociation is the inverse of the radiative capture process [@Baur].
Recently an experiment at RIKEN measured the ${}^8{\rm B}+
{}^{208}{\rm Pb} \rightarrow p+{}^7{\rm Be} +{}^{208}{\rm Pb}$ dissociation cross section at the high incident energy of 46.5 MeV/u [@Gai]. Using the semi-classical formulas of Ref. [@Baur], the Coulomb dissociation cross section was translated into $S$-factors for the ${}^7{\rm Be} (p,\gamma) {}^8{\rm B}$ radiative capture process. From this it was concluded that the ${}^7{\rm Be}
(p,\gamma) {}^8{\rm B}$ $S$-factor at solar energies is likely to be smaller than 20 eV-b, supporting the lower [@Filippone] of the two direct ${}^7{\rm Be} (p,\gamma) {}^8{\rm B}$ measurements.
In Ref. [@Gai] the Coulomb dissociation was analyzed as a pure $E1$ break-up process, ignoring possible $E2$ contributions. This assumption is certainly valid for the radiative capture reaction, in which the $E1$ cross section is estimated to dominate $E2$ captures by nearly 3 orders of magnitude at low energies [@Kim]. However, as the number of virtual photons strongly favors $E2$ transitions, the ratio of $E2$-to-$E1$ Coulomb dissociation cross sections ($\sigma^{cd}_{E2}/ \sigma^{cd}_{E1}$) is significantly different, relatively enhancing the importance of $E2$ transitions. As has been shown in studies of the ${}^6{\rm Li} +{}^{208}{\rm Pb} \rightarrow
D+\alpha +{}^{208}{\rm Pb}$ [@Typel], ${}^7{\rm Li} +{}^{208}{\rm
Pb} \rightarrow t+\alpha +{}^{208}{\rm Pb}$ [@Typel1], and ${}^{16}{\rm O} +{}^{208}{\rm Pb} \rightarrow \alpha+{}^{12}{\rm C}
+{}^{208}{\rm Pb}$ [@Shoppa] reactions, this enhancement can amount to more than two orders of magnitude, depending on the kinematics of the break-up process.
In the following we will estimate the $E2$ contribution to the ${}^8{\rm B}+ {}^{208}{\rm Pb} \rightarrow p+{}^7{\rm Be}
+{}^{208}{\rm Pb}$ cross section at the kinematics used in the RIKEN experiment. As in the analysis of Ref. [@Gai] we will use the semi-classical formalism of Baur [*et al.*]{} [@Baur] to connect the break-up cross section to the radiative capture cross section. We adopt the $E1$ and $E2$ capture cross sections from the ${}^7{\rm Be}
(p,\gamma) {}^8{\rm B}$ potential model calculation of Kim [*et al.*]{} [@Kim], which has also served as a theoretical guideline in Ref. [@Gai].
The RIKEN experiment [@Gai] has measured the double differential cross section for the ${}^8{\rm B}+ {}^{208}{\rm Pb} \rightarrow
p+{}^7{\rm Be} +{}^{208}{\rm Pb}$ Coulomb dissociation reaction as a function of the Rutherford scattering angle $\theta_R$ and the center-of-mass energy in the $p+{}^7{\rm Be}$ system, $E_{17}$. One has [@Baur] $$\frac{d^2\sigma}{d\Omega_R dE_{17}}=
\sum_{J_f,\lambda}
\left({\frac{Z_{Pb} e}{\hbar v_i}}\right)^2
a^{-2\lambda+2} B(E\lambda, J_i\rightarrow J_f, E_{17})
\frac{df_{E \lambda} (\theta_R , \xi)}{d \Omega_R}\;,$$ where $$a = \frac{Z_{B} Z_{Pb} e^2}{\mu v_i v_f}$$ is the half-distance of closest approach and $$\xi = \frac{Z_{B} Z_{Pb} e^2}{\hbar}
(\frac{1}{v_f} - \frac{1}{v_i})\;.$$ is the adiabaticity parameter. Here, $v_i, v_f$ denote the relative velocities between projectile and target in the initial and final channels, while $Z_k$ is the atomic number of the fragment $k$. The reduced mass $\mu$ is defined between the ${}^8$B and the ${}^{208}$Pb nuclei. The quantity $\frac{df_{E \lambda} (\theta_R,
\xi)} {d \Omega_R}$ can be calculated in the straight-line approximation from the formulae given in Ref. [@Alder]. Finally, the $B(E\lambda)$ matrix elements are related to the respective partial ${}^7{\rm Be} (p,\gamma) {}^8{\rm B}$ cross sections via $$\sigma_{E \lambda}^{J_f \rightarrow J_i} (p+{}^7{\rm Be}
\rightarrow{}^8{\rm B}+\gamma)=$$ $$\frac{16 \pi^3 (\lambda +1)}{\lambda [(2\lambda +1)!!]^2}
\left({\frac{E_{\gamma}}{\hbar c}} \right)^{2 \lambda +1}
\frac{\hbar^2}{2\mu_{17} E_{17}}
B(E \lambda, J_i \rightarrow J_f, E_{17} )\;,$$ where $J_i, J_f$ are the total angular momenta of the initial and final states in the Coulomb dissociation reaction, $\mu_{17}$ is the reduced mass of the $p+{}^7{\rm Be}$ system and $E_{\gamma}$ denotes the photon energy. We have calculated the $B(E\lambda)$ matrix elements from the partial ${}^7{\rm Be} (p,\gamma) {}^8{\rm B}$ $E1$ and $E2$ cross sections as given in Ref. [@Kim]. This $E1$ cross section agrees well with the measured ${}^7{\rm Be} (p,\gamma)
{}^8{\rm B}$ data. Due to the lack of better experimental constraints, the initial scattering states for the $E2$ cross section have been calculated by using the same $l$-independent radial optical potential fitted to the $M1$ resonance at 633 keV. It should be noted that the $E2$ cross section is not tested directly against experimental data and might thus be viewed as somewhat uncertain. Nevertheless the potential model estimate given in Ref. [@Kim] is probably accurate enough to determine whether $E2$ contributions can be ignored in the ${}^8{\rm B}+ {}^{208}{\rm Pb} \rightarrow
p+{}^7{\rm Be} +{}^{208}{\rm Pb}$ cross sections.
The authors of Ref. [@Gai] have studied the ${}^8{\rm B}+
{}^{208}{\rm Pb} \rightarrow p+{}^7{\rm Be} +{}^{208}{\rm Pb}$ reaction at various relative energies $E_{17}$ between 600 keV and about 2 MeV and at Rutherford scattering angles $\theta_R \leq
6^{\circ}$. In Fig. 1 we show the ratio of virtual photon numbers for $E2$ and $E1$ transitions in the ${}^8{\rm B}+ {}^{208}{\rm Pb}
\rightarrow p+{}^7{\rm Be} +{}^{208}{\rm Pb}$ reaction covering the experimental energy range and at some typical $\theta_R$-values. We observe that the $E2/E1$ enhancement increases with angles, while it decreases with relative energy. While the enhancement is smaller than 100 at all experimentally relevant energies at the smallest angles data have been taken, it already amounts to more than 100 at $\theta_R =2^{\circ}$ for the astrophysically important energy range $E_{17}\leq 1$ MeV. Considering that the ratio of partial $E1$ to $E2$ ${}^7{\rm Be} (p,\gamma) {}^8{\rm B}$ cross sections is estimated [@Kim] to be less than about a factor 1000, we expect that the $E2$ contribution to the ${}^8{\rm B}+ {}^{208}{\rm Pb}
\rightarrow p+{}^7{\rm Be} +{}^{208}{\rm Pb}$ cross section cannot be ignored at angles $\theta_R\geq 2^{\circ}$ and energies $E_{17}\leq1$ MeV. This conjecture is confirmed in Fig. 1 where we have plotted ($\sigma^{cd}_{E2}/ \sigma^{cd}_{E1}$). The maximum of this ratio at around $E_{17} = 633$ keV is related to the lowest $1^+$ resonance in ${}^8$B. The main electromagnetic decay of this state is by $M1$ transition to the ${}^8$B ground state with $J^{\pi}=2^+$. While an $E1$ Coulomb excitation of this resonance is forbidden by parity, an $E2$ excitation is allowed leading to a particularly large $E2$ contribution around the resonance energy. With the partial $E1$ and $E2$ cross sections of Ref. [@Kim], one finds that the $E2$ process dominates the total ${}^8{\rm B}+
{}^{208}{\rm Pb} \rightarrow p+{}^7{\rm Be} +{}^{208}{\rm Pb}$ cross section at angles $\theta_R\geq 4^{\circ}$.
Despite possible uncertainties in the potential model calculation, the $E2$ contribution will contribute significantly to the total Coulomb break-up cross section in the vicinity of the resonance and has to be taken into account in the data analysis. A precise measurement of the Coulomb dissociation cross section at the resonance energy and at angles $\theta_R>2^{\circ}$ will determine the strength of the partial $E2$ capture cross section at this energy and thus place an important constraint on the theoretical modeling of this cross section. Of course, it would be desirable to measure the triple-differential Coulomb dissociation cross section $\frac{d^3
\sigma} {d \Omega_R d \Omega_{17} d E_{17}}$, where $\Omega_{17}$ defines the angle between the proton and the ${}^7$Be nucleus out of the scattering plane. This quantity is sensitive to the interference of $E1$ and $E2$ Coulomb break-up transitions [@Weber].
In Ref. [@Gai] the ${}^7{\rm Be} (p,\gamma) {}^8{\rm B}$ $S$-factors at different relative energies (binned into intervals of 200 keV width) have been determined by fitting the double-differential ${}^8{\rm B}+ {}^{208}{\rm Pb} \rightarrow
p+{}^7{\rm Be} +{}^{208}{\rm Pb}$ yields for fixed energy as a function of Rutherford scattering angle (binned into intervals of width 1 degree). As mentioned above, only $E1$ Coulomb break-up has been considered. We will now discuss how significantly $E2$ break-up might contribute to the data of Ref. [@Gai]. As we do not know the detector efficiency function, a direct calculation of the yields is not possible. Assuming that the detector efficiency is the same for $E1$ and $E2$ contributions, we take the yield curves in Fig. 2 of Ref. [@Gai] and multiply by $(\sigma^{cd}_{E1}+
\sigma^{cd}_{E2})/ \sigma^{cd}_{E1}$. Here we have averaged the cross sections over the same angular and energy bins as in Ref. [@Gai]. We find that the ratio is rather robust against this averaging. The relative importance of the $E2$ contribution can be seen as the difference between the dashed ($E1+E2$) and dotted ($E1$) curves in Fig. 2. As expected, $E2$ Coulomb break-up is most important at the energy interval centered around $E_{17}=0.6$ MeV, which covers the $1^+$ resonance at 633 keV. Here we find a noticeable change of the yield curve in both magnitude and shape. At the higher energies, the effect of the $E2$ break-up is less pronounced than at the resonance energy leading to no significant change in the yield pattern.
As the $E1$ and $E2$ break-up parts add in the double-differential cross section (1), the presence of the $E2$ component in the data will reduce the partial $E1$ cross section compared to the one deduced in Ref. [@Gai], which ignored possible $E2$ contributions. We have fitted the data of Ref. [@Gai] to our ($E1+E2$) yield curves by multiplying the calculated yields with a parameter $\alpha(E_{17})$ which has been determined by $\chi^2$-minimization. As our yields are normalized to the $E1$ yields of Ref. [@Gai], the partial $E1$ ${}^7{\rm Be} (p,\gamma)
{}^8{\rm B}$ $S$-factor extracted from the data scales by the same parameter $\alpha$. We find that at the resonance ($E_{17}=0.6$ MeV) the data agree noticeably better with our ($E1+E2$) yield curve than with a pure $E1$ pattern (Fig. 2); the $\chi^2$ between the two fits is reduced by 30%. Thus, the experimental data at this energy show the presence of the $1^+$ resonance. We obtain a best-fit value of $\alpha(0.6) =0.66\pm0.08$. At the two other energies our fit procedure results in $\alpha(0.8)= 0.82\pm0.16$ and $\alpha(1.0)
=0.77\pm0.17$, while the $\chi^2$-values are about the same for pure $E1$ and our ($E1+E2$) yields pattern. The values of the parameter $\alpha(E_{17})$ translates into the partial $E1$ ${}^7{\rm Be}
(p,\gamma) {}^8{\rm B}$ $S$-factors of $11.2\pm1$ eV-b, $11.5\pm2.5$ eV-b, and $12.3\pm3$ eV-b at $E_{17}=0.6$, 0.8, and 1.0 MeV, respectively. Using the rather reliably known energy dependence of the ${}^7{\rm Be} (p,\gamma) {}^8{\rm B}$ $S$-factor [@Kim; @Descouvemont], these values extrapolate to $S(20~{\rm keV})
=12\pm3$ eV-b. This value is about 25% smaller than the $S$-factor derived from the same data in Ref. [@Gai], and it is only 55% (62%) of the $S$-factor adopted in the most recent version of Bahcall’s [@Bahcall1] (Turck-Chieze’s [@Turck]) Standard Solar Model. We note that such a low value of $S(20~{\rm keV})$ brings the predicted flux of high-energy neutrinos in agreement with the observation of Kamiokande III [@Kamio]. Thus, it is obviously very important to determine the role the $E2$ Coulomb break-up plays in the ${}^8{\rm B}+ {}^{208}{\rm Pb} \rightarrow p+{}^7{\rm Be}
+{}^{208}{\rm Pb}$ dissociation process at low energies.
The $S$-factor extracted here from the ${}^8{\rm B}+ {}^{208}{\rm Pb}
\rightarrow p+{}^7{\rm Be} +{}^{208}{\rm Pb}$ data is noticeably smaller and incompatible (within 2 standard deviations) with the one recently derived from the various direct measurements of the ${}^7{\rm Be} (p,\gamma) {}^8{\rm B}$ reaction [@Johnson]. As it is important to resolve this apparent difference between the two methods, a precise direct capture experiment at one energy to pin down the overall normalization of the direct capture results is highly desirable. A confirmation of the Coulomb dissociation data and a verification of its assumed relation to the capture cross section is also desirable.
In summary, we have shown that the $E2$ component in the ${}^8{\rm
B}+ {}^{208}{\rm Pb} \rightarrow p+{}^7{\rm Be} +{}^{208}{\rm Pb}$ break-up can have a sizeable effect at low energies, in contrast to the assumption of a previous analysis of ${}^8{\rm B}+ {}^{208}{\rm
Pb} \rightarrow p+{}^7{\rm Be} +{}^{208}{\rm Pb}$ data, which ignored the $E2$ contributions [@Gai]. If our conjecture is confirmed, the data of Ref. [@Gai] result in a ${}^7{\rm Be} (p,\gamma)
{}^8{\rm B}$ $S$-factor at solar energies of $12\pm3$ eV-b. This value is noticeably smaller than the $S$-factors obtained in direct capture measurements [@Johnson; @Kavanagh; @Filippone] and, if correct, will obviously have important consequences for the understanding of the solar neutrino puzzle. Our present estimate for the $E2$ cross section is based on a simple potential model and clearly calls for an improved treatment. A more reliable microscopic calculation based on the framework of the multichannel resonating group model is currently in progress [@Csoto]. However, due its potential importance for the solar neutrino problem, an experimental determination of the $E2$ contribution is indispensable. This can be done by measuring the triple-differential cross section $\frac{d^3
\sigma}{d \Omega_R d \Omega_{17} d E_{17}} $, which is sensitive to the interference of $E1$ and $E2$ components and should show sizeable effects of the $E2$ break-up amplitudes, even it is somewhat smaller than estimated in the presently adopted potential model.
The authors thank B. W. Filippone and S. E. Koonin for helpful comments on the manuscript. This work has been supported in part by the National Science Foundation (Grants No. PHY90-13248 and PHY91-15574).
J. N. Bahcall, [*Neutrino Astrophysics*]{} (Cambridge University Press, New York, 1989). C. W. Johnson, E. Kolbe, S. E. Koonin, and K. Langanke, Ap. J. [**392**]{}, 320 (1992). R. W. Kavanagh, T. A. Tombrello, J. M. Mosher, and D. R. Goosman, Bull. Am. Phys. Soc. [**14**]{}, 1209 (1969). B. W. Filippone, S. J. Elwyn, C. N. Davids, and D. D. Koetke, Phys. Rev. Lett. [**50**]{}, 142 (1983); Phys. Rev. C [**28**]{}, 2222 (1983). G. Baur, C. A. Bertulani, and H. Rebel, Nucl. Phys. [**A458**]{}, 188 (1986). T. Motobayashi [*et al.*]{}, RIKEN preprint Rikkyo RUP 94-2, submitted to Phys. Rev. Lett. K. H. Kim, M. H. Park, and B. T. Kim, Phys. Rev. C [**35**]{}, 363 (1987). S. Typel, G. Blüge, and K. Langanke, Z. Phys. A [**339**]{}, 335 (1991). S. Typel, in [*Nuclei in the Cosmos*]{}, edited by F. Käppeler and K. Wisshak, (Institute of Physics Publishing, Bristol, 1993). T. D. Shoppa and S. E. Koonin, Phys. Rev. C [**46**]{}, 382 (1992). K. Alder and A. Winther, Mat. Fys. Medd. Dan. Vid. Selsk. [**31**]{}, 1 (1956). G. Baur and M. Weber, Nucl. Phys. [**A504**]{}, 352 (1989). P. Descouvemont and D. Baye, Nucl. Phys. (to be published). J. N. Bahcall and M. H. Pinsonneault, Rev. Mod. Phys. [**64**]{}, 885 (1992). S. Turck-Chieze [*et al.*]{}, Phys. Rep. [**230**]{}, 57 (1993). K. Inoue, in [*Electroweak Interaction and Unified Theories*]{}, Proceedings of the XXVIIIth Rencontres de Moriond, Les Arcs, Savoyen, France (1993). A. Csótó, Phys. Rev. C (to be published).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The purpose of this paper is to present a universal approach to the study of controllability/observability problems for infinite dimensional systems governed by some stochastic/deterministic partial differential equations. The crucial analytic tool is a class of fundamental weighted identities for stochastic/deterministic partial differential operators, via which one can derive the desired global Carleman estimates. This method can also give a unified treatment of the stabilization, global unique continuation, and inverse problems for some stochastic/deterministic partial differential equations.'
author:
- 'Xu Zhang [^1]'
title: 'A unified controllability/observability theory for some stochastic and deterministic partial differential equations'
---
Primary 93B05; Secondary 35Q93, 93B07.
Controllability, observability, parabolic equations, hyperbolic equations, weighted identity.
Introduction
============
We begin with the following controlled system governed by a linear Ordinary Differential Equation (ODE for short): { =Ay(t)+Bu(t),t>0,\
y(0)=y\_0. . In (\[ols1\]), $A\in \mathbb{R}^{n\times n}$, $B\in
\mathbb{R}^{n\times m}$ ($n, m\in\mathbb{N}$), $y(\cd)$ is the [*state variable*]{}, $u(\cd)$ is the [*control variable*]{}, $\mathbb{R}^n$ and $\mathbb{R}^m$ are the [*state space and control space*]{}, respectively. System (\[ols1\]) is said to be exactly controllable at a time $T>0$ if for any initial state $y_0\in\mathbb{R}^n$ and any final state $y_1\in\mathbb{R}^n$, there is a control $u(\cd)\in L^2(0,T;\mathbb{R}^m)$ such that the solution $y(\cd)$ of (\[ols1\]) satisfies $ y(T)=y_1$.
The above definition of controllability can be easily extended to abstract evolution equations. In the general setting, it may happen that the requirement $y(T)=y_1$ has to be relaxed in one way or another. This leads to the approximate controllability, null controllability, and partial controllability, etc. Roughly speaking, the controllability problem for an evolution process is driving the state of the system to a prescribed final target state (exactly or in some approximate way) at a finite time. Also, the above $B$ can be unbounded for general controlled systems.
The controllability/observability theory for finite dimensional linear systems was introduced by R.E. Kalman ([@15]). It is by now the basis of the whole control theory. Note that a finite dimensional system is usually an approximation of some infinite dimensional system. Therefore, stimulated by Kalman’s work, many mathematicians devoted to extend it to more general systems including infinite dimensional systems, and its nonlinear and stochastic counterparts. However, compared with Kalman’s classical theory, the extended theories are not very mature.
Let us review rapidly the main results of Kalman’s theory. First of all, it is shown that: [*System (\[ols1\]) is exactly controllable at a time $T$ if and only if*]{} $ \hbox{\rm
rank$\,$}[B,AB, \cdots, A^{n-1} B] = n$. However, this criterion is not applicable for general infinite dimensional systems. Instead, in the general setting, one uses another method which reduces the controllability problem for a controlled system to an observability problem for its dual system. The dual system of (\[ols1\]) reads: { =-A\^\*w,t(0,T),\
w(T)=z\_0. . It is shown that: [*System (\[ols1\]) is exactly controllable at some time $T$ if and only if the following observability inequality (or estimate) holds*]{} |z\_0|\^2C\_0\^T|B\^\*w(t)|\^2dt, z\_0\^n. Here and henceforth, $C$ denotes a generic positive constant, which may be different from one place to another. We remark that similar results remain true in the infinite dimensional setting, where the theme of the controllability/observability theory is to establish suitable observability estimates through various approaches.
Systems governed by Partial Differential Equations (PDEs for short) are typically infinite dimensional. There exists many works on controllability/observability of PDEs. Contributions by D.L. Russell ([@Russell]) and by J.L. Lions ([@Lions]) are classical in this field. In particular, since it stimulated many in-depth researches on related problems in PDEs, J.L. Lions’s paper [@Lions] triggered extensive works addressing the controllability/observability of infinite dimensional controlled system. After [@Lions], important works in this field can be found in [@AI; @BLR; @Coron; @FGIP; @Fu1; @FI; @KoL; @LR; @TTLi; @TW; @Yama; @Zua1; @Zua2]. For other related works, we refer to [@Isakov; @LY] and so on.
The controllability/observability of PDEs depends strongly on the nature of the underlying system, such as time reversibility or not, and propagation speed of solutions, etc. The wave equation and the heat equation are typical examples. Now it is clear that essential differences exist between the controllability/observability theories for these two equations. Naturally, one expects to know whether some relationship exist between the controllability/observability theories for these two equations of different nature. Especially, it would be quite interesting to establish, in some sense and to some extend, a unified controllability/observability theory for parabolic equations and hyperbolic equations. This problem was initially studied by D.L. Russell ([@28]).
The main purpose of this paper is to present the author’s and his collaborators’ works with an effort towards a unified controllability/observability theory for stochastic/deterministic PDEs. The crucial analytic tool we employ is a class of elementary pointwise weighted identities for partial differential operators. Starting from these identities, we develop a unified approach, based on global Carleman estimate. This universal approach not only deduces the known controllability/observability results (that have been derived before via Carleman estimates) for the linear parabolic, hyperbolic, Schrödinger and plate equations, but also provides new/sharp results on controllability/observability, global unique continuation, stabilization and inverse problems for some stochastic/deterministic linear/nonlinear PDEs.
The rest of this paper is organized as follows. Section \[s2\] analyzes the main differences between the existing controllability/observability theories for parabolic equations and hyperbolic equations. Sections \[s3\] and \[s4\] address, among others, the unified treatment of the controllability/observability problem for deterministic PDEs and stochastic PDEs, respectively.
Main differences between the known theories {#s2}
===========================================
In the sequel, unless otherwise indicated, $G$ stands for a bounded domain (in $\mathbb{R}^n$) with a boundary $\G\in C^2$, $G_0$ denotes an open non-empty subset of $G$, and $T$ is a given positive number. Put $Q=(0,T)\times G$, $Q_{G_0}=(0,T)\times G_0$ and $\Si=(0,T)\times\G$.
We begin with a controlled heat equation: \[he\]{ y\_[t]{}-y=\_[G\_0]{}(x)u(t,x) & Q,\
y=0 & ,\
y(0)=y\_[0]{} & G . and a controlled wave equation: \[we\]{ y\_[tt]{}-y=\_[G\_0]{}(x)u(t,x) & Q,\
y=0 & ,\
y(0)=y\_[0]{}, y\_[t]{}(0)=y\_[1]{} & G. .In (\[he\]), $y$ and $u$ are the [*state variable*]{} and [*control variable*]{}, the [*state space*]{} and [*control space*]{} are chosen to be $L^{2}(G)$ and $L^{2}(Q_{G_0})$, respectively; while in (\[we\]), $(y,y_t)$ and $u$ are the [*state variable*]{} and [*control variable*]{}, $H^{1}_{0}(G)\times L^{2}(G)$ and $L^{2}(Q_{G_0})$ are respectively the [*state space*]{} and [*control space*]{}. System (\[he\]) is said to be null controllable (approximately controllable) in $L^2(G)$ if for any given $y_0\in L^2(G)$ (for any given $\e>0$, $y_0, y_1\in L^2(G)$), one can find a control $u\in L^2(Q_{G_0})$ such that the weak solution $y(\cd)\in C([0,T];L^2(G))\cap C((0,T];H_0^1(G))$ of (\[he\]) satisfies $y(T)=0$ ($|y(T)-y_1|_{L^2(G)}\le \e$). In the case of null controllability, the corresponding control $u$ is called a null-control (with initial state $y_0$). Note that, due to the smoothing effect of solutions to the heat equation, the exact controllability for (\[he\]) is impossible, i.e., the above $\e$ cannot be zero. On the other hand, since one can rewrite system (\[we\]) as an evolution equation in a form like (\[ols1\]), it is easy to define the exact controllability of this system. The dual systems of (\[he\]) and (\[we\]) read respectively $$\begin{aligned}
\label{olwsystem1}
\left\{
\begin{array}{lll}
\ds \psi_t+ \D \psi = 0
&\mbox{ in } Q,\\
\psi = 0 & \mbox{ on } \Si, \\
\psi(T) = \psi_0&\mbox{ in } G
\end{array}
\right.\end{aligned}$$ and $$\begin{aligned}
\label{olhsystem1}
\left\{
\begin{array}{lll}
\ds \psi_{tt} - \D \psi = 0
&\mbox{ in } Q,\\
\psi = 0 & \mbox{ on } \Si, \\
\psi(T) = \psi_0, \q\psi_t(T) = \psi_1 &\mbox{ in } G.
\end{array}
\right.\end{aligned}$$
The controllability/observability theories for parabolic equations and hyperbolic equations turns out to be quite different. First of all, we recall the related result for the heat equation.
\[2t1\] [([@LR])]{} Let $G$ be a bounded domain of class $C^{\infty}$. Then: i) System (\[he\]) is null controllable and approximately controllable in $L^2(G)$ at time $T$; ii) Solutions of equation (\[olwsystem1\]) satisfy |(0)|\_[L\^2(G)]{}C||\_[L\^2(Q\_[G\_0]{})]{},\_0L\^2(G).
Since solutions to the heat equation have an infinite propagation speed, the “waiting" time $T$ can be chosen as small as one likes, and the control domain $G_0$ dose not need to satisfy any geometric condition but being open and non-empty. On the other hand, due to the time irreversibility and the strong dissipativity of (\[olwsystem1\]), one cannot replace $|\psi(0)|_{L^2(G)}$ in inequality (\[icm1\]) by $|\psi_0|_{L^2(G)}$.
Denote by $\{\mu_i\}^\infty_{i=1}$ the eigenvalues of the homogenous Dirichlet Laplacian on $G$, and $\{\f_i\}^{\infty}_{i=1}$ the corresponding eigenvectors satisfying $|\f_i|_{L^2(G)} = 1$. The proof of Theorem \[2t1\] is based on the following observability estimate on sums of eigenfunctions for the Laplacian ([@LR]):
\[icm2t1\] Under the assumption of Theorem \[2t1\], for any $r
> 0$, it holds $$\label{lr}
\sum_{\mu_{i} \leq r}|a_{i}|^{2} \leq C
e^{C\sqrt{r}}{\int_{G_0}}\bigg| \sum_{\mu_{i} \leq r}a_{i}\f_{i}(x)
\bigg|^2dx,\ \ \forall\, \{a_{i}\}_{\mu_{i} \leq r }\hb{ with }a_{i}
\in \mathbb{C}.$$
Note that Theorem \[icm2t1\] has some other applications in control problems of PDEs ([@Lopez; @Miller; @Wang; @33; @Zua1; @Zua2]). Besides, to prove Theorem \[2t1\], one needs to utilize a time iteration method ([@LR]), which uses essentially the Fourier decomposition of solutions to (\[olwsystem1\]) and especially, the strong dissipativity of this equation. Hence, this method cannot be applied to conservative systems (say, system (\[we\])) or the system that the underlined equation is time-dependent.
As for the controllability/observability for the wave equation, we need to introduce the following notations. Fix any $x_0\in
\mathbb{R}^n$, put \[boundary\] \_0={x| (x-x\_0)(x)>0}, where $\nu(x)$ is the unit outward normal vector of $G$ at $x\in\Gamma$. For any set $S\in \mathbb{R}^n$ and $\e>0$, put ${\cal
O}_{\epsilon}(S)=\big\{y\in \mathbb{R}^n\ \big|\
|y-x|<\e\hbox { for some }x\in S\big\}$.
The exact controllability of system (\[we\]) is equivalent to the following [*observability estimate*]{} for system (\[olhsystem1\]): $$\label{oinlhsystem1}
|(\psi_0,\psi_1)|_{L^2(G)\t H^{-1}(G)} \1n\leq\1n
C|\psi|_{L^2(Q_{G_0})}, \ \ \ \forall\, (\psi_0,\psi_1)\in L^2(G)\t
H^{-1}(G).$$ Note that the left hand side of (\[oinlhsystem1\]) can be replaced by $|(\psi(0),\psi_t(0))|^2_{L^2(G)\t H^{-1}(G)}$ (because (\[olhsystem1\]) is conservative). The following classical result can be found in [@Lions].
\[thlr\] Assume $G_0=O_{\e}(\G_0)\cap G$ and $\ds
T_0=2\sup_{x\in G\setminus G_0}|x-x_0|$. Then, inequality (\[oinlhsystem1\]) holds for any time $T>T_0$.
The proof of Theorem \[thlr\] is based on a classical Rellich-type multiplier method. Indeed, it is a consequence of the following identity (e.g. [@31]):
Let $h\=(h^1, \cdots, h^n):\
\mathbb{R}\t\mathbb{R}^{n}\to\mathbb{R}^n$ be a vector field of class $C^1$. Then for any $z\in C^2(\mathbb{R}\t\mathbb{R}^n)$, it holds that $$\ba{ll}
%
\ds\n\cd\Big\{2(h\cd\n z)(\n z)+h\Big[z_t^2-\sum_{i=1}^nz_{x_i}^2\Big]\Big\}\\
\ns\ds
=-2(z_{tt}-\D z)h\cd\n z+(2z_th\cd\n z)_t-2z_th_t\cd\n z\\
\ns\ds\q+(\n\cd
h)\Big[z_t^2-\sum_{i=1}^nz_{x_i}^2\Big]
+2\sum_{i,j=1}^n\Big(\frac{\pa h^j}{\pa x_i}z_{x_i}z_{x_j}\Big).
\ea$$
The observability time $T$ in Theorem \[thlr\] should be large enough. This is due to the finite propagation speed of solutions to the wave equation (except when the control is acting in the whole domain $G$). On the other hand, it is shown in [@BLR] that exact controllability of (\[we\]) is impossible without geometric conditions on $G_0$. Note also that, the multiplier method rarely provides the optimal control/observation domain and minimal controll/observation time except for some very special geometries. These restrictions are weakened by the microlocal analysis ([@BLR]). In [@BLR; @Burq; @BG], the authors proved that, roughly speaking, inequality (\[oinlhsystem1\]) holds if and only if every ray of Geometric Optics that propagates in $G$ and is reflected on its boundary $\G$ enters $G_0$ at time less than $T$.
The above discussion indicates that the results and methods for the controllability/observability of the heat equation differ from those of the wave equation. As we mentioned before, this leads to the problem of establishing a unified theory for the controllability/observability of parabolic equations and hyperbolic equations. The first result in this direction was given in [@28], which showed that [*the exact controllability of the wave equation implies the null controllability of the heat equation with the same controller but in a short time*]{}. Further results were obtained in [@Lopez; @33], in which organic connections were established for the controllability theories between parabolic equations and hyperbolic equations. More precisely, it has been shown that: [*i) By taking the singular limit of some exactly controllable hyperbolic equations, one gives the null controllability of some parabolic equations ([@Lopez]); and ii) Controllability results of the heat equation can be derived from the exact controllability of some hyperbolic equations ([@33])*]{}. Other interesting related works can be found in [@Miller; @Phung; @TW]. In the sequel, we shall focus mainly on a unified treatment of the controllability/observability for both deterministic PDEs and stochastic PDEs, from the methodology point of view.
The deterministic case {#s3}
======================
The key to solve controllability/observability problems for PDEs is the obtention of suitable observability inequalities for the underlying homogeneous systems. Nevertheless, as we see in Section \[s2\], the techniques that have been developed to obtain such estimates depend heavily on the nature of the equations, especially when one expects to obtain sharp results for time-invariant equations. As for the time-variant case, in principle one needs to employ Carleman estimates, see [@FI] for the parabolic equation and [@31] for the hyperbolic equation. The Carleman estimate is simply a weighted energy method. However, at least formally, the Carleman estimate used to derive the observability inequality for parabolic equations is quite different from that for hyperbolic ones. The main purpose of this section is to present a universal approach for the controllability/observability of some deterministic PDEs. Our approach is based on global Carleman estimates via a fundamental pointwise weighted identity for partial differential operators of second order (It was established in [@Fu1; @fx]. See [@[20]] for an earlier result). This approach is stimulated by [@LRS; @KK], both of which are addressed for ill-posed problems.
A stimulating example
---------------------
The basic idea of Carleman estimates is available in proving the stability of ODEs ([@[20]]). Indeed, consider an ODE in $\mathbb{R}^{n}$: \[ODE\] { x\_t(t)=a(t)x(t), t,\
x(0)=x\_[0]{}, . where $a\in L^{\i}(0,T)$. A well-known simple result reads: [*Solutions of (\[ODE\]) satisfy \_[t]{}|x(t)|C|x\_0|,x\_0\^n.* ]{}
[*A Carleman-type Proof of (\[3e1\])*]{}. For any $\l\in
\mathbb{R}$, by (\[ODE\]), one obtains \[CD\] $e^{-\l t}|x(t)|^{2}$=-łe\^[-łt]{}|x(t)|\^[2]{}+2e\^[-łt]{} x\_t(t)x(t) =(2a(t)-ł)e\^[-łt]{}|x(t)|\^[2]{}. Choosing $\l$ large enough so that $2a(t)-\l\le 0$ for a.e. $t\in
(0,T)$, we find that $$|x(t)|\leq e^{\l T/2}|x_{0}|,\quad t\in[0,T],$$ which proves (\[3e1\]).
By (\[CD\]), we see the following pointwise identity: 2e\^[-łt]{} x\_t(t)x(t)=$e^{-\l t}|x(t)|^{2}$+łe\^[-łt]{}|x(t)|\^[2]{}. Note that $ x_t(t)$ is the principal operator of the first equation in (\[ODE\]). The main idea of (\[3e2\]) is to establish a pointwise identity (and/or estimate) on the principal operator $
x_t(t)$ in terms of the sum of a “divergence" term $\frac{d}{dt}(e^{-\l t}|x(t)|^{2})$ and an “energy" term $\l e^{-\l
t}|x(t)|^{2}$. As we see in the above proof, one chooses $\l$ to be big enough to absorb the undesired terms. This is the key of all Carleman-type estimates. In the sequel, we use exactly the same method, i.e., the method of Carleman estimate via pointwise estimate, to derive observability inequalities for both parabolic equations and hyperbolic equations.
Pointwise weighted identity
---------------------------
We now show a fundamental pointwise weighted identity for general partial differential operator of second order. Fix real functions $\a,\;\b\in C^1(\mathbb{R}^{1+m})$ and $b^{jk}\in
C^1(\mathbb{R}^{1+m})$ satisfying $
b^{jk}=b^{kj}$ ($j,k=1,2,\cdots,m$). Define a formal differential operator of second order: $\ds \cP
z\=(\a+i\b)z_t+\sum_{j,k=1}^m\big(b^{jk}z_{x_j}\big)_{x_k}$, $i=\sqrt{-1}$. The following identity was established in [@Fu1; @fx]:
\[uTheorem\] Let $z\in C^2(\mathbb{R}^{1+m}; \;\mathbb{C})$ and $\ell\in
C^2(\mathbb{R}^{1+m};\mathbb{R})$. Put $\th=e^\ell$ and $v=\th z$. Let $a, b,\l\in\mathbb{R}$ be parameters. Then (z+ I\_1)+M\_t+\_[k=1]{}\^m\_[x\_k]{}V\^k\
=2|I\_1|\^2+\_[j,k,j’,k’=1]{}\^m$$2(b^{j'k}\ell_{x_{j'}})_{x_{k'}}b^{jk'}-(b^{jk}b^{j'k'}\ell_{x_{j'}})_{x_{k'}}+\frac{1}{2}(\a
b^{jk})_t\\
\ns\ds\q
-ab^{jk}b^{j'k'}\ell_{x_{j'}x_{k'}}$$(v\_[x\_k]{}\_[x\_j]{}+\_[x\_k]{} v\_[x\_j]{})+$$-
\sum_{j,k=1}^mb^{jk}_{x_k}\ell_{x_j}+b\l$$(I\_1+v)\
+i\_[j,k=1]{}\^m{\[(b\^[jk]{}\_[x\_j]{})\_t+b\^[jk]{}(\_t)\_[x\_j]{}\](\_[x\_k]{}v-v\_[x\_k]{})\
+\[(b\^[jk]{}\_[x\_j]{})\_[x\_k]{}+ab\^[jk]{}\_[x\_jx\_k]{}\](v\_t-v\_t)}-\_[j,k=1]{}\^mb\^[jk]{}\_[x\_k]{}(v\_[x\_j]{}\_t+\_[x\_j]{}v\_t)\
-a\_[j,k,j’,k’=1]{}\^mb\^[jk]{}(b\^[j’k’]{}\_[x\_[j’]{}x\_[k’]{}]{})\_[x\_k]{}(\_[x\_j]{}v+v\_[x\_j]{})+B|v|\^2, where $$\left\{
\ba{ll}
\ds I_1\=i\b v_t-\a\ell_tv+\sum_{j,k=1}^m(b^{jk}v_{x_j})_{x_k}+Av,\\
\ns \ds
A\=\sum_{j,k=1}^mb^{jk}\ell_{x_j}\ell_{x_k}-(1+a)\sum_{j,k=1}^mb^{jk}\ell_{x_jx_k}-b\l,
\\
\ns \ds B\=(\a^2\ell_t+\b^2\ell_{t}-\a A)_t\\ \ns\ds
\qq+2\sum_{j,k=1}^m\[(b^{jk}\ell_{x_j}A)_{x_k}-(\a
b^{jk}\ell_{x_j}\ell_{t})_{x_k}+a(A-\a\ell_t)b^{jk}\ell_{{x_j}{x_k}}\],
\\
\ns \ds M\=\[(\a^2+\b^2)\ell_t-\a A\]
|v|^2+\a\sum_{j,k=1}^mb^{jk}v_{x_j}\ov_{x_k}\\ \ns\ds
\qq+i\b\sum_{j,k=1}^mb^{jk}\ell_{x_j}(\ov_{x_k}v-v_{x_k}\ov),
\\
\ns \ds V^k\=\sum_{j,j',k'=1}^m\Big\{-i
\b\[b^{jk}\ell_{x_j}(v\ov_t-\ov
v_t)+b^{jk}\ell_t(v_{x_j}\ov-\ov_{x_j}v)\]\\
\ns \ds\qq\qq\qq\q -\a b^{jk}(v_{x_j}\ov_t+\ov_{x_j}v_t)\\
\ns \ds\qq\qq\qq\q+(2b^{jk'}b^{j'k}-b^{jk}b^{j'k'})\ell_{x_j}(v_{x_{j'}}\ov_{x_{k'}}+\ov_{x_{j'}}v_{x_{k'}}) \\
\ns \ds\qq\qq\qq\q -ab^{j'k'}\ell_{x_{j'}x_{k'}}
b^{jk}(v_{x_j}\ov+\ov_{x_j}v)+2b^{jk}(A\ell_{x_j}-\a\ell_{x_j}\ell_t)|v|^2\Big\}.
\ea\right.$$
As we shall see later, Theorem \[uTheorem\] can be applied to study the controllability/observability as well as the stabilization of parabolic equations and hyperbolic equations. Also, as pointed by [@Fu1], starting from Theorem \[uTheorem\], one can deduce the controllability/observability for the Schrödinger equation and plate equation appeared in [@LTZ2] and [@32], respectively. Note also that, Theorem \[uTheorem\] can be applied to study the controllability of the linear/nonlinear complex Ginzburg-Landau equation (see [@Fu1; @fx; @[26]]).
Controllability/Observability of Linear PDEs
--------------------------------------------
In this subsection, we show that, starting from Theorem \[uTheorem\], one can establish sharp observability/controllability results for both parabolic systems and hyperbolic systems.
We need to introduce the following assumptions.
\[condition of d-0\] Matrix-valued function $\big(p^{ij}\big)_{1\le i,j\le n}\in C^{1}(\overline{Q};
\mathbb{R}^{n\times n})$ is uniformly positive definite.
\[condition of d-1\] Matrix-valued function $\big(h^{ij}\big)_{1\le i,j\le n}\in C^{1}(\cl{G};
\mathbb{R}^{n\times n})$ is uniformly positive definite.
Also, for any $N\in \mathbb{N}$, we introduce the following
\[condition of d-2\] Matrix-valued functions $a\in
L^{\infty}(0,T;L^{p}(G;\mathbb{R}^{N\t N}))$ for some $p \in
[n,\infty]$, and $a_1^1, \cdots, a_1^n, a_2 \in
L^{\infty}(Q;\mathbb{R}^{N\t N})$.
Let us consider first the following parabolic system: $$\begin{aligned}
\label{system4}
\left\{
\begin{array}{lll}
\ds\varphi_t - \sum_{i,j=1}^n (p^{ij}\varphi_{x_i})_{x_j} =
a\varphi + \sum_{k=1}^n a_1^k \varphi_{x_k}, &\mbox{ in } Q,\\
\ns\ds \varphi = 0, &\mbox{ on } \Si,\\
\ns\ds \varphi(0) = \varphi^0, &\mbox{ in } G,
\end{array}
\right.\end{aligned}$$ where $\varphi$ takes values in $\mathbb{R}^{N}$. By choosing $\a =
1$ and $\b = 0$ in Theorem \[uTheorem\], one obtains a weighted identity for the parabolic operator. Along with [@[20]], this identity leads to the existing controllability/observability result for parabolic equations ([@Doubova1; @FI]). One can go a little further to show the following result ([@DZZ]):
\[oTheorem4\] Let Conditions \[condition of d-0\] and \[condition of d-2\] hold. Then, solutions of (\[system4\]) satisfy \[oest3\] n |(T)|\_[(L\^2(G))\^N]{}\
{ C }||\_[(L\^2(Q\_[G\_0]{}))\^N]{}, \^0 (L\^2(G))\^N.
Note that (\[oest3\]) provides the observability inequality for the parabolic system (\[system4\]) with an explicit estimate on the observability constant, depending on the observation time $T$, the potential $a$ and $a_1^k$. Earlier result in this respect can be found in [@Doubova1] and the references cited therein. Inequality (\[oest3\]) will play a key role in the study of the null controllability problem for semilinear parabolic equations, as we shall see later.
\[icmrem1\] It is shown in [@DZZ] that when $n\ge 2$, $N\ge 2$ and $\big(p^{ij}\big)_{1\leq i,j\leq n} = I$, the exponent $\frac{2}{3}$ in $|a|_{L^{\infty}(0,T;L^{p}(G;\mathbb{R}^{N\t N}))}^{\frac{2}{3}}$ (for the case that $p=\infty$ in the inequality (\[oest3\])) is sharp. In [@DZZ], it is also proved that the quadratic dependence on $\ds\sum_{k=1}^N |a_i^k|_{L^{\infty}(Q;\mathbb{R}^{N\t
N})}$ is sharp under the same assumptions. However, it is not clear whether the exponent $\frac{3}{2}-\frac{n}{p}$ in $|a|_{L^{\infty}(0,T;L^{p}(G;\mathbb{R}^{N\t
N}))}^{\frac{1}{\frac{3}{2}-\frac{n}{p}}}$ is optimal when $p<\infty$.
Next, we consider the following hyperbolic system: $$\begin{aligned}
\label{system5}
\left\{
\begin{array}{lll}
\ds v_{tt} - \sum_{i,j=1}^n (h^{ij}v_{x_i})_{x_j} =
av + \sum_{k=1}^n a_1^k v_{x_k} + a_2 v_t, &\mbox{ in } Q,\\
\ns\ds v = 0, &\mbox{ on } \Si,\\
\ns\ds v(0) = v^0, \q v_t(0) = v^1, &\mbox{ in } G,
\end{array}
\right.\end{aligned}$$ where $v$ takes values in $\mathbb{R}^{N}$.
Compared with the parabolic case, one needs more assumptions on the coefficient matrix $\big(h^{ij}\big)_{1\le i,j\le n}$ as follows ([@DZZ; @Fu-Yong-Zhang]):
\[condition of d\] There is a positive function $d(\cdot) \in
C^2(\overline{G})$ satisfying
i\) For some constant $\mu_0 \ge 4$, it holds $$\ba{ll}\ds
\sum_{i,j=1}^n\Big\{ \sum_{i',j'=1}^n\Big[
2h^{ij'}(h^{i'j}d_{x_{i'}})_{x_{j'}} -
h^{ij}_{x_{j'}}h^{i'j'}d_{x_{i'}} \Big] \Big\}\xi^{i}\xi^{j} \geq
\mu_0
\sum_{i,j=1}^n h^{ij}\xi^{i}\xi^{j}, \\
\ns
\ds\qq\qq\qq\qq\qq\qq\qq\qq\qq\qq
\forall\; (x,\xi^{1},\cdots,\xi^{n}) \in \overline{G} \t
\mathbb{R}^n; \ea$$
ii\) There is no critical point of $d(\cdot)$ in $\overline{G}$, i.e., $\ds\min_{x\in \overline{G}}|\nabla d(x)| > 0$;
iii\) $\q\ds\frac{1}{4}\sum_{i,j=1}^n h^{ij}(x)d_{x_i}(x)d_{x_j}(x)
\geq \max_{x\in\overline{G}} d(x), \q \forall x\in \overline{G}$.
We put $$\label{tstar}
T^* = 2\max_{x\in
\overline{G}}\sqrt{d(x)}, \qq
\G^* \= \Big\{x\in\G \;\Big|\;
\sum_{i,j=1}^n h^{ij}(x)d_{x_i}(x)\nu_j(x)>0 \Big\}.$$ By choosing $b^{jk}(t,x) \equiv h^{jk}(x)$ and $\a = \b = 0$ in Theorem \[uTheorem\] (and noting that only the symmetry condition is assumed for $b^{jk}$ in this theorem), one obtains the fundamental identity derived in [@Fu-Yong-Zhang] to establish the controllability/observability of the general hyperbolic equations. One can go a little further to show the following result ([@DZZ]).
\[oTheorem5\] Let Conditions \[condition of d-1\], \[condition of d-2\] and \[condition of d\] hold, $T>T^*$ and $G_0=G\cap \cO_{\e}(\G^*)$ for some $\e>0$. Then one has the following conclusions:
ß
1\) For any $(v^0,v^1) \in (H_0^1(G))^N \t (L^2(G))^N$, the corresponding weak solution $ v \in C([0,T];(H_0^1(G))^N)\bigcap $ $C^1([0,T];(L^2(G))^N) $ of system (\[system5\]) satisfies |v\^0|\_[H\_0\^1(G))\^N]{} + |v\^1|\_[(L\^2(G))\^N]{}\
| |\_[(L\^2((0,T)\^\*))\^N]{}.
ß
2\) If $a_1^k \equiv 0$ ($k = 1,\cdots,n$) and $a_2 \equiv 0$, then for any $(v^0,v^1) \in (L^2(G))^N \t (H^{-1}(G))^N$, the weak solution $ v \in C([0,T];(L^2(G))^N)\bigcap C^1([0,T];(H^{-1}(G))^N)
$ of system (\[system5\]) satisfies |v\^0|\_[(L\^2(G))\^N]{} + |v\^1|\_[H\^[-1]{}(G))\^N]{}\
|v|\_[(L\^2(Q\_[G\_0]{}))\^N]{}.
As we shall see in the next subsection, inequality (\[oest5\]) plays a crucial role in the study of the exact controllability problem for semilinear hyperbolic equations.
\[ICMr1\] As in the parabolic case, it is shown in [@DZZ] that the exponent $\ds\frac{2}{3}$ in the estimate $|a|_{L^{\infty}(0,T;L^{p}(G;\mathbb{R}^{N\t N}))}^{\frac{2}{3}}$ in (\[oest5\]) (for the special case $p =\infty$) is sharp for $n\ge
2$ and $N\ge 2$. Also, the exponent $2$ in the term $\ds\Big(\sum_{k=1}^N |a_i^k|_{L^{\infty}(Q;\mathbb{R}^{N\t
N})}+|a_2|_{L^{\infty}(Q;\mathbb{R}^{N\t N})}\Big)^2$ in (\[oest4\]) is sharp. However, it is unknown whether the estimate is optimal for the case that $p<\infty$.
By the standard duality argument, Theorems \[oTheorem4\] and \[oTheorem5\] can be applied to deduce the controllability results for parabolic systems and hyperbolic systems, respectively. We omit the details.
Controllability of Semi-linear PDEs
-----------------------------------
The study of exact/null controllability problems for semi-linear PDEs began in the 1960s. Early works in this respect were mainly devoted to the local controllability problem. By the local controllability of a system, we mean that the controllability property holds under some smallness assumptions on the initial data and/or the final target, or the Lipschitz constant of the nonlinearity.
In this subsection we shall present some global controllability results for both semilinear parabolic equations and hyperbolic equations. These results can be deduced from Theorems \[oTheorem4\] and \[oTheorem5\], respectively.
Consider first the following controlled semi-linear parabolic equation: $$\begin{aligned}
\label{slheat1}
\left\{
\begin{array}{lll}
\ds y_t - \sum_{i,j=1}^n (p^{ij}y_{x_i})_{x_j} + f(y,\nabla y) =\chi_{G_0} u, &\mbox{ in } Q,\\
\ns\ds y = 0, &\mbox{ on } \Si,\\
\ns\ds y(0) = y_0, &\mbox{ in }G.
\end{array}
\right.\end{aligned}$$ For system (\[slheat1\]), the state variable and control variable, state space and control space, controllability, are chosen/defined in a similar way as for system (\[he\]). Concerning the nonlinearity $f(\cd,\cd)$, we introduce the following assumption ([@Doubova1]).
\[condition of d4\] Function $f(\cdot,\cdot) \in C(\mathbb{R}^{1+n})$ is locally Lipschitz-continuous. It satisfies $f(0,0) = 0$ and $$\begin{aligned}
\label{f1}
\left\{
\begin{array}{ll}
\ds \lim_{|(s,p)|\to\infty}\frac{\int_0^1
f_s (\tau s, \tau p)d\tau}{\ln^{\frac{3}{2}}(1+ |s| + |p|)} =
0,\\
\ns\ds \lim_{|(s,p)|\to\infty}\frac{|(\int_0^1 f_{p_1} (\tau s, \tau
p)d\tau,\cdots,\int_0^1 f_{p_n} (\tau s, \tau
p)d\tau)|}{\ln^{\frac{1}{2}}(1+ |s| + |p|)} = 0,
\end{array}
\right.\end{aligned}$$ where $p = (p_1,\cdots, p_n)$.
As shown in [@Doubova1] (See [@barbu] and the references therein for earlier results), linearizing the equation, estimating the cost of the control in terms of the size of the potential entering in the system (thanks to Theorem \[oTheorem4\]), and using the classical fixed point argument, one can show the following result.
\[3.1\] Assume that Conditions \[condition of d-0\] and \[condition of d4\] hold. Then system (\[slheat1\]) is null controllable.
In particular, Theorem \[3.1\] provides the possibility of controlling some blowing-up equations. More precisely, assume that $f(s,p)\equiv f(s)$ in system (\[slheat1\]) has the form $$\label{eq4connections}%4.4
f(s)=-s\ln^r(1+|s|),\qq r\ge 0.$$ When $r>1$, solutions of (\[slheat1\]), in the absence of control, i.e. with $u\equiv0$, blow-up in finite time. According to Theorem \[3.1\] the process can be controlled, and, in particular, the blow-up can be avoided when $1<r\le 3/2$. By the contrary, it is proved in [@barbu; @FZ] that for some nonlinearities $f$ satisfying $$\label{zzeq3connections}%4.3
\ds\lim_{\mid s\mid\to\infty}\frac{\mid f(s)\mid}{s\ln^r(1+\mid
s\mid)}=0,$$ where $r>2$, the corresponding system is not controllable. The reason is that the controls cannot help the system to avoid blow-up.
It is still an unsolved problem whether the controllability holds for system (\[slheat1\]) in which the nonlinear function $f(\cd)$ satisfies (\[zzeq3connections\]) with $3/2\leq r\leq 2$. Note that, the growth condition in (\[f1\]) comes from the observability inequality (\[oest3\]). Indeed, the logarithmic function in (\[f1\]) is precisely the inverse of the exponential one in (\[oest3\]). According to Remark \[icmrem1\], the estimate (\[oest3\]) cannot be improved, and therefore, the usual linearization approach cannot lead to any improvement of the growth condition (\[f1\]).
Next, we consider the following controlled semi-linear hyperbolic equation: $$\begin{aligned}
\label{system3}
\left\{
\begin{array}{lll}
\ds y_{tt} - \sum_{i,j=1}^n (h^{ij}y_{x_i})_{x_j} = h(y) +
\chi_{G_0}u &\mbox{ in } Q,\\
\ns\ds y = 0 & \mbox{ on } \Si,\\
\ns\ds y(0) = y_0, \q y_t(0) = y_1 & \mbox { in } G.
\end{array}
\right.\end{aligned}$$ For system (\[system3\]), the state variable and control variable, state space and control space, controllability, are chosen/defined in a similar way as that for system (\[we\]). Concerning the nonlinearity $h(\cd)$, we need the following assumption ([@DZZ]).
\[condition of d5\] Function $h(\cdot) \in C(\mathbb{R})$ is locally Lipschitz-continuous, and for some $r\in [0,{\frac{3}{2}})$, it satisfies that $$\begin{aligned}
\label{h}
\lim_{|s|\to\infty}\frac{\int_0^1 h_s (\tau s)d\tau}{\ln^r(1+ |s|)}
= 0.\end{aligned}$$
As mentioned in [@DZZ], proceeding as in the proof of [@Fu-Yong-Zhang Theorem 2.2], i.e., by the linearization approach (thanks to the second conclusion in Theorem \[oTheorem5\]), noting that the embedding $H_0^1(G)\hookrightarrow
L^2(G)$ is compact, and using the fixed point technique, one can show the following result.
Assume that Conditions \[condition of d-1\], \[condition of d\] and \[condition of d5\] are satisfied, and $T$ and $G_0$ are given as in Theorem \[oTheorem5\]. Then system (\[slheat1\]) is exactly controllable.
Due to the blow-up and the finite propagation speed of solutions to hyperbolic equations, one cannot expect exact controllability of system (\[slheat1\]) for nonlinearities of the form (\[h\]) with $r>2$. One could expect the system to be controllable for $r\leq2$. However, in view of Remark \[ICMr1\], the usual fixed point method cannot be applied for $r\ge3/2$. Therefore, when $n\ge 2$, the controllability problem for system (\[system3\]) is open for $3/2\le r\leq2$.
Note that the above “$3/2$ logarithmic growth" phenomenon (arising in the global exact controllability for nonlinear PDEs) does not occur in the pure PDE problem, and therefore the study of nonlinear controllability is of independent interest. More precisely, this means that for the controllability problem of nonlinear systems, there exist some extra difficulties.
Controllability of Quasilinear PDEs
-----------------------------------
In this subsection, we consider the controllability of quasilinear parabolic/hyperbolic equations.
We begin with the following controlled quasilinear hyperbolic equation: $$\begin{aligned}
\label{qhsystem1}
\left\{
\begin{array}{lll}
\ds y_{tt} - \sum_{i,j=1}^n (h^{ij}y_{x_i})_{x_j} =
F(t,x,y,\nabla_{t,x}y,\nabla^2_{t,x}y) +qy+ \phi_{G_0}u, &\mbox{ in
}
Q,\\
\ns\ds y = 0, &\mbox{ on } \Si,\\
\ns\ds y(0) = y_0, \q y_t(0) = y_1, &\mbox{ in } G.
\end{array}
\right.\end{aligned}$$ Here, $\big(h^{i j}\big)_{1\le i,j\le n}\in H^{s+1}(G;
\mathbb{R}^{n\times n})$ and $\ q\in H^{s}(Q)$ with $\ds s
> \frac{n}{2}+1$, and similar to [@ZL], the nonlinear term $F(\cdot)$ has the form $$F(t,x,y,\nabla_{t,x}y,\nabla^2_{t,x}y) = \sum_{i=1}^n \sum_{\a=0}^n
f_{i\a}(t,x,\nabla_{t,x}y)y_{x_i x_\a} + O(|y|^2 +
|\nabla_{t,x}y|^2),$$ where $f_{i\a}(t,x,0)=0$ and $x_0 = t$, $\phi_{G_0}$ is a nonnegative smooth function defined on $\overline{G}$ and satisfying $\ds\min_{x\in\cl{G_0}}\phi(x)>0$. In system (\[qhsystem1\]), as before, $(y,y_t)$ is the [*state variable*]{} and $u$ is the [*control variable*]{}. However, as we shall see later, the state space and control space have to be chosen in a different way from those used in the linear/semilinear setting.
The controllability of quasilinear hyperbolic equations is well understood in one space dimension ([@TTLi]). With regard to the multidimensional case, we introduce the following assumption.
\[condition of d6\] The linear part in (\[qhsystem1\]), i.e., hyperbolic equation $$\begin{aligned}
\left\{
\begin{array}{lll}
\ds y_{tt} - \sum_{i,j=1}^n (h^{ij}y_{x_i})_{x_j} =qy+
\chi_{G_0}u, &\mbox{ in } Q,\\
\ns\ds y=0, &\mbox{ in }\Si,\\
\ns\ds y(0)=y_0,\ y_t(0)=y_1, &\mbox{ in }G
\end{array}
\right.\end{aligned}$$ is exactly controllable in $H_0^1(G)\t L^2(G)$ at some time $T$.
Theorem \[oTheorem5\] provides a sufficient condition to guarantee Condition \[condition of d6\] is satisfied. The following result is a slight generalization of that shown in [@35].
\[qth1\] Assume Condition \[condition of d6\] holds. Then, there is a sufficiently small $\e_0>0$ such that for any $(y_0,y_1),
(z_0,z_1)\in \big(H^{s+1}(G)\bigcap H^1_0(G)\big)\t H^s(G)$ satisfying $|(y_0,y_1)|_{ H^{s+1}(G)\t H^s(G)}<\e_0$, $|(z_0,z_1)|_{
H^{s+1}(G)\t H^s(G)}<\e_0$ and the compatibility condition, one can find a control $\ds u\in \bigcap_{k=0}^{s-2}C^k([0,T];H^{s-k}(G)$ such that the corresponding solution of system (\[qhsystem1\]) verifies $y(T)=z_0$ and $y_t(T)=z_1$ in $G$.
The key in the proof of Theorem \[qth1\] is to reduce the local exact controllability of quasilinear equations to the exact controllability of the linear equation by means of a new unbounded perturbation technique (developed in [@35]), which is a universal approach to solve the local controllability problem for a large class of quasilinear time-reversible evolution equations.
Note however that the above approach does not apply to the controllability problem for quasilinear time-irreversible evolution equations, such as the following controlled quasilinear parabolic equation: $$\begin{aligned}
\label{system1}
\left\{
\begin{array}{lll}
y_t-\displaystyle\sum^{n}_{i,j=1}(a^{i j}(y) y_{x_i})_{x_j}=\chi_{G_0} u \ \ \ \ &\mbox{ in }Q,\\
\ns
y=0 &\mbox{ on }\Sigma,\\
\ns
y(0)=y_0 &\mbox{ in }G.
\end{array}
\right.\end{aligned}$$ In (\[system1\]), $y$ is the [*state variable*]{} and $u$ is the [*control variable*]{}, the nonlinear matrix-valued function $\big(a^{i j}\big)_{1\le i,j\le n}\in C^2(\mathbb{R};
\mathbb{R}^{n\times n})$ is locally positive definite. One can find very limited papers on the controllability of quasilinear parabolic-type equations ([@23] and the references therein). One of the main difficulty to solve this problem is to show the “good enough" regularity for solutions of system (\[system1\]) with a desired control.
We introduce the dual system of the linearized equation of (\[system1\]). $$\begin{aligned}
\label{lsystem1}
\left\{
\begin{array}{lll}
p_t-\displaystyle\sum^{n}_{i,j=1}(p^{i j} p_{x_i})_{x_j}=0 \ \ \ &\mbox{ in }Q,\\
\ns
p=0 &\mbox{ on }\Sigma,\\
\ns
p(0)=p_0 &\mbox{ in }G,
\end{array}
\right.\end{aligned}$$ where $\big(p^{ij}\big)_{1\le i,j\le n}$ is assumed to satisfy Condition \[condition of d-0\]. Put $ B=1+\sum\limits^{n}_{i,j=1}
|p^{ij}|^2_{C^1(\overline{Q})}$. Starting from Theorem \[uTheorem\], one can show the following observability result ([@23]).
\[oTheorem1\] There exist suitable real functions $\alpha$ and $\varphi$, and a constant $C_0=C_0(\rho, n, G, T)>0$, such that for any $\lambda\geq
C_0
e^{C_0B}$, solutions of (\[lsystem1\]) satisfy | p(T)|\_[L\^2(G)]{} C e\^[e\^[C B]{}]{} |e\^||\^[3/2]{} p|\_[L\^2(Q\_[G\_0]{})]{}, p\_0L\^2(G).
In Theorem \[oTheorem1\], the observability constant in (\[opw\]) is obtained explicitly in the form of $C e^{e^{C B}}$ in terms of the $C^1$-norms of the coefficients in the principal operator appeared in the first equation of (\[lsystem1\]). This is the key in the argument of fixed point technique to show the following local controllability of system (\[system1\]) ([@23]).
\[Theorem1\] There is a constant $\gamma>0$ such that, for any initial value $y_0\in C^{2+\frac{1}{2}}(\overline{G})$ satisfying $|y_0|_{C^{2+\frac{1}{2}}(\overline{G})}\leq \gamma$ and the first order compatibility condition, one can find a control $u\in
C^{\frac{1}{2}, \frac{1}{4}}(\overline{Q})$ with $\supp
u\subseteq[0,T]\times G_0$ such that the solution $y$ of system (\[system1\]) satisfies $y(T)=0$ in $G$.
From Theorem \[Theorem1\], it is easy to see that the [*state space*]{} and [*control space*]{} for system (\[system1\]) are chosen to be $C^{2+\frac{1}{2}}(\overline{G})$ and $C^{\frac{1}{2},
\frac{1}{4}}(\overline{Q})$, respectively. The key observation in [@23] is that, thanks to an idea in [@barbu], for smooth initial data, the regularity of the null-control function for the linearized system can be improved, and therefore, the fixed point method is applicable.
Stabilization of hyperbolic equations and further comments
----------------------------------------------------------
In this subsection, we give more applications of Theorem \[uTheorem\] to the stabilization of hyperbolic equations and comment other applications of this theorem and some related open problems.
One of the main motivation to introduce the controllability/obseervability theory is to design the feedback regulator ([@15]). Stimulated by [@Lions], there exist a lot of works addressing the stabilization problem of PDEs from the control point of view. To begin with, we fix a nonnegative function $a\in L^{\i}(\G)$ such that $\big\{x\in\G\;\big|\;a(x)>0\big\}\neq\emptyset$, and consider the following hyperbolic equation with a boundary damping: { u\_[tt]{}-\_[j,k=1]{}\^n(h\^[jk]{}u\_[x\_j]{})\_[x\_k]{}=0 & (0,)G,\
\_[j,k=1]{}\^nh\^[jk]{}u\_[x\_j]{}\_k+a(x)u\_t=0& (0,),\
u(0)=u\^0, u\_t(0)=u\^1& G. .
Put $
H\=\left\{(f,g)\in H^1(G)\t L^2(G)\;\left|\;\int_G fdx=0\right\}\right.$, which is a Hilbert space with the canonic norm. Define an unbounded operator $\cA: \;
H\to H$ by $$\left\{\ba{ll}
\3n \cA\=\left(\ba{cc}
0&I\\ \ns\ds\sum_{j,k=1}^n\pa_{x_k}(h^{jk}\pa_{x_j})&0
\ea
\right),\\
\ns\ds \3n D(\cA)\=\Big\{u=(u^0,u^1)\in H\;\Big|\ \cA u\in H,\ \(\sum_{j,k=1}^nh^{jk}u^0_{x_j}\nu_k+au^1\)\|_{\G}=0\Big\}. \ea\right.$$ It is easy to show that $\cA$ generates an $C_0$-semigroup $\{e^{t\cA}\}_{t\in \mathbb{R}}$ on $H$. Hence, system (\[0a1\]) is well-posed in $H$. Clearly, $H$ is the [*finite energy space*]{} of system (\[0a1\]). One can show that the energy of any solution of (\[0a1\]) tends to zero as $t\to\infty$ (There is no any geometric conditions on $\G$).
Starting from Theorem \[uTheorem\], one can show the following result, which is a slight improvement of the main result in [@Fu2]: \[0t1\] Assume Conditions \[condition of d-1\] holds. Then solutions $u\in
C([0,\infty);$ $ D(\cA))\bigcap C^1([0,\infty);\;H)$ of system (\[0a1\]) satisfy ||(u,u\_t)||\_[H]{}||(u\^0,u\^1)||\_[D()]{},(u\^0,u\^1)D(), t>0.
Next, we consider a semilinear hyperbolic equation with a local damping: { u\_[tt]{}-\_[j,k=1]{}\^n(h\^[jk]{}u\_[x\_j]{})\_[x\_k]{}+f(u)+b(x)g(u\_t,u)=0 & (0,)G,\
u=0& (0,),\
u(0)=u\^0, u\_t(0)=u\^1& G. .In (\[icm0a1\]), $h^{jk}$ is supposed to satisfy Conditions \[condition of d-1\] and \[condition of d\]; $f:\mathbb{R}\to
\mathbb{R}$ is a differentiable function satisfying $f(0) = 0$, $sf(s)\ge 0$ and $|f'(s)| \le C (1 + |s|^q)$ for any $s\in\mathbb{R}$, where $q \ge 0$ and $(n -2)q \le 2$; $b$ is a nonnegative function satisfying $\ds\min_{x\in G_0} b(x)>0$, where $G_0$ is given in Theorem \[oTheorem5\]; and $g :
\mathbb{R}^{n+1}\to \mathbb{R}$ is a globally Lipschitz function satisfying $g(0, w) = 0$, $|g(r,w) - g(r_1, w_1|\le C (|r - r_1| +
|w - w_1|)$ and $g(r,w)r \ge c_0r^2$ for some $c_0
> 0 $, any $w, w_1\in\mathbb{R}^n$ and any $r, r_1\in
\mathbb{R}$.
Define the energy of any solution $u$ to (\[icm0a1\]) by setting $$E(t)=\frac12\int_G\Big[|u_t|^2+\sum_{j,k=1}^nh^{jk}u_{x_j}u_{x_k}\Big]dx+\int_G\int_0^uf(s)dsdx.$$ Starting from Theorem \[uTheorem\], one can show the following stabilization result for system (\[icm0a1\]) ([@Tebou]).
Let $(u^0,u^1)\in H_0^1(G)\times L^2(G)$. Then there exist positive constants $M$ and $r$, possibly depending on $E(0)$, such that the energy $E(t)$ of the solution of (\[icm0a1\]) satisfies $E(t)\le
Me^{-rt}E(0)$ for any $t\ge 0$.
Several comments are in order.
\[icmrem2\] In [@LR], the authors need $C^\infty$-regularity for the data to establish Theorem \[icm2t1\]. Recently, based on Theorem \[uTheorem\], this result was extended in [@25] as follows: Denote by $\{\l_i\}^\infty_{i=1}$ the eigenvalues of any general elliptic operator of second order (with $C^{1}$-principal part coefficients) on $\O$ (of class $C^{2}$) with Dirichlet or Robin boundary condition, and $\{e_i\}^{\infty}_{i=1}$ the corresponding eigenvectors satisfying $|e_i|_{L^2(\O)} = 1$. Then, for any $r
> 0$, it holds $$\sum_{\l_{i} \leq r}|a_{i}|^{2} \leq C
e^{C\sqrt{r}}{\int_{G_0}}\bigg| \sum_{\l_{i} \leq r}a_{i}e_{i}(x)
\bigg|^2dx,\ \ \forall\, \{a_{i}\}_{\l_{i} \leq r }\hb{ with }a_{i}
\in \mathbb{C}.$$
As indicated in [@Fu1; @LTZ; @LTZ2], Theorem \[uTheorem\] can be employed to study the global unique continuation and inverse problems for some PDEs. Note also that this Carleman estimate based approach can be applied to solve some optimal control problems ([@WW]).
In practice, constrained controllability is more realizable. It is shown in [@[40]] that the study of this problem is unexpectedly difficult even for the $1-d$ wave equation and heat equation. We refer to [@22] for an interesting example showing that this problem is nontrivial even if the control is effective everywhere in the domain in which the system is evolved.
Note that the above mentioned approach applies mainly to the controllability, observability and stabilization of second order non-degenerate PDEs. It is quite interesting to extend it to the coupled and/or higher order systems, or degenerate systems but in general, this is nontrivial even for linear problems ([@CMV; @ZZ1]).
Similar to other nonlinear problems, nonlinear controllability problems are usually quite difficult. It seems that there is no satisfactory controllability results published for nonlinear hyperbolic-parabolic coupled equations. Also, there exists no controllability results for fully nonlinear PDEs. In the general case, of course, one could expect only local results. Therefore, the following three problems deserve deep studies: 1) The characterization of the controllability subspace; 2) Controllability problem with (sharp) lower regularity for the data; 3) The problem that cannot be linearized. Of course, all of these problems are usually challenging.
The stochastic case {#s4}
===================
In this section, we extend some of the results/approaches in Section \[s3\] to the stochastic case. As we shall see later, the stochastic counterpart is far from satisfactory, compared to the deterministic setting.
In what follows, we fix a complete filtered probability space $(\O,\cF,\{\cF_t\}_{t\ge 0},P)$ on which a one dimensional standard Brownian motion $\{B(t)\}_{t\ge0}$ is defined. Let $H$ be a Fréchet space. Denote by $L_{\cF}^2(0,T;H)$ the Fréchet space consisting of all $H$-valued $\{\cF_t\}_{t\ge 0}$-adapted processes $X(\cd)$ such that $\mathbb{E}(|X(\cd)|_{L^2(0,T;H)}^2)<\i$, with the canonical quasi-norms; by $L_{\cF}^\i(0,T;H)$ the Fréchet space consisting of all $H$-valued $\{\cF_t\}_{t\ge 0}$-adapted bounded processes, with the canonical quasi-norms; and by $L_{\cF}^2(\O; C([0,T];H))$ the Fréchet space consisting of all $H$-valued $\{\cF_t\}_{t\ge 0}$-adapted continuous processes $X(\cd)$ such that $ \mathbb{E}(|X(\cd)|_{C([0,T];H)}^2)<\i$, with the canonical quasi-norms (similarly, one can define $L^{2}_{\cal
F}(\O;C^{k}([0,T];H))$ for $k\in \mathbb{N}$).
Stochastic Parabolic Equations
------------------------------
We begin with the following stochastic parabolic equation: { dz-\_[i,j=1]{}\^n(p\^[ij]{}z\_[x\_i]{})\_[x\_j]{}dt=\[a,z+bz\]dt+cz dB(t)&Q,\
z=0&,\
z(0)=z\_0&G . with suitable coefficients $a,b$ and $c$, where $p^{ij}\in
C^2(\cl{Q})$ is assumed to satisfy Condition \[condition of d-0\] (Note that, technically we need here more regularity for $p^{ij}$ than the deterministic case). We are concerned with an observability estimate for system (\[hh6.1\]), i.e., to find a constant $\cC=\cC(a,b,c,T)>0$ such that solutions of (\[hh6.1\]) satisfy |z(T)|\_[L\^2(Ø,\_T,P;L\^2(G))]{}|z|\_[L\^2\_(0,T;L\^2(G\_0))]{},z\_0L\^2(Ø,\_0,P;L\^2(G)). Similar to Theorem \[uTheorem\], we have the following weighted identity ([@Tang-Zhang1]).
\[icmc1t1\] Let $m\in\mathbb{N}$, $b^{ij}=b^{ji}\in
L_{\cF}^2(\O;C^1([0,T];W^{2,\infty}(\mathbb{R}^m)))$ ($i,j=1,2,\cdots,m$), $\ell\in C^{1,3}((0,T)\t\mathbb{R}^m)$ and $\Psi\in C^{1,2}((0,T)\t\mathbb{R}^m)$. Assume $u$ is an $H^2(\mathbb{R}^m)$-valued continuous semi-martingale. Set $\th=e^{\ell }$ and $v=\th u$. Then for $\ae\2n$ $x\in \mathbb{R}^m$ and $P$-$\as\2n$ $\o\in\O$, $$\ba{ll}
\displaystyle
2\int_0^T\th\[-\sum_{i,j=1}^m (b^{ij}v_{x_i})_{x_j}+Av\]\[du-\sum_{i,j=1}^m(b^{ij}u_{x_i})_{x_j}dt\]+2\int_0^T\sum_{i,j=1}^m(b^{ij}v_{x_i}dv)_{x_j}\\
\noalign{\ss}
\displaystyle\q+ 2\int_0^T\sum_{i,j=1}^m\[\sum_{i',j'=1}^m\(2b^{ij} b^{i'j'}\ell_{x_{i'}}v_{x_i}v_{x_{j'}}
-b^{ij}b^{i'j'}\ell_{x_i}v_{x_{i'}}v_{x_{j'}}\)\\
\noalign{\ss}
\displaystyle\q+\Psi
b^{ij}v_{x_i}v- b^{ij}\(A\ell_{x_i}+\frac{\Psi_{x_i}}{2}\)v^2\]_{x_j}dt\\
\noalign{\ss}
\displaystyle
=2\int_0^T\sum_{i,j=1}^m \Big\{\sum_{i',j'=1}^m\[2b^{ij'}\(b^{i'j}\ell_{x_{i'}}\)_{x_{j'}} -
\(b^{ij}b^{i'j'}\ell_{x_{i'}}\)_{x_{j'}}\]-\frac{b_t^{ij}}{2}+\Psi b^{ij}
\Big\}v_{x_i}v_{x_j}dt\\
\noalign{\ss}
\displaystyle\q
+\int_0^TBv^2dt+2\int_0^T\[-\sum_{i,j=1}^m (b^{ij}v_{x_i})_{x_j}+Av\]\[-\sum_{i,j=1}^m (b^{ij}v_{x_i})_{x_j}+(A-\ell_t)v\]dt\\
\noalign{\ss}
\displaystyle\q+\(\sum_{i,j=1}^m
b^{ij}v_{x_i}v_{x_j}+Av^2\)\Big|_0^T\\
\noalign{\ss}
\displaystyle\q-\int_0^T\th^2\sum_{i,j=1}^m
b^{ij}[(du_{x_i}+\ell_{x_i}du)(du_{x_j}+\ell_{x_j}du)]-\int_0^T\th^2A(du)^2,
\ea$$ where $$%
\left\{
\ba{ll}
\ds A\=-\sum_{i,j=1}^m
(b^{ij}\ell_{x_i}\ell_{x_j}-b^{ij}_{x_j}\ell_{x_i}
-b^{ij}\ell_{x_ix_j})-\Psi,\\
\ns
\ds
B\=2\[A\Psi-
\sum_{i,j=1}^m(Ab^{ij}\ell_{x_i})_{x_j}\] -A_t-\sum_{i,j=1}^m (b^{ij}\Psi_{x_j})_{x_i}.
\ea
\right.
%$$
Note that, in Theorem \[icmc1t1\], we assume only the symmetry for matrix $\big(b^{ij}\big)_{1\le i,j\le n}$ (without assuming the positive definiteness). Hence, this theorem can be applied to study not only the observability/controllability of stochastic parabolic equations, but also similar problems for deterministic parabolic and hyperbolic equations, as indicated in Section \[s3\]. In this way, we give a unified treatment of controllability/observability problems for some stochastic and deterministic PDEs of second order.
Starting from Theorem \[icmc1t1\], one can show the following observability result in [@Tang-Zhang1] (See [@barbu1] and the references therein for some earlier results).
\[1t4\] Assume that $$a\in L^{\i}_{\cF}(0,T;L^{\i}(G;\dbR^n)), \ \ b\in
L^{\i}_{\cF}(0,T;L^{n^*}(G)), \ \ c\in
L^{\i}_{\cF}(0,T;W^{1,\i}(G)),$$ where $n^*\ge 2$ if $n=1$; $n^*> 2$ if $n=2$; $n^*\ge n$ if $n\ge
3$. Then there is a constant $\cC=\cC(a,b,c,T)>0$ such that all solutions $z$ of system (\[hh6.1\]) satisfy (\[12e2\]). Moreover, the observability constant $\cC$ may be bounded as $$\cC(a,b,c,T)=Ce^{C[T^{-4}(1+\tau^2)+T\tau^2]},$$ with $\tau\=|a|_{L^{\i}_{\cF}(0,T;L^{\i}(G;\dbR^n))}+|b|_{L^{\i}_{\cF}(0,T;L^{n^*}(G))}+|c|_{
L^{\i}_{\cF}(0,T;W^{1,\i}(G))}$.
As a consequence of Theorem \[1t4\], one can deduce a controllability result for backward stochastic parabolic equations. Unlike the deterministic case, the study of controllability problems for forward stochastic differential equations is much more difficult than that for the backward ones. We refer to [@Peng] for some important observation in this respect. It deserves to mention that, as far as I know, there exists no satisfactory controllability result published for forward stochastic parabolic equations. Note however that, as a consequence of Theorem \[icm2t1\] and its generalization (see Remark \[icmrem2\]), one can deduce a null controllability result for forward stochastic parabolic equations with time-invariant coefficients ([@25]).
Theorem \[icmc1t1\] has another application in global unique continuation of stochastic PDEs. To see this, we consider the following stochastic parabolic equation: $$\label{icmhh6.1}
\cF z\equiv dz-\sum_{i,j=1}^n(f^{ij}z_{x_i})_{x_j}dt=[\lan a_1,\n z\ran+b_1z]dt+c_1z
dB(t)\q\hb{ in }Q,$$ where $f^{ij}\in C^{1,2}([0,T]\times G)$ satisfy $f^{ij}=f^{ji}$ ($i,j=1,2,\cdots,n$) and for any open subset $G_1$ of $G$, there is a constant $s_0=s_0(G_1)>0$ so that $\ds\sum_{i,j=1}^nf^{ij}\xi^{i}\xi^{j}
\geq s_0|\xi|^{2}$ for all $(t,x,\xi)\equiv (t,x,\xi^{1},\cdots, \xi^{n})
\in(0,T)\times G_1\t \dbR^{n}$; $a_1\in
L^{\i}_{\cF}(0,T;L_{loc}^{\i}(G;\dbR^n))$, $b_1\in
L^{\i}_{\cF}(0,T;L_{loc}^{\i}(G))$, and $c_1\in L^{\i}_{\cF}(0,T;$ $W_{loc}^{1,\i}(G))$.
Starting from Theorem \[icmc1t1\], one can show the following result ([@[39]]).
\[1t3\] Any solution $ z\in
L_{\cF}^2(\O;C([0,T];L_{loc}^2(G))) \bigcap $ $
L_{\cF}^2(0,T;H_{loc}^1(G)) $ of $(\ref{icmhh6.1})$ vanishes identically in $Q\t\O, \as dP$ provided that $z=0$ in $Q_{G_0}\t\O,
\as dP$.
Note that the solution of a stochastic equation is generally [*non-analytic in time*]{} even if all coefficients of the equation are constants. Therefore, one cannot expect a Holmgren-type uniqueness theorem for stochastic equations except for some very special cases. On the other hand, the usual approach to employ Carleman-type estimate for the unique continuation needs to localize the problem. The difficulty of our present stochastic problem consists in the fact that [*one cannot simply localize the problem as usual*]{} because the usual localization technique may change the adaptedness of solutions, which is a key feature in the stochastic setting. In equation (\[icmhh6.1\]), for the space variable $x$, we may proceed as in the classical argument. However, for the time variable $t$, due to the adaptedness requirement, we will have to treat it separately and globally. We need to introduce [*partial global Carleman estimate*]{} (indeed, global in time) even for local unique continuation for stochastic parabolic equation. Note that this idea comes from the study of controllability problem even though unique continuation itself is purely an PDE problem.
Stochastic Hyperbolic Equations
-------------------------------
We consider now the following stochastic wave equation: $$\begin{aligned}
\label{ssystem3}
\left\{
\begin{array}{lll}\ds
\ds dz_{t} - \D z dt =(a_1 z_t + \lan a_2,\nabla z\ran + a_3 z + f
)dt + (a_4 z + g)dB(t) & {\mbox { in }} Q,
\\
\ns\ds z = 0 & \mbox{ on } \Si, \\
\ns\ds z(0) = z_0, \q z_{t}(0) = z_1 & \mbox{ in } G,
\end{array}
\right.\end{aligned}$$ where $a_1\in L^{\infty}_{\cF}(0,T;L^{\infty}(G))$, $a_2\in
L^{\infty}_{\cF}(0,T;L^{\infty}(G;\mathbb{R}^n))$, $a_3\in L^{\infty}_{\cF}(0,T;L^n(G))$, $a_4\in
L^{\infty}_{\cF}(0,T;L^\infty(G))$, $f\in L^2_{\cF}(0,T;L^2(G))$, $g\in L^2_{\cF}(0,T;L^2(G))$ and $(z_0, z_1) \in L^2(\O,{\cal F}_0, P; H^1_0(G) \t L^2(G))$. We shall derive an observability estimate for (\[icmhh6.1\]), i.e., find a constant $\cC(a_1,a_2,a_3,a_4)>0$ such that solutions of system (\[icmhh6.1\]) satisfy |(y(T),y\_t(T))|\_[L\^2(Ø,\_T,P;H\_0\^1(G)L\^2(G))]{}\
n(a\_1,a\_2,a\_3,a\_4)nn,\
(y\_0,y\_1)L\^2(Ø,\_0,P;H\_0\^1(G) L\^2(G)). where $\G_0$ is given by (\[boundary\]) for some $x_0\in\mathbb{R}^d\setminus \cl G$.
It is clear that, $\ds 0<R_0\=\min_{x\in G}|x-x_0|< R_1\=\max_{x\in
G}|x-x_0|$. We choose a sufficiently small constant $c\in (0,1)$ so that $\frac{(4+5c)R_0^2}{9c}>R_1^2$. In what follows, we take $T$ sufficiently large such that $\frac{4(4+5c)R_0^2}{9c}>c^2T^2>4R_1^2$. Our observability estimate for system (\[icmhh6.1\]) is stated as follows ([@Zhangxu3]).
\[t1\] Solutions of system $(\ref{ssystem3})$ satisfy $(\ref{icm12e2})$ with $$\ba{ll}
\cC(a_1,a_2,a_3,a_4)\\
\ns
\ds=\2n C\exp\1n\left\{\1n C\1n \left[
|(a_1,a_4)|_{L^{\infty}_{\cF}(0,T;(L^\infty(G))^2)}^2+| a_2|_{
L^{\infty}_{\cF}(0,T;L^{\infty}(G;\mathbb{R}^n))}^2 + |a_3|_{L^{\infty}_{\cF}(0,T;L^n(G))}^2\1n\right]\1n\right\}.
\ea$$
Surprisingly, Theorem \[t1\] was improved in [@25] by replacing the left hand side of (\[icm12e2\]) by $
|(y_0,y_1)|_{L^2(\O,\cF_0,P;H_0^1(G)\times L^2(G))}$, exactly in a way of the deterministic setting. This is highly nontrivial by considering the very fact that the stochastic wave equation is time-irreversible.
The proof of Theorem \[t1\] (and its improvement in [@25]) is based on the following identity for a stochastic hyperbolic-like operator, which is in the spirit of Theorems \[uTheorem\] and \[icmc1t1\].
\[c1t1\] Let $b^{ij}\in C^1((0,T)\t\mathbb{R}^n)$ satisfy $b^{ij}=b^{ji}$ ($i,j=1,2,\ldots,n$), $\ell,\ \Psi\in
C^2((0,T)\t\mathbb{R}^n)$. Assume $u$ is an $H^2_{loc}(\mathbb{R}^n)$-valued $\{\cF_t\}_{t\ge 0}$-adapted process such that $u_t$ is an $L^2_{loc}(\mathbb{R}^n)$-valued semimartingale. Set $\th=e^{\ell }$ and $v=\th u$. Then, for a.e. $x\in \mathbb{R}^n$ and $P$-a.s. $\o\in\O$, $$\ba{ll}
\displaystyle
\th\Big(-2\ell_tv_t+2\sum_{i,j=1}^nb^{ij}\ell_{x_i}v_{x_j}+\Psi v\Big)\Big[du_t-\sum_{i,j=1}^n(b^{ij}u_{x_i})_{x_j}dt\Big]\\
\ns
\displaystyle\q+
\sum_{i,j=1}^n\biggl[\sum_{i',j'=1}^n\left(2b^{ij} b^{i'j'}\ell_{x_{i'}}v_{x_i}v_{x_{j'}}
-b^{ij}b^{i'j'}\ell_{x_i}v_{x_{i'}}v_{x_{j'}}\right)-2b^{ij}\ell_tv_{x_i}v_t
+b^{ij}\ell_{x_i}v_t^2\\
\ns
\displaystyle\qq\qq\qq+\,\Psi b^{ij}v_{x_i}v-\Big(A\ell_{x_i}+\frac{\Psi_{x_i}}{2}\Big)b^{ij} v^2\biggr]_{x_j}dt\\
\ns
\displaystyle\q+\,d\Big[\sum_{i,j=1}^n
b^{ij}\ell_tv_{x_i}v_{x_j}-2\sum_{i,j=1}^nb^{ij}\ell_{x_i}v_{x_j}v_t+\ell_tv_t^2-\Psi
v_tv+\left(A\ell_t+\frac{\Psi_t}{2}\right)v^2\Big]\\
\ns
\displaystyle
=\Big\{\Big[\ell_{tt}+\sum_{i,j=1}^n(b^{ij}
\ell_{x_i})_{x_j}-\Psi\Big]v_t^2-2\sum_{i,j=1}^n\left[(b^{ij}\ell_{x_j})_t+b^{ij}\ell_{tx_j}\right]v_{x_i}v_t\\
\ns
\displaystyle\q+\sum_{i,j=1}^n\Big[(b^{ij}\ell_t)_t+\sum_{i',j'=1}^n\left(2b^{ij'}(b^{i'j}\ell_{x_{i'}})_{x_{j'}} -
(b^{ij}b^{i'j'}\ell_{x_{i'}})_{x_{j'}}\right)+\Psi b^{ij}
\Big]v_{x_i}v_{x_j}\\
\ns
\displaystyle\q
+\,{\it Bv}^2+\Big(-2\ell_tv_t+2\sum_{i,j=1}^nb^{ij}\ell_{x_i}v_{x_j}+\Psi v\Big)^2\Big\}dt+\th^2\ell_t(du_t)^2,
\ea$$ where $(du_t)^2$ denotes the quadratic variation process of $u_t$, $$%
\left\{
\ba{ll}
\ds A\=(\ell_t^2-\ell_{tt})-\sum_{i,j=1}^n
(b^{ij}\ell_{x_i}\ell_{x_j}-b^{ij}_{x_j}\ell_{x_i}
-b^{ij}\ell_{x_ix_j})-\Psi,\\
\ns
\ds
B\=A\Psi+(A\ell_t)_t-
\sum_{i,j=1}^n(Ab^{ij}\ell_{x_i})_{x_j} +\frac{1}{2}\Big[\Psi_{tt}-\sum_{i,j=1}^n (b^{ij}\Psi_{x_i})_{x_j}\Big].
\ea
\right.$$
Further comments
----------------
Compared to the deterministic case, the controllability/observability of stochastic PDEs is in its “enfant" stage. Therefore, the main concern of the controllability/observability theory in the near future should be that for stochastic PDEs. Some most relevant open problems are listed below.
1. [**Controllability of forward stochastic PDEs**]{}. Very little is known although there are some significant progress in the recent work [@25]. Also, it would be quite interesting to extend the result in [@BLR] to the stochastic setting but this seems to be highly nontrivial.
2. [**Controllability of nonlinear stochastic PDEs**]{}. Almost nothing is known in this direction although there are some papers addressing the problem in abstract setting by imposing some assumption which is usually very difficult to check for the nontrivial case.
3. [**Stabilization and inverse problems for stochastic PDEs**]{}. Almost nothing is known in this respect.
Acknowledgement {#acknowledgement .unnumbered}
===============
This paper is a summary of part of the work I have done in close collaboration with my colleagues, coworkers and students. I am grateful to all of them. In particular, I would like to express my sincerely gratitude to my mentors, Xunjing Li, Jiongmin Yong and Enrique Zuazua, for their continuous encouragement and for so many fruitful discussions and suggestions that have been extremely influential on the formulation and solution of most of problems addressed here. Also, I would like to mention some of my colleagues with who I had the opportunity to develop part of the theory and learn so many things and, in particular, Xiaoyu Fu, Xu Liu, Qi Lü, Kim Dang Phung, Shanjian Tang, Gengsheng Wang, Jiongmin Yong and Enrique Zuazua. Finally, I thank Viorel Barbu, Piermarco Cannarsa, Xiaoyu Fu, Yamilet Quintana, Louis Tébou, Marius Tucsnak and Gengsheng Wang for their useful comments on the first version of this paper that allowed to improve its presentation and to avoid some inaccuracies.
[a]{}
S.A. Avdonin and S.A. Ivanov, *Families of exponentials. The method of moments in controllability problems for distributed parameter systems*, Cambridge University Press, Cambridge, UK, 1995.
V. Barbu, *Controllability of parabolic and Navier-Stokes equations*, Sci. Math. Jpn., **56** (2002), 143–211.
V. Barbu, A. R$\breve{\rm a}$scanu and G. Tessitore, *Carleman estimate and controllability of linear stochastic heat equatons*, Appl. Math. Optim., **47** (2003), 97–120.
C. Bardos, G. Lebeau and J. Rauch, *Sharp sufficient conditions for the observation, control and stabilizion of waves from the boundary*, SIAM J. Control Optim., **30** (1992), 1024–1065.
N. Burq, *Contrôle de l’équation des ondes dans des ouverts peu réguliers*, Asymptot. Anal., **14** (1997), 157–191.
N. Burq and P. Gérard, *Condition nécessaire et suffisante pour la contrôlabilité exacte des ondes*, C. R. Acad. Sci. Paris Sér. I Math., **325** (1997), 749–752.
P. Cannarsa, P. Martinez and J. Vancostenoble, *Carleman estimates for a class of degenerate parabolic operators*, SIAM J. Control Optim. **47** (2008), 1–19.
J.M. Coron, *Control and nonlinearity*, Mathematical Surveys and Monographs, vol. 136, American Mathematical Society, Providence, RI, 2007.
A. Doubova, E. Fernández-Cara, M. Gouzález-Burgos and E. Zuazua, *On the controllability of parabolic systems with a nonlinear term involving the state and the gradient*, SIAM J. Control Optim., **41** (2002), 798–819.
T. Duyckaerts, X. Zhang and E. Zuazua, *On the optimality of the observability inequalities for parabolic and hyperbolic systems with potentials*, Ann. Inst. H. Poincaré Anal. Non Linéaire, **25** (2008), 1–41.
E. Fernández-Cara, S. Guerrero, O.Yu. Imanuvilov and J.-P. Puel, *Local exact controllability of the Navier-Stokes system*, J. Math. Pures Appl., **83** (2004), 1501–1542.
E. Fernández-Cara and E. Zuazua, *Null and approximate controllability for weakly blowing-up semilinear heat equations*, Ann. Inst. H. Poincaré Anal. Non Linéaire, **17** (2000), 583–616.
X. Fu, *A weighted identity for partial differential operators of second order and its applications*, C. R. Math. Acad. Sci. Paris, **342** (2006), 579–584.
X. Fu, *Logarithmic decay of hyperbolic equations with arbitrary small boundary damping*, Comm. Partial Differential Equations, **34** (2009), 957–975.
X. Fu, X. Liu and X. Zhang, *Well-posedness and local controllability for quasilinear complex Ginzburg-Landau equations*, Preprint.
X. Fu, J. Yong and X. Zhang, *Exact controllability for multidimensional semilinear hyperbolic equations*, SIAM J. Control Optim., **46** (2007), 1578–1614.
A.V. Fursikov and O.Yu. Imanuvilov, *Controllability of evolution equations*, Lecture Notes Series, vol. 34, Research Institute of Mathematics, Seoul National University, Seoul, Korea, 1996.
V. Isakov, *Inverse problems for partial differential equations. Second edition*, Applied Mathematical Sciences, vol. 127, Springer-Verlag, New York, 2006.
R.E. Kalman, *On the general theory of control systems*, Proc. 1st IFAC Congress, Moscow, 1960, vol. 1, Butterworth, London, 1961, 481–492.
M. Kazemi and M.V. Klibanov, *Stability estimates for ill-posed Cauchy problem involving hyperbolic equations and inequalities*, Appl. Anal., **50** (1993), 93–102.
V. Komornik and P. Loreti, *Fourier series in control theory*, Springer Monographs in Mathematics, Springer-Verlag, New York, 2005.
I. Lasiecka, R. Triggiani and X. Zhang, *Nonconservative wave equations with purely Neumann B.C.: Global uniqueness and observability in one shot*, Contemp. Math., **268** (2000), 227–326.
I. Lasiecka, R. Triggiani and X. Zhang, *Global uniqueness, observability and stabilization of nonconservative Schrödinger equations via pointwise Carleman estimates: Part I. $H^1$-estimates*, J. Inv. Ill-posed Problems, **11** (2004), 43–123.
M. M. Lavrentév, V. G. Romanov and S. P. Shishat$\cdot$skii, *Ill-posed problems of mathematical physics and analysis*, Translations of Mathematical Monographs, vol. 64, American Mathematical Society, Providence, RI, 1986.
G. Lebeau and L. Robbiano, *Contrôle exact de l’équation de la chaleur*, Comm. Partial Differential Equations, **20** (1995), 335–336.
T.T. Li, *Controllability and observability for quasilinear hyperbolic systems*, AIMS Series on Applied Mathematics, vol. 3, American Institute of Mathematical Sciences (AIMS), Springfield, MO, 2010.
W. Li and X. Zhang, *Controllability of parabolic and hyperbolic equations: Towards a unified theory*, Control Theory of Partial Differential Equations, Lect. Notes Pure Appl. Math., vol. 242, Chapman $\&$ Hall/CRC, Boca Raton, FL, 2005, 157–174.
X. Li and J. Yong, *Optimal control theory for infinite-dimensional systems*, Systems & Control: Foundations & Applications, Birkhäuser Boston, Inc., Boston, MA, 1995.
J.-L. Lions, *Contrôlabilité exacte, perturbations et stabilisation de systèmes distribués, Tome 1*, Recherches en Mathématiques Appliquées, vol. 8, Masson, Paris, 1988.
X. Liu and H. Gao, *Controllability of a class of Newtonian filtration equations with control and state constraints*, SIAM J. Control Optim., **46** (2007), 2256–2279.
X. Liu and X. Zhang, *On the local controllability of a class of multidimensional quasilinear parabolic equations*, C. R. Math. Acad. Sci. Paris, **347** (2009), 1379–1384.
A. López, X. Zhang and E. Zuazua, *Null controllability of the heat equation as singular limit of the exact controllability of dissipative wave equations*, J. Math. Pures Appl., **79** (2000), 741–808.
Q. Lü, *Conrol and observation of stochastic PDEs*, PhD Thesis, Sichuan University, 2010.
L. Miller, *The control transmutation method and the cost of fast controls*, SIAM J. Control Optim., **45** (2006), 762–772.
S. Peng, *Backward stochastic differential equation and exact controllability of stochastic control systems*, Progr. Natur. Sci. (English Ed.), **4** (1994), 274–284.
K.D. Phung, *Observability and control of Schrödinger equations*, SIAM J. Control Optim., **40** (2001), 211–230.
K.D. Phung, G. Wang and X. Zhang, *On the existence of time optimal control of some linear evolution equations*, Discrete Contin. Dyn. Syst. B, **8** (2007), 925–941.
L. Rosier and B.Y. Zhang, *Null Controllability of the complex Ginzburg-Landau equation*, Ann. Inst. H. Poincaré Anal. Non Linéaire, **26** (2009), 649–673.
D.L. Russell, *A unified boundary controllability theory for hyperbolic and parabolic partial differential equations*, Stud. Appl. Math., **52** (1973), 189–221.
D.L. Russell, *Controllability and stabilizability theory for linear partial differential equations: recent progress and open problems*, SIAM Rev., **20** (1978), 639–739.
S. Tang and X. Zhang, *Null controllability for forward and backward stochastic parabolic equations*, SIAM J. Control Optim., **48** (2009), 2191–2216.
L. Tébou, *A Carleman estimates based approach for the stabilization of some locally damped semilinear hyperbolic equations*, ESAIM Control Optim. Calc. Var., **14** (2008), 561–574.
M. Tucsnak and G. Weiss, *Observation and control for operator semigroups*, Birkhäuser Advanced Texts: Basler Lehrbücher, Birkhäuser Verlag, Basel, 2009.
G. Wang, *$L^{\infty}$-null controllability for the heat equation and its consequences for the time optimal control problem*, SIAM J. Control Optim., **48** (2008), 1701–1720.
G. Wang and L. Wang, *The Carleman inequality and its application to periodic optimal control governed by semilinear parabolic differential equations*, J. Optim. Theory Appl., **118** (2003), 249–461.
M. Yamamoto, *Carleman estimates for parabolic equations and applications*, Inverse Problems, **25** (2009), 123013.
X. Zhang, *Explicit observability estimate for the wave equation with potential and its application*, R. Soc. Lond. Proc. Ser. A Math. Phys. Eng. Sci., **456** (2000), 1101–1115.
X. Zhang, *Exact controllability of the semilinear plate equations*, Asymptot. Anal., **27** (2001), 95–125.
X. Zhang, *A remark on null exact controllability of the heat equation*, SIAM J. Control Optim., **40** (2001), 39–53.
X. Zhang, *Unique continuation for stochastic parabolic equations*, Differential Integral Equations, **21** (2008), 81–93.
X. Zhang, *Carleman and observability estimates for stochastic wave equations*, SIAM J. Math. Anal., **40** (2008), 851–868.
X. Zhang, *Remarks on the controllability of some quasilinear equations*, Preprint (see http://arxiv.org/abs/0904.2427v1).
X. Zhang and E. Zuazua, *Polynomial decay and control of a $1-d$ hyperbolic-parabolic coupled system*, J. Differential Equations, **204** (2004), 380–438.
Y. Zhou and Z. Lei, *Local exact boundary controllability for nonlinear wave equations*, SIAM J. Control Optim., [**46**]{} (2007), 1022–1051.
E. Zuazua, *Propagation, observation, and control of waves approximated by finite difference methods*, SIAM Rev., **47** (2005), 197–243.
E. Zuazua, *Controllability and observability of partial differential equations: some results and open problems*, Handbook of Differential Equations: Evolutionary Differential Equations, vol. 3, Elsevier Science, 2006, 527–621.
[^1]: This work is supported by the NSFC under grants 10831007, 60821091 and 60974035, and the project MTM2008-03541 of the Spanish Ministry of Science and Innovation.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The magnetic behavior of thin ferromagnetic itinerant-electron films is investigated within the strongly correlated single-band Hubbard model. For its approximate solution we apply a generalization of the modified alloy analogy (MAA) to deal with the modifications due to the reduced translational symmetry. The theory is based on exact results in the limit of strong Coulomb interaction which are important for a reliable description of ferromagnetism. Within the MAA the actual type of the alloy analogy is determined selfconsistently. The MAA allows, in particular, the investigation of quasiparticle lifetime effects in the paramagnetic as well as the ferromagnetic phase. For thin fcc(100) and fcc(111) films the layer magnetizations are discussed as a function of temperature as well as film thickness. The magnetization at the surface-layer is found to be reduced compared to the inner layers. This reduction is stronger in fcc(100) than in fcc(111) films. The magnetic behavior can be microscopically understood by means of the layer-dependent spectral density and the quasiparticle density of states. The quasiparticle lifetime that corresponds to the width of the quasiparticle peaks in the spectral density is found to be strongly spin- and temperature-dependent.'
address: 'Humboldt-Universität zu Berlin, Institut für Physik, Invalidenstr. 110, 10115 Berlin, Germany'
author:
- 'T. Herrmann and W. Nolting'
title: 'Ferromagnetism and the temperature-dependent electronic structure in thin Hubbard films'
---
Introduction {#sec_intro}
============
Remarkable advances in thin film technology have recently led to active interest in the nature of magnetism in ultrathin films, at surfaces and multilayer structures. The influence of the reduced dimensionality on the magnetic behavior of $3d$ transition metals has been extensively studied both experimentally [@All94; @Bab96; @Elm95; @GBBS97; @DTP+89; @PKH90; @FPA+97b] and theoretically [@FF87; @KMS94; @AB97; @LH96; @SH97; @RHGF95; @GM94; @PN97b; @HPN98]. On the experimental side it was shown that ultrathin transition metal films can display long range ferromagnetic order from a monolayer on [@DTP+89]. Although the Mermin-Wagner theorem [@MW66] requires the transition temperature to vanish for perfectly isotropic two dimensional systems, it was shown theoretically that even a small amount of anisotropy may lead to magnetic order with a substantial transition temperature [@BM88; @DVT96; @SN98]. In real materials magnetic anisotropy is always present by virtue of either the dipole interaction or the spin-orbit coupling.
Theoretically, the $T=0\,$K properties of thin transition metal films have been addressed by ab initio calculations within the density functional theory in the local density approximation [@FF87; @KMS94; @AB97; @LH96]. However, these approaches are strictly based on a Stoner-type model of ferromagnetism and, therefore, treat electron correlation effects which are responsible for the spontaneous magnetic order on a low level. In addition they are restricted to ground state properties only. To overcome this restriction, for example, a generalization of the fluctuating local moment method has been used [@RHGF95] to calculate the temperature-dependent electronic structure of thin ferromagnetic films. However, the layer magnetizations at finite temperatures and the magnetic short range order are needed as an input. In Ref. [@SH97] magnetic phase transitions in thin films are investigated via a mapping of the ab initio results onto an effective Ising model. Hasegawa calculates the finite temperature properties of thin Cu/Ni/Cu sandwiches by use of the single-site spin-fluctuation theory [@Has87].
For the understanding of the thermodynamical properties of thin film magnetism theoretical investigations on rather idealized model systems have proven to be a good starting point. In this context several authors have focused on localized spin models like the Heisenberg model [@HBCC72; @EM91a; @JDB92; @SY92]. For example, the mechanism that leads to the experimentally observed temperature induced reorientation of the direction of magnetization in thin Fe and Ni films [@PKH90; @FPA+97b] was investigated in great detail [@HU97; @JB98]. On the other hand it is by no means clear to what extent the results obtained by localized spin models are applicable to transition metal films, where the magnetically active electrons are itinerant.
The aim of the present paper is to study the interplay between strong electron correlations and the reduced translational symmetry due to the film geometry within an itinerant electron model system. In particular we are interested in the influence of the reduced dimensionality on spontaneous ferromagnetism and the spin-, layer- and temperature-dependent electronic structure. For this purpose we restrict ourselves, at present, to the investigation of the single-band Hubbard model [@Hub63], which includes the minimum set of terms necessary for the description of itinerant-electron magnetism. The Hubbard model was originally introduced to explain band magnetism in transition metals and has become a standard model to study the essential physics of strongly correlated electron systems over the years. It is clear that a realistic and quantitative description of ferromagnetism in transition metals requires the inclusion of the degeneracy of the 3d-bands [@Hub64a; @VBH+97; @HV98; @MK98; @NBDF89; @HCO+94; @DDP97]. Although the band-degeneracy is neglected in our model study, we believe that a treatment of electron correlation effects well beyond Hartree-Fock theory will provide important insight into generic properties of thin film ferromagnetism. For example, contrary to the expectation on the basis of the well-known Stoner criterion, the magnetic order at the film surface may be reduced and less stable compared to the inner layers if electron correlations are taken into account properly [@Has92].
Despite its apparent simplicity no general solution exists until now for the Hubbard model. However, recently exact results have been obtained by finite temperature quantum Monte Carlo calculations in the limit of infinite dimensions [@Ulm98; @VBH+97] which prove the existence of ferromagnetic solutions for intermediate to strong Coulomb interaction $U$. In addition the decisive importance of the lattice geometry, i.e. the dispersion and distribution of spectral weight in the non-interacting (Bloch) density of states (BDOS), on the magnetic stability was stressed by several authors [@VBH+97; @Ulm98; @Uhr96; @HUM97; @WBS+98; @OPK97; @HN97b]. A reasonable treatment of electron correlation effects led to an argument for the stability of ferromagnetism which is decisively more restrictive the well known Stoner criterion. A BDOS with large spectral weight near one of the band edges is an essential ingredient for ferromagnetism. The thermal stability of ferromagnetic solutions is favored by a strong asymmetry in the BDOS [@HUM97; @WBS+98; @PHWN98]. This behavior of the BDOS is found, for example, in non-bipartite lattices like the fcc lattice.
Due to the broken translational symmetry even more complications are introduced to the highly non-trivial many-body problem of the Hubbard model. Thus we require an approximation scheme which is simple enough to allow for an extended study of magnetic phase transitions and electronic correlations in thin films. On the other hand it should be clearly beyond Hartree-Fock (Stoner) theory which has been applied previously [@GM94], since we believe a reasonable treatment of electron correlation effects to be vital for a proper description of ferromagnetism especially for non-zero temperatures.
In this context interpolating theories which are essentially based on exact results obtained by the $1/U$ perturbation theory first introduced by Harris and Lange [@HL67; @EO94] have proven to be a good starting point [@PHWN98]. A theory that reproduces the rigorous strong coupling results in a conceptual clear and straightforward manner is given by the spectral density approach (SDA) which has been discussed with respect to spontaneous magnetic order for various three-dimensional [@NB89; @HN97a; @HN97b] as well as infinite dimensional [@HN97b; @PHWN98] lattices. A similar approach applied to a multiband Hubbard model led to surprisingly accurate results for the magnetic key quantities of the prototype band ferromagnets Fe, Co, and Ni [@NBDF89]. A generalization of the SDA to systems with reduced translational symmetry has recently been given in Refs. [@PN97b; @HPN98], which led, for example, to the description of the temperature-driven reorientation transition within an itinerant-electron film [@HPN98]. However, a severe limitation of the SDA results from the fact that quasiparticle damping is neglected completely. To tackle this problem a modified alloy analogy has been proposed [@HN96; @NH98] which is closely related to the SDA but includes quasiparticle damping effects in a natural way. For bulk systems it was found that the magnetic region in the phase diagram is significantly reduced by the inclusion of damping effects. By comparison [@PHWN98; @PHN98] with exact results for the fcc lattice in the limit of infinite dimension and intermediate Coulomb interaction it is clear that the Curie temperatures are somewhat overestimated within the MAA. However, the qualitative behavior of the ferromagnetic solutions and in particular the dependence of the Curie temperature on the band occupation is found to be in good agreement with the exact results.
In the present work we want to apply the MAA to systems with reduced translational symmetry. For this purpose the paper is organized in the following way: In the next section we will give a short introduction to the underlying many-body problem. The concept of an alloy analogy for the Hubbard film is developed in Sect. \[sec\_alloy\]. In Sect. \[sec\_maa\] we will generalize the MAA to systems with reduced translational symmetry. The results of the numerical evaluations will be discussed in Sect. \[sec\_results\] in terms of temperature- and layer-dependent magnetizations, the quasiparticle bandstructure and the quasiparticle densities of states. We will end with a conclusion in Sect. \[sec\_conclusion\].
The many-body problem of the Hubbard film {#sec_many_body}
=========================================
Let us first introduce the notation used to deal with the film geometry. Each lattice vector of the film system is decomposed into two parts according to: $${{\bf R}}_{i\alpha}={{\bf R}}_i+{{\bf r}}_\alpha.$$ ${{\bf R}}_i$ denotes a lattice vector of the underlying two-dimensional Bravais lattice with $N$ sites. To each of theses lattice sites there is associated a $d$-atom basis ${{\bf r}}_\alpha$ ($\alpha=1,\dots,d$) which refers to the $d$ layers of the film. The same labeling with Latin and Greek indices applies for all quantities related to the film geometry. Within each layer we assume translational invariance. Then a two-dimensional Fourier transformation with respect to the Bravais lattice can be applied.
Using this notation the Hamiltonian for the single-band Hubbard film reads: $$\label{hub_op}
{\cal H}=\sum_{i,j,\alpha,\beta,\sigma}
(T_{ij}^{\alpha\beta}-\mu\delta_{ij}^{\alpha\beta})
c_{i\alpha\sigma}^{\dagger}c_{j\beta\sigma}+
\frac{U}{2}\sum_{i,\alpha,\sigma}n_{i\alpha\sigma} n_{i\alpha-\sigma}.$$ Here $c_{i\alpha\sigma}$ ($c_{i\alpha\sigma}^{\dagger}$) stands for the annihilation (creation) operator of an electron with spin $\sigma$ at the lattice site ${{\bf R}}_{i\alpha}$, $n_{i\alpha\sigma}=c_{i\alpha\sigma}^{\dagger}c_{i\alpha\sigma}$ is the number operator. $U$ denotes the on-site Coulomb matrix element and $\mu$ the chemical potential. $T_{ij}^{\alpha\beta}$ is the hopping integral between the lattice sites ${{\bf R}}_{i\alpha}$ and ${{\bf R}}_{j\beta}$. A two-dimensional Fourier transformation yields the corresponding dispersions $$\label{hopping}
T_{{\bf k}}^{\alpha\beta} =\frac{1}{N}\sum_{ij} T_{ij}^{\alpha\beta}
e^{-i{{\bf k}}({{\bf R}}_{i}-{{\bf R}}_{j})}.$$ Here and in the following ${{\bf k}}$ denotes a wave-vector from the underlying two-dimensional (surface) Brillouin zone. Further we define $T_{0\alpha}=T_{ii}^{\alpha\alpha}=
\frac{1}{N}\sum_{{{\bf k}}}T_{{{\bf k}}}^{\alpha\alpha}=\textrm{const.}$ which gives the center of gravity of the $\alpha$-th layer in the BDOS.
The basic quantity to be calculated is the retarded single-electron Green function $$\label{green}
G_{ij\sigma}^{\alpha\beta}(E)=\langle\langle c_{i\alpha\sigma};
c_{j\beta\sigma}^{\dagger}\rangle\rangle_E.$$ From $G_{ij\sigma}^{\alpha\beta}(E)$ we can obtain all relevant information on the system. After a two-dimensional Fourier transformation one obtains from $G_{ij\sigma}^{\alpha\beta}(E)$ the spectral density $$S_{{{\bf k}}\sigma}^{\alpha\beta}(E)=-\frac{1}{\pi}
\textrm{Im} G_{{{\bf k}}\sigma}^{\alpha\beta}(E),$$ which represents the bare lineshape of a (direct, inverse) photoemission experiment. The diagonal elements of the Green function determine the spin- and layer-dependent quasiparticle density of states (QDOS): $$\label{eq:qdos}
\rho_{\alpha\sigma}(E)=
\frac{1}{N}\sum_{{{\bf k}}}S_{{{\bf k}}\sigma}^{\alpha\alpha}(E-\mu)
=-\frac{1}{\pi} \textrm{Im}G_{ii\sigma}^{\alpha\alpha}(E-\mu).$$ Via an energy integration one immediately gets from $\rho_{\alpha\sigma}(E)$ the band occupations $$\label{nas}
n_{\alpha\sigma}\equiv\langle n_{i\alpha\sigma}\rangle=
\int\limits_{-\infty}^{\infty} dE f_{-}(E) \rho_{\alpha\sigma}(E).$$ $\langle\dots\rangle$ denotes the grand-canonical average and $f_{-}(E)$ is the Fermi function. Here the site index $i$ has been omitted due to the assumed translational invariance within the layers. Ferromagnetism is indicated by a spin-asymmetry in the band occupations $n_{\alpha\sigma}$ leading to non-zero layer magnetizations $m_\alpha=n_{\alpha\uparrow}-n_{\alpha\downarrow}$. The mean band occupation $n$ and the mean magnetization $m$ are given by $n=\frac{1}{d}\sum_{\alpha\sigma}n_{\alpha\sigma}$ and $m=\frac{1}{d}\sum_{\alpha}m_\alpha$, respectively.
The equation of motion for the single-electron Green function reads: $$\label{eq_motion}
\sum_{l\gamma}\left[(E+\mu)\delta_{il}^{\alpha\gamma}-T_{il}^{\alpha\gamma}
-\Sigma_{il\sigma}^{\alpha\gamma}(E)\right]
G_{lj\sigma}^{\gamma\beta}(E) =\hbar\delta_{ij}^{\alpha\beta}.$$ Here we have introduced the electronic self-energy $\Sigma_{ij\sigma}^{\alpha\beta}(E)$ which incorporates all effects of electron correlations.
For later use we want to define the moments of the Green function $$\label{mom1}
M_{ij\sigma}^{(m)\alpha\beta}=-\frac{1}{\pi}\textrm{Im}
\int\limits_{-\infty}^{\infty}\! dE \, E^{m} \,G_{ij\sigma}^{\alpha\beta}(E).$$ The usefulness of the moments $M_{ij\sigma}^{(m)\alpha\beta}$ ($m=0,1,2,\dots$) results from the fact that an alternative but equivalent representation can be derived by use of the Heisenberg representation of the creation and annihilation operators. Thus $M_{ij\sigma}^{(m)\alpha\beta}$ can be calculated up to the desired order $m$ directly from the Hamiltonian (\[hub\_op\]) itself [@NB89; @PN96]: $$\label{mom2}
M_{ij\sigma}^{(m)\alpha\beta}=
\bigg\langle\!\Big[
\underbrace{\big[...
[c_{i\alpha\sigma},{\cal H}]_{\!-}... ,
{\cal H}\big]_{\!-}}_{m-\textrm{times}},
c_{j\beta\sigma}^{\dagger}
\Big]_{+}\!\bigg\rangle.$$ Here $[\dots,\dots ]_{-(+)}$ denotes the commutator (anticommutator). Eqs. (\[mom1\]) and (\[mom2\]) impose rigorous sum rules on the Green function and the self-energy which have been recognized to state important guidelines when constructing approximate solutions for the Hubbard model [@PHWN98]. For example, the high energy expansion of the Green function is directly determined by the moments $M_{ij\sigma}^{(m)\alpha\beta}$. It has been shown [@PHWN98] that the sum rules are especially important in the limit of strong Coulomb interaction: Being consistent with the sum rules up to the order $m=3$ states a necessary condition in order to reproduce the exact results of the $1/U$-perturbation theory [@HL67; @EO94]. Furthermore, the $m=3$ sum rule turns out to be of particular importance what concerns the stability of ferromagnetic solutions in the Hubbard model [@PHWN98].
The sum rules up to order $m=3$ will be exploited in Sect. \[sec\_maa\] for the construction of a modified alloy analogy (MAA) to the Hubbard film. First we want to introduce the concept of the alloy analogy approach for systems with reduced translational symmetry.
The alloy analogy concept for the Hubbard film {#sec_alloy}
==============================================
The main idea of the conventional alloy analogy approach [@Hub64b] is to consider, for the moment, the $-\sigma$-electrons to be “frozen” and to be randomly distributed over the sites of the lattice. Then a propagating $\sigma$-electron encounters a situation which is equivalent to a fictitious alloy: At empty lattice sites it finds the atomic energy $E_{1\sigma}$, at sites with a $-\sigma$-electron present the atomic energy $E_{2\sigma}$. These energy levels are randomly distributed over the lattice with concentrations $x_{1\sigma}$ and $x_{2\sigma}$ which correspond to the probabilities for the $\sigma$-electron to meet these respective situations. Note that at this point it is not at all clear what choice of the energy levels and concentrations gives the best approximation for the initial Hamiltonian. However, an “optimal” choice of the alloy analogy parameters should by some means account for the itineracy of the $-\sigma$-electrons (see Sect. \[sec\_maa\]). In the present film system the energy levels and concentrations may, in addition, exhibit a layer-dependence. Thus the alloy analogy for the Hubbard film is described by $4\cdot d$ a priori unknown parameters $$\label{ap}
E_{1\sigma}^{(\alpha)},\,\,x_{1\sigma}^{(\alpha)},\,\,
E_{2\sigma}^{(\alpha)},\,\,x_{2\sigma}^{(\alpha)}
\qquad \alpha=1,\,\dots,\,d$$
For the solution of the fictitious alloy problem given by (\[ap\]) the coherent potential approximation (CPA) [@VKE68] provides a well known method. The CPA has been realized to be the rigorous solution of the alloy problem in the limit of infinite dimensions [@VV92] where the single-site aspect used in the derivation of the CPA becomes exact. In this sense the CPA can be termed to be the best single-site approximation to the alloy problem. Due to the single-site aspect and the assumed translational invariance within the layers we have $\Sigma_{ij\sigma}^{\alpha\beta}(E)=
\delta_{ij}^{\alpha\beta}\Sigma_{\alpha\sigma}(E)$. The implicit CPA equation [@VKE68] for the self-energy is readily formulated via an effective medium approach similar to the one discussed in Ref. [@PN96]: $$\label{cpa}
0=\sum_{p=1,2} x_{p\sigma}^{(\alpha)}
\frac{E_{p\sigma}^{(\alpha)}-\Sigma_{\alpha\sigma}(E)-T_{0\alpha}}
{1-\frac{1}{\hbar}G_{ii\sigma}^{\alpha\alpha}
[E_{p\sigma}^{(\alpha)}-\Sigma_{\alpha\sigma}(E)-T_{0\alpha}]}.$$ In addition the self-energy appears implicitly in the expression for the local Green function which is given by matrix inversion from (\[eq\_motion\]) after applying a two-dimensional Fourier transformation: $$\label{local_g}
G_{ii\sigma}^{\alpha\beta}(E) =\frac{\hbar}{N}\!\sum_{{{\bf k}}}\!\!
\mbox{\footnotesize
$\displaystyle
\left(\begin{array}{ccc}
\!E\!+\!\mu-T_{{{\bf k}}}^{11}-\Sigma_{1\sigma}(E)&
-T_{{{\bf k}}}^{12}&\dots\\
-T_{{{\bf k}}}^{21}&\!\!\!\!\!E\!+\!\mu-T_{{{\bf k}}}^{22}
-\Sigma_{2\sigma}(E)&\ddots\\
\vdots&\ddots&\ddots\\
\end{array}\right)$}^{-1}_{\alpha\beta} \\[0.5cm]$$ Eqs. (\[cpa\]) and (\[local\_g\]) have to be solved selfconsistently to obtain $G_{ii\sigma}^{\alpha\beta}(E)$ and $\Sigma_{\alpha\sigma}(E)$.
The modified alloy analogy {#sec_maa}
==========================
Up to now nothing has been said about the actual choice of the atomic energies $E_{p\sigma}^{(\alpha)}$ and the corresponding concentrations $x_{p\sigma}^{(\alpha)}$ ($p=1,2$). In the conventional alloy analogy (AA) [@Hub64b] the alloy parameters (\[ap\]) are taken from the zero-bandwidth limit which directly corresponds to the assumption of strictly “frozen” $-\sigma$-electrons: $$\begin{aligned}
&\tilde E_{1\sigma}^{(\alpha)}=T_{0\alpha}\qquad
&\tilde E_{2\sigma}^{(\alpha)}=T_{0\alpha}+U \nonumber\\
&\tilde x_{1\sigma}^{(\alpha)}=1-n_{\alpha-\sigma}\qquad
&\tilde x_{2\sigma}^{(\alpha)}=n_{\alpha-\sigma}.\label{aa}\end{aligned}$$ However, it was soon realized that the AA is not able to describe itinerant ferromagnetism[@SD75]. This is closely related to the fact that the energy levels $\tilde E_{p\sigma}^{(\alpha)}$ are rigid and, in particular, spin-independent quantities within the AA. Further it is known [@HN96; @NH98; @PHN98] that the AA fulfills the sum rules (\[mom1\]), (\[mom2\]) up to the order $m=2$ only and fails to reproduce the correct strong coupling behavior. Note that within the AA the energy levels $\tilde E_{p\sigma}^{(\alpha)}$ are layer-independent (for uniform $T_{0\alpha}$) which is a crude approximation since a possible layer-dependence in the quasiparticle spectrum is suppressed almost completely.
The basic idea of the MAA is to exploit the information provided by the non-trivial but exact results in the limit of strong Coulomb interaction ($U/t\gg1$) [@HL67; @EO94] to determine the energy levels and concentrations (\[ap\]). This can most elegantly be achieved by imposing the sum rules (\[mom1\]), (\[mom2\]) on the CPA equation (\[cpa\]) [@PHN98; @PHWN98]: By inserting the high energy expansion of the self-energy $\Sigma_{\alpha\sigma}(E)$ and the local Green function $G_{ii}^{\alpha\alpha}(E)$, which are determined by the sum rules, the CPA equation (\[cpa\]) can be expanded in powers of $1/E$. Taking into account the sum rules up to the order $m=3$ unambiguously determines the parameters $E_{p\sigma}^{(\alpha)}$, $x_{p\sigma}^{(\alpha)}$. Then the exact strong coupling results are reproduced automatically [@PHWN98]. Note that due to the single site aspect of the CPA only the local terms of the $1/U$ perturbation theory are reproduced. On the other hand the MAA is not restricted solely to the strong coupling limit but is also applicable for intermediate interaction strengths where it has an interpolating character [@HN96; @NH98]. Following this procedure yields the energy levels and concentrations of the MAA for the Hubbard film: $$\begin{aligned}
E_{p\sigma}^{(\alpha)}&=&
{1\over 2} \bigg[T_{0\alpha}+U+B_{\alpha-\sigma}(\pm)^p\nonumber\\*
&&\sqrt{(U+B_{\alpha-\sigma}-T_{0\alpha})^2+
4 U n_{\alpha-\sigma}(T_{0\alpha}-B_{\alpha-\sigma})}\bigg],
\nonumber\\*
x_{1\sigma}^{(\alpha)}&=&
\frac{B_{\alpha-\sigma}+U(1- n_{\alpha-\sigma})-E_{1\sigma}^{(\alpha)}}
{E_{2\sigma}^{(\alpha)}-E_{1\sigma}^{(\alpha)}}\label{maa}\\
x_{2\sigma}^{(\alpha)}&=&1-x_{1\sigma}^{(\alpha)}\nonumber\end{aligned}$$ An alternative derivation of the MAA for bulk systems which is based on physical arguments can be found in Refs. [@HN96; @NH98]. Note that the expressions for $E_{p\sigma}^{(\alpha)}$ and $x_{p\sigma}^{(\alpha)}$ in (\[maa\]) are directly related to the position and the weight of the two poles of the spectral density within the SDA. Eqs. (\[maa\]) are obtained from the SDA results [@HN96; @NH98] if the electron dispersion is replaced by the center of gravity of the non-interacting band. The energy levels and concentrations (\[maa\]) are not only dependent on the model parameters $T_{0\alpha}$ and $U$ but also on the band occupations $n_{\alpha-\sigma}$ and the so-called bandshift $B_{\alpha-\sigma}$. The bandshift that is introduced via the fourth moment $M_{ij-\sigma}^{(3)\alpha\beta}$ consists of higher correlation functions: $$\begin{aligned}
B_{\alpha-\sigma}&=&\,T_{0\alpha}\!+\!
\frac{1}{n_{\alpha-\sigma}(1\!-\!n_{\alpha-\sigma})}
\sum_{j,\beta}^{j\beta\neq i\alpha}
T_{ij}^{\alpha\beta}\langle c_{i\alpha-\sigma}^{\dagger}
c_{j\beta-\sigma}(2n_{i\alpha\sigma}\!-\!1)\rangle.\end{aligned}$$ Nevertheless $B_{\alpha-\sigma}$ can exactly be calculated [@NB89; @HN97a] by use of the local Green function and the self-energy: $$\begin{aligned}
B_{\alpha-\sigma}&=&T_{0\alpha}+
\frac{1}{n_{\alpha-\sigma}(1-n_{\alpha-\sigma})}
\frac{1}{\hbar}\,\textrm{Im}\!\!\!\int\limits_{-\infty}^{+\infty}\!\!\!dE\,
f_{-}(E)
\left(\frac{2}{U}\Sigma_{\alpha-\sigma}(E-\mu)-1\right)
\times\nonumber\\
&&[(E-\Sigma_{\alpha-\sigma}(E-\mu)-T_{0\alpha})
G_{ii-\sigma}^{\alpha\alpha}(E-\mu)-\hbar]
\label{bas}\end{aligned}$$
In the strict zero-bandwidth limit $B_{\alpha-\sigma}$ is identical to $T_{0\alpha}$ and the MAA (\[maa\]) reduces to the conventional alloy analogy (\[aa\]). However, as soon as the hopping is switched on, the bandshift $B_{\alpha-\sigma}$, which is for strong Coulomb interaction proportional to the kinetic energy of the $-\sigma$-electrons in the $\alpha$-th layer [@HN97a], has to be calculated selfconsistently by iteration. Thus, via (\[maa\]) the type of the underlying alloy changes in each step of the iteration process. In this sense $B_{\alpha-\sigma}$ accounts for the itineracy of the $-\sigma$-electrons. In the paramagnetic phase there are only minor differences in the quasiparticle spectrum between MAA and AA. However, the bandshift may get a real spin-dependence for special parameter constellations. Thus $B_{\alpha-\sigma}$ may generate and stabilize ferromagnetic solutions which are excluded within the AA. It is worth to stress that the energy levels and concentrations are implicitly temperature-dependent via $n_{\alpha-\sigma}$ and $B_{\alpha-\sigma}$ leading, therefore, to a temperature dependent electronic structure.
The evaluation of the MAA requires the solution of two nested selfconsistency cycles. One starts with an initial guess for the band occupations $n_{\alpha-\sigma}$ and the bandshift $B_{\alpha-\sigma}$ which determine the energy levels and concentrations (\[maa\]). Via the CPA-equation (\[cpa\]) and (\[local\_g\]) the corresponding self-energy and Green function can be calculated selfconsistently. With this solution new values for $n_{\alpha-\sigma}$ and $B_{\alpha-\sigma}$ are obtained via (\[nas\]) and (\[bas\]). This procedure is iterated until convergence is achieved. For efficiency reasons the numerical evaluations of the integrals in (\[nas\]) and (\[bas\]) are performed via discrete Matsubara sums on the imaginary energy axis [@PHN98]. Only the spectral density and the quasiparticle density of states are calculated on the real axis at the end of each selfconsistency procedure.
Results and Discussion {#sec_results}
======================
For the numerical evaluations we consider in the present work thin fcc films with an (100) as well as an (111) surface and a film thickness up to $d=15$. The hopping integral between nearest neighbor sites is chosen to be uniform throughout the film and is set to $t=-0.25\,$eV. All other hopping integrals as well as $T_{0\alpha}$ are set to zero. For an fcc bulk system this yields a total bandwidth $W^{\textrm{\scriptsize bulk}}=4\,$eV of the non-interacting system. Further, we keep the on-site Coulomb interaction fixed at $U=50\,$eV which clearly refers to the strong coupling regime. In all calculations the total band occupation is kept fixed at the representative value $n=1.6$. Bulk calculations within the MAA have shown [@NH98] that for the fcc lattice ferromagnetic order is possible for all band occupations above half filling $n>1$.
---------------- ------- ------- ---
(100) (111)
\[0.3ex\] +1 4 3
\[0.3ex\] n.n. 0 4 6
\[0.3ex\] -1 4 3
\[0.3ex\] 8 9
-24 -30
0.667 0.750
0.5 0.625
---------------- ------- ------- ---
: Number of nearest neighbors (n.n.) for the fcc lattice as well as the fcc(100) and the fcc(111) film structure. In addition the moments $\Delta_\alpha^{(m)}$ (see Eq. (19)) of the BDOS are given. $\Delta_s^{(m)}$ refers to the surface layer, while the moments of all other layers ($2\le\alpha\le d-1$) are equal to the respective bulk values $\Delta_b^{(m)}$.[]{data-label="tab_one"}
For both film structures considered here, fcc(100) and fcc(111), all nearest neighbors are placed in the same or in the adjacent layer. The number of nearest neighbors in these two film geometries are given in Tab. \[tab\_one\]. The corresponding dispersions $T_{{{\bf k}}}^{\alpha\beta}$ can then be written as: $$T_{{{\bf k}}}^{\alpha\beta}=\left\{
\begin{array}{ll}
\,T_{0\alpha}+t\gamma_{||}({{\bf k}}),\,\, & \alpha=\beta\\
\,t\gamma_{\perp}({{\bf k}}),\,\,& \alpha=\beta-1\\
\,t(\gamma_{\perp}({{\bf k}}))^{\star},\,\,&
\alpha=\beta+1.\\
\end{array}
\right.$$ According to (\[hopping\]) one gets $$\begin{aligned}
\gamma_{||}^{(100)}({{\bf k}})\,\,\,&=&
2\Big[\cos(\frac{k_x+k_y}{2})+\cos(\frac{k_x-k_y}{2})\Big]\\
\gamma_{\perp}^{(100)}({{\bf k}})&=&
1+e^{-i\frac{k_x+k_y}{2}}+e^{-i\frac{k_x-k_y}{2}}
+e^{-i k_x}\end{aligned}$$ for the fcc(100) film geometry and $$\begin{aligned}
\gamma_{||}^{(111)}({{\bf k}})\,\,\,&=&
2\Big[\cos(\sqrt{3/8}k_x+\sqrt{1/8}k_y)+
\cos(\sqrt{3/8}k_x-\sqrt{1/8}k_y)+\nonumber\\
&&\cos(\sqrt{1/2}k_y)\Big]\\
\gamma_{\perp}^{(111)}({{\bf k}})&=&
1+e^{-i(\sqrt{3/8}k_x+\sqrt{1/8}k_y)}
+e^{-i(\sqrt{3/8}k_x-\sqrt{1/8}k_y)} \end{aligned}$$ for fcc(111). Here the lattice constant is set to $a=1$. The layer-dependent Bloch density of states $\rho_{0\alpha}(E)=\rho_{\alpha\sigma}(E)|_{U=0}$ for a five layer film is plotted in Fig. \[fig\_bdos\] for both film structures considered. The BDOS is strongly asymmetric and shows a distinct layer dependence. Considering the moments $$\label{bdos_mom}
\Delta_\alpha^{(m)}=\frac{1}{t^m}
\int\limits_{-\infty}^{\infty} dE (E-T_{0\alpha})^m \rho_{0\alpha}(E)$$ of the BDOS yields that the variance $\Delta_\alpha^{(2)}$ as well as the skewness $\Delta_\alpha^{(3)}$ are reduced at the surface layer compared to the inner layers due to the reduced coordination number at the surface (see Tab. \[tab\_one\]).
The charge distributions $n_\alpha$ as well as the layer magnetizations $m_\alpha$ are determined by the selfconsistently calculated QDOS (\[eq:qdos\]) via (\[nas\]). The chemical potential $\mu$ and the band centers $T_{0\alpha}$ are assumed to be uniform throughout the film, allowing, therefore, for charge transfer between the layers. However, in the actual calculation the difference in the occupation numbers $n_\alpha$ turns out to be very small ($<3\%$).
The layer-dependent magnetizations $m_\alpha$ together with the mean magnetization $m$ for a five layer film are plotted in Fig. \[fig\_m\_t\] as a function of temperature for both film structures. With respect to the overall shape the magnetization curves show the usual Brillouin-type behavior. However, the surface magnetization is found to be reduced compared to the inner layers for all temperatures. The reduction is particularly strong for the fcc(100) film geometry and leads to a non-saturated groundstate whereas the fcc(111) film is fully polarized at $T=0$. The enhanced surface effects in the fcc(100) structure are related to the higher percent of missing nearest neighbors at the surface layer which is $1/3$ for fcc(100) and $1/4$ for fcc(111).
We want to emphasize that the finding of a reduced surface magnetization cannot be explained by the well-known Stoner criterion of ferromagnetism. Since the variance of the BDOS is reduced at the surface layer ($\Delta_s^{(2)}<\Delta_b^{(2)}$, see Tab. \[tab\_one\]) due to the reduced coordination number one might intuitively expect the magnetization at the surface to be more robust than in the bulk. However, as discussed in Sect. \[sec\_intro\], intensive investigations of strongly correlated electron systems well beyond Hartree-Fock (Stoner) theory clearly point out the importance of a large skewness $\Delta^{(3)}$ for the stability of ferromagnetism [@HUM97; @WBS+98; @PHWN98]. Since the skewness of the BDOS is strongly reduced at the surface (see Tab. \[tab\_one\]) this explains the trend of a reduced surface magnetization. The above argument can be checked by considering the BDOS of the surface layer (see. Fig. \[fig\_bdos\]) as an input for an additional MAA calculation. Doing so we find that a “separated” surface layer would be ferromagnetic for the fcc(111) but paramagnetic for the fcc(100) film structure. In this sense the surface layer of an fcc(100) film is magnetized only because of the effective field induced by the ferromagnetically ordered inner layers.
The Curie temperature is found to be unique for the whole film. Note, that although the mean magnetization is reduced for the fcc(100) film with respect to fcc(111) the corresponding Curie temperature is enhanced ($T_C^{(100)}(d=5)=1140\,$K, $T_C^{(111)}(d=5)=1050\,$K). The inner layers that are fully polarized at $T=0\,$K for both film structures appear to be magnetically more stable for fcc(100) compared to fcc(111). Again, this trend can also be seen in an additional MAA calculation for the BDOS of the respective central layers. The Curie temperatures converge to the corresponding bulk value ($T_c^{\textrm{\scriptsize bulk}}=1050\,$K) for $d^{(111)}\approx 3$ and $d^{(100)}\approx 6$.
In Fig. \[fig\_m\_d\] the surface-, center-, and mean magnetization ($m_1$, $m_c$ and $m$) are shown as a function of the film thickness. The surface magnetization is reduced compared to the mean magnetization. The reduction is weak for fcc(111) films but very pronounced in the case of the fcc(100) structure. This holds not only for thin films where some oscillations are present due to the finite film thickness, but also extends to the limit $d\rightarrow\infty$ where the two surfaces are well separated and do not interact. The oscillations as a function of $d$ which are present for the fcc(100) structure get damped for higher temperatures. One can see from Fig. \[fig\_m\_d\] that the center layer magnetization $m_c$ for thick films ($d>10$) is in good agreement with the corresponding fcc bulk calculation [@NH98].
The magnetization profile for both film geometries is plotted for $d=10$ in Fig. \[fig\_m\_alpha\]. Here again the magnetizations of the fcc(100) film show a pronounced layer dependence while they are very close to the bulk value from the second layer on in the case of the fcc(111) film geometry. The magnetization profiles are similar to the ones obtained in [@Has87] for Cu/Ni/Cu sandwiches calculated within a single-site spin-fluctuation theory. However, within the present approach the deviation from the bulk magnetization is enhanced close to the Curie temperature for the fcc(100) structure (Fig. \[fig\_m\_alpha\]). Note that a similar trend to a reduced surface magnetization is also found within localized spin models. However, for the uniform Heisenberg model without a layer-dependent anisotropy contribution, the layer magnetizations necessarily increase [*monotonously*]{} from the surface to the central layer [@HBCC72; @SN98].
To understand the magnetic behavior on a microscopic basis we will, in the following, discuss the temperature-dependent electronic structure of the thin film systems. For a five layer fcc(100) film the spin- and layer-dependent spectral density at the gamma $\bar\Gamma$ point and the quasiparticle density of states are plotted in Fig. \[fig\_sk\_g\_qdos\]. There appear two correlation induced band-splittings in the quasiparticle spectrum: Due to the strong Coulomb interaction the spectrum splits into a low and a high energy subband (“Hubbard bands”) which are separated by an energy of the order $U$. Besides this so-called “Hubbard splitting” that is present for all temperatures there is an additional exchange splitting in majority ($\sigma=\uparrow$) and minority ($\sigma=\downarrow$) spin direction for temperatures below $T_C$. In the lower subband the electron mainly hops over empty sites, while in the upper subband it hops over lattice sites that are already occupied by another electron with opposite spin. The corresponding weights of the subbands scale with the probability of the realization of these two situations. In the strong coupling limit the scaling is given by $(1-n_{\alpha-\sigma})$ and $n_{\alpha-\sigma}$ for the lower and upper subband, respectively. Since the total band occupation ($n=1.6$) considered here is above half filling, the chemical potential $\mu$ is located in the upper subband while the lower subband is completely filled.
Starting from the Curie temperature the evolution of the quasiparticle spectrum with decreasing temperature is dominated by two distinct correlation effects. Both are driven by an increasing spin-asymmetry in the bandshift $B_{\alpha-\sigma}$. Firstly the centers of gravity of the majority and minority subbands move apart with decreasing temperature (Stoner-type behavior). Secondly there is a strong spin-dependent transfer of spectral weight between the lower and the upper subbands according to the above mentioned scaling, which results in spin- and temperature-dependent widths of the respective subbands. This behavior can also be seen in detail in Fig. \[fig\_sk\_qdos\_100\] and Fig. \[fig\_sk\_qdos\_111\] where the quasiparticle bandstructure of the surface and central layer is plotted for a five layer fcc(100) and fcc(111) film, respectively. Here, only the upper subbands are shown. While the centers of gravity of the upper $\sigma=\downarrow$ subbands are shifted to higher energies for decreasing temperatures, the lowest excitation peak in the spectral density at $\bar\Gamma$ is even lowered due to the increasing bandwidth. On the other hand the width of the upper $\sigma=\uparrow$ subband decreases. The interplay of these two correlation effects leads to an inverse exchange splitting at the lower edge of the upper subband near to the $\bar\Gamma$ point. The corresponding quasiparticle density of states is, however, very small. Note, that for the same reason the position of the central peak of the upper subband is almost spin- and temperature-independent. This behavior holds for both film structures for ${{\bf k}}$-vectors not too far away from $\bar\Gamma$.
With help of the quasiparticle bandstructure given in Fig. \[fig\_sk\_qdos\_100\] and Fig. \[fig\_sk\_qdos\_111\] the exchange splitting between majority and minority spin direction can be analyzed in more detail. For both film structures the exchange splitting is wavevector-dependent. It is strongest near $\bar M$ for fcc(100) and between $\bar M$ and $\bar K$ for fcc(111). Contrary to the fcc(100) structure all layers are fully polarized at $T=0\,$K for the fcc(111) film. In the case of ferromagnetic saturation the exchange splitting can be estimated [@HN96] to be at most $(1-n_{\alpha\downarrow})[4t-B_{\alpha\downarrow}]$, where $(1-n_{\alpha\downarrow})B_{\alpha\downarrow}$ is the effective bandshift between the centers of gravity of the upper quasiparticle subbands. For strong Coulomb interaction $n_{\alpha\downarrow}(1-n_{\alpha\downarrow})B_{\alpha\downarrow}$ is proportional to the kinetic energy [@HN97a] of the $\sigma=\downarrow$ electrons in the $\alpha$-th layer. Note that the kinetic energy of the $\sigma=\uparrow$ electrons vanishes for the ferromagnetic saturated state since the $\sigma=\uparrow$ band is completely filled. We want to point out that these results strongly contrast the findings of Hartree-Fock theory where the exchange splitting is wavevector-independent and proportional to $m_\alpha U$ leading to substantially higher Curie temperatures compared to the MAA. The temperature-dependence of the electronic structure within the MAA is completely different to the Stoner picture of ferromagnetism.
Let us discuss the quasiparticle lifetime which corresponds to the width of the quasiparticle peaks. From the spectral density (Figs. \[fig\_sk\_g\_qdos\], \[fig\_sk\_qdos\_100\], \[fig\_sk\_qdos\_111\]) one can clearly read off that the lifetime of the quasiparticles is strongly spin- and temperature-dependent. For low temperatures the upper minority spectrum is sharply peaked which indicates long living quasiparticles. This is due to the fact that in the ferromagnetic saturated state a $\sigma=\downarrow$ electron does meet a $\sigma=\uparrow$ electron at any lattice site and thus effectively does not perform any scattering process. The width of the $\sigma=\uparrow$ quasiparticle peaks, however, is broadened for decreasing temperature. Thus in the majority spectrum the quasiparticle lifetime decreases for increasing magnetization. What concerns the lower subbands (see Fig. \[fig\_sk\_g\_qdos\]) the respective spectrum is strongly damped and the different excitations due to the five layer structure are almost indistinguishable.
For given spin and wavevector the positions of the quasiparticle peaks are layer-independent. In principle, their number corresponds to the number of layers of the film. However, due to symmetry some peaks are left out for certain layers. For thicker films the different peaks move closer together as their number increases until they build a continuum for $d\rightarrow\infty$ which corresponds to the projection of the three dimensional bandstructure onto the surface Brillouin zone. Between $\bar M$ and $\bar X$ for fcc(100) and at $\bar K$ for fcc(111) the different peaks merge together due to vanishing interlayer hopping ($\gamma_{\perp}({{\bf k}})=0$).
In Figs. \[fig\_sk\_qdos\_100\], \[fig\_sk\_qdos\_111\] the QDOS of the surface and central layer are shown additionally. The Van Hove singularities resulting from the different branches of the quasiparticle dispersion are clearly visible. There are sharp Van Hove singularities in the minority spectrum while they are broadened for the majority spin direction because of the finite widths of the $\sigma=\uparrow$ quasiparticle peaks due to the quasiparticle damping.
Finally we want to stress that the results presented above do not depend on the size of the Coulomb interaction $U$ as long as $U$ is chosen from the strong-coupling region ($U\gg W$). Contrary to Hartree-Fock theory, all magnetic key quantities like the Curie temperature and the exchange splitting saturate as a function of $U$. On the other hand, although the MAA was optimized with respect to the strong coupling limit we believe that, at least qualitatively, the correlation effects in the spin-, layer-, and temperature-dependent electronic structure are valid down to intermediate Coulomb interaction as well.
Conclusion {#sec_conclusion}
==========
For the investigation of spontaneous ferromagnetism and electron correlation effects in thin itinerant-electron films we have applied a generalization of the modified alloy analogy (MAA) to the single-band Hubbard model with reduced translational symmetry. The MAA is based on the alloy analogy concept and is optimized with respect to correct strong coupling behavior [@HL67; @EO94]. Within the MAA the actual type of the underlying alloy is not predetermined but has to be determined selfconsistently. In this sense the MAA is able to account for the itineracy of the $-\sigma$ electrons which are considered as strictly “frozen” in the conventional alloy analogy (AA). In the paramagnetic phase MAA and AA are almost identical. However, contrary to the AA spontaneous ferromagnetic order is possible for special parameter constellations within the MAA. With help of the MAA the interplay of magnetism and quasiparticle damping effects can be studied in a natural way.
For an fcc(100) and an fcc(111) film geometry the layer-dependent magnetizations have been discussed as a function of temperature as well as film thickness. The magnetization in the surface layer is found to be reduced with respect to the inner layers for all thicknesses and temperatures considered. While this reduction is weak for fcc(111) films it is pronounced in the case of an fcc(100) geometry. The effect of the surface is considerably stronger for fcc(100) films due to the higher percent of missing nearest neighbor atoms. The reduction of the surface layer magnetization is not to be expected within an Hartree-Fock type approach (Stoner criterion) to the Hubbard film being, therefore, a genuine effect induced by strong electron correlations.
The magnetic behavior of the thin film systems can be microscopically understood by means of the spin- layer- and temperature- dependent quasiparticle bandstructure and the corresponding quasiparticle density of states. There appear two correlation induced band splittings in the quasiparticle spectrum. Besides the Hubbard splitting there is an additional exchange splitting for temperatures below $T_C$. The demagnetization process as a function of temperature is dominated by two distinct correlation effects: A Stoner-like shift in the centers of gravity of the majority and minority subbands together with a strong spin-dependent transfer of spectral weight between the upper and lower subbands. An interplay of these two effects results in Curie temperatures far below the corresponding Hartree-Fock values. The exchange splitting is found to be strongly wavevector-dependent and is substantially different for the various quasiparticle branches in the bandstructure. The widths of the quasiparticle peaks that correspond to the quasiparticle lifetime exhibit a strong spin- and temperature-dependence. For $T=0\,$K the minority-spin quasiparticle peaks are sharply peaked while the majority-spin spectrum is substantially broadened.
Clearly the degeneracy of the 3d-bands has to be included if a direct comparison to the experiment is intended. Within the present scheme this could be achieved by a similar approach as presented in [@NBDF89] which is planed for the future. However we believe the correlation effects found here to be important within a generalized Hubbard model as well. In this work we have exclusively focused on purely ferromagnetic films. In addition one can examine within the same theory a phase with antiferromagnetic order between the layers. We expect such a situation to exist close to half-filling ($n=1$) and for intermediate values of the Coulomb interaction. Further the influence of a non-magnetic top layer on the magnetic behavior of thin films can be investigated.
This work has been done within the Sonderforschungsbereich 290 (“Metallische dünne Filme: Struktur, Magnetismus und elektronische Eigenschaften”) of the Deutsche Forschungsgemeinschaft.
References {#references .unnumbered}
==========
[10]{}
R. Allenspach, J. Magn. Magn. Mat. [**129**]{}, 160 (1994).
K. Baberschke, Appl. Phys. A [**62**]{}, 417 (1996).
H. J. Elmers, Int. J. Mod. Phys. B [**9**]{}, 3115 (1995).
M. Getzlaff, J. Bansmann, J. Braun, and G. Schönhense, Z. Phys. B [**104**]{}, 11 (1997).
W. Dürr, M. Taborelli, O. Paul, R. Germar, W. Gudat, D. Pescia, and M. Landolt, Phys. Rev. Lett. [**62**]{}, 206 (1989).
D. P. Pappas, K.-P. Kämper, and H. Hopster, Phys. Rev. Lett. [**64**]{}, 3179 (1990).
M. Farle, W. Platow, A. N. Anisimov, P. Poulopoulos, and K. Baberschke, Phys. Rev. B [**56**]{}, 5100 (1997).
C. L. Fu and A. J. Freeman, Phys. Rev. B [**35**]{}, 925 (1987).
T. Kraft, P. M. Marcus, and M. Scheffler, Phys. Rev. B [**49**]{}, 11511 (1995).
T. Asada and S. Blügel, Phys. Rev. Lett. [**79**]{}, 507 (1997).
R. Lorenz and J. Hafner, Phys. Rev. B [**54**]{}, 15937 (1996).
D. Spišák and J. Hafner, Phys. Rev. B [**56**]{}, 2646 (1997).
D. Reiser, J. Henk, H. Gollisch, and R. Feder, Solid State Commun. [**93**]{}, 231 (1995).
M. P. Gokhale and D. L. Mills, Phys. Rev. B [**49**]{}, 3880 (1994); M. Plihal and D. L. Mills, Phys. Rev. B [**52**]{}, 12813 (1995).
M. Potthoff and W. Nolting, Surf. Sci. [**377-379**]{}, 457 (1997).
T. Herrmann, M. Potthoff, and W. Nolting, Phys. Rev. B [**58**]{}, 831 (1998).
N. D. Mermin and H. Wagner, Phys. Rev. Lett. [**17**]{}, 1133 (1966).
M. Bander and D. L. Mills, Phys. Rev. B [**38**]{}, 12015 (1988).
A.-M. Daré and Y. M. Vilk anf A.-M. S. Tremblay, Phys. Rev. B [**53**]{}, 14236 (1996).
R. Schiller and W. Nolting, to be published (1998).
H. Hasegawa, Surf. Sci. [**182**]{}, 591 (1987).
W. Haubenreisser, W. Brodkorb, A. Corciovei, and G. Costache, phys. stat. sol. (b) [**53**]{}, 9 (1972); D. T. Hung, J.C. S. Levy, and O. Nagai, phys. stat. sol. (b) [**93**]{}, 351 (1979).
R. P. Erickson and D. L. Mills, Phys. Rev. B [**43**]{}, 10715 (1991); R. P. Erickson and D. L. Mills, Phys. Rev. B [**44**]{}, 11825 (1991).
P. J. Jensen, H. Dreyssé, and K. H. Bennemann, Surf. Sci. [**269/270**]{}, 627 (1991).
Long-Pei Shi and Wei-Gang Yang, J. Phys.: Condens. Matter [**4**]{}, 7997 (1992).
A. Hucht and K. D. Usadel, Phys. Rev. B [**55**]{}, 12309 (1997).
P. J. Jensen and K. H. Bennemann, Solid State Commun. [**105**]{}, 577 (1998).
J. Hubbard, Proc. R. Soc. London, Ser. A [**276**]{}, 238 (1963); M. C. Gutzwiller, Phys. Rev. Lett. [**10**]{}, 159 (1963); J. Kanamori, Prog. Theor. Phys. (Kyoto) [**30**]{}, 275 (1963).
J. Hubbard, Proc. R. Soc. London, Ser. A [**277**]{}, 237 (1964).
D. Vollhardt, N. Blümer, K. Held, J. Schlipf, and M. Ulmke, Z. Phys. B [**103**]{}, 283 (1997).
K. Held and D. Vollhardt, to appear in Eur. Phys. J. B (1998), cond–mat/9803182.
T. Momoi and K. Kubo, Phys. Rev. B [**58**]{}, R567 (1998).
W. Nolting, W. Borgie[ł]{}, V. Dose, and Th. Fauster, Phys. Rev. B [**40**]{}, 5015 (1989); W. Nolting, A. Vega, and Th. Fauster, Z. Phys. B [**96**]{}, 357 (1995); A. Vega and W.Nolting, phys. stat. sol. (b) [**193**]{}, 177 (1996).
A. Haroun, A. Chouairi, S. Ouannasser, H. Dreyssé, G. Fabricius, and A.M. Llois, Surf. Sci. [**307-309**]{}, 1087 (1994).
J. Dorantes-Dávila, H. Dreyssé, and G. M. Pastor, Phys. Rev. B [**55**]{}, 15033 (1997).
H. Hasegawa, J. Phys.: Condens. Matter [**4**]{}, 1047 (1992).
M. Ulmke, Eur. Phys. J. B [**1**]{}, 301 (1998).
G. S. Uhrig, Phys. Rev. Lett. [**77**]{}, 3629 (1996).
Th. Hanisch and G. Uhrig. E. Müller-Hartmann, Phys. Rev. B [**56**]{}, 13960 (1997).
J. Wahle, N. Blümer, J. Schlipf, K. Held, and D. Vollhardt, Phys. Rev. B [**58**]{}, 12749 (1998).
T. Obermaier, T. Pruschke, and J. Keller, Phys. Rev. B [**56**]{}, 8479 (1997).
T. Herrmann and W. Nolting, Solid State Commun. [**103**]{}, 351 (1997).
M. Potthoff, T. Herrmann, T. Wegner, and W. Nolting, phys. stat. sol. (b) [**210**]{}, 199 (1998).
A. B. Harris and R. V. Lange, Phys. Rev. [**157**]{}, 295 (1967).
H. Eskes and A. M. Oleś, Phys. Rev. Lett. [**73**]{}, 1279 (1994); H. Eskes, A. M. Oleś, M. B. J. Meinders, and W. Stephan, Phys. Rev. B [**50**]{}, 17980 (1994).
W. Nolting and W. Borgieł, Phys. Rev. B [**39**]{}, 6962 (1989).
T. Herrmann and W. Nolting, J. Magn. Magn. Mat. [**170**]{}, 253 (1997).
T. Herrmann and W. Nolting, Phys. Rev. B [**53**]{}, 10579 (1996).
W. Nolting and T. Herrmann, Condensed Matter Theories [**13**]{}, 169 (1998).
M. Potthoff, T. Herrmann, and W. Nolting, Eur. Phys. J. B [**4**]{}, 485 (1998).
M. Potthoff and W. Nolting, J. Phys.: Condens. Matter [**8**]{}, 4937 (1996).
J. Hubbard, Proc. R. Soc. London, Ser. A [**281**]{}, 401 (1964).
B. Velický, S. Kirkpatrick, and H. Ehrenreich, Phys. Rev. [**175**]{}, 747 (1968).
R. Vlamming and D. Vollhardt, Phys. Rev. B [**45**]{}, 4637 (1992).
J. Schneider and V. Drchal, phys. stat. sol. (b) [**68**]{}, 207 (1975).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We consider the problem of verifying stochastic models of biochemical networks against behavioral properties expressed in temporal logic terms. Exact probabilistic verification approaches such as, for example, CSL/PCTL model checking, are undermined by a huge computational demand which rule them out for most real case studies. Less demanding approaches, such as statistical model checking, estimate the likelihood that a property is satisfied by sampling *executions* out of the stochastic model. We propose a methodology for efficiently estimating the likelihood that a LTL property $\phi$ holds of a stochastic model of a biochemical network. As with other statistical verification techniques, the methodology we propose uses a stochastic simulation algorithm for generating execution samples, however there are three key aspects that improve the efficiency: first, the sample generation is driven by on-the-fly verification of $\phi$ which results in optimal overall simulation time. Second, the confidence interval estimation for the probability of $\phi$ to hold is based on an efficient variant of the Wilson method which ensures a faster convergence. Third, the whole methodology is designed according to a parallel fashion and a prototype software tool has been implemented that performs the sampling/verification process in parallel over an HPC architecture.'
author:
- 'P. Ballarini M. Forlin T. Mazza D. Prandi'
bibliography:
- 'PDMC09\_EPTCS\_BIBLIO.bib'
title: Efficient Parallel Statistical Model Checking of Biochemical Networks
---
Introduction {#sec:Intro}
============
Systems biology [@Kitano02] is concerned with developing detailed models of complex biological networks, which then need to be validated and analysed. A model of a biological system essentially describes the dynamics of a population of $n$ interacting biochemical species $S_1, S_2, \ldots S_n$. Analysing the behaviour of such models entails looking for the occurrence of biologically relevant events during the evolution of the system, a problem whose complexity is exponentially proportional to the population of the considered model.
In this paper we consider discrete stochastic modelling of biological systems. In the discrete-stochastic setting, biochemical species are enumerable quantities representing the number of molecules of a given substance, and the evolution of the system is probabilistic, rather than deterministic, leading to Continuous Time Markov Chain (CTMC) models.
Classical *transient-state* and *steady-state* analysis [@Stewart94] allows to understand important features of a CTMC model. Unfortunately numerical solution of Markov models is undermined by storage requirements which explode with the dimension of the model. Although techniques for the efficient storing of CTMC’s matrix/vector have been widely studied (see for example [@MiCi99; @ScBo98]), solving CTMC models remains unfeasible, in most realistic case studies.
In recent times, query based verification methods (i.e. model checking [@Clarke99]) proved to be valuable instruments for a more expressive analysis of stochastic biological models [@KNP09e; @ballarini2009studying; @Barnat:1211142]. Biologically relevant events can be formally characterised as temporal logic formulae that can then be automatically checked against a discrete state model. *Exact* probabilistic PCTL [@HansJons89] and stochastic CSL [@BHHK03; @aziz00] model checking suffer, as well, of the state-space explosion problem which limits their accessibility specially in systems biology where very large models are just common. *Statistical model checking* [@YKNP06; @Clarke08; @Donaldson08] has been proposed as an alternative to exact probabilistic model checking that allows for getting an estimate of the likelihood of a condition to hold of a CTMC model.
The *statistical model checking* paradigm essentially is a combination of efficient stochastic simulation with model checking procedure. It comprises three ingredients: (*i*) a stochastic engine that generates trajectories of the underlying state space; (*ii*) a model checking algorithm capable of analyzing a single trajectory; (*iii*) a statistical support for estimating the accuracy of the answer. The advantage of statistical model checking is that, contrary to exact model checking, it does not need to build (i.e. store) the state space of the model as it only explores a limited number of (finite) trajectories. The cost paid for such space saving is in terms of precision of the calculated measure.
We propose a methodology that given a CTMC model $M$, a property $\phi$ and a desired *level of confidence* estimates the probability of $\phi$ to be satisfied by $M$ with the estimated measure meeting the desired confidence. The method we propose is based on three key aspects. Firstly the trajectory generation is controlled by on-the-fly verification of the considered formula which means that simulation halts as soon as a state which verifies (falsifies) the formula is reached. Secondly, we use an efficient statistical method (i.e. a variant of the *Wilson score interval method*) which results in smaller samples (i.e fewer simulation runs) in order to meet the desired confidence. Finally the whole simulation/verification framework has been designed and tested on a parallel prototype, which is based on independent simulation/verification engines generation, and a MPI client/server parallel computation architecture.
The reminder of the paper is organised as follows: in Section \[sec:BioNetworks\] we introduce the basics about CTMC modelling and stochastic simulation algorithms. In Section \[sec:statLTLMC\] we describe the core of our statistical model checking procedure; in Section \[sec:tool\] the architecture of the prototype which exploits the methodology is described. In Section \[sec:Experiments\] we provide evidence of application of statistical verification of relevant properties to an abstract model of the budding yeast cell cycle. Concluding remarks are given in the final section of the paper.
Stochastic Simulation of Biochemical Networks {#sec:BioNetworks}
=============================================
The time evolution of a well stirred biochemical reacting system can be described as a Continuous-time Markov Chain (CTMC) whose states are vectors $\vec{X}(t)$ of discrete random variables $X_i(t)$, that give the amount of a molecule $i$ at time $t$. The associated joint probability distribution $\mathcal{P}(\vec{X},t)$, often called *Chemical Master Equation* (CME), gives a precise description of a biochemical network [@gillespie1992rigorous]. Unfortunately, as soon as the system becomes non-trivial, it is nearly impossible to solve CME (i.e. the corresponding CTMC), neither analytically nor numerically. The Stochastic Simulation Algorithm [@Gillespie77] is a computer program that takes a biological reacting system and produces a trace in the space of the CME. In the following we first introduce the basics about CTMCs and then we describe how the stochastic simulation algorithm works.
Continuous Time Markov Chains {#sec:CTMC}
-----------------------------
A Continuous-time Markov Chain (CTMC) for biological modelling is a tuple $M=(S,s_0,Q,Var,Eval)$ where $S$ is a finite set of states, $s_0$ is the initial state, $Q$ is the *infinitesimal generator matrix* of the chain, $Var$ is the set of *state-variables* and $Eval:S\times Var \to \mathbb{R}_{\geq 0}$ is the *evaluation function* that, given the variable $x$ and state $s$, returns the value of $x$ in $s$.
A path of a CTMC $M$ is a possibly infinite sequence of $S\times\mathbb{R}_{> 0}$ pairs $\sigma\equiv (s_0,t_0), (s_1,t_1) \ldots (s_n,t_n)\ldots$ describing a trajectory of $M$, also denoted: $$s_0\stackrel{t_0}\longrightarrow s_1\stackrel{t_1}\longrightarrow s_2\ldots\stackrel{t_{n-1}}\longrightarrow s_n\stackrel{t_n}\longrightarrow\ldots$$ where $\forall i\in\mathbb{N}$ $s_i\in S$ and $Q(s_i,s_{i+1})>0$ and $t_i\in\mathbb{R}_{>0}$. A point $(s_i,t_i)$ of $\sigma$ indicates that at the $i$-th step of the execution $\sigma$ the system was in state $s_i$ and that it stayed in $s_i$ for $t_i$. We adopt the following notation: $\sigma[i]$ denotes the $i$-th state of $\sigma$; $\sigma^i$ denotes the $i$-th suffix of $\sigma$; $\delta(\sigma, i)$ denotes the time spent in the $i$-th state of $\sigma$ whereas $\sigma@ t$ denotes which state $\sigma$ is in at time $t$.
The state-variables $Var$ of a CTMC *biochemical* model will coincide with $ \{ S_1, \ldots, S_N \}$, the set of biochemical species.
Stochastic Simulation Algorithm {#sec:SSA}
-------------------------------
The Stochastic Simulation Algorithm (SSA) supposes a well-stirred set $ \{ S_1, \ldots, S_N \}$ of biochemical species reacting through $M \geq 1$ reaction channels (reactions for short) $ \{ R_1, \ldots, R_M \}$. A reaction channel $R_j$ is characterized by:
- the probability $a_j(\vec{x}) dt$ that, given $\vec{X}(t) = \vec{x}$, one reaction $R_j$ will occur in the next infinitesimal interval $[ t, t + dt )$;
- the change $v_{ji}$ of the number of molecules of the specie $S_i$ produced or consumed by a reaction $R_j$.
The physical and mathematical explanation of $a_j(\vec{x}) dt$ is described in [@gillespie1992rigorous]. Basically, $a_j(\vec{x})$ is a function of the number of possible active instances of reaction $R_j$. Consider, for example, a reaction $R_1: S_1 + S_2 \rightarrow S_3$, then $a_1(\vec{x}) = c_1 x_1 x_2 $, where $x_1 x_2$ is the number of active $R_1$ in the current state $\vec{x}$, and $c_1$ is a constant that depends on the physical characteristics of $S_1$ and $S_2$.
Given $a_j(\vec{x})$, the evolution of a biochemical network is described by the *next reaction density function* $p(\tau, j | \vec{x}, t )$, i.e. the probability, given $\vec{X}(t) = \vec{x}$, that the next reaction in the system will occur in the infinitesimal time interval $[ t + \tau, t + \tau + dt )$ and will be on channel $R_j$: $$p(\tau, j | \vec{x}, t ) = a_j( \vec{x} ) e^{ - a_0(\vec{x}) \tau } \label{eq:nextreaction}$$ where $a_0 = \sum_{j=1}^{M} a_j(\vec{x})$. By conditional probability, Eq. \[eq:nextreaction\] becomes $p(\tau, j | \vec{x}, t ) = p_1(\tau | \vec{x}, t)p_2(j | \tau, \vec{x}, t)$ where
$$p_1(\tau | \vec{x}, t) = a_0(\vec{x}) e^{ - a_0(\vec{x}) \tau } \qquad\ \qquad (\tau \geq 0) \label{eq:time}$$
$$p_2(j | \tau, \vec{x}, t) = \frac{a_j(\vec{x})}{a_0(\vec{x})} \label{eq:reaction}$$
meaning that, $\tau$ is a sample from an exponential random variable with rate $a_0({\vec{x}})$, and the selected reaction $j$ is independently taken from a discrete random variable with values in \[1,M\] and probabilities $\frac{a_j(\vec{x})}{a_0(\vec{x})}$.
Standard Monte Carlo methods are then used to select time consumed and next reaction according to Eq. \[eq:time\] and Eq. \[eq:reaction\], to produce a trajectory in the discrete state-space of the CME. Several variants of the SSA exist [@li2008algorithms] that differ in how the next reaction is selected and in the data structures used, but all of them are based on the following common template:
1
: Compute $a_0$
2
: Randomly select a reaction $j$ according to Equation \[eq:reaction\]
3
: Randomly select a time duration $\tau$ according to Equation \[eq:time\]
4
: Update state vector as $\vec{x} \gets \vec{x} + v_{j}$ and total time as $t \gets t + \tau$
5
: Go to step 1 or Terminate
On-the-fly statistical BLTLc Model Checking {#sec:statLTLMC}
===========================================
Statistical verification of a CTMC model $M$ is based on the simple principle of collecting $N$ sample realisations $\sigma_i$ ($i\in\{1,\ldots ,N\}$) of $M$ and verifying each of them against a given property $\phi$. The estimate of the likelihood of $\phi$ to hold true of $M$ is obtained as the frequency $\hat{p}_{\phi}=\frac{po}{N}$ of positive outcomes ($po$) of the verification of $\phi$ versus $\sigma_i$. In the following we introduce the temporal logic we refer to as the language for stating properties of simulated trajectories, namely the BLTLc logic. We then provide details of a procedure for on-the-fly simulation-verification of a CTMC model which combines the random generation of a trajectory, by means of stochastic simulation (see Section \[sec:BioNetworks\]), with a verification procedure based on the BLTLc semantics. Finally we describe a statistical procedure for the efficient estimation of the likelihood of a formula. This procedure iteratively works out the number of simulation runs needed in order to meet the desired level of confidence.
Bounded Linear-time Temporal Logic with numerical constraints {#sec:LTL}
-------------------------------------------------------------
We introduce the Bounded Linear-time Temporal Logic with numerical constraints (BLTLc) a logic for stating properties referred to timed-trajectories resulting from stochastic simulation of a CTMC model. BLTLc combines the Constraint-LTL [@Fages06; @DBLP:conf/cmsb/FagesR07] ($LTLc$ or $LTL(\mathbb{R})$) for *arithmetic-constrained* LTL formulae with the Bounded-LTL (BLTL) [@Clarke08] for time-bounded LTL formulae. Both LTLc and BLTL are based on classical LTL [@Pnueli77] temporal operators. However while LTLc allows for using (complex) arithmetical conditions between state variables, it does not allow for expressing time bounded conditions. On the other hand with BLTL time bounded LTL expressions can be formed but based on simple non-arithmetical conditions rather than on a grammar for arithmetic expressions as it is the case with LTLc. The syntax of the BLTLc logic is given by the following grammar: $$\begin{aligned}
\phi &:= val \trianglelefteq val \mid \lnot\phi \mid\ \phi\lor\phi \mid\ \phi \land\phi \mid X^I\ \phi\mid \phi\ U^I\phi\\ \\
val &:= x \mid val \sim val \mid func \mid Int \mid Real\end{aligned}$$ where $I\subseteq\mathbb{R}_{\geq 0}$, $\trianglelefteq{\in}\{<,\leq,\geq,>,=,\neq\}$, $\sim{\in}\{+,-,*,/\}$ and $\mathit{func}$ denotes common functions such as $pow(),sqrt(),\ldots $. Note that unbounded formulae correspond to $I{=}[0,+\infty)$ (for simplicity $ [0,+\infty)$ is usually omitted for unbounded formulae). The operators $\lnot$, $\lor$ and $\land$ are the standard boolean logic connectives *not*, *or* and *and* respectively, whereas $X^I$ and $U^I$ denote the *temporal connectives*, *next* and *until*, respectively. The *next* operator refers to the notion of *true in the next state* within time $I$ whereas the *until* operator ($\phi\ U^I \psi$) indicates that a future state where its second argument ($\psi$) holds is reached within time $I$ and while its first argument ($\phi$) continuously holds. The two popular *Finally* operator ($F^I (\phi)$, which refers to the notion of a condition holding true at some point in the future and within time $I$), and *Globally* operator ($G^I (\phi)$, which refers to the notion of a condition to continuously holding true within time $I$), are also supported by BLTLc relying on the well-know equivalences $F^I(\phi)\equiv[ tt\ U^I \phi]$ and $G^I(\phi)\equiv\neg F^I(\neg\phi)$.
BLTLc formulae are evaluated against timed-paths resulting from simulation of a CTMC model. The formal semantics of BLTLc formulae, expressed in terms of the $\models$ relation, is given below, where $\sigma$ is a timed-path of a CTMC model.
- $\sigma\models (val'\trianglelefteq val'')$ if and only if $Eval(\sigma[0],val')\trianglelefteq Eval(\sigma[0],val'')$
- $\sigma\models \lnot \phi $ if and only if $\sigma\not\models\phi$
- $\sigma\models \phi'\lor\phi'' $ if and only if $\sigma\models\phi'$ or $\sigma\models\phi''$
- $\sigma\models \phi'\land\phi'' $ if and only if $\sigma\models\phi'$ and $\sigma\models\phi''$
- $\sigma\models X^I \phi $ if and only if $\sigma^1\models\phi$ and $\delta(\sigma,0){\in}I$
- $\sigma \models \phi' \ U^I\ \phi''$ if and only if $\exists i{\in}\mathbb{N}:
\sigma^i\models\phi''$ and $\sum_{j<i} \delta(\sigma,j) {\in}I$ and $\forall j<i$, $ \sigma^j\models\phi' $
$Eval(\sigma[i],val)$ is the *evaluation function* that assigns a numerical value to the expression $val$ by looking up at the value that each variable $x{\in}Var(val)$ has in state $\sigma[i]$ of path $\sigma$ (see Section \[sec:CTMC\]).
As an example of the expressiveness of the BLTLc logic consider the following formula $\phi\equiv[(X_1<Sqrt(X_2))\ U\ (X_2\geq 10+X_3)]$ which states that the concentration of $X_1$ shall be less than the square root of that of $X_2$ until that of $X_2$ exceeds $X_3$ by at least 10.
// The procedure returns a (Result,Trace) pair, where Result is the boolean
// result of verification and Trace is the simulation trace generated.
(Result, Trace) simulate_verify(i, sigmabuff, M ,phi, tmax)
(s,t) = sigmabuff[i];
case phi of
propositional formula: return (Eval(psi, M, s), sigmabuff);
neg psi: (r, tr) = simulate_verify(i, sigmabuff, M, phi, tmax);
return (not r, tr);
psi1 and psi2: (r, tr) = simulate_verify(i, sigmabuff, M, phi, tmax);
(rprime, trprime) = simulate_verify(i, tr, M, phi, tmax);
return (r and rprime, trprime);
nextI psi: if (t notin I) then return (false, sigmabuff);
if (i == max_index(sigmabuff))
{ (snext, tnext) = Next((s,t), M);
sigmabuff += (snext, tnext); }
return simulate_verify(i+1, sigmabuff, M, phi, tmax);
psi1 untilI psi2: if (t notin I) return (false, sigmabuff);
(r, tr) = simulate_verify(i, sigmabuff, M, psi2, tmax);
if ( last(tr).time > tmax) return (false, tr);
if ( r ) then return (true, tr);
(rprime, trprime) = simulate_verify(i, tr, M, psi1, tmax);
if ( neg rprime ) return (false, trprime);
if (i == max_index(trprime))
{ (snext, tnext) = Next((s,t), M);
sigmabuff += (snext, tnext); }
return simulate_verify(i+1, trprime, M, phi, tmax);
Finally we stress that differently from the original BLTL [@Clarke08], the association of [probabilistic bounds]{} to BLTLc formulae is not supported. This is because the aim of the statistical verification procedure we introduce is to estimate the actual measure of probability of a BLTLc formula rather than to estimate whether such measure is below/above a certain threshold.
On-the-fly verification of simulated trajectory {#sec:onthefly}
-----------------------------------------------
We define a procedure that given a CTMC model $M$ and a BLTLc formula $\phi$ allows for stochastically generating trajectories of $M$ while verifying whether the generated trajectories satisfy or falsify the considered formula $\phi$. The verification of $\phi$ is performed *on-the-fly* meaning that simulation proceeds with the generation of the next state only if $\phi$ is neither satisfied nor falsified in the current one (and if the simulation time limit has not been reached). The pseudocode for the verification procedure is outlined in Figure \[fig:verification\]. Function *simulate\_verify()* takes few input parameters: a BLTLc formula $\phi$, a trace $\sigma_{buff}$ (containing the *pre-viewed* trace resulting from recursive calls corresponding to the verification of sub-formulae) the current position $i$ in the buffered trace, a CTMC $M$ and a maximum simulation time $t_{max}$. The algorithm works in the following manner: propositional formulae are verified trivially relying on the function *Eval*$(\phi, M, s)$, that gives the value of a non temporal formula $\phi$ in a state $s$. On the other hand temporal formulae (may) require the generation of a simulation-trace, obtained through a call to the the stochastic engine function *Next*$((s,t),M)$. Verification of temporal formulae may result in passing of the already-generated trace (i.e. the buffered trace $\sigma_{buff}$ that results from verification of a sub-formula) from the inner-most sub-formulae to the outer-most ones. This is required for two reasons: first to avoid possible branching during the generation of a single trace (if verification requires to roll-back to a previously generated state then re-application of the SSA may result in the generation of a different successor state, else in branching, which is wrong in the LTL context), secondly for efficiency (the already generated traces should be re-used whenever feasible). For these reasons a call to the *simulate\_verify()* returns a pair $(r,tr)$ where $r$ is the boolean result of the verification of the considered formula $\phi$ and $tr$ is the simulation trace generated up until verification of $\phi$. The buffer trace $\sigma_{buff}$ in the initial call will contain a single element $\sigma_{buff}=\{(s_0,t_0)\}$, representing the initial state and time of simulation. Finally, for time-unbounded formulae, we adopt the following *pessimistic* approach: if the simulation max time $t_{max}$ is reached and the formula is neither verified nor falsified then the algorithm returns $\mathit{false}$. Thus the exact probability of time-unbounded formulae is an upper bound of the estimated one.
Estimating the probability of a property {#sec:EstimateProb}
----------------------------------------
Checking of a BLTLc property on a simulated trace corresponds to a Bernoulli experiment, where the outcome can be either positive or negative. Thus the number of successes that results from reiterated checking of the same property on $n$ independent simulations, represents a random variable $X$ with a Binomial distribution. The point estimation of the unknown probability of success $p$ out of $n$ independent trials, is given by the well known maximum likelihood estimator $\hat{p}=po/n$, where $po$ represents the number of successes. Clearly the reliability of such estimate is highly affected by the number of simulations performed, i.e. by the sample size $n$. As a consequence the point estimate $\hat{p}$ is usually associated with a confidence interval, expressed in terms of a real value $\alpha{\in}(0,1)$, which represents the range within which the actual value of the unknown parameter $\theta$ (i.e. the actual probability of the considered formula to hold against the simulated model) shall fall $(1-\alpha)\%$ times[^1].
The standard approach to compute the confidence interval for the probability of success of a binomial distribution uses the normal approximation, producing the so-called Wald interval. The hypothesis-testing-based statistical model checking approach of Younes *et al.* [@YKNP06], for example, uses the Wald interval method for validating the hypothesis that the probability of a certain CSL formula is below (above) a given threshold.
Brown *et al.* [@Brown01intervalestimation],[@Brown02confidenceintervals], have studied the coverage characteristics of different types of binomial proportion confidence intervals, and they showed that the Wald interval present unstable coverage characteristics also for large $n$, suggesting thus the use of other types of confidence intervals. Among the interval discussed, the Wilson score interval [@wilson27] have shown good coverage characteristics also for small $n$ and extreme probabilities. Wilson score confidence interval is calculated by means of $Wilson\_interval(\hat{p},n,\alpha)=[L ,U]$ with, $$\label{eq:wilson_limits}
[L, U]\!=\!\frac{\hat{p}+\frac{1}{2n}z^2_{1-\alpha/2}\mp z_{1-\alpha/2} \sqrt{\frac{\hat{p}(1-\hat{p})}{n}+\frac{z^2_{1-\alpha/2}}{4n^2}}}{1+\frac{1}{n}z^2_{1-\alpha/2}}$$ where $\hat{p}$ is the estimated probability from the statistical sample, $\alpha$ is the confidence level, $z_{1-\alpha/2}$ is the $1-\alpha/2$ percentile of a standard normal distribution, and $n$ is the sample size. As the confidence interval is a function of $n$ we may reverse the problem and ask which is the proper sample size to obtain a confidence interval of a given width at a specific confidence level $\alpha$. This is extremely useful for establishing how long a simulative experiments has to be (i.e. how many runs are needed) in order to meet a desired reliability for the measure we estimate.
### An iterative algorithm for sample size determination {#sec:Wilson}
The sample size required for the Wilson interval of width $2\epsilon$ at $1-\alpha$ confidence level can be obtained simply by solving for $n$ the Wilson score limits in equation (\[eq:wilson\_limits\]) [@Piegorsch2004]: formally this is given by function $Wilson\_sample(p,\epsilon,\alpha)=N$ with, $$\label{eq:wilson_sample}
N \geq z^2_{1-\alpha/2}\frac{\hat{p}(1-\hat{p}) - 2\epsilon^2+ \sqrt{\hat{p}^2(1-\hat{p})^2 + 4\epsilon^2(\hat{p}-0.5)^2}}{2\epsilon^2}$$ where $\hat{p}$ is the frequency of positive outcomes of a re-iterated Bernoulli experiment. As we usually do not have a guess of the probability $\hat{p}$ to be used in (\[eq:wilson\_sample\]), the standard approach is to take a conservative estimate, by considering $\hat{p}=0.5$ which is the estimate with maximum variance, and, as such, produces the highest sample size.
The method we present here consists in adopting a different approach in determining the sample size. By iterating (\[eq:wilson\_sample\]) with successive estimates of $\hat{p}$ we are able to drastically reduce the number of samples required when the true probability $p$ is far from 0.5. More specifically, given a confidence interval width 2$\epsilon$ and a confidence level $1-\alpha$, the algorithm starts by calculating the sample size required for an initial estimate $\hat{p}=1$ (or equivalently $\hat{p}=0$) and returns the minimum number $N$ of simulations to be performed. After computing the proportion of successes the new estimate $\hat{p}$ is rounded by adding or subtracting the quantity $\epsilon$ if $\hat{p}\leq0.5$ or $\hat{p}>0.5$ respectively. The rounded estimate $p^\prime$ is then used to recalculate the sample size resulting in $N^\prime$. If we have already performed a cumulative $N_{tot}\geq N^\prime$ simulations the algorithm stops. Conversely, we iterate the process again by launching $N^\prime-N_{tot}$ simulations.
%[numberfirstline=false,numbers=left]
Wilson Procedure
Ntot = 0;
N = Wilson_Sample(1, epsilon, alpha);
Perform N experiments;
Ntot = Ntot + N; pest = yess;
if pest <= 0.5
pprime = pest + epsilon;
else pprime = pest - epsilon;
Nprime = Wilson_sample(pprime, epsilon, alpha); Nnew = Nprime - Ntot;
if Nnew > 0
N = Nnew; goto 3;
else return (pest, Wilson_interval(pest, Ntot, alpha));
The $\hat{p}$ rounding step is crucial and ensures that a successive sample size calculation would avoid undersized samples due to erratic estimates. If the current estimated probability of success drifts from the true unknown one, towards extreme probabilities, of more than $\epsilon$, this would produce an undersized sample which would produce a confidence interval not covering the parameter at $1-\alpha$ level.
### Performance of the iterative sample size determination
The iterative method for the determination of sample size for the Wilson score interval drastically reduces the number of required samples with respect to the conservative approach that starts with $\hat{p}=0.5$. Of course this gain is greater when the actual $p$ is close to extreme values $p=0$ and $p=1$, while using the same sample size for $p$ close to 0.5.
![Average sample size for CI-width$=0.05$ at $99\%$ confidence lelvel. Comparison of sample size required with i) conservative approach, ii) Iterative Wilson, iii) Minimum sample size with known $p$[]{data-label="fig:ComparisonSampleSize"}](PIC/SampleSize2.png){width="4in"}
Figure \[fig:ComparisonSampleSize\] represents the sample size required to obtain a confidence interval of a given width $2\epsilon=0.05$ at a confidence level $1-\alpha=0.99$ for values of $p$ ranging from 0 to 1 by 0.1. Bars from light gray to dark gray represent respectively: **I**. Conservative approach (using always $\hat{p}$=0.5); **II**. Iterative Wilson; **III**. Minimum sample size if the unknown $p$ would be known (i.e. $\hat{p}=p$). On the right side of the plot we reported the percentage of reduction in sample size by using the Iterative Wilson approach with respect to the Conservative approach. As it can be easily seen, we have the strongest reduction in sample size when the true $p$ is close to 0 or 1. By using the Iterative Wilson method we presented here, we can reduce the sample size required by up to 88% with respect to the Conservative approach. This reduction at extreme values is explained by the fact that at those probabilities the variance of the binomial distribution is smaller, so we need less samples to obtain good estimates.
Prototype architecture {#sec:tool}
======================
What so far illustrated has been implemented within a distributed software architecture, capable to get the better out of the MRiP computational policy (refer to [@ballarini2009BIB] for a wide overview of this topic). It is actually made of two distinct modules: a graphical front-end and a remote simulation engine. The front-end part acts as a server and is in charge of drawing the computational graph relative to the loaded [BlenX]{} [@dematte2008blenx] models. [BlenX]{} is a new programming language intended to develop executable biological models starting from the composition of the description of the molecules involved in the system. The prototype collects any information from a [BlenX]{} model and serializes them in a proprietary, xml-based, data format along with all the simulation information manually inputted by the user (see Figure \[fig:statediagram\] - State `A`). Further information about the logical formula to be checked, the $\alpha$ and $\epsilon$ values are required only whenever one wants to automatically calculate the number of replicated simulations (or “replicas") needed to reach the required confidence threshold. However, both in the case that the number of replicas is user-defined and that it is automatically computed, a random number generator is instantiated and used to make a stream of initial seeds, one for each simulation (see Figure \[fig:statediagram\] - State `B`).
As soon as the simulation task is invoked by the user, a number of independent simulation engines is instantiated (see Figure \[fig:activitydiagram\] - Activity `A`). Among them, one is entitled to be master. The master handles both the inter-process and the client-server communications. In the former case, it takes care of scattering and dispatching the initial seeds to the slave processes (and to itself) and of gathering the results (see Figure \[fig:activitydiagram\] - Activities `C`). `MPI` is broadly used for these purposes. In the latter case, the master node is responsible for counting the computed `YES` and for its sending to the server (see Figure \[fig:activitydiagram\] - Activities `C`). This is accomplished by means of a socket-based interface. Hence, each process simulates independently (see Figure \[fig:statediagram\] - State `C` and Figure \[fig:activitydiagram\] - Activities $C$) and evaluates on-the-fly a logical formula, giving a boolean answer. The summation of the positive answers is sent to the server, which recomputes the Wilson method and returns a new number of simulation replicas to be performed (see Figure \[fig:statediagram\] - State `D` and Figure \[fig:activitydiagram\] - Activity `B`). This loop halts only when no more replicas are required to be performed.
Policies of communications
--------------------------
The prototype initializes as many independent processes as the replicas number estimated by the Wilson procedure. In agreement with the SIMD (Single Instruction, Multiple Data) computational policy, they share the same code area, namely they execute the same instructions in a parallel fashion on different instances of the same input model. In particular, two important tasks are executed in parallel: the import of the input model and the actual simulation. The former task consists just in loading a [BlenX]{} model and in extracting some important information from it. This can happen essentially because all the processes address a chunk of shared filesystem area where the input model lies in. After that, every process starts its independent simulation according to the parameters entered by the user.
Therefore, we set two synchronization points. One occurs before starting the simulation task, exactly when the starting random seeds are computed. Actually, this would not be strictly necessary because, in principle, every process could generate its own starting random seed. But, to tackle the problem of avoiding correlation among the individual trajectories of random numbers [@tian2005parallel], we empowered the master process (the process with the lowest `MPI rank`) to generate as many random numbers as the number of processes and to split and distribute it to each other process (through the `MPI_Scatter` procedure). Contrarily, a very critical synchronization point concerns the Wilson step. It occurs whenever all the processes end their simulation and must communicate their `YES` number as well as must write the simulation traces to filesystem. Here, two kind of communications take place. One happens among the processes themselves. Anyone send the master its computed simulation trace along with a boolean answer to the formula (via the `MPI_Gather` procedure), both encapsulated in an ad-hoc serializable data structure. Thus, the master node prints the traces to file and computes the probability $\hat{p}$, as described before. Finally, it sends back $\hat{p}$ to the server side via a private socket for the Wilson re-computation.
Experiments and Performances {#sec:Experiments}
============================
We consider a stochastic model of (a part) of the regulatory network that controls the buddying yeast cell cycle [@novakTyson] and we present some preliminary experiments. We adopt the following protocol: first we identify few BLTLc formulae characterising relevant aspects of the cell-cycle behaviour. We then run our statistical verification tool to estimate the probability of the considered formulae. To assess the accuracy of the statistical procedure we compare the estimates obtained through our statistical model checker with the exact values calculated through numerical model checking, namely by means of the PRISM model checker [@KNP09a]. As the state-space dimension corresponding to the original cell-cycle model is too large to be handled through numerical model checkers we consider a “scaled-down” version of the model for validating the statistical model checkin approach against the numerical one.
Finally, for assessing the performances of the improved Wilson method we compare the number of simulations performed by running every experiment twice: once with the initial point estimate $\alpha=1$, and then, following the original “conservative" approach, with $\alpha=0.5$.
The cell cycle is the a concatenation of biochemical and morphological events that lead to the reproduction (duplication) of a cell. The “standard” model considers a loop of four phases: *G1*, growth and preparation of the chromosomes for replication; *S*, synthesis and duplication of DNA; *G2*, synthesis of significant protein for mitosis; *M*, mitosis, i.e., cell division. The cell cycle is regulated by a network of biochemical reactions centered around complexes of cyclin dependent kinases (Cdk’s) and their regulatory partners. Active complex Cdk/CycB induces cell cycle phase changes by activating or inhibiting target molecules [@Mor95; @Mor06].
The simplified model we consider is sketched in Figure \[subfig:modelcartoon\]. It consists of three species, $x$ (Cdk/CycB complex), $y$ (activated APC/Cdh1 complex) and $a$ (activated Cdc20), and nine molecular reactions listed in Table \[tab:modelreactions\] and with parameters in Table \[tab:ParameterValuesCellCycleToyModel\].
![Budding yeast cell-cycle cartoon and simulation[]{data-label="fig:modelcartoon"}](PIC/modelcartoon.png "fig:") ![Budding yeast cell-cycle cartoon and simulation[]{data-label="fig:modelcartoon"}](PIC/simulation.pdf "fig:")
Complex $x$ is synthesized and degraded by reactions $R_{1x}$ and $R_{2x}$, respectively. Complex $y$ speeds up $x$ production by means of reaction $R_{3x}$. At the same time, $x$ deactivates $y$ by $R_{3y}$. Also, $y$ turns active by itself with $R_{1y}$ and with the help of $a$ in reaction $R_{2y}$. Finally, $a$ is produced and consumed by $R_{1a}$ and $R_{2a}$ and regulated by $x$ in reaction $R_{2a}$. Such system behaves has a bistable switch with two stable states: with low $x$ and high $y$, and *S*/*G2*/*M* with high $x$ and low $y$, as shown in the plot depicting a typical simulation outcome in Figure \[subfig:simulation\].
-- ------------- --------------------------------- ------------- ---------- -------------- --------------------------- -------------- ---------- ------------- ----------------------------- -------------
$\emptyset$ $\xrightarrow{k_1 \alpha}$ $x$ $R_{1y}$ $y_{in}$ $\xrightarrow{k_3^*}$ $y$ $R_{1a}$ $\emptyset$ $\xrightarrow{k'_5 \alpha}$ $a$
$x$ $\xrightarrow{k_2'}$ $\emptyset$ $R_{2y}$ $y_{in} + a$ $\xrightarrow{k^{'''}_3}$ $y + a$ $R_{2a}$ $x$ $\xrightarrow{k^*_5}$ $x + a$
$x + y$ $\xrightarrow{k^{''}_2 \alpha}$ $y$ $R_{3y}$ $x + y$ $\xrightarrow{k^{*}_4}$ $x + y_{in}$ $R_{3a}$ $a$ $\xrightarrow{k_6}$ $\emptyset$
-- ------------- --------------------------------- ------------- ---------- -------------- --------------------------- -------------- ---------- ------------- ----------------------------- -------------
: Budding yeast cell-cycle reactions[]{data-label="tab:modelreactions"}
[**Component**]{} [**Rate Constant**]{} [**Dimensionless constants**]{}
------------------- ----------------------------------------------------------------------------------------------------------------------------------- ---------------------------------
Cdk/CycB $k_1=0.04$, $k'_2=0.04$, $k''_2=1$, $k'''_2=1$
APC/Cdh1 $k'_3=1$, $k''_3=10$, $k'_4=2$, $k_4=35$ $J_3 = 0.04, J_4 = 0.04$
$k^{*}_3 = \frac{k_4 m \alpha x y}{J_4 + ( \alpha y) }$ $k^{'''}_3 = \frac{ k^{''}_3 \alpha a y_{in} }{J_3 + ( \alpha y_{in}}$ $J_5 = 0.3$
$k^*_4 = \frac{k'_3 y_{in}}{ J3 + ( \alpha y_{in})}$ $m = 0.80 $
Cdc20 $k'_5=0.005$, $k''_5=0.2$, $k_6=0.1$, $k_4=35$ $\alpha = 0.00236012$
$k^*_5 = \frac{k^{''}}{\alpha} / \frac{J_5}{m \alpha x} $
: Parameter Values: Cell Cycle toy model[]{data-label="tab:ParameterValuesCellCycleToyModel"}
#### Studying the role of Cdc20 in the *S*/*G2*/*M* transition.
We target the experiments of this section to the study of the so-called *S*/*G2*/*M* transition which begins in states with low level of activated APC, high concentration of Cdk/CycB and (initially) low level of Cdc20. By looking at the topology of the network in Figure \[subfig:modelcartoon\], and at the form of the corresponding equations (Table \[tab:modelreactions\]), it is evident that $a$ (i.e. Cdc20) plays a fundamental part in the activation of $y$ hence in the controlling the *S*/*G2*/*M* transition. Specifically the progressive growth of $a$ results in the (initially slow) activation of $y$ which then, in turns, is responsible for the degradation of $x$. The influence of $a$ on $y$ can be studied through BLTLc formulae of the following type: $$\phi_1\equiv (a \leq i)\ U\ (y\geq j), \hspace{4ex}
\phi_2\equiv (a \leq i)\ U^{\leq t} \ (y\geq j)$$ Formula $\phi_1$ represents the possibility that $y$ grows above threshold $j$ while $a$ does not exceed threshold $i$. Since $y$ gets abruptly activated only after $a$ has reached high concentration (see Figure \[subfig:simulation\]) then, for $i<j$ and $\delta=j-i$, we expect a low probability of $\phi_1$ for large $\delta$, and a higher probability of $\phi_1$ for high $i$ and small $\delta$[^2].
Figure \[fig:experiment1\] compares exact versus estimated probability measure for the time-bounded formula $\phi_2$, verified with respect to different time points ($t\in[0.2,1.6]$ step $0.2$). The (cross marked) point estimates (depicted in Figure \[fig:experiment1\] together with their confidence interval), have been calculated with $99.99\%$ confidence and $0.005$ interval semi-amplitude ($\epsilon=0.005$). The exact values computed with PRISM (red plot in Figure \[fig:experiment1\]) fall within the confidence interval of each point estimates, confirming the accuracy of the statistical verification method we have realized.
Further data regarding application of statistical verification to formulae $\phi_1$ and $\phi_2$ (and variants) are reported in Table \[table:experiments\]. The calculated estimate ($\hat{p}$), confidence interval ($[L,U]$) corresponding to the chosen confidence level ($conf$) and interval width ($\epsilon$) are depicted together with the exact measure of probability $p$ (calculated with PRISM). For each experiment we also calculate the sample size ($N$) as well as the average path length $\overline{|\sigma|}$. Results in Table \[table:experiments\] indicate a good accuracy of the obtained statistics and confirm the performance gain (i.e. in terms of sample size) of the sequential algorithm of Figure \[fig:wilson\] versus the conservative approach corresponding to $\alpha=0.5$. As argued before such gain is greater with a larger distance of the estimate from the median.
![Exact vs Estimated probability of the time-bounded Until formula: $(a \leq 4) U^{(0,t]} (y\geq 5)$, estimates with $99.99\%$ confidence and $\epsilon=0.005$ semi-interval amplitude.[]{data-label="fig:experiment1"}](PIC/timeUntilExperiment1.pdf)
formula $\hat{p}$ $[L,U]$ conf $\epsilon$ $\alpha$ $p$ $N$ $\overline{|\sigma|}$
-------------------------------------- ----------- ---------------------- ------ ------------ ---------- --------- ------ -----------------------
$(a \leq 4)$ U $(y\geq 5)$ 0.10384 \[0.080518,0.12700\] 99 $0.025$ 1 0.10458 1124 84.74
$(a \leq 4)$ U $(y\geq 5)$ 0,09554 \[0,081823,0,11128\] 99 $0.025$ 0.5 0.10458 2648 84,61
$(a \leq 20)$ U $(y\geq 35)$ 0.00534 \[0.080518,0.12700\] 99 $0.025$ 1 0.00571 187 1274.65
$(a \leq 20)$ U $(y\geq 35)$ 0,00717 \[0,00401,0,01280\] 99 $0.025$ 0.5 0.00571 2648 1296,72
$(a \leq 40)$ U $(40\leq y\leq 42)$ 0.62004 \[0.59472,0.64472\] 99 $0.025$ 1 0.6312 2495 4597.67
$(a \leq 40)$ U $ (40\leq y\leq 42)$ 0.63085 \[0.60064,0.65064 \] 99 $0.025$ 0.5 0.6312 2648 4721,63
: Estimated probability of cell-cycle properties vs numerical evaluation.
\[table:experiments\]
Conclusions {#sec:Conclusion}
===========
The stochastic modeling framework have been demonstrated very important in systems biology. Unfortunately the complexity of living systems most often results in models which are so large that methods based on numerical approaches such as, for example, transient/steady-state analysis and/or *exact* (stochastic) model checking are simply not feasible. For this reason, stochastic models of biological systems are usually studied by means of simulation-based approaches. In this paper we have presented a statistical model checking approach targeted to the verification of biological models. A novel temporal logic language, namely the BLTLc language, has been introduced: it allows to formally capture complex features of a biological system’s dynamics. The methodology we proposed employes a statistical engine to estimate the probability of a BLTLc formula to hold true of a stochastic model $M$. Such estimates is obtained by *on-the-fly* verification of the considered BTLTc formula $\phi$ against *simulated executions* of the $M$. The resulting estimates is given by the frequency of the positive answers (i.e. the number of TRUEs resulting from the verification of $\phi$ against each simulation trace) out of the total of the simulated traces. The statistical engine we introduced is based on an efficient variant of the so-called Wilson score-interval method, which improves the performances, i.e. it requires a smaller number of trajectories in order to meet the chosen confidence level, with respect to more popular statistical engines such as those based on the so-called Wald-interval method. We have implemented the proposed algorithms for analysis and statistical testing as a prototype tool, which employs MPI technology to distribute verification engines. The current implementation exploits multi-core processors, in particular, our tests have been performed on a quad core machine Intel Q9300 CPU with 4G of RAM under Widows XP. A cluster and a GRID version of our tool is under development, that will maximize the parallelism of the proposed methodology. Moreover we are planning a complete suite of performance tests against tools with similar features, as those stated above.
**Related work.** Techniques for the verification of temporal logic property against probabilistic/stochastic models, can be either *exact* or *approximate*. Exact approaches work by constructing a complete representation of a finite state space model and, because of this, their application to complex systems is unfeasible. PRISM [@KNP09a] and MRMC [@KatoenKZ_QEST05] are two popular probabilistic model checking tools that support both exact and approximated CSL verification. Approximated verification can be one of two different types. If the considered problem is to establishing whether the likelihood $p$ of a formula is $p \trianglelefteq b$ where $b{\in}[0,1]$ is a threshold and $\trianglelefteq\{<,\leq,\geq,>\}$ (i.e. *model checking problem*) then the outcome of verification is boolean and is determined based on Hypothesis testing. On the other hand if the problem is one of determining an estimate for $p$ then this is achieved through confidence interval based techniques. PRISM approximated verification belong to the latter type: the size of the sample is determined statically as a function of the chosen level of confidence and the desired approximation, rather than being calculated iteratively as function of intermediate estimates, as is the case with our method, whereas paths generation is controlled by on-the-fly checking of the considered formula. Furthermore although PRISM has been recently added with support for (exact) probabilistic LTL model checking, at the best of our knowledge, it currently supports statistical verification only for CSL (and not for LTL). The YMER [@YKNP06] and MRMC tool, on other hand, features approximated (hypothesis testing based) model checking which uses on-the-fly verification of sampled path in order to decide whether the probability of formula is below/above a threshold. The Monte Carlo Model Checker MC2(PLTLc) [@Donaldson08] computes a point estimate of a Probabilistic LTL logic (with numerical constraints) formula to hold of model. MC2(PLTLc) does not include any simulation engine but works *offline* by taking a set of sampled trajectories generated by any simulation or ODE solver software. Besides MC2(PLTLc) calculates also the probabilistic domain of satisfaction for any free variable of PLTLc formula. Finally the APMC tool [@DBLP:conf/qest/HeraultLP06] features confidence interval based estimates of the probability of Probabilistic LTL and PCTL formulae to hold of either DTMC and CTMC models.\
**Acknowledgments.** The authors would like to thank Alida Palmisano for her valuable advises.
[^1]: There exists a strong connection between confidence intervals and hypothesis testing: all the values $\theta_0$ for the unknown parameter $\theta$ external to a $1-\alpha$ confidence interval would end in the rejection of the two sided hypothesis testing (i.e. Null Hypothesis $H_0: \theta=\theta_0$) at the $\alpha$ level.
[^2]: $\phi_2$ allows also to study the dependence on time of such an attitude.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
We analyze linkage strategies for a set ${{{\mathcal{I}}}}$ of webpages for which the webmaster wants to maximize the sum of Google’s PageRank scores. The webmaster can only choose the hyperlinks *starting* from the webpages of ${{{\mathcal{I}}}}$ and has no control on the hyperlinks from other webpages. We provide an optimal linkage strategy under some reasonable assumptions.
*Keywords:* PageRank, Google matrix, Markov chain, Perron vector, Optimal linkage strategy
*AMS classification:* 15A18, 15A48, 15A51, 60J15, 68U35
author:
- Cristobald de Kerchove
- Laure Ninove
- Paul Van Dooren
title: Maximizing PageRank via outlinks
---
Introduction
============
PageRank, a measure of webpages’ relevance introduced by Brin and Page, is at the heart of the well known search engine Google [@BP98; @PBMW98]. Google classifies the webpages according to the pertinence scores given by PageRank, which are computed from the graph structure of the Web. A page with a high PageRank will appear among the first items in the list of pages corresponding to a particular query.
If we look at the popularity of Google, it is not surprising that some webmasters want to increase the PageRank of their webpages in order to get more visits from websurfers to their website. Since PageRank is based on the link structure of the Web, it is therefore useful to understand how addition or deletion of hyperlinks influence it.
Mathematical analysis of PageRank’s sensitivity with respect to perturbations of the matrix describing the webgraph is a topical subject of interest (see for instance [@AL06; @BGS05; @Kir06; @LM04; @LM06book; @LM05] and the references therein). Normwise and componentwise conditioning bounds [@Kir06] as well as the derivative [@LM04; @LM06book] are used to understand the sensitivity of the PageRank vector. It appears that the PageRank vector is relatively insensitive to small changes in the graph structure, at least when these changes concern webpages with a low PageRank score [@BGS05; @LM04]. One could think therefore that trying to modify its PageRank via changes in the link structure of the Web is a waste of time. However, what is important for webmasters is not the values of the PageRank vector but the *ranking* that ensues from it. Lempel and Morel [@LM05] showed that PageRank is not rank-stable, i.e. small modifications in the link structure of the webgraph may cause dramatic changes in the ranking of the webpages. Therefore, the question of how the PageRank of a particular page or set of pages could be increased–even slightly–by adding or removing links to the webgraph remains of interest.
As it is well known [@AL04; @IW06], if a hyperlink from a page $i$ to a page $j$ is added, without no other modification in the Web, then the PageRank of $j$ will increase. But in general, you do not have control on the *inlinks* of your webpage unless you pay another webmaster to add a hyperlink from his/her page to your or you make an *alliance* with him/her by trading a link for a link [@BYCL05; @GGM05]. But it is natural to ask how you could modify your PageRank by yourself. This leads to analyze how the choice of the *outlinks* of a page can influence its own PageRank. Sydow [@Syd05] showed via numerical simulations that adding well chosen outlinks to a webpage may increase significantly its PageRank ranking. Avrachenkov and Litvak [@AL06] analyzed theoretically the possible effect of new outlinks on the PageRank of a page and its neighbors. Supposing that a webpage has control only on its outlinks, they gave the optimal linkage strategy for this single page. Bianchini et al. [@BGS05] as well as Avrachenkov and Litvak in [@AL04] consider the impact of links between web communities (websites or sets of related webpages), respectively on the sum of the PageRanks and on the individual PageRank scores of the pages of some community. They give general rules in order to have a PageRank as high as possible but they do not provide an optimal link structure for a website.
Our aim in this paper is to find a generalization of Avrachenkov–Litvak’s optimal linkage strategy [@AL06] to the case of *a website with several pages*. [We consider a given set of pages and suppose we have only control on the *outlinks* of these pages. We are interested in the problem of *maximizing the sum of the PageRanks* of these pages.]{}
Suppose $\mathcal{G}=({\mathcal{N}},{\mathcal{E}})$ be the webgraph, with a set of nodes ${\mathcal{N}}={\{1,\dots,n\}}$ and a set of links ${\mathcal{E}}\subseteq{\mathcal{N}}\times {\mathcal{N}}$. For a subset of nodes ${{{\mathcal{I}}}}\subseteq{\mathcal{N}}$, we define $$\begin{aligned}
{\mathcal{E}}_{{{\mathcal{I}}}}&={\{(i,j)\in{\mathcal{E}}\colon i,j\in{{{\mathcal{I}}}}\}} \text{ the set of internal links},\\
{\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}&={\{(i,j)\in{\mathcal{E}}\colon i\in{{{\mathcal{I}}}},j\notin{{{\mathcal{I}}}}\}}
\text{ the set of external outlinks},\\
{\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}&={\{(i,j)\in{\mathcal{E}}\colon i\notin{{{\mathcal{I}}}},j\in{{{\mathcal{I}}}}\}}
\text{ the set of external inlinks},\\
{\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}&={\{(i,j)\in{\mathcal{E}}\colon i,j\notin{{{\mathcal{I}}}}\}} \text{ the set of external links}.\end{aligned}$$
[If we do not impose any condition on ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$, the problem of maximizing the sum of the PageRanks of pages of ${{{\mathcal{I}}}}$ is quite trivial and does not have much interest (see the discussion in Section \[sec:optimal-linkage-strategy\]). Therefore, when characterizing optimal link structures, we will make the following *accessibility assumption*: every page of the website must have an access to the rest of the Web.]{}
Our first main result concerns the *optimal outlink structure* for a given website. In the case where the subgraph corresponding to the website is strongly connected, Theorem \[thm:out-opt\] can be particularized as follows.
Let ${\mathcal{E}}_{{{\mathcal{I}}}}$, ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given. Suppose that the subgraph $({{{\mathcal{I}}}},{\mathcal{E}}_{{{\mathcal{I}}}})$ is strongly connected and ${\mathcal{E}}_{{{\mathcal{I}}}}\neq\emptyset$. Then every optimal outlink structure ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ is to have only one outlink to a particular page outside of ${{{\mathcal{I}}}}$.
We [are]{} also [interested in]{} the optimal *internal* link structure for a website. In the case where there is a unique leaking node in the website, that is only one node linking to the rest of the web, Theorem \[thm:intr-opt\] can be particularized as follows.
Let ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$, ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given. Suppose that there is only one leaking node in ${{{\mathcal{I}}}}$. Then every optimal internal link structure ${\mathcal{E}}_{{{\mathcal{I}}}}$ is composed of together with every possible backward link.
Putting together Theorems \[thm:out-opt\] and \[thm:intr-opt\], we get in Theorem \[thm:website-opt\] the *optimal link structure* for a website. This optimal structure is illustrated in Figure \[fig:optimal-link-structure\].
Let ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given. Then, for every optimal link structure, ${\mathcal{E}}_{{{\mathcal{I}}}}$ is composed of a forward chain of links together with every possible backward link, and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ consists of a unique outlink, starting from the last node of the chain.
![ Every optimal linkage strategy for a set ${{{\mathcal{I}}}}$ of five pages must have this structure.[]{data-label="fig:optimal-link-structure"}](figuresopts.4)
This paper is organized as follows. In the following preliminary section, we recall some graph concepts as well as the definition of the PageRank, and we introduce some notations. In Section \[sec:PR-set\], we develop tools for analysing the PageRank of a set of pages ${{{\mathcal{I}}}}$. Then we come to the main part of this paper: in Section \[sec:optimal-linkage-strategy\] we provide the optimal linkage strategy for a set of nodes. In Section \[sec:extensions-variants\], we give some extensions and variants of the main theorems. We end this paper with some concluding remarks.
Graphs and PageRank {#sec:preliminaries}
===================
Let $\mathcal G=({\mathcal{N}},{\mathcal{E}})$ be a directed graph representing the Web. The webpages are represented by the set of nodes ${\mathcal{N}}={\{1,\dots,n\}}$ and the hyperlinks are represented by the set of directed links ${\mathcal{E}}\subseteq{\mathcal{N}}\times {\mathcal{N}}$. That means that $(i,j)\in{\mathcal{E}}$ if and only if there exists a hyperlink linking page $i$ to page $j$.
Let us first briefly recall some usual concepts about directed graphs (see for instance [@BP94]). A link $(i,j)$ is said to be an [*outlink*]{} for node $i$ and an [*inlink*]{} for node $j$. If $(i,j)\in{\mathcal{E}}$, node $i$ is called a [*parent*]{} of node $j$. By $$j{\gets{i}},$$ we mean that $j$ belongs to the set of [*children*]{} of $i$, that is $j\in{\{k\in{\mathcal{N}}\colon(i,k)\in{\mathcal{E}}\}}$. The [*outdegree*]{} ${d_{i}}$ of a node $i$ is its number of children, that is $${d_{i}}={|{\{j\in{\mathcal{N}}\colon (i,j)\in{\mathcal{E}}\}}|}.$$ A [*path*]{} from $i_0$ to $i_s$ is a sequence of nodes $\langle i_0,i_1,\dots,i_s\rangle$ such that $(i_k,i_{k+1})\in{\mathcal{E}}$ for every $k=0,1,\dots,s-1$. A node $i$ [*has an access to a node*]{} $j$ if there exists a path from $i$ to $j$. In this paper, we will also say that a node $i$ [*has an access to a set*]{} ${\mathcal{J}}$ if $i$ has an access to at least one node $j\in{\mathcal{J}}$. The graph ${\mathcal{G}}$ is *strongly connected* if every node [of]{} ${\mathcal{N}}$ has an access to every other node [of]{} ${\mathcal{N}}$. [A set of nodes ${\mathcal{F}}\subseteq{\mathcal{N}}$ is a [*final class*]{} of the graph $\mathcal G=({\mathcal{N}},{\mathcal{E}})$ if the subgraph $({\mathcal{F}},{\mathcal{E}}_{{\mathcal{F}}})$ is strongly connected and moreover ${\mathcal{E}}_{\mathrm{out}({\mathcal{F}})}=\emptyset$ (i.e. nodes of ${\mathcal{F}}$ do not have an access to ${\mathcal{N}}\setminus{\mathcal{F}}$).]{}
Let us now briefly introduce the PageRank score (see [@BGS05; @BP98; @LM04; @LM06book; @PBMW98] for background). Without loss of generality (please refer to the book of Langville and Meyer [@LM06book] or the survey of Bianchini et al. [@BGS05] for details), we can make the assumption that *each node has at least one outlink*, i.e. ${d_{i}}\neq0 \text{ for every } i\in{\mathcal{N}}$. Therefore the $n\times n$ stochastic matrix $P=[P_{ij}]_{i,j\in{\mathcal{N}}}$ given by $$P_{ij}=\begin{cases}
{{d_{i}}}^{-1} &\text{if }(i,j)\in{\mathcal{E}},\\
0 &\text{otherwise},
\end{cases}$$ is well defined and is a scaling of the adjacency matrix of $\mathcal G$. Let also $0<c<1$ be a [*damping factor*]{} and ${\boldsymbol{z}}$ be a positive stochastic [*personalization vector*]{}, i.e. [${\boldsymbol{z}}_i>0$ for all $i=1,\dots,n$ and]{} ${\boldsymbol{z}}^T{{\boldsymbol{1}}}=1$, where ${{\boldsymbol{1}}}$ denotes the vector of all ones. The [*Google matrix*]{} is then defined as $$G=cP+(1-c){{\boldsymbol{1}}}{\boldsymbol{z}}^T.$$ Since ${\boldsymbol{z}}>0$ and $c<1$, this stochastic matrix is positive, [i.e. $G_{ij}>0$ for all $i,j$]{}. The [*PageRank vector*]{} ${{\boldsymbol{\pi}}}$ is then defined as the unique invariant measure of the matrix $G$, [that is]{} the unique left Perron vector of $G$, $$\label{eq:PR-definition}
\begin{split}
{{\boldsymbol{\pi}}}^T &={{\boldsymbol{\pi}}}^TG,\\
{{\boldsymbol{\pi}}}^T{{\boldsymbol{1}}}&=1.
\end{split}$$ The [*PageRank of a node*]{} $i$ is the $i^\text{th}$ entry ${{\boldsymbol{\pi}}}_i={{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_i$ of the PageRank vector.
The PageRank vector is usually interpreted as the stationary distribution of the following Markov chain (see for instance [@LM06book])[: a]{} random surfer moves on the webgraph, using hyperlinks between pages with a probability $c$ and [*zapping*]{} to some new page according to the personalization vector with a probability $(1-c)$. The Google matrix $G$ is the probability transition matrix of this random walk. In this stochastic interpretation, the PageRank of a node is equal to the inverse of its mean return time, that is [${{\boldsymbol{\pi}}}_i^{-1}$]{} is the mean number of steps a random surfer starting in node $i$ will take for coming back to $i$ (see [@CM00; @KS60]).
PageRank of a website {#sec:PR-set}
=====================
We are interested in characterizing the [*PageRank of a set*]{} ${{{\mathcal{I}}}}$. We define this as the sum $${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}=\sum_{i\in{{{\mathcal{I}}}}}{{\boldsymbol{\pi}}}_i,$$ where ${\boldsymbol{e}}_{{{\mathcal{I}}}}$ denotes the vector with a $1$ in the entries of ${{{\mathcal{I}}}}$ and $0$ elsewhere. Note that the PageRank of a set corresponds to the notion of energy of a community in [@BGS05].
Let ${{{\mathcal{I}}}}\subseteq{\mathcal{N}}$ be a subset of the nodes of the graph. The PageRank of ${{{\mathcal{I}}}}$ can be expressed as ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}=(1-c){\boldsymbol{z}}^T(I-cP)^{-1}{\boldsymbol{e}}_{{{\mathcal{I}}}}$ from PageRank equations . Let us then define the vector $$\label{eq:v-definition} {\boldsymbol{v}} = (I-cP)^{-1}{\boldsymbol{e}}_{{{\mathcal{I}}}}.$$ With this, we have the following expression for the PageRank of the set ${{{\mathcal{I}}}}$: $$\label{eq:PR-zv} {{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}=(1-c){\boldsymbol{z}}^T{\boldsymbol{v}}.$$
The vector ${\boldsymbol{v}}$ will play a crucial role throughout this paper. In this section, we will first present a probabilistic interpretation for this vector and prove some of its properties. We will then show how it can be used in order to analyze the influence of some page $i\in{{{\mathcal{I}}}}$ on the PageRank of the set ${{{\mathcal{I}}}}$. We will end this section by briefly introducing the concept of [basic]{} absorbing graph, which will be useful in order to analyze optimal linkage strategies under some assumptions.
Mean number of visits before zapping
------------------------------------
Let us first see how the entries of the vector ${\boldsymbol{v}}=(I-cP)^{-1}{\boldsymbol{e}}_{{{\mathcal{I}}}}$ can be interpreted. Let us consider a random surfer on the webgraph ${\mathcal{G}}$ that, as described in Section \[sec:preliminaries\], follows the hyperlinks of the webgraph with a probability $c$. But, instead of zapping to some page of ${\mathcal{G}}$ with a probability $(1-c)$, he *stops* his walk with probability $(1-c)$ at each step of time. This is equivalent to consider a random walk on the extended graph $\mathcal{G}_e=({\mathcal{N}}\cup{\{n+1\}},{\mathcal{E}}\cup{\{(i,n+1)\colon i\in{\mathcal{N}}\}})$ with a transition probability matrix $$P_e = \begin{pmatrix}
cP&(1-c){{\boldsymbol{1}}}\\0&1
\end{pmatrix}.$$ At each step of time, with probability $1-c$, the random surfer can *disappear* from the original graph, that is he can reach the absorbing node $n+1$.
The nonnegative matrix $(I-cP)^{-1}$ is commonly called the fundamental matrix of the absorbing Markov chain defined by $P_e$ (see for instance [@KS60; @Sen81]). In the extended graph $\mathcal{G}_e$, the entry $[(I-cP)^{-1}]_{ij}$ is the expected number of visits to node $j$ before reaching the absorbing node $n+1$ when starting from node $i$. From the point of view of the standard random surfer described in Section \[sec:preliminaries\], the entry $[(I-cP)^{-1}]_{ij}$ is the expected number of visits to node $j$ before zapping for the first time when starting from node $i$.
Therefore, the vector ${\boldsymbol{v}}$ defined in equation has the following probabilistic interpretation. The entry ${\boldsymbol{v}}_i$ is the *expected number of visits to the set ${\mathcal{I}}$ before zapping* for the first time when the random surfer starts his walk in node $i$.
Now, let us first prove some simple properties about this vector.
\[lem:minmaxv\] Let ${\boldsymbol{v}}\in{\mathbb{R}}^n_{\ge0}$ be defined by ${\boldsymbol{v}}=cP{\boldsymbol{v}}+{\boldsymbol{e}}_{{{\mathcal{I}}}}$. Then,
- $\max_{i\notin{\mathcal{I}}}{\boldsymbol{v}}_i \le c\,\max_{i\in{\mathcal{I}}} {\boldsymbol{v}}_i$,
- ${\boldsymbol{v}}_i\le 1+c\,{\boldsymbol{v}}_i$ for all $i\in{\mathcal{N}}$; with equality if and only if the node $i$ does not have an access to ${{\overline{{{{\mathcal{I}}}}}}}$,
- ${\boldsymbol{v}}_i\ge\min_{j{\gets{i}}}{\boldsymbol{v}}_j$ for all $i\in{{{\mathcal{I}}}}$; with equality if and only if the node $i$ does not have an access to ${{\overline{{{{\mathcal{I}}}}}}}$;
<!-- -->
- Since $c<1$, for all $i\notin{{{\mathcal{I}}}}$, $$\max_{i\notin{{{\mathcal{I}}}}}{\boldsymbol{v}}_i
=\max_{i\notin{{{\mathcal{I}}}}}\bigg(c\sum_{j{\gets{i}}}\frac{{\boldsymbol{v}}_j}{{d_{i}}}\bigg)
\le c\max_j{\boldsymbol{v}}_j.$$ Since $c<1$, it then follows that $\max_{j} {\boldsymbol{v}}_j = \max_{i\in{{{\mathcal{I}}}}} {\boldsymbol{v}}_i$.
- The inequality ${\boldsymbol{v}}_i\le \frac{1}{1-c}$ follows directly from $$\max_i{\boldsymbol{v}}_i
\le\max_i\bigg(1+c\sum_{j{\gets{i}}}\frac{{\boldsymbol{v}}_j}{{d_{i}}}\bigg)
\le 1+c\max_j{\boldsymbol{v}}_j.$$ From $(a)$ it then also follows that ${\boldsymbol{v}}_i\le \frac{c}{1-c}$ for all $i\notin{{{\mathcal{I}}}}$. Now, let $i\in{\mathcal{N}}$ such that ${\boldsymbol{v}}_i= \frac{1}{1-c}$. Then $i\in{{{\mathcal{I}}}}$. Moreover, $$1+c\,{\boldsymbol{v}}_i={\boldsymbol{v}}_i=1+c\sum_{j{\gets{i}}}\frac{{\boldsymbol{v}}_j}{{d_{i}}},$$ that is ${\boldsymbol{v}}_j= \frac{1}{1-c}$ for every $j{\gets{i}}$. Hence node $j$ must also belong to ${{{\mathcal{I}}}}$. By induction, every node $k$ such that $i$ has an access to $k$ must belong to ${{{\mathcal{I}}}}$.
- Let $i\in{{{\mathcal{I}}}}$. Then, by $(b)$ $$1+c\,{\boldsymbol{v}}_i\ge{\boldsymbol{v}}_i=1+c\sum_{j{\gets{i}}}\frac{{\boldsymbol{v}}_j}{{d_{i}}}
\ge 1+c\min_{j{\gets{i}}}{}{{\boldsymbol{v}}_j},$$ so ${\boldsymbol{v}}_i\ge\min_{j{\gets{i}}}{\boldsymbol{v}}_j$ for all $i\in{{{\mathcal{I}}}}$. If ${\boldsymbol{v}}_i=\min_{j{\gets{i}}}{\boldsymbol{v}}_j$ then also $1+c\,{\boldsymbol{v}}_i={\boldsymbol{v}}_i$ and hence, by $(b)$, the node $i$ does not have an access to ${{\overline{{{{\mathcal{I}}}}}}}$.
Let us denote the set of nodes [of]{} ${{\overline{{{{\mathcal{I}}}}}}}$ which on average give the most visits [to]{} ${{{\mathcal{I}}}}$ before zapping by $${{\mathcal{V}}}=\operatorname*{argmax}_{j\in{{\overline{{{{\mathcal{I}}}}}}}}{\boldsymbol{v}}_j .$$ Then the following lemma is quite intuitive. It says that, among the nodes [of]{} ${{\overline{{{{\mathcal{I}}}}}}}$, those which provide the higher mean number of visits [to ${{{\mathcal{I}}}}$ are parents of ${{{\mathcal{I}}}}$, i.e. parents of some node of ${{{\mathcal{I}}}}$]{}.
\[lem:vmax-parents\] If ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}\neq\emptyset$, then $${{\mathcal{V}}}\subseteq
{\{j\in{{\overline{{{{\mathcal{I}}}}}}}\colon
\text{ there exists } \ell\in{{{\mathcal{I}}}}\text{ such that } (j,\ell)\in{\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}\}}.$$ If ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}=\emptyset$, then ${\boldsymbol{v}}_j=0$ for every $j\in{{\overline{{{{\mathcal{I}}}}}}}$.
Suppose first that ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}\neq\emptyset$. Let $k\in{{\mathcal{V}}}$ with ${\boldsymbol{v}}=(I-cP)^{-1}{\boldsymbol{e}}_{{{\mathcal{I}}}}$. If we supposed that there does not exist $\ell\in{{{\mathcal{I}}}}$ such that $(k,\ell)\in{\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$, then we would have, since ${\boldsymbol{v}}_k>0$, $${\boldsymbol{v}}_k=c\sum_{j{\gets{k}}}\frac{{\boldsymbol{v}}_j}{{d_{k}}}
\le c\max_{j\notin{{{\mathcal{I}}}}}{\boldsymbol{v}}_j=c{\boldsymbol{v}}_k<{\boldsymbol{v}}_k,$$ which is a contradiction. Now, if ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}=\emptyset$, then [there is no access to ${{{\mathcal{I}}}}$ from ${{\overline{{{{\mathcal{I}}}}}}}$, so]{} clearly ${\boldsymbol{v}}_j=0$ for every $j\in{{\overline{{{{\mathcal{I}}}}}}}$.
Lemma \[lem:vmax-parents\] shows that the nodes $j\in{{\overline{{{{\mathcal{I}}}}}}}$ which provide the higher value of ${\boldsymbol{v}}_j$ must belong to the set of [parents of ${{{\mathcal{I}}}}$]{}. The converse is not true, as we will see in the following example: some parents of ${{{\mathcal{I}}}}$ can provide a lower mean number of visits [to]{} ${{{\mathcal{I}}}}$ that other nodes which are not parents of ${{{\mathcal{I}}}}$. In other word, Lemma \[lem:vmax-parents\] gives a necessary but not sufficient condition in order to maximize the entry ${\boldsymbol{v}}_j$ for some $j\in{{\overline{{{{\mathcal{I}}}}}}}$.
\[ex:parents-not-sufficient\] Let us see on an example that having $(j,i)\in{\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ for some $i\in{{{\mathcal{I}}}}$ is not sufficient to have $j\in{{\mathcal{V}}}$. Consider the graph in Figure \[fig:parents-not-sufficient\]. Let ${{{\mathcal{I}}}}={\{1\}}$ and take a damping factor $c=0.85$. For ${\boldsymbol{v}} = (I-cP)^{-1}{\boldsymbol{e}}_1$, we have $$ {\boldsymbol{v}}_2={\boldsymbol{v}}_3={\boldsymbol{v}}_4 =4.359> {\boldsymbol{v}}_5= 3.521
>{\boldsymbol{v}}_6=3.492>{\boldsymbol{v}}_{7}>\cdots>{\boldsymbol{v}}_{11},$$ so ${{\mathcal{V}}}={\{2,3,4\}}$. As ensured by Lemma \[lem:vmax-parents\], every node [of]{} the set ${{\mathcal{V}}}$ is a parent of node $1$. But here, ${{\mathcal{V}}}$ does not contain all parents of [node]{} $1$. Indeed, the node $6\notin{{\mathcal{V}}}$ while it is a parent of $1$ and is moreover its parent with the [lowest]{} outdegree. Moreover, we see in this example that [node]{} $5$, which is a not a parent of node $1$ but a parent of [node]{} $6$, gives a higher value of the expected number of visits [to]{} ${{{\mathcal{I}}}}$ before zapping, than node $6$, parent of $1$. Let us try to get some intuition about that. When starting from node $6$, a random surfer has probability one half to reach node $1$ in only one step. But he has also a probability one half to move to node $11$ and to be send far away from node $1$. On the other side, when starting from node $5$, the random surfer can not reach node $1$ in only one step. But with probability $3/4$ he will reach one of the nodes $2$, $3$ or $4$ in one step. And from these nodes, the websurfer [stays]{} very near to node $1$ and can not be [sent]{} far away from it.
![The node $6\notin{{\mathcal{V}}}$ [and yet]{} it is a parent of ${{{\mathcal{I}}}}={\{1\}}$ (see Example \[ex:parents-not-sufficient\]).[]{data-label="fig:parents-not-sufficient"}](figures.1)
In the next lemma, we show that from some node $i\in{{{\mathcal{I}}}}$ which [has]{} an access to ${{\overline{{{{\mathcal{I}}}}}}}$, there always exists what we call a [*decreasing path*]{} to ${{\overline{{{{\mathcal{I}}}}}}}$. That is, we can find a path such that the mean number of visits [to]{} ${{{\mathcal{I}}}}$ is higher when starting from some node of the path than when starting from the successor of this node in the path.
\[lem:decreasing-path\] For every $i_0\in{{{\mathcal{I}}}}$ which has an access to $\overline{{\mathcal{I}}}$, there exists a path $\langle i_0,i_1,\dots,i_s\rangle$ with $i_1,\dots,i_{s-1}\in{{{\mathcal{I}}}}$ and $i_s\in{{\overline{{{{\mathcal{I}}}}}}}$ such that $${\boldsymbol{v}}_{i_0}>{\boldsymbol{v}}_{i_1}>...>{\boldsymbol{v}}_{i_s}.$$
Let us simply construct a decreasing path recursively by $$i_{k+1}\in\operatorname*{argmin}_{j{\gets{i_k}}}{\boldsymbol{v}}_j,$$ as long as $i_k\in{{{\mathcal{I}}}}$. If $i_k$ has an access to ${{\overline{{{{\mathcal{I}}}}}}}$, then ${\boldsymbol{v}}_{i_{k+1}}<{\boldsymbol{v}}_{i_k}<\frac{1}{1-c}$ by Lemma \[lem:minmaxv\]$(b)$ and $(c)$, so the node $i_{k+1}$ has also an access to ${{\overline{{{{\mathcal{I}}}}}}}$. By [assumption]{}, $i_0$ has an access to ${{\overline{{{{\mathcal{I}}}}}}}$. Moreover, the set ${{{\mathcal{I}}}}$ has a finite number of elements, so there must exist [an]{} $s$ such that $i_s\in{{\overline{{{{\mathcal{I}}}}}}}$.
Influence of the outlinks of a node
-----------------------------------
We will now see how a modification of the outlinks of some node $i\in{\mathcal{N}}$ can change the PageRank of a subset of nodes ${{{\mathcal{I}}}}\subseteq{\mathcal{N}}$. So we will compare two graphs on ${\mathcal{N}}$ defined by their set of links, ${\mathcal{E}}$ and $\widetilde{{\mathcal{E}}}$ respectively.
Every [item]{} corresponding to the graph defined by the set of links $\widetilde{{\mathcal{E}}}$ will be written with a tilde symbol. So $\widetilde{P}$ denotes its scaled adjacency matrix, ${\widetilde{{{\boldsymbol{\pi}}}}}$ the corresponding PageRank vector, ${\widetilde{d}_{i}}={|{\{j\colon (i,j)\in\widetilde{{\mathcal{E}}}\}}|}$ the outdegree of some node $i$ in this graph, $\widetilde{{\boldsymbol{v}}}=(I-c\widetilde P)^{-1}{\boldsymbol{e}}_{{{\mathcal{I}}}}$ and ${\widetilde{{\mathcal{V}}}}=\operatorname*{argmax}_{j\in{{\overline{{{{\mathcal{I}}}}}}}}\widetilde{{\boldsymbol{v}}}_j$. Finally, by $j{\widetilde\gets{i}}$ we mean $j\in{\{k\colon(i,k)\in\widetilde{{\mathcal{E}}}\}}$.
So, let us consider two graphs defined respectively by their set of links ${\mathcal{E}}$ and $\widetilde{{\mathcal{E}}}$. Suppose that they differ only in the links starting from some given node $i$, that is ${\{j\colon (k,j)\in{\mathcal{E}}\}}={\{j\colon (k,j)\in\widetilde{{\mathcal{E}}}\}}$ for all $k\neq i$. Then their scaled adjacency matrices $P$ and $\widetilde P$ are linked by a rank one correction. Let us then define the vector $${\boldsymbol{\delta}}=\sum_{j{\widetilde\gets{i}}}\frac{{\boldsymbol{e}}_j}{{\widetilde{d}_{i}}}
-\sum_{j{\gets{i}}}\frac{{\boldsymbol{e}}_j}{{d_{i}}},$$ which gives the correction to apply to the line $i$ of the matrix $P$ in order to get $\widetilde P$.
Now let us first express the difference between the PageRank of ${{{\mathcal{I}}}}$ for two configurations differing only in the links starting from some node $i$. Note [that]{} in the following lemma the personalization vector ${\boldsymbol{z}}$ does not appear explicitly in the expression of ${\widetilde{{{\boldsymbol{\pi}}}}}$.
\[lem:pt-pi\] Let two graphs defined respectively by ${\mathcal{E}}$ and $\widetilde{{\mathcal{E}}}$ and let $i\in{\mathcal{N}}$ such that for all $k\neq i$, ${\{j\colon (k,j)\in{\mathcal{E}}\}}={\{j\colon (k,j)\in\widetilde{{\mathcal{E}}}\}}$. Then $${\widetilde{{{\boldsymbol{\pi}}}}}^T {\boldsymbol{e}}_{{{\mathcal{I}}}}={{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}+ c\,{{\boldsymbol{\pi}}}_i\,
\frac{{\boldsymbol{\delta}}^T {\boldsymbol{v}} }{1-c\,{\boldsymbol{\delta}}^T(I-cP)^{-1}{\boldsymbol{e}}_i}.$$
Clearly, the scaled adjacency matrices are linked by $\widetilde P = P + {\boldsymbol{e}}_i\, {\boldsymbol{\delta}}^T$. Since $c<1$, the matrix $(I-cP)^{-1}$ exists and the PageRank vectors can be expressed as $$\begin{aligned}
{{\boldsymbol{\pi}}}^T&=(1-c){\boldsymbol{z}}^T(I-cP)^{-1},\\
{\widetilde{{{\boldsymbol{\pi}}}}}^T&=(1-c){\boldsymbol{z}}^T(I-c\,(P+{\boldsymbol{e}}_i{\boldsymbol{\delta}}^T))^{-1}.
\end{aligned}$$ Applying the Sherman–Morrison formula [to]{} $((I-cP)-c{\boldsymbol{e}}_i{\boldsymbol{\delta}}^T)^{-1}$, we get $${\widetilde{{{\boldsymbol{\pi}}}}}^T =(1-c){\boldsymbol{z}}^T(I-cP)^{-1}
+ (1-c){\boldsymbol{z}}^T(I-cP)^{-1}{\boldsymbol{e}}_i\,
\frac{c{\boldsymbol{\delta}}^T (I-cP)^{-1}}{1-c{\boldsymbol{\delta}}^T(I-cP)^{-1}{\boldsymbol{e}}_i},$$ and the result follows immediately.
Let us now give an equivalent condition in order to increase the PageRank of ${{{\mathcal{I}}}}$ by changing outlinks of some node $i$. The PageRank of ${{{\mathcal{I}}}}$ increases [essentially]{} when the new set of links favors nodes giving a higher mean number of visits [to]{} ${{{\mathcal{I}}}}$ before zapping.
\[thm:p>p-d>0\] Let two graphs defined respectively by ${\mathcal{E}}$ and $\widetilde{{\mathcal{E}}}$ and let $i\in{\mathcal{N}}$ such that for all $k\neq i$, ${\{j\colon (k,j)\in{\mathcal{E}}\}}={\{j\colon (k,j)\in\widetilde{{\mathcal{E}}}\}}$. Then $${\widetilde{{{\boldsymbol{\pi}}}}}^T {\boldsymbol{e}}_{{{\mathcal{I}}}}>{{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}\quad \text{ if and only if } \quad
{\boldsymbol{\delta}}^T{\boldsymbol{v}}>0$$ and ${\widetilde{{{\boldsymbol{\pi}}}}}^T {\boldsymbol{e}}_{{{\mathcal{I}}}}={{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ if and only if ${\boldsymbol{\delta}}^T{\boldsymbol{v}}=0$.
Let us first show that ${\boldsymbol{\delta}}^T(I-cP)^{-1}{\boldsymbol{e}}_i\le1$ is always verified. Let ${\boldsymbol{u}} =(I-cP)^{-1}{\boldsymbol{e}}_i$. Then ${\boldsymbol{u}}-cP{\boldsymbol{u}}={\boldsymbol{e}}_i$ and, by Lemma \[lem:minmaxv\]$(a)$, ${\boldsymbol{u}}_j\le{\boldsymbol{u}}_i$ for all $j$. So $${\boldsymbol{\delta}}^T{\boldsymbol{u}}
= \sum_{j{\widetilde\gets{i}}}\frac{{\boldsymbol{u}}_j}{{\widetilde{d}_{i}}}
-\sum_{j{\gets{i}}}\frac{{\boldsymbol{u}}_j}{{d_{i}}}
\le {\boldsymbol{u}}_i - \sum_{j{\gets{i}}}\frac{{\boldsymbol{u}}_j}{{d_{i}}}
\le {\boldsymbol{u}}_i - c\sum_{j{\gets{i}}}\frac{{\boldsymbol{u}}_j}{{d_{i}}} =1.$$ Now, since $c<1$ and ${{\boldsymbol{\pi}}}>0$, the conclusion follows by Lemma \[lem:pt-pi\].
The following Proposition \[prop:add-link\] shows how to [add a new]{} link $(i,j)$ starting from a given node $i$ in order to increase the PageRank of the set ${{{\mathcal{I}}}}$. The PageRank of ${{{\mathcal{I}}}}$ increases as soon as a node $i\in{{{\mathcal{I}}}}$ adds a link to a node [$j$ with a larger or equal expected number of visits to ${{{\mathcal{I}}}}$ before zapping]{}.
\[prop:add-link\] Let $i\in{{{\mathcal{I}}}}$ and let $j\in{\mathcal{N}}$ [be]{} such that $(i,j)\notin{\mathcal{E}}$ and ${\boldsymbol{v}}_i\le {\boldsymbol{v}}_j$. Let $\widetilde{{\mathcal{E}}}={\mathcal{E}}\cup{\{(i,j)\}}$. Then $${\widetilde{{{\boldsymbol{\pi}}}}}^T{\boldsymbol{e}}_{{{{\mathcal{I}}}}} \ge {{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{{\mathcal{I}}}}}$$ with equality if and only [if]{} the node $i$ does not have an access to ${{\overline{{{{\mathcal{I}}}}}}}$.
Let $i\in{{{\mathcal{I}}}}$ and let $j\in{\mathcal{N}}$ [be]{} such that $(i,j)\notin{\mathcal{E}}$ and ${\boldsymbol{v}}_i\le {\boldsymbol{v}}_j$. Then $$1+c\sum_{k{\gets{i}}}\frac{{\boldsymbol{v}}_k}{{d_{i}}}={\boldsymbol{v}}_i\le1+c{\boldsymbol{v}}_i\le 1+c{\boldsymbol{v}}_j,$$ with equality if and only if $i$ does not have an access to ${{\overline{{{{\mathcal{I}}}}}}}$ by Lemma \[lem:minmaxv\]$(b)$. Let $\widetilde{{\mathcal{E}}}={\mathcal{E}}\cup{\{(i,j)\}}$. Then $${\boldsymbol{\delta}}^T{\boldsymbol{v}}
= \frac{1}{{d_{i}}+1}
\bigg( {\boldsymbol{v}}_j-\sum_{k{\gets{i}}}\frac{{\boldsymbol{v}}_k}{{d_{i}}}\bigg)
\ge0,$$ with equality if and only [if]{} $i$ does not have an access to ${{\overline{{{{\mathcal{I}}}}}}}$. [The conclusion follows from]{} Theorem \[thm:p>p-d>0\].
Now let us see how to [remove]{} a link $(i,j)$ starting from a given node $i$ in order to increase the PageRank of the set ${{{\mathcal{I}}}}$. If a node $i\in{\mathcal{N}}$ removes a link to its worst child from the point of view of the expected number of visits to ${{{\mathcal{I}}}}$ before zapping, then the PageRank of ${{{\mathcal{I}}}}$ increases.
\[prop:remove-link\] Let $i\in{\mathcal{N}}$ and let $j\in\operatorname*{argmin}_{k{\gets{i}}} {\boldsymbol{v}}_k$. Let $\widetilde{{\mathcal{E}}}={\mathcal{E}}\setminus{\{(i,j)\}}$. Then $${\widetilde{{{\boldsymbol{\pi}}}}}^T{\boldsymbol{e}}_{{{{\mathcal{I}}}}} \ge {{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{{\mathcal{I}}}}}$$ with equality if and only [if]{} ${\boldsymbol{v}}_k={\boldsymbol{v}}_j$ for every $k$ such that $(i,k)\in{\mathcal{E}}$.
Let $i\in{\mathcal{N}}$ and let $j\in\operatorname*{argmin}_{k{\gets{i}}} {\boldsymbol{v}}_k$. Let $\widetilde{{\mathcal{E}}}={\mathcal{E}}\setminus{\{(i,j)\}}$. Then $${\boldsymbol{\delta}}^T{\boldsymbol{v}}
= \sum_{k{\gets{i}}}\frac{{\boldsymbol{v}}_k-{\boldsymbol{v}}_j}{{d_{i}}({d_{i}}-1)}
\ge0,$$ with equality if and only if ${\boldsymbol{v}}_k={\boldsymbol{v}}_j$ for all $k{\gets{i}}$. The conclusion follows by Theorem \[thm:p>p-d>0\].
[In order to increase the PageRank of ${\mathcal{I}}$ with a new link $(i,j)$, Proposition \[prop:add-link\] only requires that ${\boldsymbol{v}}_j\le{\boldsymbol{v}}_i$. On the other side, Proposition \[prop:remove-link\] requires that ${\boldsymbol{v}}_j=\min_{k{\gets{i}}}{\boldsymbol{v}}_k$ in order to increase the PageRank of ${\mathcal{I}}$ by deleting link $(i,j)$. One could wonder whether or not this condition could be weakened to ${\boldsymbol{v}}_j<{\boldsymbol{v}}_i$, so as to have symmetric conditions for the addition or deletion of links. In fact, this can not be done as shown in the following example.]{}
\[ex:v<v-not-sufficient\] Let us see by an example that the condition $j\in\operatorname*{argmin}_{k{\gets{i}}} {\boldsymbol{v}}_k$ in Proposition \[prop:remove-link\] can not be weakened [to]{} ${\boldsymbol{v}}_j<{\boldsymbol{v}}_i$. Consider the graph in Figure \[fig:v<v-not-sufficient\] and take a damping factor $c=0.85$. Let ${{{\mathcal{I}}}}={\{1,2,3\}}$. We have $${\boldsymbol{v}}_1=2.63>{\boldsymbol{v}}_2=2.303>{\boldsymbol{v}}_3 =1.533.$$ As ensured by Proposition \[prop:remove-link\], if we remove the link $(1,3)$, the PageRank of ${{{\mathcal{I}}}}$ increases (e.g. from $0.199$ to $0.22$ with a uniform personalization vector ${\boldsymbol{z}}=\frac{1}{n}{{\boldsymbol{1}}}$), since $3\in\operatorname*{argmin}_{k{\gets{1}}}{\boldsymbol{v}}_k$. But, if we remove instead the link $(1,2)$, the PageRank of ${{{\mathcal{I}}}}$ decreases (from $0.199$ to $0.179$ with ${\boldsymbol{z}}$ uniform) even if ${\boldsymbol{v}}_2<{\boldsymbol{v}}_1$.
![For ${{{\mathcal{I}}}}={\{1,2,3\}}$, removing link $(1,2)$ gives ${\widetilde{{{\boldsymbol{\pi}}}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}<{{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$, even if ${\boldsymbol{v}}_1>{\boldsymbol{v}}_2$ (see Example \[ex:v<v-not-sufficient\]).[]{data-label="fig:v<v-not-sufficient"}](figures.2)
Let us note that, if the node $i$ does not have an access to the set ${{\overline{{{{\mathcal{I}}}}}}}$, then for every [ *deletion* of a]{} link starting from $i$, the PageRank of ${{{\mathcal{I}}}}$ will not be modified. Indeed, in this case ${\boldsymbol{\delta}}^T{\boldsymbol{v}}=0$ since by Lemma \[lem:minmaxv\]$(b)$, ${\boldsymbol{v}}_j=\frac{1}{1-c}$ for every $j{\gets{i}}$.
Basic absorbing graph
---------------------
Now, let us introduce briefly the notion of basic absorbing graph (see Chapter III about absorbing Markov chains in Kemeny and Snell’s book [@KS60]).
For a given graph $({\mathcal{N}},{\mathcal{E}})$ and a specified subset of nodes ${{{\mathcal{I}}}}\subseteq{\mathcal{N}}$, the [*basic absorbing graph*]{} is the graph $({\mathcal{N}},{\mathcal{E}}^0)$ defined by ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}^0=\emptyset$, ${\mathcal{E}}_{{{\mathcal{I}}}}^0={\{(i,i)\colon i\in{{{\mathcal{I}}}}\}}$, ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}^0={\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}^0={\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$. In other words, the basic absorbing graph $({\mathcal{N}},{\mathcal{E}}^0)$ is a graph constructed from $({\mathcal{N}},{\mathcal{E}})$, keeping the same sets of external inlinks and external links [${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}},{\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$]{}, removing the external outlinks [${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$]{} and changing the internal link structure [${\mathcal{E}}_{{{\mathcal{I}}}}$]{} in order to have only self-links for nodes of ${{{\mathcal{I}}}}$.
Like in the previous subsection, every [item]{} corresponding to the basic absorbing graph will [have]{} a zero symbol. For instance, we will write ${{\boldsymbol{\pi}}}_0$ for the PageRank vector corresponding to the basic absorbing graph and ${{\mathcal{V}}}_0=\operatorname*{argmax}_{j\in{{\overline{{{{\mathcal{I}}}}}}}}[(I-cP_0)^{-1}{\boldsymbol{e}}_{{{\mathcal{I}}}}]_j$.
\[prop:out0-best\] Let a graph defined by a set of links ${\mathcal{E}}$ and let ${{{\mathcal{I}}}}\subseteq{\mathcal{N}}$. Then $${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}\le{{\boldsymbol{\pi}}}_0^T{\boldsymbol{e}}_{{{\mathcal{I}}}},$$ with equality if and only if ${{\mathcal{E}}}_{\mathrm{out({{{\mathcal{I}}}})}}=\emptyset$.
Up to a permutation of the indices, equation can be written as $$\begin{pmatrix}I-cP_{{{{\mathcal{I}}}}}&-cP_{{\mathrm{out({{{\mathcal{I}}}})}}}\\-cP_{{\mathrm{in({{{\mathcal{I}}}})}}}&I-cP_{{{\overline{{{{\mathcal{I}}}}}}}}\end{pmatrix}
\begin{pmatrix}{\boldsymbol{v}}_{{{\mathcal{I}}}}\\{\boldsymbol{v}}_{{\overline{{{{\mathcal{I}}}}}}}\end{pmatrix}
=\begin{pmatrix}{{\boldsymbol{1}}}\\0\end{pmatrix},$$ so we get $$\label{eq:vintr-vext}
{\boldsymbol{v}} = \begin{pmatrix}{\boldsymbol{v}}_{{{\mathcal{I}}}}\\c (I-cP_{{{\overline{{{{\mathcal{I}}}}}}}})^{-1}P_{\mathrm{in({{{\mathcal{I}}}})}}{\boldsymbol{v}}_{{{\mathcal{I}}}}\end{pmatrix}.$$ By Lemma \[lem:minmaxv\](b) and since $(I-cP_{{{\overline{{{{\mathcal{I}}}}}}}})^{-1}$ is a nonnegative matrix (see for instance the chapter on $M$-matrices in Berman and Plemmons’s book [@BP94]), we then have $${\boldsymbol{v}}
\le \begin{pmatrix}\frac{1}{1-c}\,{{\boldsymbol{1}}}\\
\frac{c}{1-c}\, (I-cP_{{{\overline{{{{\mathcal{I}}}}}}}})^{-1}P_{\mathrm{in({{{\mathcal{I}}}})}}{{\boldsymbol{1}}}\end{pmatrix}={\boldsymbol{v}}_0,$$ with equality if and only if no node of ${{{\mathcal{I}}}}$ has an access to ${{\overline{{{{\mathcal{I}}}}}}}$, that is ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}=\emptyset$. The conclusion now follows from equation and ${\boldsymbol{z}}>0$.
Let us finally prove a nice property of the set ${{\mathcal{V}}}$ when ${{{\mathcal{I}}}}={\{i\}}$ is a singleton: it is independent of the outlinks of $i$. In particular, it can be found from the basic absorbing graph.
\[lem:V=V0-single\] Let a graph defined by a set of links ${\mathcal{E}}$ and let ${{{\mathcal{I}}}}={\{i\}}$ Then there exists [an]{} $\alpha\neq0$ such that $(I-cP)^{-1}{\boldsymbol{e}}_i=\alpha(I-cP_0)^{-1}{\boldsymbol{e}}_i$. As a consequence, $${{\mathcal{V}}}={{\mathcal{V}}}_0 .$$
Let ${{{\mathcal{I}}}}={\{i\}}$. Since ${\boldsymbol{v}}_{{{\mathcal{I}}}}={\boldsymbol{v}}_i$ is a scalar, it follows from equation that the direction of the vector ${\boldsymbol{v}}$ does not depend on ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ but only on ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$.
Optimal linkage strategy for a website {#sec:optimal-linkage-strategy}
======================================
In this section, we consider a set of nodes ${{{\mathcal{I}}}}$. For this set, we want to choose the sets of internal links ${\mathcal{E}}_{{{{\mathcal{I}}}}}\subseteq{{{\mathcal{I}}}}\times {{{\mathcal{I}}}}$ and external outlinks ${\mathcal{E}}_{{\mathrm{out({{{\mathcal{I}}}})}}}\subseteq{{{\mathcal{I}}}}\times {{\overline{{{{\mathcal{I}}}}}}}$ in order to maximize the PageRank score of ${{{\mathcal{I}}}}$, that is ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$.
Let us first discuss about the constraints on ${\mathcal{E}}$ we will consider. If we do not impose any condition on ${\mathcal{E}}$, the problem of maximizing ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ is quite trivial. As shown by Proposition \[prop:out0-best\], you should take in this case ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}=\emptyset$ and ${\mathcal{E}}_{{{\mathcal{I}}}}$ an arbitrary subset of ${{{\mathcal{I}}}}\times {{{\mathcal{I}}}}$ such that each node has at least one outlink. You just try to [lure the random walker to]{} your pages, not allowing him to leave ${{{\mathcal{I}}}}$ except by zapping according to the preference vector. Therefore, it seems sensible to impose that *${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ must be nonempty*.
[Now, let us show that, in order to avoid trivial solutions to our maximization problem, it is not enough to assume that ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ must be nonempty. Indeed, with this single constraint, in order to lose as few as possible visits from the random walker, you should take a unique leaking node $k\in{{{\mathcal{I}}}}$ (i.e. ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}={\{(k,\ell)\}}$ for some $\ell\in{{\overline{{{{\mathcal{I}}}}}}}$) and isolate it from the rest of the set ${{{\mathcal{I}}}}$ (i.e. ${\{i\in{{{\mathcal{I}}}}\colon(i,k)\in{\mathcal{E}}_{{{\mathcal{I}}}}\}}=\emptyset$).]{}
Moreover, it seems reasonable to imagine that Google penalizes (or at least [tries]{} to penalize) such behavior in the context of spam alliances [@GGM05].
All this discussion leads us to make the following assumption.
\[ass:access-to-ext\] Every node [of]{} ${{{\mathcal{I}}}}$ has an access to at least one node [of]{} ${{\overline{{{{\mathcal{I}}}}}}}$.
Let us now explain the basic ideas we will use in order to determine an optimal linkage strategy for a set of webpages ${{{\mathcal{I}}}}$. We determine some forbidden patterns for an optimal linkage strategy and deduce the only possible structure an optimal strategy can have. In other words, we assume that we have a configuration which gives an optimal PageRank ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$. Then we prove that if some particular pattern appeared in this optimal structure, then we could construct another graph for which the PageRank ${\widetilde{{{\boldsymbol{\pi}}}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ is strictly higher than ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$.
We will firstly determine the shape of an optimal external outlink structure ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$, when the internal link structure ${\mathcal{E}}_{{{\mathcal{I}}}}$ is given, in Theorem \[thm:out-opt\]. Then, given the external outlink structure ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ we will determine the possible optimal internal link structure ${\mathcal{E}}_{{{\mathcal{I}}}}$ in Theorem \[thm:intr-opt\]. Finally, we will put both results together in Theorem \[thm:website-opt\] in order to get the general shape of an optimal linkage strategy for a set ${{{\mathcal{I}}}}$ when ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ are given.
Proofs of this section will be illustrated by several figures [for which we take the following drawing convention]{}.
[When nodes are [drawn]{} from left to right on the same horizontal line, they are arranged by decreasing value of ${\boldsymbol{v}}_j$. Links are represented by continuous arrows and paths by dashed arrows.]{}
The first result of this section concerns the optimal *outlink* structure ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ for the set ${{{\mathcal{I}}}}$, while its internal structure ${\mathcal{E}}_{{{\mathcal{I}}}}$ is given. An example of optimal outlink structure is given [after the theorem]{}.
\[thm:out-opt\] Let ${\mathcal{E}}_{{{\mathcal{I}}}}$, ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given. Let ${\mathcal{F}}_1, \dots,{\mathcal{F}}_r$ be the final classes [of the subgraph]{} $({{{\mathcal{I}}}},{\mathcal{E}}_{{{\mathcal{I}}}})$. Let ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ such that the PageRank ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ is maximal under Assumption A. Then ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ has the following structure: $${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}= {\mathcal{E}}_{\mathrm{out}({\mathcal{F}}_1)} \cup\cdots\cup{\mathcal{E}}_{\mathrm{out}({\mathcal{F}}_r)},$$ where for every $s=1,\dots,r$, $${\mathcal{E}}_{\mathrm{out}({\mathcal{F}}_s)}
\subseteq{\{(i,j)\colon i\in\operatorname*{argmin}_{k\in{\mathcal{F}}_s}{\boldsymbol{v}}_k \text{ and } j\in{{\mathcal{V}}}\}}.$$ Moreover for every $s=1,\dots,r$, if ${\mathcal{E}}_{{\mathcal{F}}_s}\neq\emptyset$, then ${|{\mathcal{E}}_{\mathrm{out}({\mathcal{F}}_s)}|}=1$.
Let ${\mathcal{E}}_{{{\mathcal{I}}}}$, ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given. Suppose ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ is such that ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ is maximal under Assumption A.
[We will determine the possible leaking nodes of ${{{\mathcal{I}}}}$ by analyzing three different cases.]{}
[Firstly, let us consider some node $i\in{{{\mathcal{I}}}}$ such that $i$ does not have children in ${{{\mathcal{I}}}}$, i.e. ${\{k\in{{{\mathcal{I}}}}\colon (i,k)\in{\mathcal{E}}_{{{\mathcal{I}}}}\}}=\emptyset$. Then clearly we have ${\{i\}}={\mathcal{F}}_s$ for some $s=1,\dots,r$, with $i\in\operatorname*{argmin}_{k\in{\mathcal{F}}_s}{\boldsymbol{v}}_k$ and ${\mathcal{E}}_{{\mathcal{F}}_s}=\emptyset$. From Assumption A, we have ${\mathcal{E}}_{\mathrm{out}({\mathcal{F}}_s)}\neq\emptyset$, and from Theorem \[thm:p>p-d>0\] and the optimality assumption, we have ${\mathcal{E}}_{\mathrm{out}({\mathcal{F}}_s)}\subseteq{\{(i,j)\colon j\in{{\mathcal{V}}}\}}$ (see Figure \[fig:out-opt-fig2\]).]{}
Secondly, let us consider some $i\in{{{\mathcal{I}}}}$ such that [$i$ has children in ${{{\mathcal{I}}}}$, i.e. ${\{k\in{{{\mathcal{I}}}}\colon (i,k)\in{\mathcal{E}}_{{{\mathcal{I}}}}\}}\neq\emptyset$ and]{} $${}{ {\boldsymbol{v}}_i\le\min_{\substack{k{\gets{i}}\\k\in{{{\mathcal{I}}}}}}{\boldsymbol{v}}_k}.$$ Let $j\in\operatorname*{argmin}_{k{\gets{i}}}{\boldsymbol{v}}_k$. Then $j\in{{\overline{{{{\mathcal{I}}}}}}}$ and ${\boldsymbol{v}}_j<{\boldsymbol{v}}_i$ by Lemma \[lem:minmaxv\]$(c)$. Suppose by contradiction that the node $i$ would keep an access to ${{\overline{{{{\mathcal{I}}}}}}}$ if we took $\widetilde{{\mathcal{E}}}_{\mathrm{out({{{\mathcal{I}}}})}}={\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}\setminus{\{(i,j)\}}$ instead of ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$. Then, by Proposition \[prop:remove-link\], considering $\widetilde{{\mathcal{E}}}_{\mathrm{out({{{\mathcal{I}}}})}}$ instead of ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ would increase strictly the PageRank of ${{{\mathcal{I}}}}$ while Assumption A remains satisfied (see Figure \[fig:out-opt-fig1\]).
This would contradict the optimality [assumption]{} for ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$. From this, we conclude that
- the node $i$ belongs to final class ${\mathcal{F}}_s$ of [the subgraph]{} $({{{\mathcal{I}}}},{\mathcal{E}}_{{{\mathcal{I}}}})$ with ${\mathcal{E}}_{{\mathcal{F}}_s}\neq\emptyset$ for some $s=1,\dots,r$;
- there does not exist another $\ell\in{{\overline{{{{\mathcal{I}}}}}}}$, $\ell\neq j$ such that $(i,\ell)\in{\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$;
- there does not exist another $k$ in the same final class ${\mathcal{F}}_s$, $k\neq i$ such that such that $(k,\ell)\in{\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ for some $\ell\in{{\overline{{{{\mathcal{I}}}}}}}$.
[Again, by Theorem \[thm:p>p-d>0\] and the optimality assumption, we have $j\in{\mathcal{V}}$ (see Figure \[fig:out-opt-fig2\]).]{}
[Let us now notice that]{} $$\label{eq:maxbarIv<minIv}
\max_{k\in{{\overline{{{{\mathcal{I}}}}}}}} {\boldsymbol{v}}_k<\min_{k\in{{{\mathcal{I}}}}}{\boldsymbol{v}}_k.$$ [Indeed, with $i\in\operatorname*{argmin}_{k\in{{{\mathcal{I}}}}}{\boldsymbol{v}}_k$, we are in one of the two cases analyzed above for which we have seen that ${\boldsymbol{v}}_i>{\boldsymbol{v}}_j=\operatorname*{argmax}_{k\in{{\overline{{{{\mathcal{I}}}}}}}} {\boldsymbol{v}}_k$. ]{}
Finally, consider a node $i\in{{{\mathcal{I}}}}$ that does not belong to any of the final classes of [the subgraph]{} $({{{\mathcal{I}}}},{\mathcal{E}}_{{{\mathcal{I}}}})$. Suppose by contradiction that there exists $j\in{{\overline{{{{\mathcal{I}}}}}}}$ such that $(i,j)\in{\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$. Let $\ell\in\operatorname*{argmin}_{k{\gets{i}}}{\boldsymbol{v}}_k$. Then it follows from inequality that $\ell\in{{\overline{{{{\mathcal{I}}}}}}}$. But the same argument as above shows that the link $(i,\ell)\in{\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ must be removed since ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ is supposed to be optimal (see Figure \[fig:out-opt-fig1\] again). So, there does not exist $j\in{{\overline{{{{\mathcal{I}}}}}}}$ such that $(i,j)\in{\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ for a node $i\in{{{\mathcal{I}}}}$ which does not belong to any of the final classes ${\mathcal{F}}_1,\dots,{\mathcal{F}}_r$.
\[ex:out-opt-ex\] [Let us consider the graph given in Figure \[fig:out-opt-ex\]. The internal link structure ${\mathcal{E}}_{{{\mathcal{I}}}}$, as well as ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ are given. The subgraph $({{{\mathcal{I}}}},{\mathcal{E}}_{{{\mathcal{I}}}})$ has two final classes ${\mathcal{F}}_1$ and ${\mathcal{F}}_2$. With $c=0.85$ and ${\boldsymbol{z}}$ the uniform probability vector, this configuration has six optimal outlink structures (one of these solutions is represented by bold arrows in Figure \[fig:out-opt-ex\]). Each one can be written as ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}={\mathcal{E}}_{\mathrm{out}({\mathcal{F}}_1)}\cup{\mathcal{E}}_{\mathrm{out}({\mathcal{F}}_2)}$, with ${\mathcal{E}}_{\mathrm{out}({\mathcal{F}}_1)}={\{(4,6)\}}$ or ${\mathcal{E}}_{\mathrm{out}({\mathcal{F}}_1)}={\{(4,7)\}}$ and $\emptyset\neq{\mathcal{E}}_{\mathrm{out}({\mathcal{F}}_2)}\subseteq{\{(5,6),(5,7)\}}$. Indeed, since ${\mathcal{E}}_{{\mathcal{F}}_1}\neq\emptyset$, as stated by Theorem \[thm:out-opt\], the final class ${\mathcal{F}}_1$ has exactly one external outlink in every optimal outlink structure. On the other hand, the final class ${\mathcal{F}}_2$ may have several external outlinks, since it is composed of a unique node and moreover this node does not have a self-link. Note that ${\mathcal{V}}={\{6,7\}}$ in each of these six optimal configurations, but this set ${\mathcal{V}}$ can not be determined a priori since it depends on the chosen outlink structure.]{}
Now, let us determine the optimal *internal* link structure ${\mathcal{E}}_{{{\mathcal{I}}}}$ for the set ${{{\mathcal{I}}}}$, while its outlink structure ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ is given. Examples of optimal internal structure are given [after the proof of the theorem]{}.
\[thm:intr-opt\] Let ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$, ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given. Let ${\mathcal{L}}={\{i\in {{{\mathcal{I}}}}\colon (i,j)\in{\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}\text{ for some }j\in{{\overline{{{{\mathcal{I}}}}}}}\}}$ be the set of leaking nodes [of]{} ${{{\mathcal{I}}}}$ and let $n_{{\mathcal{L}}}={|{\mathcal{L}}|}$ [be]{} the number of leaking nodes. Let ${\mathcal{E}}_{{{\mathcal{I}}}}$ such that the PageRank ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ is maximal under Assumption A. Then there exists a permutation of the indices such that ${{{\mathcal{I}}}}={\{1,2,\dots,n_{{{\mathcal{I}}}}\}}$, ${\mathcal{L}}={\{n_{{{\mathcal{I}}}}-n_{{\mathcal{L}}}+1,\dots,n_{{{\mathcal{I}}}}\}}$, $${\boldsymbol{v}}_1>\cdots>{\boldsymbol{v}}_{{n_{{{\mathcal{I}}}}}-n_{{\mathcal{L}}}}
>{\boldsymbol{v}}_{{n_{{{\mathcal{I}}}}}-n_{{\mathcal{L}}}+1}\ge\cdots\ge{\boldsymbol{v}}_{{n_{{{\mathcal{I}}}}}},$$ and ${\mathcal{E}}_{{{\mathcal{I}}}}$ has the following structure: $${\mathcal{E}}_{{{\mathcal{I}}}}^L\subseteq{\mathcal{E}}_{{{\mathcal{I}}}}\subseteq{\mathcal{E}}_{{{\mathcal{I}}}}^U,$$ where $$\begin{aligned}
{\mathcal{E}}_{{{\mathcal{I}}}}^L&={\{(i,j)\in{{{\mathcal{I}}}}\times{{{\mathcal{I}}}}\colon j\le i\}}
\cup {\{(i,j)\in({{{\mathcal{I}}}}\setminus{\mathcal{L}})\times{{{\mathcal{I}}}}\colon j=i+1\}},\\
{\mathcal{E}}_{{{\mathcal{I}}}}^U&={\mathcal{E}}_{{{\mathcal{I}}}}^L
\cup{\{(i,j)\in{\mathcal{L}}\times{\mathcal{L}}\colon i<j\}}.
\end{aligned}$$
Let ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$, ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given. Suppose ${\mathcal{E}}_{{{\mathcal{I}}}}$ is such that ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ is maximal under Assumption A.
Firstly, by Proposition \[prop:add-link\] and since every node [of]{} ${{{\mathcal{I}}}}$ has an access to ${{\overline{{{{\mathcal{I}}}}}}}$, every node $i\in{{{\mathcal{I}}}}$ links to every node $j\in{{{\mathcal{I}}}}$ such that ${\boldsymbol{v}}_j\ge {\boldsymbol{v}}_i$ (see Figure \[fig:intr-opt-fig1\]), that is $$\label{eq:thm-all-pred}
{\{(i,j)\in{\mathcal{E}}_{{{\mathcal{I}}}}\colon {\boldsymbol{v}}_i\le {\boldsymbol{v}}_j\}}
={\{(i,j)\in{{{\mathcal{I}}}}\times{{{\mathcal{I}}}}\colon {\boldsymbol{v}}_i\le {\boldsymbol{v}}_j\}}.$$
Secondly, let $(k,i)\in{\mathcal{E}}_{{{\mathcal{I}}}}$ such that $k\neq i$ and $k\in{{{\mathcal{I}}}}\setminus{\mathcal{L}}$. Let us prove that, if the node $i$ has an access to ${{\overline{{{{\mathcal{I}}}}}}}$ by a path $\langle i,i_1,\dots,i_s\rangle$ such that $i_j\neq k$ for all $j=1,\dots,s$ and $i_s\in{{\overline{{{{\mathcal{I}}}}}}}$, then ${\boldsymbol{v}}_i<{\boldsymbol{v}}_k$ (see Figure \[fig:intr-opt-fig3\]).
Indeed, if we had ${\boldsymbol{v}}_k\le {\boldsymbol{v}}_i$ then, by Lemma \[lem:minmaxv\]$(c)$, there would exists $\ell\in{{{\mathcal{I}}}}$ such that $(k,\ell)\in{\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\boldsymbol{v}}_\ell=\min_{j\gets k}{\boldsymbol{v}}_j<{\boldsymbol{v}}_i\le{\boldsymbol{v}}_k$. But, with $\widetilde{{\mathcal{E}}}_{{{\mathcal{I}}}}={\mathcal{E}}_{{{\mathcal{I}}}}\setminus{\{(k,\ell)\}}$, we would have ${\widetilde{{{\boldsymbol{\pi}}}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}>{{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ by Proposition \[prop:remove-link\] while Assumption A remains satisfied since the node $k$ would keep access to ${{\overline{{{{\mathcal{I}}}}}}}$ via the node $i$ (see Figure \[fig:intr-opt-fig2\]).
That contradicts the optimality [assumption]{}. This leads us to the conclusion that ${\boldsymbol{v}}_k>{\boldsymbol{v}}_i$ for every $k\in{{{\mathcal{I}}}}\setminus{\mathcal{L}}$ and $i\in{\mathcal{L}}$. Moreover ${\boldsymbol{v}}_i\neq {\boldsymbol{v}}_k$ for every $i,k\in{{{\mathcal{I}}}}\setminus{\mathcal{L}}$, $i\neq k$. Indeed, if we had ${\boldsymbol{v}}_i={\boldsymbol{v}}_k$, then $(k,i)\in{\mathcal{E}}_{{{\mathcal{I}}}}$ by while by Lemma \[lem:decreasing-path\], the node $i$ would have an access to ${{\overline{{{{\mathcal{I}}}}}}}$ by a path independant from $k$. So we should have ${\boldsymbol{v}}_i<{\boldsymbol{v}}_k$.
We conclude from this that we can relabel the nodes of ${\mathcal{N}}$ such that ${{{\mathcal{I}}}}={\{1,2,\dots n_{{{\mathcal{I}}}}\}}$, ${\mathcal{L}}={\{n_{{{\mathcal{I}}}}-n_{{\mathcal{L}}}+1,\dots,n_{{{\mathcal{I}}}}\}}$ and $$\label{eq:vI>vF}
{\boldsymbol{v}}_1>{\boldsymbol{v}}_2>\cdots>{\boldsymbol{v}}_{{n_{{{\mathcal{I}}}}}-n_{{\mathcal{L}}}}>
{\boldsymbol{v}}_{{n_{{{\mathcal{I}}}}}-n_{{\mathcal{L}}}+1}\ge\cdots\ge{\boldsymbol{v}}_{{n_{{{\mathcal{I}}}}}}.$$
It follows also that, for $i\in{{{\mathcal{I}}}}\setminus{\mathcal{L}}$ and $j>i$, $(i,j)\in{\mathcal{E}}_{{{\mathcal{I}}}}\text{ if and only if } j=i+1$. Indeed, suppose first $i<n_{{{\mathcal{I}}}}-n_{{\mathcal{L}}}$. Then, we cannot have $(i,j)\in{\mathcal{E}}_{{{\mathcal{I}}}}$ with $j>i+1$ since in this case we would contradict the ordering of the nodes given by equation (see Figure \[fig:intr-opt-fig3\] again with $k=i+1$ and remember that by Lemma \[lem:decreasing-path\], node $j$ has an access to ${{\overline{{{{\mathcal{I}}}}}}}$ by a decreasing path). Moreover, node $i$ must link to some node $j>i$ in order to satisfy Assumption A, so $(i,i+1)$ must belong to ${\mathcal{E}}_{{{\mathcal{I}}}}$. Now, consider the case $i=n_{{{\mathcal{I}}}}-n_{{\mathcal{L}}}$. Suppose we had $(i,j)\in{\mathcal{E}}_{{{\mathcal{I}}}}$ with $j>i+1$. Let us first note that there can not exist two or more different links $(i,\ell)$ with $\ell\in{\mathcal{L}}$ since in this case we could remove one of these links and increase strictly the PageRank of the set ${{{\mathcal{I}}}}$. If ${\boldsymbol{v}}_j={\boldsymbol{v}}_{i+1}$, we could relabel the nodes by permuting these two indices. If ${\boldsymbol{v}}_j<{\boldsymbol{v}}_{i+1}$, then with $\widetilde{{\mathcal{E}}}_{{{\mathcal{I}}}}={\mathcal{E}}_{{{\mathcal{I}}}}\cup{\{(i,i+1)\}}\setminus{\{(i,j)\}}$, we would have ${\widetilde{{{\boldsymbol{\pi}}}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}>{{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ by Theorem \[thm:p>p-d>0\] while Assumption A remains satisfied since the $i$ would keep access to ${{\overline{{{{\mathcal{I}}}}}}}$ via node $i+1$. That contradicts the optimality [assumption]{}. So we have proved that $$\label{eq:thm-one-succ}
{\{(i,j)\in{\mathcal{E}}_{{{\mathcal{I}}}}\colon i<j \text{ and } i\in{{{\mathcal{I}}}}\setminus{\mathcal{L}}\}}
={\{(i,i+1)\colon i\in{{{\mathcal{I}}}}\setminus{\mathcal{L}}\}}.$$
Thirdly, it is obvious that $$\label{eq:thm-F-F}
{\{(i,j)\in{\mathcal{E}}_{{{\mathcal{I}}}}\colon i<j \text{ and } i\in{\mathcal{L}}\}}
\subseteq{\{(i,j)\in{\mathcal{L}}\times{\mathcal{L}}\colon i<j\}}.$$
The announced structure for a set ${\mathcal{E}}_{{{\mathcal{I}}}}$ giving a maximal PageRank score ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ under Assumption A now follows directly from equations , and .
\[ex:intr-opt-ex\] [Let us consider the graphs given in Figure \[fig:intr-opt-ex\]. For both cases, the external outlink structure ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ with two leaking nodes, as well as ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ are given. With $c=0.85$ and ${\boldsymbol{z}}$ the uniform probability vector, the optimal internal link structure for configuration (a) is given by ${\mathcal{E}}_{{{\mathcal{I}}}}={\mathcal{E}}_{{{\mathcal{I}}}}^L$, while in configuration (b) we have ${\mathcal{E}}_{{{\mathcal{I}}}}={\mathcal{E}}_{{{\mathcal{I}}}}^U$ (bold arrows), with ${\mathcal{E}}_{{{\mathcal{I}}}}^L$ and ${\mathcal{E}}_{{{\mathcal{I}}}}^U$ defined in Theorem \[thm:intr-opt\].]{}
Finally, combining the optimal outlink structure and the optimal internal link structure described in Theorems \[thm:out-opt\] and \[thm:intr-opt\], we find the *optimal linkage strategy* for a set of webpages. Let us note that, since we have here control on both ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$, there [are]{} no more cases of several final classes or several leaking nodes to consider. For an example of optimal link structure, see Figure \[fig:optimal-link-structure\].
\[thm:website-opt\] Let ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given. Let ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ such that ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ is maximal under Assumption A. Then there exists a permutation of the indices such that ${{{\mathcal{I}}}}={\{1,2,\dots,n_{{{\mathcal{I}}}}\}}$, $${\boldsymbol{v}}_1>\cdots>{\boldsymbol{v}}_{{n_{{{\mathcal{I}}}}}}>{\boldsymbol{v}}_{{n_{{{\mathcal{I}}}}}+1}\ge\cdots\ge{\boldsymbol{v}}_n,$$ and ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ have the following structure: $$\begin{aligned}
{\mathcal{E}}_{{{\mathcal{I}}}}&={\{(i,j)\in{{{\mathcal{I}}}}\times{{{\mathcal{I}}}}\colon j\le i \text{ or } j=i+1\}}, \\
{\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}&={\{(n_{{{\mathcal{I}}}},n_{{{\mathcal{I}}}}+1)\}}.
\end{aligned}$$
Let ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given and suppose ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ are such that ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ is maximal under Assumption A. [Let us relabel the nodes of ${\mathcal{N}}$ such that ${{{\mathcal{I}}}}={\{1,2,\dots,n_{{{\mathcal{I}}}}\}}$ and ${\boldsymbol{v}}_1\ge\cdots\ge{\boldsymbol{v}}_{n_{{{\mathcal{I}}}}}>{\boldsymbol{v}}_{n_{{{\mathcal{I}}}}+1}=\max_{j\in{{\overline{{{{\mathcal{I}}}}}}}}{\boldsymbol{v}}_j$. By Theorem \[thm:intr-opt\], $(i,j)\in{\mathcal{E}}_{{{\mathcal{I}}}}$ for every nodes $i,j\in{{{\mathcal{I}}}}$ such that $j\le i$. In particular, every node [of]{} ${{{\mathcal{I}}}}$ has an access to node $1$. Therefore, there is a unique final class ${\mathcal{F}}_1\subseteq{{{\mathcal{I}}}}$ in the subgraph $({{{\mathcal{I}}}},{\mathcal{E}}_{{{\mathcal{I}}}})$. So, by Theorem \[thm:out-opt\], ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}={\{(k,\ell)\}}$ for some $k\in{\mathcal{F}}_1$ and $\ell\in{{\overline{{{{\mathcal{I}}}}}}}$. Without loss of generality, we can suppose that $\ell=n_{{{\mathcal{I}}}}+1$. By Theorem \[thm:intr-opt\] again, the leaking node $k=n_{{{\mathcal{I}}}}$]{} and therefore $(i,i+1)\in{\mathcal{E}}_{{{\mathcal{I}}}}$ for every node $i\in{\{1,\dots,n_{{{\mathcal{I}}}}-1\}}$.
Let us note that having a structure like described in Theorem \[thm:website-opt\] is a *necessary but not sufficient* condition in order to have a maximal PageRank.
\[ex:optimal-link-structure-conterexample\] Let us show by an example that the graph structure given in Theorem \[thm:website-opt\] is not sufficient to have a maximal PageRank. Consider for instance the graphs in Figure \[fig:optimal-link-structure-conterexample\]. Let $c=0.85$ and a uniform personalization vector ${\boldsymbol{z}} = \frac{1}{n}{{\boldsymbol{1}}}$. Both graphs have the link structure required Theorem \[thm:website-opt\] in order to have a maximal PageRank, with ${\boldsymbol{v}}_{(a)}=\begin{pmatrix}6.484&6.42&6.224&5.457\end{pmatrix}^T$ and ${\boldsymbol{v}}_{(b)}=\begin{pmatrix}6.432&6.494&6.247&5.52\end{pmatrix}^T$. But the configuration (a) is not optimal since in this case, the PageRank ${{\boldsymbol{\pi}}}_{(a)}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}= 0.922$ is strictly less than the PageRank ${{\boldsymbol{\pi}}}_{(b)}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}= 0.926$ obtained by the configuration (b). Let us nevertheless note that, with a non uniform personalization vector ${\boldsymbol{z}}=\begin{pmatrix}0.7&0.1&0.1&0.1\end{pmatrix}^T$, the link structure (a) would be optimal.
Extensions and variants {#sec:extensions-variants}
=======================
Let us now present some extensions and variants of the results of the previous section. We will first emphasize the role of parents of ${{{\mathcal{I}}}}$. Secondly, we will briefly talk about Avrachenkov–Litvak’s optimal link structure for the case where ${{{\mathcal{I}}}}$ is a singleton. Then we will give variants of Theorem \[thm:website-opt\] when self-links are forbidden or when a minimal number of external outlinks is required. Finally, we will make some comments of the influence of external *inlinks* on the PageRank of ${{{\mathcal{I}}}}$.
Linking to parents
------------------
If some node of ${{{\mathcal{I}}}}$ has at least one parent in ${{\overline{{{{\mathcal{I}}}}}}}$ then the optimal linkage strategy for ${{{\mathcal{I}}}}$ is to have an internal link structure like described in Theorem \[thm:website-opt\] together with [a single link to one of the parents]{} [of]{} ${{{\mathcal{I}}}}$.
Let ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}{}{\neq\emptyset}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given. Let ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ such that ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ is maximal under Assumption A. Then ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}= {\{(i,j)\}}$, for some $i\in{{{\mathcal{I}}}}$ and $j\in{{\overline{{{{\mathcal{I}}}}}}}$ such that $(j,k)\in{\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ for some $k\in{{{\mathcal{I}}}}$.
This is a direct consequence of Lemma \[lem:vmax-parents\] and Theorem \[thm:website-opt\].
Let us nevertheless remember that not every parent of nodes of ${{{\mathcal{I}}}}$ will give an optimal link structure, as we have already discussed in Example \[ex:parents-not-sufficient\] and we develop now.
\[ex:parents-not-sufficient-single\] Let us continue Example \[ex:parents-not-sufficient\]. We consider the graph in Figure \[fig:parents-not-sufficient\] as basic absorbing graph for ${{{\mathcal{I}}}}={\{1\}}$, that is ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ are given. We take $c=0.85$ as damping factor and a uniform personalization vector ${\boldsymbol{z}}=\frac{1}{n}{{\boldsymbol{1}}}$. We have seen in Example \[ex:parents-not-sufficient\] than ${{\mathcal{V}}}_0={\{2,3,4\}}$. Let us consider the value of the PageRank ${{\boldsymbol{\pi}}}_1$ for different sets ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$:
[c|ccccc]{}\
&$\emptyset$&${\{(1,2)\}}$& ${\{(1,5)\}}$&${\{(1,6)\}}$&${\{(1,2),(1,3)\}}$\
${\mathcal{E}}_{{{\mathcal{I}}}}=\emptyset$&$\diagup$&$0.1739$&$0.1402$&$0.1392$&$0.1739$\
${\mathcal{E}}_{{{\mathcal{I}}}}={\{(1,1)\}}$&$0.5150$&$0.2600$&$0.2204$&$0.2192$&$0.2231$
As expected from Corollary \[cor:opt-single\], the optimal linkage strategy for ${{{\mathcal{I}}}}={\{1\}}$ is to have a self-link and a link to one of the nodes $2$, $3$ or $4$. We note also that a link to node $6$, which is a parent of node $1$ provides a lower PageRank that a link to node $5$, which is not parent of $1$. Finally, if we suppose self-links are forbidden (see below), then the optimal linkage strategy is to link to one *or more* of the nodes $2,3,4$.
In the case where no node of ${{{\mathcal{I}}}}$ has a parent in ${{\overline{{{{\mathcal{I}}}}}}}$, then *every* structure like described in Theorem \[thm:website-opt\] will give an optimal link structure.
Let ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given. Suppose that ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}=\emptyset$. Then the PageRank ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ is maximal under Assumption A if and only if $$\begin{aligned}
{\mathcal{E}}_{{{\mathcal{I}}}}&={\{(i,j)\in{{{\mathcal{I}}}}\times{{{\mathcal{I}}}}\colon j\le i \text{ or } j=i+1\}}, \\
{\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}&={\{(n_{{{\mathcal{I}}}},n_{{{\mathcal{I}}}}+1)\}}.
\end{aligned}$$ for some permutation of the indices such that ${{{\mathcal{I}}}}={\{1,2,\dots,n_{{{\mathcal{I}}}}\}}$.
This follows directly from ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}=(1-c){\boldsymbol{z}}^T{\boldsymbol{v}}$ and the fact that, if ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}=\emptyset$, $${\boldsymbol{v}}=(I-cP)^{-1}{\boldsymbol{e}}_{{{\mathcal{I}}}}=
\begin{pmatrix}(I-cP_{{{\mathcal{I}}}})^{-1}{{\boldsymbol{1}}}\\0\end{pmatrix},$$ up to a permutation of the indices.
Optimal linkage strategy for a singleton
----------------------------------------
The optimal outlink structure for a [single webpage]{} has already been given by Avrachenkov and Litvak in [@AL06]. Their result becomes a particular case of Theorem \[thm:website-opt\]. Note that in the case of a single node, the possible choices for ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ can be found a priori by considering the basic absorbing graph, since ${{\mathcal{V}}}={{\mathcal{V}}}_0$.
\[cor:opt-single\] Let ${{{\mathcal{I}}}}={\{i\}}$ and let ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given. Then the PageRank ${{\boldsymbol{\pi}}}_i$ is maximal under Assumption A if and only if ${\mathcal{E}}_{{{\mathcal{I}}}}={\{(i,i)\}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}={\{(i,j)\}}$ for some $j\in{{\mathcal{V}}}_0$.
This follows directly from Lemma \[lem:V=V0-single\] and Theorem \[thm:website-opt\].
Optimal linkage strategy under additional assumptions
-----------------------------------------------------
Let us consider the problem of maximizing the PageRank ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ when *self-links are forbidden*. Indeed, it seems to be often supposed that Google’s PageRank algorithm does not take self-links [into]{} account. In this case, Theorem \[thm:website-opt\] can be adapted readily for the case where ${|{{{\mathcal{I}}}}|}\ge2$. When ${{{\mathcal{I}}}}$ is a singleton, we must have ${\mathcal{E}}_{{{\mathcal{I}}}}=\emptyset$, so ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ can contain *several* links, as stated in Theorem \[thm:out-opt\].
Suppose ${|{{{\mathcal{I}}}}|}\ge2$. Let ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given. Let ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ such that ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ is maximal under Assumption A and assumption that there does not exist $i\in{{{\mathcal{I}}}}$ such that ${\{(i,i)\}}\in{\mathcal{E}}_{{{\mathcal{I}}}}$. Then there exists a permutation of the indices such that ${{{\mathcal{I}}}}={\{1,2,\dots,n_{{{\mathcal{I}}}}\}}$, ${\boldsymbol{v}}_1>\cdots>{\boldsymbol{v}}_{{n_{{{\mathcal{I}}}}}}>{\boldsymbol{v}}_{{n_{{{\mathcal{I}}}}}+1}\ge\cdots\ge{\boldsymbol{v}}_n$, and ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ have the following structure: $$\begin{aligned}
{\mathcal{E}}_{{{\mathcal{I}}}}&={\{(i,j)\in{{{\mathcal{I}}}}\times{{{\mathcal{I}}}}\colon j< i \text{ or } j=i+1\}}, \\
{\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}&={\{(n_{{{\mathcal{I}}}},n_{{{\mathcal{I}}}}+1)\}}.
\end{aligned}$$
Suppose ${{{\mathcal{I}}}}={\{i\}}$. Let ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given. Suppose ${\mathcal{E}}_{{{\mathcal{I}}}}=\emptyset$. Then the PageRank ${{\boldsymbol{\pi}}}_i$ is maximal under Assumption A if and only if $\emptyset\neq{\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}\subseteq{{\mathcal{V}}}_0$.
Let us now consider the problem of maximizing the PageRank ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ when *several external outlinks are required*. Then the proof of Theorem \[thm:out-opt\] can be adapted readily in order to have the following variant of Theorem \[thm:website-opt\].
Let ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given. Let ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ such that ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ is maximal under Assumption A and assumption that ${|{\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}|}\ge r$. Then there exists a permutation of the indices such that ${{{\mathcal{I}}}}={\{1,2,\dots,n_{{{\mathcal{I}}}}\}}$, ${\boldsymbol{v}}_1>\cdots>{\boldsymbol{v}}_{{n_{{{\mathcal{I}}}}}}>{\boldsymbol{v}}_{{n_{{{\mathcal{I}}}}}+1}\ge\cdots\ge{\boldsymbol{v}}_n$, and ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ have the following structure: $$\begin{aligned}
{\mathcal{E}}_{{{\mathcal{I}}}}&={\{(i,j)\in{{{\mathcal{I}}}}\times{{{\mathcal{I}}}}\colon j< i \text{ or } j=i+1\}}, \\
{\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}&={\{(n_{{{\mathcal{I}}}},j_k)\colon j_k\in{{\mathcal{V}}}\text{ for } k=1,\dots, r\}}.
\end{aligned}$$
External inlinks
----------------
Finally, let us make some comments about the addition of [external inlinks]{} to the set ${{{\mathcal{I}}}}$. It is well known that adding an inlink to a particular page always increases the PageRank of this page [@AL04; @IW06]. This can be viewed as a direct consequence of Theorem \[thm:p>p-d>0\] and Lemma \[lem:minmaxv\]. The case of a set of several pages ${{{\mathcal{I}}}}$ is not so simple. We prove in the following theorem that, if the set ${{{\mathcal{I}}}}$ has a link structure as described in Theorem \[thm:website-opt\] then adding an inlink to a page of ${{{\mathcal{I}}}}$ from a page $j\in{{\overline{{{{\mathcal{I}}}}}}}$ which is *not* a parent of some node of ${{{\mathcal{I}}}}$ will increase the PageRank of ${{{\mathcal{I}}}}$. But in general, adding an inlink to some page of ${{{\mathcal{I}}}}$ from ${{\overline{{{{\mathcal{I}}}}}}}$ may *decrease* the PageRank of the set ${{{\mathcal{I}}}}$, as shown in Examples \[ex:add-external-inlinks-conterexample1\] and \[ex:add-external-inlinks-conterexample2\].
Let ${{{\mathcal{I}}}}\subseteq{\mathcal{N}}$ and a graph defined by a set of links ${\mathcal{E}}$. If $$\min_{i\in{{{\mathcal{I}}}}}{\boldsymbol{v}}_i>\max_{j\notin{{{\mathcal{I}}}}}{\boldsymbol{v}}_j,$$ then, for every $j\in{{\overline{{{{\mathcal{I}}}}}}}$ [which is not a parent of ${{{\mathcal{I}}}}$]{}, and for every $i\in{{{\mathcal{I}}}}$, the graph defined by $\widetilde{{\mathcal{E}}}={\mathcal{E}}\cup{\{(j,i)\}}$ gives ${\widetilde{{{\boldsymbol{\pi}}}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}>{{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$.
This follows directly from Theorem \[thm:p>p-d>0\].
\[ex:add-external-inlinks-conterexample1\] Let us show by an example that a new external inlink is not always profitable for a set ${{{\mathcal{I}}}}$ in order to improve its PageRank, even if ${{{\mathcal{I}}}}$ has an optimal linkage strategy. Consider for instance the graph in Figure \[fig:add-external-inlinks-conterexample1\]. With $c=0.85$ and ${\boldsymbol{z}}$ uniform, we have ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}=0.8481$. But if we consider the graph defined by $\widetilde{{\mathcal{E}}}_{\mathrm{in({{{\mathcal{I}}}})}}={{\mathcal{E}}}_{\mathrm{in({{{\mathcal{I}}}})}}\cup{\{(3,2)\}}$, then we have ${\widetilde{{{\boldsymbol{\pi}}}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}=0.8321<{{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$.
\[ex:add-external-inlinks-conterexample2\] A new external inlink does not not always increase the PageRank of a set ${{{\mathcal{I}}}}$ in even if this new inlink comes from a page which is not already a parent of some node of ${{{\mathcal{I}}}}$. Consider for instance the graph in Figure \[fig:add-external-inlinks-conterexample2\]. With $c=0.85$ and ${\boldsymbol{z}}$ uniform, we have ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}=0.6$. But if we consider the graph defined by $\widetilde{{\mathcal{E}}}_{\mathrm{in({{{\mathcal{I}}}})}}={{\mathcal{E}}}_{\mathrm{in({{{\mathcal{I}}}})}}\cup{\{(4,3)\}}$, then we have ${\widetilde{{{\boldsymbol{\pi}}}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}=0.5897<{{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$.
Conclusions
===========
In this paper we provide the general shape of an optimal link structure for a website in order to maximize its PageRank. This structure with a forward chain and every possible backward links may be not intuitive. At our knowledge, it [has never been mentioned, while topologies]{} like a clique, a ring or a star are considered in the literature on collusion and alliance between pages [@BYCL05; @GGM05]. Moreover, this optimal structure gives new insight into the affirmation of Bianchini et al. [@BGS05] that, in order to maximize the PageRank of a website, hyperlinks to the rest of the webgraph “should be in pages with a small PageRank and that have many internal hyperlinks”. More precisely, we have seen that the leaking pages must be choosen with respect to the mean number of visits before zapping they give to the website, rather than their PageRank.
Let us now present some possible directions for future work.
We have noticed in Example \[ex:optimal-link-structure-conterexample\] that the first node of ${{{\mathcal{I}}}}$ in the forward chain of an optimal link structure is not [necessarily]{} *a child* of some node of ${{\overline{{{{\mathcal{I}}}}}}}$. In the example we gave, the personalization vector was not uniform. We wonder if this could occur with a uniform personalization vector and make the following conjecture.
Let ${\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}{}{\neq\emptyset}$ and ${\mathcal{E}}_{{\overline{{{{\mathcal{I}}}}}}}$ be given. Let ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ such that ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{{\mathcal{I}}}}$ is maximal under Assumption A. If ${\boldsymbol{z}}=\frac{1}{n}{{\boldsymbol{1}}}$, then there exists $j\in{{\overline{{{{\mathcal{I}}}}}}}$ such that $(j,i)\in{\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$, where $i\in\operatorname*{argmax}_{k}{\boldsymbol{v}}_k$.
If this conjecture was true we could also ask if the node $j\in{{\overline{{{{\mathcal{I}}}}}}}$ such that $(j,i)\in{\mathcal{E}}_{\mathrm{in({{{\mathcal{I}}}})}}$ where $i\in\operatorname*{argmax}_{k}{\boldsymbol{v}}_k$ belongs to ${{\mathcal{V}}}$.
Another question concerns the optimal linkage strategy in order to maximize an arbitrary linear combination of the PageRanks of the nodes of ${{{\mathcal{I}}}}$. In particular, we could want to maximize the PageRank ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{{\mathcal{S}}}$ of *a target subset* ${\mathcal{S}}\subseteq{{{\mathcal{I}}}}$ by choosing ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ as usual. A general shape for an optimal link structure seems difficult to find, as shown in the following example.
\[ex:target-set-examples\] Consider the graphs in Figure \[fig:target-set-examples\]. In both cases, let $c=0.85$ and ${\boldsymbol{z}}=\frac{1}{n}{{\boldsymbol{1}}}$. Let ${{{\mathcal{I}}}}={\{1,2,3\}}$ and let ${\mathcal{S}}={\{1,2\}}$ be the target set. In the configuration (a), the optimal sets of links ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ for maximizing ${{\boldsymbol{\pi}}}^T{\boldsymbol{e}}_{\mathcal{S}}$ has the link structure described in Theorem \[thm:website-opt\]. But in (a), the optimal ${\mathcal{E}}_{{{\mathcal{I}}}}$ and ${\mathcal{E}}_{\mathrm{out({{{\mathcal{I}}}})}}$ do not have this structure. Let us note nevertheless that, by Theorem \[thm:website-opt\], the subsets ${\mathcal{E}}_{{\mathcal{S}}}$ and ${\mathcal{E}}_{\mathrm{out}({\mathcal{S}})}$ must have the link structure described in Theorem \[thm:website-opt\].
Acknowledgements {#acknowledgements .unnumbered}
================
This paper presents research supported by a grant “Actions de recherche concertées – Large Graphs and Networks” of the “Communauté Française de Belgique” and by the Belgian Network DYSCO (Dynamical Systems, Control, and Optimization), funded by the Interuniversity Attraction Poles Programme, initiated by the Belgian State, Science Policy Office. The second author was supported by a research fellow grant of the “Fonds de la Recherche Scientifique – FNRS” (Belgium). The scientific responsibility rests with the authors.
[10]{}
Konstantin Avrachenkov and Nelly Litvak, *Decomposition of the [G]{}oogle [P]{}age[R]{}ank and optimal linking strategy*, Tech. report, INRIA, 2004, `http://www.inria.fr/rrrt/rr-5101.html`.
[to3em]{}, *The effect of new links on [G]{}oogle [P]{}age[R]{}ank*, Stoch. Models **22** (2006), no. 2, 319–331.
Ricardo Baeza-Yates, Carlos Castillo, and Vicente L[ó]{}pez, *[P]{}age[R]{}ank increase under different collusion topologies*, First International Workshop on Adversarial Information Retrieval on the Web, 2005, `http://airweb.cse.lehigh.edu/2005/baeza-yates.pdf`.
Abraham Berman and Robert J. Plemmons, *Nonnegative matrices in the mathematical sciences*, Classics in Applied Mathematics, vol. 9, Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 1994.
Monica Bianchini, Marco Gori, and Franco Scarselli, *Inside [P]{}age[R]{}ank*, ACM Trans. Inter. Tech. **5** (2005), no. 1, 92–128.
Sergey Brin and Lawrence Page, *The anatomy of a large-scale hypertextual web search engine*, Computer Networks and ISDN Systems **30** (1998), no. 1–7, 107–117, Proceedings of the Seventh International World Wide Web Conference, April 1998.
Grace E. Cho and Carl D. Meyer, *Markov chain sensitivity measured by mean first passage times*, Linear Algebra Appl. **316** (2000), no. 1-3, 21–28, Conference Celebrating the 60th Birthday of Robert J. Plemmons (Winston-Salem, NC, 1999).
Zoltán Gyöngyi and Hector Garcia-Molina, *Link spam alliances*, VLDB ’05: Proceedings of the 31st international conference on Very large data bases, VLDB Endowment, 2005, `http://portal.acm.org/citation.cfm?id=1083654`, pp. 517–528.
Ilse C. F. Ipsen and Rebecca S. Wills, *Mathematical properties and analysis of [G]{}oogle’s [P]{}age[R]{}ank*, Bol. Soc. Esp. Mat. Apl. **34** (2006), 191–196.
John G. Kemeny and J. Laurie Snell, *Finite [M]{}arkov chains*, The University Series in Undergraduate Mathematics, D. Van Nostrand Co., Inc., Princeton, N.J.-Toronto-London-New York, 1960.
Steve Kirkland, *Conditioning of the entries in the stationary vector of a [G]{}oogle-type matrix*, Linear Algebra Appl. **418** (2006), no. 2-3, 665–681.
Amy N. Langville and Carl D. Meyer, *Deeper inside [P]{}age[R]{}ank*, Internet Math. **1** (2004), no. 3, 335–380.
[to3em]{}, *Google’s [P]{}age[R]{}ank and beyond: the science of search engine rankings*, Princeton University Press, Princeton, NJ, 2006.
Ronny Lempel and Shlomo Moran, *Rank-stability and rank-similarity of link-based web ranking algorithms in authority-connected graphs.*, Inf. Retr. **8** (2005), no. 2, 245–264.
Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd, *The [P]{}age[R]{}ank citation ranking: Bringing order to the [W]{}eb*, Tech. report, Computer Science Department, Stanford University, 1998, `http://dbpubs.stanford.edu:8090/pub/1999-66`.
Eugene Seneta, *Nonnegative matrices and [M]{}arkov chains*, second ed., Springer Series in Statistics, Springer-Verlag, New York, 1981.
Marcin Sydow, *Can one out-link change your [P]{}age[R]{}ank?*, Advances in Web Intelligence (AWIC 2005), Lecture Notes in Computer Science, vol. 3528, Springer Berlin Heidelberg, 2005, pp. 408–414.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
A sign pattern matrix is a matrix whose entries are from the set $\{+,-,0\}$. If $A$ is an $m\times n$ sign pattern matrix, the qualitative class of $A$, denoted $Q(A)$, is the set of all real $m\times n$ matrices $B=[b_{i,j}]$ with $b_{i,j}$ positive (respectively, negative, zero) if $a_{i,j}$ is $+$ (respectively, $-$, $0$). The minimum rank of a sign pattern matrix $A$, denoted ${\mbox{mr}}(A)$, is the minimum of the ranks of the real matrices in $Q(A)$. Determination of the minimum rank of a sign pattern matrix is a longstanding open problem.
For the case that the sign pattern matrix has a $1$-separation, we present a formula to compute the minimum rank of a sign pattern matrix using the minimum ranks of certain generalized sign pattern matrices associated with the $1$-separation.
author:
- |
Marina Arav\
Frank J. Hall\
Zhongshan Li\
Hein van der Holst[^1]\
Lihua Zhang\
Wenyan Zhou\
Department of Mathematics and Statistics\
Georgia State University\
Atlanta, GA 30303, USA
title: 'The minimum rank of a sign pattern matrix with a $1$-separation'
---
Introduction {#introduction .unnumbered}
============
A *sign pattern matrix* (or *sign pattern*) is a matrix whose entries are from the set $\{+,-,0\}$. If $B=[b_{i,j}]$ is a real matrix, then ${\mbox{sgn}}(B)$ is the sign pattern matrix $A=[a_{i,j}]$ with $a_{i,j}=+$ (respectively, $-$, $0$) if $b_{i,j}$ is positive (respectively, negative, zero). If $A$ is a sign pattern matrix, the *sign pattern class* of $A$, denoted $Q(A)$, is the set of all real matrices $B=[b_{i,j}]$ with ${\mbox{sgn}}(B)=A$. The *minimum rank* of a sign pattern matrix $A$, denoted ${\mbox{mr}}(A)$, is the minimum of the ranks of matrices in $Q(A)$; see [@Hall07]. Recently, Li et al. [@LivdHolst2013] obtained a characterization of sign pattern matrices $A$ with ${\mbox{mr}}(A)\leq 2$. In this paper, we present a formula to compute the minimum rank of a sign pattern matrix with a $1$-separation using the minimum ranks of certain generalized sign pattern matrices associated with the $1$-separation.
The notion of sign pattern matrix can be extended to generalized sign pattern matrices by allowing certain entries to be $\#$; see [@Hall07]. For a generalized sign pattern matrix $A$, the generalized sign pattern class of $A$, denoted $Q(A)$, is defined by allowing entries of a matrix $B=[b_{i,j}]\in Q(A)$ to be any real number if the corresponding entries of $A$ are $\#$. The minimum rank ${\mbox{mr}}(A)$ of a generalized sign pattern matrix $A$ is defined in the same way as for a sign pattern matrix: ${\mbox{mr}}(A)$ is the minimum of the ranks of matrices in $Q(A)$. If $A=[a_{i,j}]$ and $C=[c_{i,j}]$ are generalized sign pattern matrices of the same size, we write $A\leq C$ if for each entry of $A$, $a_{i,j} = c_{i,j}$ or $c_{i,j} = \#$. It is clear that if $A\leq C$, then $Q(A)\subseteq Q(C)$. For a generalized sign pattern $C$, let $\mathcal{C}$ be the set of all sign pattern matrices $A$ such that $A\leq C$. Then, clearly, $Q(C) = \cup_{A\in \mathcal{C}} Q(A)$. Hence the minimum rank of a generalized sign pattern matrix $C$ equals $\min_{A\in \mathcal{C}} {\mbox{mr}}(A)$.
We define subtraction of two elements from $\{+,-,0\}$ as follows:
1. $(+)-(0)=+,(0)-(-)=+, (+)-(-)=+$,
2. $(-)-(+)=-,(0)-(+)=-,(-)-(0)=-$,
3. $(0)-(0)=0$,
4. $(+)-(+)=\#,(-)-(-)=\#$.
The idea behind the definition of, for example, $(-)-(+)=-$ is that subtracting a positive number from a negative number gives a negative number.
Let $$M = \begin{bmatrix}
A_{1,1} & A_{1,2} & 0\\
A_{2,1} & a_{2,2} & A_{2,3}\\
0 & A_{3,2} & A_{3,3}
\end{bmatrix}$$ be a sign pattern matrix, where $A_{1,2}$ has only one column and $A_{2,1}$ only one row. We also say that the sign pattern matrix $M$ has a *$1$-separation*. For $p\in \{+,-,0\}$, let $$R_p = \begin{bmatrix}
A_{1,1} & A_{1,2}\\
A_{2,1} & p
\end{bmatrix}
\text{ and }
S_p = \begin{bmatrix}
a_{2,2}-p & A_{2,3}\\
A_{3,2} & A_{3,3}
\end{bmatrix}.$$ (Indeed $S_p$ might be a generalized sign pattern matrix.) In this paper, we prove that the following formula holds: $$\label{mainformula}
\begin{split}
{\mbox{mr}}(M) = \min \{ & {\mbox{mr}}(A_{1,1})+ {\mbox{mr}}(A_{3,3})+2,\\
& {\mbox{mr}}([A_{1,1}\;A_{1,2}])+{\mbox{mr}}([A_{3,2}\;A_{3,3}])+1,\\
& {\mbox{mr}}(\begin{bmatrix}
A_{1,1}\\
A_{2,1}
\end{bmatrix})+{\mbox{mr}}(\begin{bmatrix}
A_{2,3}\\
A_{3,3}
\end{bmatrix})+1,\\
& {\mbox{mr}}(R_+)+{\mbox{mr}}(S_+),\\
& {\mbox{mr}}(R_0)+{\mbox{mr}}(S_0),\\
& {\mbox{mr}}(R_-)+{\mbox{mr}}(S_-)
\}
\end{split}$$ In the next section, we show that each of the terms in the minimum is at least ${\mbox{mr}}(M)$. In Section \[sec:minrank1sep\], we show that at least one of the terms in the minimum attains ${\mbox{mr}}(M)$.
Formula (\[mainformula\]) is analogous to the formula for the minimum rank of $1$-sums of graphs. This formula was given by Hsieh [@Hsieh2001], and, independently, by Barioli, Fallat, and Hogben [@BarFalHog2004a]. In case the graph is permitted to have loops, a formula was given by Mikkelson [@Mikkelson:2008aa].
The reader can see Fallat and Hogben [@FalHog2007] for a survey on the minimum ranks of graphs.
Inequalities
============
Let $0\leq k\leq m,n,r,s$ and let $$A=\begin{bmatrix}
A_{1,1}&A_{1,2}\\
A_{2,1}&A_{2,2}
\end{bmatrix} \text{ and } B=\begin{bmatrix}
B_{1,1}&B_{1,2}\\
B_{2,1}&B_{2,2}
\end{bmatrix}$$ be $m\times n$ and $r\times s$ matrices with real entries, respectively, where $A_{2,2}$ and $B_{1,1}$ are $k\times k$. In [@FalJoh1999a], the *$k-$subdirect sum* of $A$ and $B$, denoted by $A\oplus_k B$, was introduced; this is the matrix $$A \oplus_{k} B=
\begin{bmatrix}
A_{1,1}&A_{1,2}&0\\
A_{2,1}&A_{2,2}+B_{1,1} & B_{1,2}\\
0 & B_{2,1} & B_{2,2}
\end{bmatrix}$$
[@MR3010007]\[lem:sum\] Let $$C = \begin{bmatrix}
C_{1,1} & C_{1,2}\\
C_{2,1} & C_{2,2}
\end{bmatrix}
\quad\text{and}\quad
D = \begin{bmatrix}
D_{1,1} & D_{1,2}\\
D_{2,1} & D_{2,2}
\end{bmatrix},$$ where $C_{2,2}$ and $D_{1,1}$ are $k\times k$ matrices. Then $\operatorname{rank}(C\oplus_k D)\leq \operatorname{rank}(C\oplus D)$.
In the next theorem, we show that each term in the minimum of Formula (\[mainformula\]) is at least ${\mbox{mr}}(M)$.
Let $$M=\begin{bmatrix}
A_{1,1}&A_{1,2}&0\\
A_{2,1}& a_{2,2} & A_{2,3}\\
0 & A_{3,2} & A_{3,3}\\
\end{bmatrix}$$ be a sign pattern matrix, where $A_{1,1}$ is $m_1\times n_1$, $A_{1,2}$ is $m_1\times 1$, $A_{2,1}$ is $1\times n_1$, $a_{2,2}$ is $1\times 1$, $A_{2,3}$ is $1\times n_2$, $A_{3, 2}$ is $m_2\times 1$ and $A_{3,3}$ is $m_2\times n_2$. Then each of the following inequalities hold:
(i) ${\mbox{mr}}(A_{1,1})+{\mbox{mr}}(A_{3,3})+2\geq {\mbox{mr}}(M)$,
(ii) ${\mbox{mr}}([A_{1,1}\;A_{1,2}])+{\mbox{mr}}([A_{3,2}\;A_{3,3}])+1\geq {\mbox{mr}}(M)$,
(iii) ${\mbox{mr}}(\begin{bmatrix}
A_{1,1}\\
A_{2,1}\\
\end{bmatrix})+{\mbox{mr}}(\begin{bmatrix}
A_{2,3}\\
A_{3,3}\\
\end{bmatrix})+1\geq {\mbox{mr}}(M)$, and
(iv) for each $p\in \{+,-,0\}$, $${\mbox{mr}}(\begin{bmatrix}
A_{1,1} & A_{1,2}\\
A_{2,1} & p
\end{bmatrix}
)+{\mbox{mr}}(
\begin{bmatrix}
a_{2,2}-p & A_{2,3}\\
A_{3,2} & A_{3,3}
\end{bmatrix}
)\geq {\mbox{mr}}(M).$$
To see that ${\mbox{mr}}(A_{1,1})+{\mbox{mr}}(A_{3,3})+2\geq {\mbox{mr}}(M)$, let $C_{1,1}\in Q(A_{1,1})$ and $C_{3,3} \in Q(A_{3,3})$ such that $\operatorname{rank}(C_{1,1}) = {\mbox{mr}}(A_{1,1})$ and $\operatorname{rank}(C_{3,3})={\mbox{mr}}(A_{3,3})$. Let $C_{1,2}\in Q(A_{1,2})$, $C_{2,1}\in Q(A_{2,1})$, $C_{2,3} \in Q(A_{2,3})$, $C_{3,2}\in Q(A_{3,2})$, and ${\mbox{sgn}}(c_{2,2}) = a_{2,2}$. Then $$\begin{split}
{\mbox{mr}}(A_{1,1}) + {\mbox{mr}}(A_{3,3})+2 & = \operatorname{rank}(\begin{bmatrix}
C_{1,1} & 0\\
0 & C_{3,3}
\end{bmatrix}) + 2\\
&\geq \operatorname{rank}(\begin{bmatrix}
C_{1,1} & C_{1,2} & 0\\
C_{2,1} & c_{2,2} & C_{2,3}\\
0 & C_{3,2} & C_{3,3}
\end{bmatrix})\\
&\geq {\mbox{mr}}(M),
\end{split}$$
To see that ${\mbox{mr}}([A_{1,1}\;A_{1,2}])+{\mbox{mr}}([A_{3,2}\;A_{3,3}])+1\geq {\mbox{mr}}(M)$, let $[C_{1,1}\;C_{1,2}]\in Q([A_{1,1}\;A_{1,2}])$ and $[C_{3,2}\;C_{3,3}]\in Q([A_{3,2}\;B_{3,3}])$ be such that $\operatorname{rank}([C_{1,1}\;C_{1,2}])={\mbox{mr}}([A_{1,1}\;A_{1,2}])$ and $\operatorname{rank}([C_{3,2}\;C_{3,3}])={\mbox{mr}}([A_{3,2}\;A_{3,3}])$. Clearly, $$\operatorname{rank}([C_{1,1}\;C_{1,2}]) + \operatorname{rank}([C_{3,2}\;C_{3,3}])\geq \operatorname{rank}(\begin{bmatrix}
C_{1,1} & C_{1,2} & 0\\
0 & C_{3,2} & C_{3,3}
\end{bmatrix}).$$ Since $$\operatorname{rank}(\begin{bmatrix}
C_{1,1} & C_{1,2} & 0\\
0 & C_{3,2} & C_{3,3}
\end{bmatrix}) + 1\geq \operatorname{rank}(
\begin{bmatrix}
C_{1,1} & C_{1,2} & 0\\
C_{2,1} & c_{2,2} & C_{2,3}\\
0 & C_{3,2} & C_{3,3}
\end{bmatrix}),$$ we obtain ${\mbox{mr}}([A_{1,1}\;A_{1,2}])+{\mbox{mr}}([A_{3,2}\;A_{3,3}])+1\geq {\mbox{mr}}(M)$.
The proof that ${\mbox{mr}}(\begin{bmatrix}
A_{1,1}\\
A_{2,1}\\
\end{bmatrix})+{\mbox{mr}}(\begin{bmatrix}
B_{1,2}\\
B_{2,2}\\
\end{bmatrix})+1\geq {\mbox{mr}}(M)$ is similar to the proof of the previous case.
Let $p\in \{+,-,0\}$. To shorten notation, let $$R_p = \begin{bmatrix}
A_{1,1} & A_{1,2}\\
A_{2,1} & p
\end{bmatrix}
\text{ and }
S_p = \begin{bmatrix}
a_{2,2}-p & A_{2,3}\\
A_{3,2} & A_{3,3}
\end{bmatrix}.$$ To see that ${\mbox{mr}}(R_p) + {\mbox{mr}}(S_p) \geq {\mbox{mr}}(M)$, let $$C = \begin{bmatrix}
C_{1,1} & C_{1,2}\\
C_{2,1} & c
\end{bmatrix}\in Q(R_p)
\text{ and }
D = \begin{bmatrix}
d & C_{2,3}\\
C_{3,2} & C_{3,3}
\end{bmatrix}\in Q(S_p)$$ be such that $\operatorname{rank}(C) = {\mbox{mr}}(R_p)$ and $\operatorname{rank}(D)={\mbox{mr}}(S_p)$.
We now do a case-checking.
Suppose first that $a_{2,2}-p=0$. Then $p=0$ and $a_{2,2}=0$. Hence $c=0$ and $d=0$. Then $C\oplus_1 D \in Q(M)$, and, by Lemma \[lem:sum\], ${\mbox{mr}}(M)\leq \operatorname{rank}(C\oplus_1 D)\leq \operatorname{rank}(C)+\operatorname{rank}(D) = {\mbox{mr}}(R_p) + {\mbox{mr}}(S_p)$.
Suppose next that $a_{2,2}-p=+$. Then one of the following holds:
1. $p=-$ and $a_{2,2}=0$,
2. $p=0$ and $a_{2,2}=+$, and
3. $p=-$ and $a_{2,2}=+$.
Suppose $p=-$ and $a_{2,2}=0$. By scaling $D$ by a positive scalar, we may assume that $d=-c$. Then $C\oplus_1 D\in Q(M)$, and, by Lemma \[lem:sum\], ${\mbox{mr}}(M) \leq \operatorname{rank}(C\oplus_1 D)\leq \operatorname{rank}(C)+\operatorname{rank}(D) = {\mbox{mr}}(R_p) + {\mbox{mr}}(S_p)$. Suppose $p=0$ and $a_{2,2}=+$. Then $C\oplus_1 D\in Q(M)$, and, by Lemma \[lem:sum\], ${\mbox{mr}}(M) \leq \operatorname{rank}(C\oplus_1 D)\leq \operatorname{rank}(C)+\operatorname{rank}(D) = {\mbox{mr}}(R_p) + {\mbox{mr}}(S_p)$. Suppose $p=-$ and $a_{2,2}=+$. By scaling $D$ by a positive scalar, we may assume that $c+d > 0$. Then $C\oplus_1 D\in Q(M)$, and, by Lemma \[lem:sum\], ${\mbox{mr}}(M)\leq \operatorname{rank}(C\oplus_1 D)\leq \operatorname{rank}(C)+\operatorname{rank}(D) = {\mbox{mr}}(R_p) + {\mbox{mr}}(S_p)$.
The case where $a_{2,2}-p=-$ is similar.
Suppose finally that $a_{2,2}-p=\#$. Then one of the following holds:
1. $p=+$ and $a_{2,2}=+$, and
2. $p=-$ and $a_{2,2}=-$.
Suppose $p=+$ and $a_{2,2}=+$. Then $C\oplus_1 D\in Q(M)$, and, by Lemma \[lem:sum\], ${\mbox{mr}}(M) \leq \operatorname{rank}(C\oplus_1 D)\leq \operatorname{rank}(C)+\operatorname{rank}(D) = {\mbox{mr}}(R_p) + {\mbox{mr}}(S_p)$. The case where $p=-$ and $a_{2,2}=-$ is similar.
Minimum Rank of Sign Pattern with 1-Separation {#sec:minrank1sep}
==============================================
In this section we finish the proof that Formula (\[mainformula\]) is correct. First we prove some lemmas.
\[lem:vertexadd\] For any $m\times n$ real matrix $B$ with $m, n\geq 1$, and any nonzero real numbers $a$ and $c$, $$\operatorname{rank}(\begin{bmatrix}
0 & a & 0\\
c & b_{1,1} & B_{1,2}\\
0 & B_{2,1} & B_{2,2}
\end{bmatrix}) = \operatorname{rank}(B_{2,2}) + 2.$$
Let $$P = \begin{bmatrix}
\frac{1}{a} & 0 & 0\\
-\frac{b_{1,1}}{2a} & 1 & 0\\
-\frac{B_{2,1}}{a} & 0 & I_{m-2}
\end{bmatrix}\quad\text{and}\quad
Q = \begin{bmatrix}
\frac{1}{c} & -\frac{b_{1,1}}{2c} & -\frac{B_{1,2}}{c}\\
0 & 1 & 0\\
0 & 0 & I_{n-2}
\end{bmatrix}.$$ Then $$P \begin{bmatrix}
0 & a & 0\\
c & b_{1,1} & B_{1,2}\\
0 & B_{2,1} & B_{2,2}
\end{bmatrix} Q =
\begin{bmatrix}
0 & 1 & 0\\
1 & 0 & 0\\
0 & 0 & B_{2,2}
\end{bmatrix}.$$ From this the lemma easily follows.
\[lem:adjoin\] Let $A = \begin{bmatrix}
A_{1,1} & A_{1,2}\\
A_{2,1} & A_{2,2}
\end{bmatrix}$ be a real matrix, where $A_{1,1}$ is $m_1\times n_1$, $A_{1,2}$ is $m_1\times n_2$, $A_{2,1}$ is $m_2\times n_1$, and $A_{2,2}$ is $m_2\times n_2$. If $x\in \ker(A_{2,2}^T)$ and $y\in \ker(A_{2,2})$, then $$\operatorname{rank}\begin{bmatrix}
0 & x^T A_{2,1} & 0\\
A_{1,2} y & A_{1,1} & A_{1,2}\\
0 & A_{2,1} & A_{2,2}
\end{bmatrix} = \operatorname{rank}A.$$
Let $$P = \begin{bmatrix}
0 & x^T\\
I_{m_1} & 0\\
0 & I_{m_2}
\end{bmatrix}\quad\text{and}\quad
Q= \begin{bmatrix}
0 & I_{n_1} & 0\\
y & 0 & I_{n_2}
\end{bmatrix}.$$ Then $$P A Q= \begin{bmatrix}
0 & x^T A_{2,1} & 0\\
A_{1,2} y & A_{1,1} & A_{1,2}\\
0 & A_{2,1} & A_{2,2}
\end{bmatrix}.$$ Hence, $\operatorname{rank}\begin{bmatrix}
0 & x^T A_{2,1} & 0\\
A_{1,2} y & A_{1,1} & A_{1,2}\\
0 & A_{2,1} & A_{2,2}
\end{bmatrix} \leq \operatorname{rank}A$. The other inequality is clear.
\[lem:decomp\] Let $A = \begin{bmatrix}
A_{1,1} & A_{1,2} & 0\\
A_{2,1} & a_{2,2} & A_{2,3}\\
0 & A_{3,2} & A_{3,3}
\end{bmatrix}$ be an $m\times n$ real matrix, where $A_{1,1}$ is $m_1\times n_1$ and $A_{3,3}$ is $m_2\times n_2$, (and so $m=m_1+m_2+1$ and $n=n_1+n_2+1$). Then at least one of the following holds:
(i) There exist vectors $v \in \mathbb{R}^{m_1}$ and $z\in \mathbb{R}^{n_1}$ such that $$\operatorname{rank}(\begin{bmatrix}
A_{1,1} & A_{1,2}\\
A_{2,1} & v^T A_{1,1} z
\end{bmatrix})+\operatorname{rank}(
\begin{bmatrix}
a_{2,2} - v^T A_{1,1} z & A_{2,3}\\
A_{3,2} & A_{3,3}
\end{bmatrix}) = \operatorname{rank}(A).$$
(ii) \[item:sumcase2\] $\operatorname{rank}(\begin{bmatrix}
A_{1,1}\\
A_{2,1}
\end{bmatrix}) + \operatorname{rank}(\begin{bmatrix}
A_{2,3}\\
A_{3,3}
\end{bmatrix})+1 = \operatorname{rank}(A)$.
(iii) \[item:sumcase3\] $\operatorname{rank}([A_{1,1}\;A_{1,2}]) + \operatorname{rank}([A_{3,2}\;A_{3,3}]) +1 = \operatorname{rank}(A)$.
(iv) $\operatorname{rank}(A_{1,1}) + \operatorname{rank}(A_{3,3}) + 2=\operatorname{rank}(A)$.
Suppose first that $[A_{2,1}\;A_{2,3}]x = 0$ for all $x\in \ker(A_{1,1}\oplus A_{3,3})$ and that $y^T \begin{bmatrix}
A_{1,2}\\
A_{3,2}
\end{bmatrix} = 0$ for all $y\in \ker((A_{1,1}\oplus A_{3,3})^T)$. Then there exist a vector $v\in \mathbb{R}^{m_1}$ such that $v^T A_{1,1} = A_{2,1}$ and a vector $z\in \mathbb{R}^{n_1}$ such that $A_{1,1} z = A_{1,2}$. Let $$P = \begin{bmatrix}
I_{m_1} & 0 & 0\\
v^T & 0 & 0\\
-v^T & 1 & 0\\
0 & 0 & I_{m_2}
\end{bmatrix}\quad\text{and}\quad
Q=\begin{bmatrix}
I_{n_1} & z & -z & 0\\
0 & 0 & 1 & 0\\
0 & 0 & 0 & I_{n_2}
\end{bmatrix}.$$ A calculation shows that $$P A Q = \begin{bmatrix}
A_{1,1} & A_{1,2}\\
A_{2,1} & v^T A_{1,1} z
\end{bmatrix}\oplus
\begin{bmatrix}
a_{2,2} - v^T A_{1,1} z & A_{2,3}\\
A_{3,2} & A_{3,3}
\end{bmatrix}.$$ Hence $$\operatorname{rank}(A) \geq \operatorname{rank}(\begin{bmatrix}
A_{1,1} & A_{1,2}\\
A_{2,1} & v^T A_{1,1} z
\end{bmatrix}) + \operatorname{rank}(
\begin{bmatrix}
a_{2,2} - v^T A_{1,1} z & A_{2,3}\\
A_{3,2} & A_{3,3}
\end{bmatrix})$$ By Lemma \[lem:sum\], the opposite inequality also holds.
Suppose next that $[A_{2,1}\;A_{2,3}]x = 0$ for all $x\in \ker(A_{1,1}\oplus A_{3,3})$ and that there exists a $y\in \ker((A_{1,1}\oplus A_{3,3})^T)$ such that $y^T \begin{bmatrix}
A_{1,2}\\
A_{3,2}
\end{bmatrix} = e \not= 0$. By Lemma \[lem:adjoin\], $$\operatorname{rank}(\begin{bmatrix}
0 & 0 & e & 0\\
0 & A_{1,1} & A_{1,2} & 0\\
0 & A_{2,1} & a_{2,2} & A_{2,3}\\
0 & 0 & A_{3,2} & A_{3,3}
\end{bmatrix}) = \operatorname{rank}(A).$$ Hence $$1 + \operatorname{rank}(
\begin{bmatrix}
A_{1,1} & 0\\
A_{2,1} & A_{2,3}\\
0 & A_{3,3}
\end{bmatrix}) = \operatorname{rank}(A).$$ Since $${\mbox{nullity }}(\begin{bmatrix}
A_{1,1} & 0\\
A_{2,1} & A_{2,3}\\
0 & A_{3,3}
\end{bmatrix}) =
{\mbox{nullity }}(
\begin{bmatrix}
A_{1,1} & 0\\
0 & A_{3,3}
\end{bmatrix}),$$ we obtain $$\operatorname{rank}(\begin{bmatrix}
A_{1,1} & 0\\
A_{2,1} & A_{2,3}\\
0 & A_{3,3}
\end{bmatrix}) =
\operatorname{rank}(\begin{bmatrix}
A_{1,1} & 0\\
0 & A_{3,3}
\end{bmatrix}).$$ Hence $\operatorname{rank}(A) = \operatorname{rank}(A_{1,1}) + \operatorname{rank}(A_{3,3})+1$. From $[A_{2,1}\;A_{2,3}]x = 0$ for all $x\in \ker(A_{1,1}\oplus A_{3,3})$, it follows that $\operatorname{rank}(\begin{bmatrix}
A_{1,1}\\
A_{1,2}
\end{bmatrix})
= \operatorname{rank}(A_{1,1})$ and $\operatorname{rank}(
\begin{bmatrix}
A_{2,3}\\
A_{3,3}
\end{bmatrix}) = \operatorname{rank}(A_{3,3})$. Thus $\operatorname{rank}(\begin{bmatrix}
A_{1,1}\\
A_{2,1}
\end{bmatrix}) + \operatorname{rank}(\begin{bmatrix}
A_{2,3}\\
A_{3,3}
\end{bmatrix})+1 = \operatorname{rank}(A)$.
The case that there exists an $x\in \ker(A_{1,1}\oplus A_{3,3})$ such that $[A_{2,1}\;A_{2,3}]x$ is nonzero and $y^T \begin{bmatrix}
A_{1,2}\\
A_{3,2}
\end{bmatrix} = 0$ for all $y\in \ker((A_{1,1}\oplus A_{3,3})^T)$ yields $\operatorname{rank}([A_{1,1}\;A_{1,2}]) + \operatorname{rank}([A_{3,2}\;A_{3,3}]) +1 = \operatorname{rank}(A)$.
Hence, we are left with the case that there exist an $x\in \ker(A_{1,1}\oplus A_{3,3})$ such that $f = [A_{2,1}\;A_{2,3}]x$ is nonzero and there exists a $y\in \ker((A_{1,1}\oplus A_{3,3})^T)$ such that $e = y^T \begin{bmatrix}
A_{1,2}\\
A_{3,2}
\end{bmatrix}$ is nonzero. Then, by Lemma \[lem:adjoin\], $$\operatorname{rank}(\begin{bmatrix}
0 & 0 & e & 0\\
0 & A_{1,1} & A_{1,2} & 0\\
f & A_{2,1} & a_{2,2} & A_{2,3}\\
0 & 0 & A_{3,2} & A_{3,3}
\end{bmatrix}) = \operatorname{rank}(A).$$ By Lemma \[lem:vertexadd\], $$\operatorname{rank}(\begin{bmatrix}
A_{1,1} & 0\\
0 & A_{3,3}
\end{bmatrix}) + 2 = \operatorname{rank}(A).$$ Thus $\operatorname{rank}(A_{1,1}) + \operatorname{rank}(A_{3,3}) + 2 = \operatorname{rank}(A)$.
The proof shows that if Case (\[item:sumcase2\]) happens, then $A_{2,1}$ belongs to the row space of $A_{1,1}$, and $A_{2,3}$ belongs to the row space of $A_{3,3}$. Similarly, if Case (\[item:sumcase3\]) happens, then $A_{1,2}$ belongs to the column space of $A_{1,1}$, and $A_{2,3}$ belongs to the column space of $A_{3,3}$.
Let $$M = \begin{bmatrix}
A_{1,1} & A_{1,2}& 0\\
A_{2,1} & a_{2,2} & A_{2,3}\\
0 & A_{3,2} & A_{3,3}
\end{bmatrix},$$ where $A_{1,2}$ is $n_1\times 1$, $A_{2,1}$ is $m_1\times 1$, and, for $p\in \{+,-,0\}$, let $$R_p = \begin{bmatrix}
A_{1,1} & A_{1,2}\\
A_{2,1} & p
\end{bmatrix}
\text{ and }
S = \begin{bmatrix}
a_{2,2}-p & A_{2,3}\\
A_{3,2} & A_{3,3}
\end{bmatrix}.$$ Then $$\begin{split}
{\mbox{mr}}(M) = \min \{ & {\mbox{mr}}(A_{1,1})+ {\mbox{mr}}(A_{3,3})+2,\\
& {\mbox{mr}}([A_{1,1}\;A_{1,2}])+{\mbox{mr}}([A_{3,2}\;A_{3,3}])+1,\\
& {\mbox{mr}}(\begin{bmatrix}
A_{1,1}\\
A_{2,1}
\end{bmatrix})+{\mbox{mr}}(\begin{bmatrix}
A_{2,3}\\
A_{3,3}
\end{bmatrix})+1,\\
& {\mbox{mr}}(R_+)+{\mbox{mr}}(S_+),\\
& {\mbox{mr}}(R_0)+{\mbox{mr}}(S_0),\\
& {\mbox{mr}}(R_-)+{\mbox{mr}}(S_-)
\}
\end{split}$$
By the previous section, $$\label{eq:formula2}
\begin{split}
{\mbox{mr}}(M) \leq \min \{ & {\mbox{mr}}(A_{1,1})+ {\mbox{mr}}(A_{3,3})+2,\\
& {\mbox{mr}}([A_{1,1}\;A_{1,2}])+{\mbox{mr}}([A_{3,2}\;A_{3,3}])+1,\\
& {\mbox{mr}}(\begin{bmatrix}
A_{1,1}\\
A_{2,1}
\end{bmatrix})+{\mbox{mr}}(\begin{bmatrix}
A_{2,3}\\
A_{3,3}
\end{bmatrix})+1,\\
& {\mbox{mr}}(R_+)+{\mbox{mr}}(S_+),\\
& {\mbox{mr}}(R_0)+{\mbox{mr}}(S_0),\\
& {\mbox{mr}}(R_-)+{\mbox{mr}}(S_-)
\}
\end{split}$$ We now show that at least one of the terms in the minimum on the right-hand side of $(\ref{eq:formula2})$ equals ${\mbox{mr}}(M)$.
Let $$C = \begin{bmatrix}
C_{1,1} & C_{1,2} & 0\\
C_{2,1} & c_{2,2} & C_{2,3}\\
0 & C_{3,2} & C_{3,3}
\end{bmatrix}\in Q(M)$$ be such that $\operatorname{rank}(C) = {\mbox{mr}}(M)$. Then, by Lemma \[lem:decomp\], at least one of the following holds:
(i) \[item1\] There exist vectors $v\in{\ensuremath{\mathbb{R}}}^{n_1}$ and $z\in{\ensuremath{\mathbb{R}}}^{m_1}$ such that $$\operatorname{rank}(\begin{bmatrix}
C_{1,1} & C_{1,2}\\
C_{2,1} & v^T C_{1,1} z
\end{bmatrix})+\operatorname{rank}(
\begin{bmatrix}
c_{2,2} - v^T C_{1,1} z & C_{2,3}\\
C_{3,2} & C_{3,3}
\end{bmatrix}) = \operatorname{rank}(C).$$
(ii) \[item2\] $\operatorname{rank}(\begin{bmatrix}
C_{1,1}\\
C_{2,1}
\end{bmatrix}) + \operatorname{rank}(\begin{bmatrix}
C_{2,3}\\
C_{3,3}
\end{bmatrix})+1 = \operatorname{rank}(C)$.
(iii) \[item3\] $\operatorname{rank}([C_{1,1}\;C_{1,2}]) + \operatorname{rank}([C_{3,2}\;C_{3,3}]) +1 = \operatorname{rank}(C)$.
(iv) \[item4\] $\operatorname{rank}(C_{1,1}) + \operatorname{rank}(C_{3,3}) + 2=\operatorname{rank}(C)$.
Suppose first that $(\ref{item2})$ holds. Then $$\begin{split}
{\mbox{mr}}(\begin{bmatrix}
A_{1,1}\\
A_{2,1}
\end{bmatrix}) + {\mbox{mr}}(\begin{bmatrix}
A_{2,3}\\
A_{3,3}
\end{bmatrix})+1 & \leq \operatorname{rank}(\begin{bmatrix}
C_{1,1}\\
C_{2,1}
\end{bmatrix}) + \operatorname{rank}(\begin{bmatrix}
C_{2,3}\\
C_{3,3}
\end{bmatrix})+1\\
& = \operatorname{rank}(C) = {\mbox{mr}}(M).
\end{split}$$
Case $(\ref{item3})$ is similar to $(\ref{item2})$.
Suppose next that $(\ref{item4})$ holds. Then $$\begin{split}
{\mbox{mr}}(A_{1,1}) + {\mbox{mr}}(A_{3,3}) + 2 & \leq \operatorname{rank}(C_{1,1}) + \operatorname{rank}(C_{3,3}) + 2 \\
& = \operatorname{rank}(C) = {\mbox{mr}}(M).
\end{split}$$
Suppose finally that $(\ref{item1})$ holds. If $v^T C_{1,1} z > 0$, then $$\begin{bmatrix}
C_{1,1} & C_{1,2}\\
C_{2,1} & v^T C_{1,1} z
\end{bmatrix}\in Q(R_+)\quad\text{and}\quad
\begin{bmatrix}
c_{2,2} - v^T C_{1,1} z & C_{2,3}\\
C_{3,2} & C_{3,3}
\end{bmatrix}\in Q(S_+).$$ Hence $$\begin{split}
{\mbox{mr}}(M) = \operatorname{rank}(C) &= \operatorname{rank}(\begin{bmatrix}
C_{1,1} & C_{1,2}\\
C_{2,1} & v^T C_{1,1} z
\end{bmatrix})+\operatorname{rank}(
\begin{bmatrix}
c_{2,2} - v^T C_{1,1} z & C_{2,3}\\
C_{3,2} & C_{3,3}
\end{bmatrix}) \\
&\geq {\mbox{mr}}(R_+) + {\mbox{mr}}(S_+).
\end{split}$$ The cases where $v^T C_{1,1} z = 0$ and $v^T C_{1,1} z < 0$ are similar.
Examples
========
We exhibit several examples of sign pattern matrices illustrating that each term in Formula (\[mainformula\]) is needed.
To see that the term ${\mbox{mr}}(A_{1,1})+{\mbox{mr}}(A_{3,3})+2$ is needed in Formula (\[mainformula\]), let $$M = \begin{bmatrix}
A_{1,1} & A_{1,2}& 0\\
A_{2,1} & a_{2,2} & A_{2,3}\\
0 & A_{3,2} & A_{3,3}
\end{bmatrix}=\begin{bmatrix}
0 & + & 0\\
+ & 0 & +\\
0 & + & 0
\end{bmatrix},$$ and, for $p\in \{+,-,0\}$, let $$R_p = \begin{bmatrix}
0 & +\\
+ & p
\end{bmatrix}
\text{ and }
S_p = \begin{bmatrix}
-p & +\\
+ & 0
\end{bmatrix}.$$ Observe that ${\mbox{mr}}(M)=2$. Note that ${\mbox{mr}}(A_{1,1})+ {\mbox{mr}}(A_{3,3})+2=0+0+2=2$, while ${\mbox{mr}}([A_{1,1}\;A_{1,2}])+{\mbox{mr}}([A_{3,2}\;A_{3,3}])+1=1+1+1=3$, ${\mbox{mr}}(\begin{bmatrix}
A_{1,1}\\
A_{2,1}
\end{bmatrix})+{\mbox{mr}}(\begin{bmatrix}
A_{2,3}\\
A_{3,3}
\end{bmatrix})+1=1+1+1=3$, ${\mbox{mr}}(R_+)+{\mbox{mr}}(S_+)=2+2=4$, ${\mbox{mr}}(R_0)+{\mbox{mr}}(S_0)=2+2=4$, and ${\mbox{mr}}(R_-)+{\mbox{mr}}(S_-)=2+2=4$.
To see that the term ${\mbox{mr}}(\begin{bmatrix}
A_{1,1} & A_{1,2}
\end{bmatrix}) +
{\mbox{mr}}(\begin{bmatrix}
A_{3,2}& A_{3,3}
\end{bmatrix})+1$ is needed in Formula (\[mainformula\]), let $$M = \begin{bmatrix}
A_{1,1} & A_{1,2} & 0\\
A_{2,1} & a_{2,2} & A_{2,3}\\
0 & A_{3,2} & A_{3,3}
\end{bmatrix} =
\begin{bmatrix}
+ & + & 0 & 0 & 0\\
+ & + & 0 & 0 & 0\\
0 & + & + & + & 0\\
0 & 0 & + & 0 & +\\
\end{bmatrix},$$ and let $$R_p = \begin{bmatrix}
A_{1,1} & A_{1,2}\\
A_{2,1} & p
\end{bmatrix}=
\begin{bmatrix}
+ & + & 0\\
+ & + & 0\\
0 & + & p
\end{bmatrix}$$ and $$S_p = \begin{bmatrix}
a_{2,2}-p & A_{2,3}\\
A_{3,2} & A_{3,3}
\end{bmatrix}=
\begin{bmatrix}
(+)-p & + & 0\\
+ & 0 & +
\end{bmatrix}.$$ Observe that ${\mbox{mr}}(M) = 3$. Note that ${\mbox{mr}}(A_{1,1}) + {\mbox{mr}}(A_{3,3}) + 2 = 1+1+2 = 4$, ${\mbox{mr}}([A_{1,1}\;A_{1,2}]) + {\mbox{mr}}([A_{3,2}\;A_{3,3}]) + 1 = 1 + 1 + 1 = 3, {\mbox{mr}}(\begin{bmatrix}
A_{1,1}\\
A_{2,1}
\end{bmatrix}) +
{\mbox{mr}}(\begin{bmatrix}
A_{2,3}\\
A_{3,3}
\end{bmatrix})+1 = 2 + 2 + 1 = 5$, ${\mbox{mr}}(R_+) + {\mbox{mr}}(S_+) = 2 + 2 = 4$, ${\mbox{mr}}(R_0) + {\mbox{mr}}(S_0) = 2 + 2 = 4$, ${\mbox{mr}}(R_-) + {\mbox{mr}}(S_-) = 2+2=4$.
Taking the transpose of $M$ in the previous example shows that the term ${\mbox{mr}}(\begin{bmatrix}
A_{1,1}\\
A_{2,1}
\end{bmatrix}) +
{\mbox{mr}}(\begin{bmatrix}
A_{2,3}\\
A_{3,3}
\end{bmatrix})+1$ is needed in Formula (\[mainformula\]).
To see that the term ${\mbox{mr}}(R_+)+{\mbox{mr}}(S_+)$ is needed in Formula (\[mainformula\]), let $$M = \begin{bmatrix}
A_{1,1} & A_{1,2}& 0\\
A_{2,1} & a_{2,2} & A_{2,3}\\
0 & A_{3,2} & A_{3,3}
\end{bmatrix}=\begin{bmatrix}
+ & + & 0\\
+ & - & -\\
0 & + & +
\end{bmatrix},$$ and let $$R_p = \begin{bmatrix}
A_{1,1} & A_{1,2}\\
A_{2,1} & p
\end{bmatrix}=
\begin{bmatrix}
+ & +\\
+ & p
\end{bmatrix}$$ and $$S_p = \begin{bmatrix}
a_{2,2}-p & A_{2,3}\\
A_{3,2} & A_{3,3}
\end{bmatrix}=
\begin{bmatrix}
(-)-p & -\\
+ & +
\end{bmatrix}.$$
Observe that ${\mbox{mr}}(M)=2$. Note that ${\mbox{mr}}(A_{1,1})+ {\mbox{mr}}(A_{3,3})+2=1+1+2=4, {\mbox{mr}}([A_{1,1}\;A_{1,2}])+{\mbox{mr}}([A_{3,2}\;A_{3,3}])+1=1+1+1=3,{\mbox{mr}}(\begin{bmatrix}
A_{1,1}\\
A_{2,1}
\end{bmatrix})+{\mbox{mr}}(\begin{bmatrix}
A_{2,3}\\
A_{3,3}
\end{bmatrix})+1=1+1+1=3, {\mbox{mr}}(R_+)+{\mbox{mr}}(S_+)=1+1=2, {\mbox{mr}}(R_0)+{\mbox{mr}}(S_0)=2+1=3,{\mbox{mr}}(R_-)+{\mbox{mr}}(S_-)=2+1=3$.
Taking $-M$ in the previous example shows that the term ${\mbox{mr}}(R_-)+{\mbox{mr}}(S_-)$ is needed in Formula (\[mainformula\]).
To see that the term ${\mbox{mr}}(R_0)+{\mbox{mr}}(S_0)$ is needed in Formula (\[mainformula\]), let $$M = \begin{bmatrix}
A_{1,1} & A_{1,2}& 0\\
A_{2,1} & a_{2,2} & A_{2,3}\\
0 & A_{3,2} & A_{3,3}
\end{bmatrix}=\begin{bmatrix}
+ & 0 & 0\\
+ & - & +\\
0 & + & -
\end{bmatrix},$$ and let $$R_p = \begin{bmatrix}
A_{1,1} & A_{1,2}\\
A_{2,1} & p
\end{bmatrix}=
\begin{bmatrix}
+ & +\\
+ & p
\end{bmatrix}$$ and $$S_p = \begin{bmatrix}
a_{2,2}-p & A_{2,3}\\
A_{3,2} & A_{3,3}
\end{bmatrix}=
\begin{bmatrix}
(-)-p & -\\
+ & +
\end{bmatrix}.$$
Observe that ${\mbox{mr}}(M)=2$. Note that ${\mbox{mr}}(A_{1,1})+ {\mbox{mr}}(A_{3,3})+2=1+1+2=4, {\mbox{mr}}([A_{1,1}\;A_{1,2}])+{\mbox{mr}}([A_{3,2}\;A_{3,3}])+1=1+1+1=3,{\mbox{mr}}(\begin{bmatrix}
A_{1,1}\\
A_{2,1}
\end{bmatrix})+{\mbox{mr}}(\begin{bmatrix}
A_{2,3}\\
A_{3,3}
\end{bmatrix})+1=1+1+1=3, {\mbox{mr}}(R_+)+{\mbox{mr}}(S_+)=2+1=3, {\mbox{mr}}(R_0)+{\mbox{mr}}(S_0)=1+1=2,{\mbox{mr}}(R_-)+{\mbox{mr}}(S_-)=2+1=3$.
[1]{}
F. Barioli, W. Barrett, S. M. Fallat, H. T. Hall, L. Hogben, B. Shader, P. van den Driessche, and H. van der Holst. Parameters related to tree-width, zero forcing, and maximum nullity of a graph. , 72(2):146–177, 2013.
F. Barioli, S. Fallat, and L. Hogben. Computation of minimal rank and path cover number for certain graphs. , 392:289–303, 2004.
S. Fallat and L. Hogben. The minimum rank of symmetric matrices described by a graph: A survey. , 426(2–3):558–582, 2007.
S. M. Fallat and C. R. Johnson. Sub-direct sums and positivity classes of matrices. , 288:149–173, February 1999.
F. J. Hall and Z. Li. Sign pattern matrices. In L. Hogben, editor, [*Handbook of Linear Algebra*]{}, chapter 33. Simon and Hall/CRC Press, 2007.
L.-Y. Hsieh. . PhD thesis, University of Wisconsin-Madison, 2001.
Z. Li, Y. Gao, M. Arav, F. Gong, W. Gao, F. Hall, and H. van der Holst. Sign patterns with minimum rank 2 and upper bounds on minimum ranks. , 61:895–908, 2013.
R. C. Mikkelson. . PhD thesis, Iowa State University, 2008.
[^1]: Corresponding author, E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'An index theorem for the anti-self-dual deformation complex on anti-self-dual orbifolds with cyclic quotient singularities is proved. We present two applications of this theorem. The first is to compute the dimension of the deformation space of the Calderbank-Singer scalar-flat Kähler toric ALE spaces. A corollary of this is that, except for the Eguchi-Hanson metric, all of these spaces admit non-toric anti-self-dual deformations, thus yielding many new examples of anti-self-dual ALE spaces. For our second application, we compute the dimension of the deformation space of the canonical Bochner-Kähler metric on any weighted projective space ${\mathbb{CP}}^2_{(r,q,p)}$ for relatively prime integers $1 < r < q < p$. A corollary of this is that, while these metrics are rigid as Bochner-Kähler metrics, infinitely many of these admit non-trival self-dual deformations, yielding a large class of new examples of self-dual orbifold metrics on certain weighted projective spaces.'
address:
- 'Department of Mathematics, University of Wisconsin, Madison, WI, 53706'
- 'Department of Mathematics, University of Wisconsin, Madison, WI, 53706'
author:
- 'Michael T. Lock'
- 'Jeff A. Viaclovsky'
bibliography:
- 'ASD\_references.bib'
date: 'May 17, 2012'
title: |
Anti-self-dual orbifolds with cyclic\
quotient singularities
---
[^1]
Introduction {#intro}
============
If $(M^{4},g)$ is an [*[oriented]{}*]{} four-dimensional Riemannian manifold, the Hodge star operator $*:\Lambda^{2} \mapsto\Lambda^{2}$ satisfies $*^2 = Id$, and induces the decomposition on the space of $2$-forms $\Lambda^{2}=\Lambda^{2}_{+} \oplus\Lambda^{2}_{-}$, where $\Lambda^{2}_{\pm}$ are the $\pm 1$ eigenspaces of $*$. The Weyl tensor can be viewed as an operator $\mathcal{W}_g: \Lambda^2 \rightarrow \Lambda^2$, so this decomposition enables us to decompose the Weyl tensor as $\mathcal{W}_g = \mathcal{W}^{+}_g + \mathcal{W}^{-}_g$, into the self-dual and anti-self-dual Weyl tensors, respectively. The metric $g$ is called [*[anti-self-dual]{}*]{} if $\mathcal{W}^+_g = 0$, and $g$ is called [*[self-dual]{}*]{} if $\mathcal{W}^-_g = 0$. Note that, by reversing orientation, a self-dual manifold is converted into an anti-self-dual manifold, and vice versa. There are now so many known examples of anti-self-dual metrics on various compact four-manifolds, that it is difficult to give a complete list here, and we refer the reader to [@ViaclovskyIndex] for a recent list of references.
The deformation theory of anti-self-dual metrics is roughly analogous to the theory of deformation of complex structures. If $(M,g)$ is an anti-self-dual four-manifold, the anti-self-dual deformation complex is given by $$\begin{aligned}
\label{thecomplex}
\Gamma(T^*M) \overset{\mathcal{K}_g}{\longrightarrow}
\Gamma(S^2_0(T^*M)) \overset{\mathcal{D}}{\longrightarrow}
\Gamma(S^2_0(\Lambda^2_+)),\end{aligned}$$ where $\mathcal{K}_g$ is the conformal Killing operator defined by $$\begin{aligned}
( \mathcal{K}_g(\omega))_{ij} = \nabla_i \omega_j + \nabla_j \omega_i -
\frac{1}{2} (\delta \omega) g, \end{aligned}$$ with $\delta \omega = \nabla^i \omega_i$, $S^2_0(T^*M)$ denotes traceless symmetric tensors, and $\mathcal{D} = (\mathcal{W}^+)_g'$ is the linearized self-dual Weyl curvature operator.
If $M$ is a compact manifold then there is a formula for the index depending only upon topological quantities. The analytical index is given by $$\begin{aligned}
Ind(M, g) = \dim( H^0(M,g)) - \dim( H^1(M,g)) + \dim( H^2(M,g)),\end{aligned}$$ where $H^i(M,g)$ is the $i$th cohomology of the complex , for $i = 0,1,2$. The index is given in terms of topology via the Atiyah-Singer index theorem $$\begin{aligned}
\label{manifoldindex}
Ind(M, g) = \frac{1}{2} ( 15 \chi(M) + 29 \tau(M)), \end{aligned}$$ where $\chi(M)$ is the Euler characteristic and $\tau(M)$ is the signature of $M$, see [@KotschickKing].
The cohomology groups of the complex yield information about the local structure of the moduli space of anti-self-dual conformal classes, which we briefly recall [@Itoh2; @KotschickKing]. There is a map $$\begin{aligned}
\Psi: H^1(M,g) \rightarrow H^2(M,g)\end{aligned}$$ called the [*[Kuranishi map]{}*]{} which is equivariant with respect to the action of $H^0$, and the moduli space of anti-self-dual conformal structures near $g$ is localy isomorphic to $\Psi^{-1}(0) / H^0$. Therefore, if $H^2 = 0$, the moduli space is locally isomorphic to $H^1 / H^0$.
In this paper, we will be concerned with orbifolds in dimension four with isolated singularities modeled on ${\mathbb{R}}^4 / \Gamma$, where $\Gamma$ is a finite subgroup of ${\rm{SO}}(4)$ acting freely on ${\mathbb{R}}^4 \setminus \{0 \}$. We will say that $(M,g)$ is a [*[Riemannian orbifold]{}*]{} if $g$ is a smooth metric away from the singular points, and at any singular point, the metric is locally the quotient of a smooth $\Gamma$-invariant metric on $B^4$ under the orbifold group $\Gamma$.
The above results regarding the Kuranishi map are also valid for anti-self-dual Riemannian orbifolds. However, the index formula does not hold without adding a correction term. In [@Kawasaki], Kawasaki proved a version of the Atiyah-Singer index theorem for orbifolds, and gave a general formula for the correction term. Our first result is an explicit formula for this correction term for the complex in the case that $\Gamma$ is an action of a cyclic group. In order to state this, we first make some definitions.
For $1 \leq q < p$ relatively prime integers, we denote by $\Gamma_{(q,p)}$ the cyclic action $$\begin{aligned}
\label{qpaction}
\left(
\begin{matrix}
\exp^{2 \pi i k / p} & 0 \\
0 & \exp^{2 \pi i k q / p } \\
\end{matrix}
\right), \ \ 0 \leq k < p,\end{aligned}$$ acting on ${\mathbb{R}}^4$, which we identify with ${\mathbb{C}}^2$ using $z_1 = x_1 + i x_2$, and $z_2 = x_3 + i x_4$. We will also refer to this action as a type $(q,p)$-action.
[*A group action $\Gamma_1: G \rightarrow {\rm{SO}(4)}$ is [*[conjugate]{}*]{} to another group action $\Gamma_2: G \rightarrow {\rm{SO}(4)}$ if there exists an element $O \in O(4)$ such that for any $g \in G$, $\Gamma_1(g) \circ O = O \circ \Gamma_2(g)$. If $O \in {\rm{SO}}(4)$, then the actions are said to be orientation-preserving conjugate, while if $O \notin {\rm{SO}}(4)$, the actions are said to be orientation-reversing conjugate.* ]{}
\[actrem\] [*We note the important fact that if $\Gamma$ is an $\rm{SO}(4)$ representation of a cyclic group, then $\Gamma$ is orientation-preserving conjugate to a $\Gamma_{(q,p)}$-action [@MCC]; we therefore only need consider the $\Gamma_{(q,p)}$-actions. Furthermore, for $1 \leq q , q' < p$, if a $\Gamma_{(q,p)}$-action is orientation-preserving conjugate to a $\Gamma_{(q',p)}$-action then $q q' \equiv 1 \mod p$. We also note that a $\Gamma_{(q,p)}$-action is orientation-reversing conjugate to a $\Gamma_{(p-q,p)}$-action.* ]{}
We will employ the following modified Euclidean algorithm. For $1 \leq q < p$ relatively prime integers, write $$\begin{aligned}
\begin{split}\label{mea}
p&=e_1q-a_1\\
q&=e_2a_1-a_2\\
&\hspace{2mm} \vdots \\
a_{k-2}&=e_ka_{k-1}-1,
\end{split}\end{aligned}$$ where $e_i \geq 2$, and $0 \leq a_i < a_{i-1}$. This can also be written as the continued fraction expansion $$\begin{aligned}
\frac{p}{q} = \cfrac{1}{e_1 - \cfrac{1}{e_2 - \cdots \cfrac{1}{e_k}}}.\end{aligned}$$ We refer to the integer $k$ as the [*[length]{}*]{} of the modified Euclidean algorithm.
Our main theorem expresses the correction term in the index theorem in terms of the $e_i$ and the length of the modified Euclidean algorithm:
\[mainthm\] Let $(M,g)$ be a compact anti-self-dual orbifold with a single orbifold point of type $(q,p)$. The index of the anti-self-dual deformation complex on $(M,g)$ is given by $$\begin{aligned}
Ind(M,g)=
\begin{cases}
\displaystyle \frac{1}{2}(15\chi_{top}+29\tau_{top})+\sum_{i=1}^{k}4e_i-12k-2 &\text{ when $q \neq p-1$}\\
\displaystyle \frac{1}{2}(15\chi_{top}+29\tau_{top}) - 4p + 4
&\text{ when $q=p-1$}.
\end{cases}\end{aligned}$$
In some other special cases, the correction term may be written directly in terms of $p$. For example, if $q = 1$, and $p > 2$, we have $$\begin{aligned}
\sum_{i=1}^{k}4e_i-12k-2 = 4p -14.\end{aligned}$$ We note that the cases $q = 1$ and $q = p-1$ were proved earlier in [@ViaclovskyIndex] using a different method.
\[srmk\] [*While Theorem \[mainthm\] is stated in the case of a single orbifold point for simplicity, if a compact anti-self-dual orbifold has several cyclic quotient orbifold points, then a similar formula holds, with the correction term simply being the sum of the corresponding correction terms for each type of orbifold point.* ]{}
Asymptotically locally Euclidean spaces
---------------------------------------
Many interesting examples of anti-self-dual metrics are complete and non-compact. Given a compact Riemannian orbifold $(\hat{X}, \hat{g})$ with non-negative scalar curvature, letting $G_p$ denote the Green’s function for the conformal Laplacian associated with any point $p$, the non-compact space $X = \hat{X} \setminus \{p\}$ with metric $g_p = G_p^{2}\hat{g}$ is a complete scalar-flat orbifold. Inverted normal coordinates in the metric $\hat{g}$ in a neighborhood of the point $p$, give rise to a coordinate system in a neighborhood of infinity of $X$, which motivates the following:
\[ALEdef\] [ *A noncompact Riemannian orbifold $(X^4,g)$ is called [*[asymptotically locally Euclidean]{}*]{} or [*[ALE]{}*]{} of order $\tau$ if there exists a finite subgroup $\Gamma \subset {\rm{SO}}(4)$ acting freely on ${\mathbb{R}}^4 \setminus \{0\}$, and a diffeomorphism $\phi : X \setminus K \rightarrow ( {\mathbb{R}}^4 \setminus B(0,R)) / \Gamma$ where $K$ is a compact subset of $X$, satisfying $(\phi_* g)_{ij} = \delta_{ij} + O( r^{-\tau})$ and $\partial^{|k|} (\phi_*g)_{ij} = O(r^{-\tau - k })$ for any partial derivative of order $k$, as $r \rightarrow \infty$, where $r$ is the distance to some fixed basepoint.* ]{}
An [*[orbifold compactification]{}*]{} of an ALE space $(X,g)$, is a choice of a conformal factor $u : X \rightarrow {\mathbb{R}}_+$ such that $u = O(r^{-2})$ as $r \rightarrow \infty$. The space $(X, u^2 g)$ then compactifies to a $C^{1,\alpha}$ orbifold. If $(X,g)$ is anti-self-dual, then there moreover exists a $C^{\infty}$-orbifold conformal compactification $(\hat{X}, \hat{g})$ [@CLW Proposition 12].
[*It is crucial to note that if $(X,g)$ is an anti-self-dual ALE space with a $\Gamma$-action at infinity, then the conformal compactification $(\hat{X}, \hat{g})$ with the anti-self-dual orientation has a $\tilde{\Gamma}$-action at the orbifold point where $\tilde{\Gamma}$ is orientation-reversing conjugate to $\Gamma$. In the case of a cyclic group, if the action at infinity of the anti-self-dual ALE space $(X,g)$ is of type $(q,p)$, then the action at the orbifold point of the compactification $(\hat{X}, \hat{g})$ with the anti-self-dual orientation is of type $(p-q,p)$.* ]{}
Many examples of anti-self-dual ALE spaces with nontrivial group at infinity have been discovered. The first non-trivial example was due to Eguchi and Hanson, who found a Ricci-flat anti-self-dual metric on $\mathcal{O}(-2)$ which is ALE with group ${\mathbb{Z}}/ 2 {\mathbb{Z}}$ at infinity [@EguchiHanson]. Gibbons-Hawking then wrote down an metric ansatz depending on the choice of $n$ monopole points in ${\mathbb{R}}^3$, giving an anti-self-dual ALE hyperkähler metric with cyclic action at infinity contained in ${\rm{SU(2)}}$, which are called multi-Eguchi-Hanson metrics [@GibbonsHawking; @Hitchin2].
Using the Joyce construction from [@Joyce1995], Calderbank and Singer produced many examples of toric ALE anti-self-dual metrics, which are moreover scalar-flat Kähler, and have cyclic groups at infinity contained in ${\rm{U}}(2)$ [@CalderbankSinger]. For a type $(q,p)$-action, the space $X$ is the minimal Hirzebruch-Jung resolution of ${\mathbb{C}}^2/ \Gamma_{(q,p)}$, with exceptional divisor given by the union of $2$-spheres $S_1 \cup \cdots \cup S_k$, with intersection matrix $$\begin{aligned}
\label{intersect}
(S_i \cdot S_j) =
\left(
\begin{matrix}
-e_1 & 1 & 0 & \cdots & 0 \\
1 & - e_2 & 1 & \cdots & 0 \\
0 & 1 & - e_3 & \cdots & 0 \\
\vdots & \vdots & \vdots & & \vdots \\
0 & 0 & 0 & \cdots & - e_k \\
\end{matrix}
\right),\end{aligned}$$ where the $e_i$ and $k$ are defined above in with $e_i \geq 2$. The Kähler scalar-flat metric on $X$ is then written down explicitly using the Joyce ansatz from [@Joyce1995]. We do not require the details of the construction here, but only note the following: For $1 < q$ the identity component of the isometry group of these metrics is a real $2$-torus, and for $q = 1$, it is ${\rm{U}}(2)$.
When $q = p - 1$, these metrics are the [*[toric]{}*]{} Gibbons-Hawking multi-Eguchi-Hanson metrics (when all monopole points are on a common line). In this case $k = p-1$ and $e_i = 2$ for $1 \leq i \leq k$. The moduli space of toric metrics in this case is of dimension $p-2$. But the moduli space of all multi-Eguchi-Hanson metrics is of dimension $3(p-2)$. So it is well-known that these metrics admit non-toric anti-self-dual deformations. When $q = 1$, these metrics agree with the LeBrun negative mass metrics on $\mathcal{O}(-p)$ discovered in [@LeBrunnegative]. In this case $k =1$ and $e_1 = p$. For $p > 2$, it was recently shown in [@HondaOn; @ViaclovskyIndex] that these spaces also admit non-toric anti-self-dual deformations. Theorem \[CScor\] will give a vast generalization of this phenomenon to the general case $1 < q < p-1$. The proof of Theorem \[CScor\] relies on the following explicit formula for the index of the complex on the conformal compactification of these metrics:
\[CSthm\] Let $(\hat{X}, \hat{g})$ be the orbifold conformal compactification of a Calderbank-Singer space $(X,g)$ with a $(q,p)$-action at infinity. Then the index of of anti-self-dual deformation complex is given by $$\begin{aligned}
Ind(\hat{X}, \hat{g}) =
\begin{cases}
\displaystyle 5k+5-\sum_{i=1}^{k}4e_i &\text{ when $q \neq 1$}\\
\displaystyle-4p+12 &\text{ when $q=1$},
\end{cases}\end{aligned}$$ where the integers $k$ and $e_i$, $1 \leq i \leq k$, are the integers occuring in the modified Euclidean algorithm defined in .
We note that if $q = p - 1$, the index simplifies to $-3 p + 8$. A consequence of the above is that the Calderbank-Singer spaces admit large families of non-toric anti-self-dual deformations, thereby yielding many new examples:
\[CScor\] Let $(X,g)$ be a Calderbank-Singer space with a $(q,p)$-action at infinity, and $(\hat{X}, \hat{g})$ be the orbifold conformal compactification. Let $\mathcal{M}_{\hat{g}}$ denote the moduli space of anti-self-dual conformal structures near $(\hat{X}, \hat{g})$. Then,
- If $q = 1$ and $p =2$, then $\hat{g}$ is rigid.
- If $q = 1$ and $p = 3$, then $\mathcal{M}_{\hat{g}}$ is locally of dimension $1$.
- If $q = 1$ and $p > 3$, then $\mathcal{M}_{\hat{g}}$ is locally of dimension $4p -12$.
- If $q = p -1$, then $\mathcal{M}_{\hat{g}}$ is locally of dimension $3p -7$.
- If $1 < q < p - 1$, then $\mathcal{M}_{\hat{g}}$ is locally of dimension at least $$\begin{aligned}
\label{csest}
\dim(H^1) - 2 = -5k-5+\sum_{i=1}^{k}4e_i. $$
Consequently, if $p > 2$, these spaces admit non-toric anti-self-dual deformations.
[*Theorem \[CSthm\] could be equivalently stated in terms of the ALE metrics rather than the compactified metrics. However, the definition of the index on an ALE space involves defining certain weighted spaces; see [@ViaclovskyIndex Proposition 3.1] for the precise formula which relates the index on the ALE space to the index on the compactification; for our purposes here, we only require the statement on the compactification. Similarly, Theorem \[CScor\] could be equivalently stated in terms of anti-self-dual ALE deformations of the ALE model.* ]{}
By a result of LeBrun-Maskit, $H^2(\hat{X}, \hat{g}) = 0$ for these metrics, so the actual moduli space is locally isomorphic to $H^1/ H^0$, [@LeBrunMaskit Theorem 4.2]. Therefore the moduli space could be of dimension $\dim(H^1)$, $\dim(H^1) - 1$, or $\dim(H^1) - 2$. This action in the toric multi-Eguchi-Hanson case $q = p -1$ is well-known; in this case for $p \geq 3$, $\dim(H^1) = 3p - 6$, and the dimension of the moduli space is equal to $\dim(H^1) -1 = 3p - 7$. In the LeBrun negative mass case $q = 1$, this action was recently completely determined by Nobuhiro Honda using arguments from twistor theory [@HondaOn]. For $ 1 < q < p-1$, further arguments are needed to determine this action explicitly; this is an interesting problem.
Weighted projective spaces
--------------------------
We first recall the definition of weighted projective spaces in real dimension four:
[*[ For relatively prime integers $1 \leq r \leq q \leq p$, the [*[weighted projective space]{}*]{} $\mathbb{CP}^2_{(r,q,p)}$ is $S^{5}/\mathbb{C}^*$, where $\mathbb{C}^*$ acts by $$\begin{aligned}
(z_0,z_1,z_2)\mapsto (e^{ir\theta}z_0,e^{iq\theta}z_1 ,e^{ip\theta}z_2),\end{aligned}$$ for $0\leq \theta <2\pi$. ]{}*]{}
The space $\mathbb{CP}^2_{(r,q,p)}$ has the structure of a compact complex orbifold. In [@Bryant], Bryant proved that every weighted projective space admits a Bochner-Kähler metric. Subsequently, David and Gauduchon gave a simple and direct construction of these metrics [@DavidGauduchon]. Using an argument due to Apostolov, they also showed that this metric is the unique Bocher-Kähler metric on a given weighted projective space [@DavidGauduchon Appendix D], and thus we will call this metric the [*[canonical]{}*]{} Bochner-Kähler metric. In complex dimension two, the Bochner tensor is the same as the anti-self-dual part of the Weyl tensor so Bochner-Kähler metrics are the same as self-dual Kähler metrics.
The work of Derdzinski [@Derdzinski] showed that a self-dual Kähler metric $g$ is conformal to a self-dual Hermitian Einstein metric on $M^*:=\{p\in M : R(p)\neq 0\}$, given by $\tilde{g}=R^{-2}g$, where $R$ is the scalar curvature. This conformal metric is not Kähler unless $R$ is constant. Conversely, Apostolov and Gauduchon [@ApostolovGauduchon] showed that any self-dual Hermitian Einstein metric that is not conformally flat is of the form $\tilde{g}$ for a unique self-dual Kähler metric $g$ with $R\neq 0$.
For a weighted projective space $\mathbb{CP}^2_{(r,q,p)}$, there are the following 3 cases:
- When $p < r+q$ the canonical Bochner-Kähler metric has $R>0$ everywhere, so it is conformal to a Hermitian Einstein metric with positive Einstein constant.
- When $p = r+q$ the canonical Bochner-Kähler metric has $R>0$ except at one point, so it is conformal to a complete Hermitian Einstein metric with vanishing Einstein constant outside this point.
- When $p > r+q$ the canonical Bochner-Kähler metric has $R$ vanishing along a hypersurface and the complement is composed of two open sets on which the metric is conformal to a Hermitian Einstein metric with negative Einstein constant.
For $x \in {\mathbb{R}}$, $\lfloor x \rfloor$ denotes the integer part of $x$, and $\{x\} = x -\lfloor x \rfloor $ denotes the fractional part of $x$. We also define the integer $\epsilon$ by $$\begin{aligned}
\epsilon =
\begin{cases}
0 & \text{if } p \not\equiv q \text{ mod } r \mbox{ and } p \not\equiv r \text{ mod } q\\
1 & \text{if } p \equiv q \text{ mod } r \mbox{ or } p \equiv r \text{ mod } q,
\mbox{ but not both}, \\
2 & \text{if } p \equiv q \text{ mod } r \mbox{ and } p \equiv r \text{ mod } q.
\\
\end{cases}\end{aligned}$$ Our main result for the index on weighted projective spaces is the following, with the answer depending upon certain number-theoretic properties of the triple $(r,q,p)$:
\[introthm\] Let $g$ be the canonical Bochner-Kähler metric with reversed orientation on $\overline{\mathbb{CP}}^2_{(r,q,p)}$. Assume that $1<r<q<p$. If $r+q\geq p$ then $$\begin{aligned}
Ind(\overline{\mathbb{CP}}^2_{(r,q,p)},g)=2.\end{aligned}$$ If $r+q<p$, then $$\begin{aligned}
Ind(\overline{\mathbb{CP}}^2_{(r,q,p)},g)=
\begin{cases}
2 +2\epsilon -4\lfloor \frac{p}{qr} \rfloor &\text{ when $\{ \frac{p}{qr}\}<\{\frac{q^{-1;r}p}{r}\}$}\\
-2 +2\epsilon -4\lfloor \frac{p}{qr} \rfloor &\text{ when $\{ \frac{p}{qr}\}>\{\frac{q^{-1;r}p}{r}\}$}.
\end{cases}\end{aligned}$$
We note that in the case $\{ \frac{p}{qr}\}<\{\frac{q^{-1;r}p}{r}\}$, the integer $\epsilon$ can only be $0$ or $1$; the integer $2$ does not actually occur in this case. Thus there are exactly $5$ cases which do in fact all occur, see Section \[wpssec\].
Theorem \[introthm\] implies the following result regarding the moduli space of anti-self-dual metrics on $\overline{\mathbb{CP}}^2_{(r,q,p)}$:
\[wpsthm\] Let $g$ be the canonical Bochner-Kähler metric with reversed orientation on $\overline{\mathbb{CP}}^2_{(r,q,p)}$. Assume that $1<r<q<p$. Then,
- If $p \leq q + r$ then $[g]$ is isolated as an anti-self-dual conformal class.
- If $p > q + r$ then the moduli space of anti-self-dual orbifold conformal classes near $g$, $\mathcal{M}_g$, is of dimension at least $$\begin{aligned}
\dim( \mathcal{M}_g) \geq
\begin{cases}
4\lfloor \frac{p}{qr} \rfloor- 2 -2\epsilon &\text{ when $\{ \frac{p}{qr}\}<\{\frac{q^{-1;r}p}{r}\}$}\\
4\lfloor \frac{p}{qr} \rfloor +2 -2\epsilon &\text{ when $\{ \frac{p}{qr}\}>\{\frac{q^{-1;r}p}{r}\}$}.
\end{cases}\end{aligned}$$
[*Since the case $p < q + r$ is conformal to an Einstein metric, it is perhaps not surprising (although not obvious) that these metrics are also isolated as self-dual metrics. But the non-trivial anti-self-dual deformations we have found in the case $p > q +r$ are quite surprising, since these metrics are rigid as Bochner-Kähler metrics.* ]{}
The proof of Theorem \[wpsthm\] also relies on the fact that $H^2(M,g) = 0$ for these metrics, see Corollary \[h2wps\] below. Then as pointed out above, the actual moduli space is locally isomorphic to $H^1/ H^0$, so the moduli space could be of dimension $\dim(H^1)$, $\dim(H^1) - 1$, or $\dim(H^1) - 2$. As in the case of the Calderbank-Singer spaces, we do not determine this action explicitly here; this is another very interesting problem.
Outline of paper
----------------
We begin in Section \[ogi\] by recalling Kawasaki’s orbifold index theorem, and apply it to the complex . Then in Section \[gcyclic\], we analyze the correction terms for cyclic group actions, culminating in the following formula for the index in terms of the following trigonometric sum when $1 < q < p -1$: $$\begin{aligned}
\begin{split}
\label{N(q,p)intro}
Ind_{\Gamma}(\hat{M})&=\frac{1}{2}(15\chi_{top}+29\tau_{top})-6+\frac{14}{p}\sum_{j=1}^{p-1}\Big[\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\Big]\\
&\phantom{==}-\frac{2}{p}\sum_{j=1}^{p-1}\Big[\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\cos(\frac{2\pi}{p}j)\cos(\frac{2\pi}{p}qj)\Big].
\end{split}\end{aligned}$$ We note that the quantity $$\begin{aligned}
\label{Dedekind}
s(q,p) = \frac{1}{4p}\sum_{j=1}^{p-1}\Big[\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\Big]\end{aligned}$$ is the well-known Dedekind sum [@Rademacher2]. This has a closed form expression in several special cases, but not in general. It is not surprising that this term appears, since Dedekind sums arise naturally in the index theorem for the signature complex [@HirzebruchZagier; @Katase1; @Zagier]. However, for the anti-self-dual deformation complex, the interaction of the Dedekind sum term with the final term in makes a huge difference. In particular, formula says that the sum of these terms must always be an integer!
For $x \in {\mathbb{R}}\setminus {\mathbb{Z}}$, we define the sawtooth function $((x)) = \{ x \} - \frac{1}{2}$. In Section \[nontop\], we show that when $1 < q < p-1$, the non-topological terms in can be rewritten as a Dedekind sum plus terms involving the sawtooth function: $$\begin{aligned}
N(q,p)=-6+\frac{12}{p}\sum_{j=1}^{p-1}\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)-4\bigg(\bigg(\frac{q^{-1;p}}{p}\bigg)\bigg) -4\bigg(\bigg(\frac{q}{p}\bigg)\bigg),\end{aligned}$$ where $q^{-1;p}$ is the inverse of $q$ modulo $p$. In Section \[explicit\] we use this, together with classical reciprocity for Dedekind sums to prove Theorem \[mainthm\]. The results dealing with the Calderbank-Singer spaces, Theorems \[CSthm\] and \[CScor\], are proved in Section \[CSindex\]. Finally, in Section \[wpssec\], we present a complete analysis of the index for the canonical Bochner-Kähler metric on a weighted projective space, and prove Theorem \[wpsthm\]. Interestingly, an important ingredient is Rademacher’s triple reciprocity formula for Dedekind sums [@Rademacher1]. We conclude the paper with some remarks on the number-theoretic condition on the triple $(r,q,p)$ which occurs in Theorem \[introthm\].
Acknowledgements
----------------
The authors would like to thank John Lott for crucial discussions about index theory. Thanks are also due to Nobuhiro Honda for many valuable discussions on the moduli space of anti-self-dual metrics.
The orbifold $\Gamma$-Index {#ogi}
===========================
For an orbifold $(M,g)$, the $\Gamma$-Index is given analytically by $$\begin{aligned}
Ind_{\Gamma}(M,g) = \dim (H^0(M,g)) - \dim (H^1(M,g)) + \dim (H^2(M,g)).\end{aligned}$$ From Kawasaki’s orbifold index theorem [@Kawasaki], it follows that we have a $\Gamma$-index formula of the form $$\begin{aligned}
\label{I-G}
Ind_{\Gamma}(M)=\frac{1}{2}(15\chi_{orb}(M)+29\tau_{orb}(M))+\frac{1}{|\Gamma|}\sum_{\gamma \neq Id}\frac{ch_{\gamma}(i^*\sigma)}{ch_{\gamma}({\lambda_{-1}N_{\mathbb{C}}})}.\end{aligned}$$ where the quantity $\chi_{orb}(M)$ is the orbifold Euler characteristic defined by $$\begin{aligned}
\chi_{orb}(M)=\frac{1}{8\pi^2}\int_{M}(|W|^2-\frac{1}{2}|Ric|^2+\frac{1}{6}R^2)dV_{g},\end{aligned}$$ the quantity $\tau_{orb}(M)$ is the orbifold signature defined by $$\begin{aligned}
\tau_{orb}(M)=\frac{1}{12\pi^2}\int_{M}(|W^+|^2-|W^-|^2)dV_{g},\end{aligned}$$ and the quantity $\frac{ch_{\gamma}(i^*\sigma)}{ch_{\gamma}(\lambda_{-1}N_{\mathbb{C}})}$ is a correction term depending upon the action of $\gamma$ on certain bundles, which we will describe in what follows.
In the next subsection, we compute the trace of the action of $\gamma$, an element in the orbifold group $\Gamma$, on the bundles $[N_{\mathbb{C}}]$, $[S^2_0(N_{\mathbb{C}})]$ and $[S^2_0(\Lambda^2_+)]$ over the fixed point set, which we then use to compute a general formula for the $\frac{ch_{\gamma}(i^*\sigma)}{ch_{\gamma}(\lambda_{-1}N_{\mathbb{C}})}$ term. Then we give the orbifold Euler characteristic and orbifold signature in terms of the topological Euler characteristic and topological signature and correction terms also depending upon the $\gamma$-action respectively. Finally, we combine this information into a formula for the orbifold $\Gamma$-Index.
Group action on bundles
-----------------------
In order to compute the $\Gamma$-Index, we first need to find the trace of the $\gamma$-action, for every $\gamma$ in $\Gamma$, on the pullback of the complexified principal symbol, $i^*\sigma$, where $$\begin{aligned}
\label{pr1}
i:p\rightarrow M\end{aligned}$$ is the inclusion map from the fixed point $p$ into the orbifold $M$. In this case $$\begin{aligned}
\label{symbol}
i^*\sigma=[N_{\mathbb{C}}]-[S^2_0(N_{\mathbb{C}})]+[S^2_0\Lambda^2_+].\end{aligned}$$ For a general $\gamma$ of the form $$\begin{aligned}
\gamma = \left( {\begin{array}{*{20}c}
\cos\theta_1 & -\sin\theta_1 & 0 & 0 \\
\sin\theta_1 & \cos\theta_1 & 0 & 0 \\
0 & 0 & \cos\theta_2 & -\sin\theta_2 \\
0 & 0 & \sin\theta_2 & \cos\theta_2
\end{array} } \right),\end{aligned}$$ fixing the point $p$, the normal bundle is trivial, so $N_\mathbb{C}:=N\otimes \mathbb{C}=\mathbb{C}^4$, and we have the following proposition.
\[traceprop\] The trace of the $\gamma$-action on the components of $i^*\sigma$ is as follows:
1. $\displaystyle tr(\gamma|_{N_{\mathbb{C}}})=2\cos (\theta_1)+2\cos(\theta_2)$,
2. $\displaystyle tr(\gamma|_{S^2_0(N_{\mathbb{C}})})=1+2\cos(\theta_1+\theta_2)+ 2\cos(-\theta_1+\theta_2)+4\cos(\theta_1+\theta_2)\cos(-\theta_1+\theta_2)$,
3. $\displaystyle tr(\gamma|_{S^2_0(\Lambda^2_+)})=2cos(\theta_1+\theta_2)+4cos^2(\theta_1+\theta_2)-1.$
The normal bundle can be written as $N=x_1\oplus \cdot \cdot \cdot \oplus x_4$ in real coordinates. After complexifying the normal bundle we can diagonalize $\gamma$ to write $$\begin{aligned}
\gamma|_{N_{\mathbb{C}}} =
\left( {\begin{array}{*{20}c}
e^{i\theta_1} & 0 & 0 & 0 \\
0 & e^{-i\theta_1} & 0 & 0 \\
0 & 0 & e^{i\theta_2}& 0 \\
0 & 0 & 0 & e^{-i\theta_2}
\end{array} } \right),\end{aligned}$$ with respect to the complex basis $\{\lambda_1\oplus \lambda_2\oplus \lambda_3 \oplus \lambda_4\}=\mathbb{C}^4$, where $$\begin{aligned}
\{2x_1, 2x_2, 2x_3, 2x_4 \} = \{
\lambda_1-i\lambda_2, i\lambda_1-\lambda_2,
\lambda_3-i\lambda_4,i\lambda_3-\lambda_4 \}.\end{aligned}$$ Formula $(1)$ follows immediately.
Next, to see how $\gamma$ acts on $S^2_0(N_g)=\Lambda^2_+ \otimes \Lambda^2_-$ we first examine how $\gamma$ acts on $\Lambda^2_+$ and $\Lambda^2_-$ independently. We use the following basis for $\Lambda^2_+$: $$\begin{aligned}
\begin{split}
\omega_1^+&=\frac{1}{2} [d\lambda_2\wedge d\lambda_1+d\lambda_4\wedge d\lambda_3],\\
\omega_2^+&=
\frac{1}{2}[d\lambda_1\wedge d\lambda_3+d\lambda_4\wedge d\lambda_2],\\
\omega_3^+
&=\frac{1}{2}[id\lambda_1\wedge d\lambda_3+id\lambda_2\wedge d\lambda_4],
\end{split}\end{aligned}$$ and the following basis for $\Lambda^2_-$: $$\begin{aligned}
\begin{split}
\omega_1^-&=\frac{1}{2} [d\lambda_2\wedge d\lambda_1-d\lambda_4\wedge d\lambda_3],\\
\omega_2^-&=\frac{1}{2}[id\lambda_3\wedge d\lambda_2+id\lambda_4\wedge d\lambda_1],\\
\omega_3^-&=\frac{1}{2}[d\lambda_2\wedge d\lambda_3+d\lambda_4\wedge d\lambda_1].
\end{split}\end{aligned}$$ So we see that $\gamma$ acts on $\Lambda_+^2$ by $$\begin{aligned}
\begin{split}
\gamma(\omega_1^+)&=\omega_1^+\\
\gamma(\omega_2^+)
&=\frac{1}{2}[e^{i(\theta_1+\theta_2)}(\omega_2^+-i\omega_3^+)+e^{-i(\theta_1+\theta_2)}(\omega_2^++i\omega_3^+)]\\
\gamma(\omega_3^+)
&=\frac{1}{2}[e^{i(\theta_1+\theta_2)}(\omega_3^++i\omega_2^+)+e^{-i(\theta_1+\theta_2)}(\omega_3^+-i\omega_2^+)],
\end{split}\end{aligned}$$ and $\gamma$ acts on $\Lambda_-^2$ by $$\begin{aligned}
\begin{split}
\gamma(\omega_1^-)&=\omega_1^-\\
\gamma(\omega_2^-)
&=\frac{1}{2}[e^{i(-\theta_1+\theta_2)} (\omega_2^--i\omega_3^-)+e^{i(\theta_1-\theta_2)}(\omega_2^-+i\omega_3^-)]\\
\gamma(\omega_3^+)
&=\frac{1}{2}[e^{i(-\theta_1+\theta_2)}(\omega_3^-+i\omega_2^-)+e^{i(\theta_1-\theta_2)}(\omega_3^--i\omega_2^-)].
\end{split}\end{aligned}$$ Therefore, we can write $$\begin{aligned}
\gamma|_{\Lambda_+^2}
= \left(
\begin{matrix}
1&0&0\\
0& \cos(\theta_1+\theta_2) &- \sin(\theta_1+\theta_2) \\
0& \sin(\theta_1+\theta_2) & \cos(\theta_1+\theta_2)
\end{matrix}
\right),\end{aligned}$$ and $$\begin{aligned}
\gamma|_{\Lambda_-^2}
= \left(
\begin{matrix}
1&0&0\\
0& \cos(-\theta_1+\theta_2) &- \sin(-\theta_1+\theta_2) \\
0& \sin(-\theta_1+\theta_2) & \cos(-\theta_1+\theta_2)
\end{matrix} \right).\end{aligned}$$ To derive $(2)$, we compute $$\begin{aligned}
\begin{split}
tr(\gamma|_{S^2_0 N_{\mathbb{C}}})&=tr(\gamma|_{\Lambda^2_+\otimes \Lambda^2_-})\\
&=tr(\gamma|_{\Lambda^2_+})\cdot tr(\gamma|_{\Lambda^2_-})\\
&=(1+ 2\cos(\theta_1+\theta_2))\cdot (1+ 2\cos(-\theta_1+\theta_2))\\
=1+2\cos(\theta_1+\theta_2)&+ 2\cos(-\theta_1+\theta_2)+4\cos(\theta_1+\theta_2)\cos(-\theta_1+\theta_2).
\end{split}\end{aligned}$$ Next, to see how $\gamma$ acts on $S^2_0(\Lambda^2_+)$, decompose $$\begin{aligned}
S^2_0\Lambda_+^2=[\mathbb{C}\otimes (\omega_2^+\oplus \omega_3^+)]\oplus S^2_0(\omega_2^+\oplus \omega_3^+)\oplus tr,\end{aligned}$$ where $ tr=2\omega_1^+-(\omega_2^++\omega_3^+)$ denotes the trace component, and write the basis of $S^2_0(\omega_2^+\oplus \omega_3^+)$ as $$\begin{aligned}
\{\omega_2^+\otimes \omega_2^+-\omega_3^+\otimes \omega_3^+, \omega_2^+\otimes \omega_3^++\omega_3^+\otimes \omega_2^+\}.
\end{aligned}$$ We see that $$\begin{aligned}
&\gamma|_{\omega_1^+\otimes (\omega_2^+ \oplus \omega_3^+)}=\left(
\begin{matrix}
\cos(\theta_1+\theta_2) &- \sin(\theta_1+\theta_2) \\
\sin(\theta_1+\theta_2) & \cos(\theta_1+\theta_2)
\end{matrix} \right), \\
&\gamma|_{S^2_0(\omega_2^+\oplus \omega_3^2)}=\left(
\begin{matrix}
\cos^2(\theta_1+\theta_2) -\sin^2(\theta_1+\theta_2) & -2\sin(\theta_1+\theta_2) \cos(\theta_1+\theta_2) \\
2\sin(\theta_1+\theta_2) \cos(\theta_1+\theta_2) & \cos^2(\theta_1+\theta_2) -\sin^2(\theta_1+\theta_2)
\end{matrix} \right),\\
&\gamma|_{tr\in S^2_0\Lambda^2_+}=1.\end{aligned}$$ Using these, we derive $(3)$ by computing $$\begin{aligned}
\begin{split}
tr(\gamma|_{S^2_0\Lambda^2_+})&=[2cos(\theta_1+\theta_2)]+[4cos^2(\theta_1+\theta_2)-2]+[1]\\
&=2cos(\theta_1+\theta_2)+4cos^2(\theta_1+\theta_2)-1.
\end{split}\end{aligned}$$
Equivariant Chern character
---------------------------
We next compute the term $\frac{ch_{\gamma}(i^*\sigma)}{ch_{\gamma}(\lambda_{-1}N_{\mathbb{C}})}$. The numerator of this term is the ${\gamma}$-equivariant Chern character of the pullback of the principal symbol, $i^*\sigma$, described in $\eqref{pr1}$ and $\eqref{symbol}$. The denominator is the ${\gamma}$-equivariant Chern character of the $K$-theoretic Thom class of the complexified normal bundle. Since the normal bundle is trivial over the fixed point, this is $$\begin{aligned}
\lambda_{-1}N_\mathbb{C}=[\Lambda^0(\mathbb{C}^4)]-[\Lambda^1(\mathbb{C}^4)]+[\Lambda^2(\mathbb{C}^4)]-[\Lambda^3(\mathbb{C}^4)]+[\Lambda^4(\mathbb{C}^4)].\end{aligned}$$ Since the ${\gamma}$-equivariant Chern character is just the ${\gamma}$-action times the Chern character of each eigenspace, using Proposition \[traceprop\], we compute $$\begin{aligned}
\begin{split}
ch_{\gamma}(i^*\sigma)&=tr({\gamma}|_{N_\mathbb{C}})-tr({\gamma}|_{S^2_0N_\mathbb{C}})+tr({\gamma}|_{S^2_0\Lambda^2_+})\\
&=[2\cos(\theta_1)+2\cos(\theta_2)]\\
&\phantom{=}-[1+2\cos(\theta_1+\theta_2)+2\cos(-\theta_1+\theta_2)+4\cos(\theta_1+\theta_2)\cos(-\theta_1+\theta_2)]\\
&\phantom{=}+[2\cos(\theta_1+\theta_2)+4\cos^2(\theta_1+\theta_2)-1]\\
&=[2\cos\theta_1+2\cos\theta_2-2-2\cos(\theta_1)\cos(\theta_2)]\\
&\phantom{=}+[-2\sin(\theta_1)\sin(\theta_2)-8\cos(\theta_1)\cos(\theta_2)\sin(\theta_1)\sin(\theta_2)+8\sin^2(\theta_1)\sin^2(\theta_2)]\\
&=[-2(\cos\theta_1-1)(\cos\theta_2-1)]+[8(1-\cos^2\theta_1)(1-\cos^2\theta_2)]\\
&\phantom{=}+[-2\sin(\theta_1)\sin(\theta_2)-8\cos(\theta_1)\cos(\theta_2)\sin(\theta_1)\sin(\theta_2)].
\end{split}\end{aligned}$$ Similarly, we compute $$\begin{aligned}
\begin{split}
ch_{\gamma}(\lambda_{-1}N_{\mathbb{C}})&=tr({\gamma}|_{[\Lambda^0(\mathbb{C}^4)]})-tr({\gamma}|_{[\Lambda^1(\mathbb{C}^4)]})+tr({\gamma}|_{[\Lambda^2(\mathbb{C}^4)]})\\
&\phantom{=}-tr({\gamma}|_{[\Lambda^3(\mathbb{C}^4)]})+tr({\gamma}|_{[\Lambda^4(\mathbb{C}^4)]})
=4(\cos\theta_1-1)(\cos\theta_2-1).
\end{split}\end{aligned}$$ Therefore $$\begin{aligned}
\begin{split}
\frac{ch_{\gamma}(i^*\sigma)}{ch_{\gamma}(\lambda_{-1}N_{\mathbb{C}})}&=\Big[-\frac{1}{2}+2(1+\cos\theta_1)(1+\cos\theta_2)\Big]\\
&\phantom{=}-\bigg[\frac{2\sin(\theta_1)\sin(\theta_2)+8\cos(\theta_1)\cos(\theta_2)\sin(\theta_1)\sin(\theta_2)}{4(\cos\theta_1-1)(\cos\theta_2-1)}\bigg].
\end{split}\end{aligned}$$ Since $\frac{\sin(\theta_1)\sin(\theta_2)}{(\cos\theta_1-1)(\cos\theta_2-1)}=\cot(\frac{\theta_1}{2})\cot(\frac{\theta_2}{2})$, we see that $$\begin{aligned}
\begin{split}
\label{ch_g}
\frac{ch_{\gamma}(i^*\sigma)}{ch_{\gamma}(\lambda_{-1}N_{\mathbb{C}})}
&=-\frac{1}{2}+2(1+\cos\theta_1)(1+\cos\theta_2)-\frac{1}{2}\cot (\frac{\theta_1}{2})\cot(\frac{\theta_2}{2})\\
&\phantom{==}-2\cot(\frac{\theta_1}{2})\cot(\frac{\theta_2}{2})\cos(\theta_1)\cos(\theta_2).
\end{split}\end{aligned}$$
The $\Gamma$-Index
------------------
For an orbifold with a single isolated singularity, we have a formula for the Euler characteristic $$\begin{aligned}
\label{euler}
\chi_{top}(M)=\chi_{orb}(M)+\frac{|\Gamma|-1}{|\Gamma|},\end{aligned}$$ and a formula for the signature $$\begin{aligned}
\label{tau}
\tau_{top}(M)=\tau_{orb}(M)-\eta(S^3/\Gamma),\end{aligned}$$ where $\Gamma \subset {\rm{SO}}(4)$ is the orbifold group around the fixed point and $\eta(S^3/\Gamma)$ is the eta-invariant, which in our case is given by $$\begin{aligned}
\eta(S^3/\Gamma)=\frac{1}{|\Gamma|}\sum_{\gamma \neq Id}\Big[\cot(\frac{\theta_1}{2}j)\cot(\frac{\theta_2}{2}j)\Big].\end{aligned}$$ See [@Hitchin] for a useful discussion of the formulas $\eqref{euler}$ and $\eqref{tau}$.
Combining formulas $\eqref{euler}$ and $\eqref{tau}$ with the formula for the $\Gamma$-Index given in $\eqref{I-G}$, we have $$\begin{aligned}
\label{Gamma-I}
Ind_{\Gamma}=\frac{1}{2}(15\chi_{top}+29\tau_{top})-\frac{15}{2}\bigg(\frac{|\Gamma|-1}{|\Gamma|}\bigg)+\frac{29}{2}\eta(S^3/\Gamma)+\frac{1}{|\Gamma|}\sum_{\gamma\neq Id}\frac{ch_{\gamma}(i^*\sigma)}{ch_{\gamma}(\lambda_{-1}N_{\mathbb{C}})},\end{aligned}$$ where the last term is given by formula $\eqref{ch_g}$.
$\Gamma$-Index for cyclic group actions {#gcyclic}
=======================================
We consider an orbifold with an isolated singularity having the group action $\Gamma_{(q,p)}$ generated by $$\begin{aligned}
\gamma =
\left( {\begin{array}{*{20}c}
\cos(\frac{2\pi}{p}) & -\sin(\frac{2\pi}{p}) & 0 & 0 \\
\sin(\frac{2\pi}{p}) & \cos(\frac{2\pi}{p}) & 0 & 0 \\
0 & 0 & \cos(\frac{2\pi}{p}q)& -\sin(\frac{2\pi}{p}q) \\
0 & 0 & \sin(\frac{2\pi}{p}q)& \cos(\frac{2\pi}{p}q)
\end{array} } \right),\end{aligned}$$ where $p$ and $q$ are relatively prime. The cases when $q=1$ and $q=p-1$ have already been resolved in [@ViaclovskyIndex], and although we are specifically interested when $1<q<p-1$, we will make use of the sum $$\begin{aligned}
\sum_{\gamma \neq Id} \frac{ch_{\gamma}(i^*\sigma)}{ch_{\gamma}(\lambda_{-1}N_{\mathbb{C}})}\end{aligned}$$ in all cases, and make our computations accordingly. We begin this section by simplifying our formula for this sum in general: $$\begin{aligned}
\begin{split}
\label{sich}
\sum_{{\gamma} \neq Id} \frac{ch_{\gamma}(i^*\sigma)}{ch_{\gamma}(\lambda_{-1}N_{\mathbb{C}})}&=\sum_{j=1}^{p-1}\Big[-\frac{1}{2}+2(1+\cos(\frac{2\pi}{p}))(1+\cos(\frac{2\pi}{p}q))-\frac{1}{2}\cot(\frac{\pi}{p})\cot(\frac{\pi}{p}q)\Big]\\
&\phantom{==}-\sum_{j=1}^{p-1}\Big[2\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\cos(\frac{2\pi}{p}j)\cos(\frac{2\pi}{p}qj)\Big]\\
&=\sum_{j=1}^{p-1}\Big[\frac{3}{2}+2\cos(\frac{2\pi}{p}j)+2\cos(\frac{2\pi}{p}qj)+ \cos(\frac{2\pi}{p}(q+1)j)\Big]\\
&\phantom{==}+\sum_{j=1}^{p-1}\Big[\cos(\frac{2\pi}{p}(q-1)j)-\frac{1}{2}\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj) \Big]\\
&\phantom{==}+\sum_{j=1}^{p-1}\Big[-2\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\cos(\frac{2\pi}{p}j)\cos(\frac{2\pi}{p}qj)\Big].
\end{split}\end{aligned}$$ Now, to further simplify our formula for the $\Gamma$-Index, it is necessary to separate into the following cases:
$\Gamma$-Index when $1<q<p-1$
-----------------------------
Using $\eqref{sich}$, we see that in this case $$\begin{aligned}
\begin{split}
\sum_{{\gamma}\neq Id} \frac{ch_{\gamma}(i^*\sigma)}{ch_{\gamma}(\lambda_{-1}N_{\mathbb{C}})}&=\Big[\frac{3}{2}p-\frac{15}{2}\Big]
-\frac{1}{2}\sum_{j=1}^{p-1}\Big[\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\Big]\\
&\phantom{==}-2\sum_{j=1}^{p-1}\Big[\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\cos(\frac{2\pi}{p}j)\cos(\frac{2\pi}{p}qj)\Big].
\end{split}\end{aligned}$$ Therefore, by combining this with formula $\eqref{Gamma-I}$ for the $\Gamma$-Index, we have $$\begin{aligned}
\begin{split}
\label{N(q,p)}
Ind_{\Gamma}(M)&=\frac{1}{2}(15\chi_{top}+29\tau_{top})-6+\frac{14}{p}\sum_{j=1}^{p-1}\Big[\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\Big]\\
&\phantom{==}-\frac{2}{p}\sum_{j=1}^{p-1}\Big[\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\cos(\frac{2\pi}{p}j)\cos(\frac{2\pi}{p}qj)\Big].
\end{split}\end{aligned}$$
$\Gamma$-Index when $q=1$ and $p=2$
-----------------------------------
Using $\eqref{sich}$, we see that in this case $$\begin{aligned}
\begin{split}
\sum_{{\gamma}\neq Id} \frac{ch_{\gamma}(i^*\sigma)}{ch_{\gamma}(\lambda_{-1}N_{\mathbb{C}})}=-\frac{1}{2}.
\end{split}\end{aligned}$$ Therefore, by combining this with formula $\eqref{Gamma-I}$ for the $\Gamma$-Index, we have $$\begin{aligned}
\begin{split}
Ind_{\Gamma}(M)&=\frac{1}{2}(15\chi_{top}+29\tau_{top})-4
\end{split}\end{aligned}$$
$\Gamma$-Index when $q=1$ and $p>2$
-----------------------------------
Using $\eqref{sich}$, we see that in this case $$\begin{aligned}
\begin{split}
\sum_{{\gamma}\neq Id} \frac{ch_{\gamma}(i^*\sigma)}{ch_{\gamma}(\lambda_{-1}N_{\mathbb{C}})}&=\frac{5}{2}p-\frac{15}{2}-\sum_{j=1}^{p-1}\Big[\frac{1}{2}\cot^2(\frac{\pi}{p}j)+2\cot^2(\frac{\pi}{p}j)\cos^2(\frac{2\pi}{p}j)\Big].
\end{split}\end{aligned}$$ Therefore, by combining this with formula $\eqref{Gamma-I}$ for the $\Gamma$-Index, and the following well-known formula for the Dedekind sum (see [@Rademacher2]): $$\begin{aligned}
\frac{1}{4p}\sum_{j=1}^{p-1}\cot^2(\frac{\pi}{p}j) = \frac{1}{12p}(p-1)(p-2),\end{aligned}$$ we have $$\begin{aligned}
\begin{split}
\label{q=1}
Ind_{\Gamma}(M)&=\frac{1}{2}(15\chi_{top}+29\tau_{top})-5+\frac{14}{p}\sum_{j=1}^{p-1}\cot^2(\frac{\pi}{p}j)-\frac{2}{p}\sum_{j=1}^{p-1}\cot^2(\frac{\pi}{p}j)\cos^2(\frac{2\pi}{p}j)\\
&=\frac{1}{2}(15\chi_{top}+29\tau_{top})-5+\frac{12}{p}\sum_{j=1}^{p-1}\cot^2(\frac{\pi}{p}j)+\frac{8}{p}\sum_{j=1}^{p-1}\cos^4(\frac{\pi}{p}j)\\
&=\frac{1}{2}(15\chi_{top}+29\tau_{top})-2-\frac{8}{p}+\frac{12}{p}\sum_{j=1}^{p-1}\cot^2(\frac{\pi}{p}j)\\
&=\frac{1}{2}(15\chi_{top}+29\tau_{top})-2-\frac{8}{p}+\frac{4}{p}(p^2-3p+2)\\
&=\frac{1}{2}(15\chi_{top}+29\tau_{top})+4p-14.
\end{split}\end{aligned}$$
$\Gamma$-Index when $q=p-1$ and $p>2$
-------------------------------------
Using $\eqref{sich}$, we see that in this case $$\begin{aligned}
\begin{split}
\sum_{{\gamma}\neq Id} \frac{ch_{\gamma}(i^*\sigma)}{ch_g(\lambda_{-1}N_{\mathbb{C}})}&=\frac{5}{2}p-\frac{15}{2}+\sum_{j=1}^{p-1}\Big[\frac{1}{2}\cot^2(\frac{\pi}{p}j)+2\cot^2(\frac{\pi}{p}j)\cos^2(\frac{2\pi}{p}j)\Big].
\end{split}\end{aligned}$$ Therefore, by combining this with formula $\eqref{Gamma-I}$ for the $\Gamma$-Index, we have $$\begin{aligned}
\begin{split}
\label{q=p-1}
Ind_{\Gamma}(M)&=\frac{1}{2}(15\chi_{top}+29\tau_{top})-5-\frac{14}{p}\sum_{j=1}^{p-1}\cot^2(\frac{\pi}{p}j)+\frac{2}{p}\sum_{j=1}^{p-1}\cot^2(\frac{\pi}{p}j)\cos^2(\frac{2\pi}{p}j)\\
&=\frac{1}{2}(15\chi_{top}+29\tau_{top})-8+\frac{8}{p}-\frac{12}{p}\sum_{j=1}^{p-1}\cot^2(\frac{\pi}{p}j)\\
&=\frac{1}{2}(15\chi_{top}+29\tau_{top})-8+\frac{8}{p}-\frac{4}{p}(p^2-3p+2)\\
&=\frac{1}{2}(15\chi_{top}+29\tau_{top})-4p+4.
\end{split}\end{aligned}$$
Non-topological terms in the $\Gamma$-Index {#nontop}
===========================================
We denote the terms in the $\Gamma$-Index not involving the topological Euler characteristic or topological signature by $N(q,p)$. Also we change our notation of the $\Gamma$-Index from $Ind_{\Gamma}$ to $Ind_{(q,p)}$ to reflect the particular group action. With this new notation we can write the index as $$\begin{aligned}
\label{indn}
Ind_{(q,p)}&=\frac{1}{2}(15\chi_{top}+29\tau_{top})+N(q,p).\end{aligned}$$ In this section we will simplify our formulas for $N(q,p)$. Also, for the remainder of the paper we will use the following notation. For two relatively prime positive integers $\alpha < \beta$, denote $\alpha$’s inverse modulo $\beta$ by $\alpha^{-1;\beta}$, and $\beta$’s inverse modulo $\alpha$ by $\beta^{-1;\alpha}$, i.e. $$\begin{aligned}
\alpha\alpha^{-1;\beta}\equiv \text{1 mod $\beta$} \text{ and }
\beta\beta^{-1;\alpha}\equiv \text{1 mod $\alpha$}.\end{aligned}$$
In the cases that $N(q,p)$ is easy to compute we see that $$\begin{aligned}
\label{etc}
N(q,p)=
\begin{cases}
4p-14 &\text{ when $1=q<p-1$}\\
-4p+4 &\text{ when } q=p-1.
\end{cases}\end{aligned}$$ Note that the case when $q=\pm 1$ and $p=2$ can be actually included in the $q=p-1$ case. It will be convenient later in paper if we also have these formulas written in terms of sawtooth functions, a cotangent sum and a constant where the sawtooth function is defined to be $$\begin{aligned}
\label{st}
((x))=
\begin{cases}
x-\lfloor x \rfloor -\frac{1}{2} &\text{ when $x\notin \mathbb{Z}$}\\
0 &\text{ when $x\in \mathbb{Z}$}.
\end{cases}\end{aligned}$$ We will include the formulas from $\eqref{etc}$, written in this way, below in Theorem $\ref{N-nonexceptional}$.
To compute $N(q,p)$ in all other cases we will employ the following proposition:
\[fsp\] $$\begin{aligned}
-\frac{1}{2p}\sum_{j=1}^{p-1}\sin(\frac{2\pi}{p}qj)\cot(\frac{\pi}{p}j)=\bigg(\bigg(\frac{q}{p}\bigg)\bigg),\end{aligned}$$ which is the sawtooth function defined in $\eqref{st}$.
This is due to Eisenstein; see [@Apostol].
Now, we have
\[N-nonexceptional\] When $q\not\equiv (p-1) \text { mod } p$ we have the formula $$\begin{aligned}
N(q,p)=-6+\frac{12}{p}\sum_{j=1}^{p-1}\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)-4\bigg(\bigg(\frac{q^{-1;p}}{p}\bigg)\bigg) -4\bigg(\bigg(\frac{q}{p}\bigg)\bigg),\end{aligned}$$ and when $q\equiv (p-1) \text{ mod } p$ we have the formula $$\begin{aligned}
N(q,p)=N(p-1,p)=-4 -\frac{12}{p}\sum_{j=1}^{p-1}cot^2(\frac{\pi}{p}j)+4\bigg(\bigg(\frac{1}{p}\bigg)\bigg)+4\bigg(\bigg(\frac{1}{p}\bigg)\bigg).\end{aligned}$$
For the $q=p-1$ case, by examining the formulas in $\eqref{q=1}$ and $\eqref{q=p-1}$, one can easily see that we can also write $N(p-1,p)=-4p+4$ in this way. Now, consider the $1\leq q<p$ case. From , we begin by computing $$\begin{aligned}
\begin{split}
N(q,p)&=-6+\frac{14}{p}\sum_{j=1}^{p-1}\Big[\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\Big]\\
&\phantom{==}-\frac{2}{p}\sum_{j=1}^{p-1}\Big[\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\cos(\frac{2\pi}{p}j)\cos(\frac{2\pi}{p}qj)\Big]\\
&=-6+\frac{2}{p}\sum_{j=1}^{p-1}\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\Big[7-\cos(\frac{2\pi}{p}j)\cos(\frac{2\pi}{p}qj)\Big],
\end{split}\end{aligned}$$ and using the identity $\cos (2x) = 1 - 2 \sin^2 (x)$ this expands to $$\begin{aligned}
\begin{split}
&=-6+\frac{2}{p}\sum_{j=1}^{p-1}\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\Big[7-(1-2\sin^2(\frac{\pi}{p}j))(1-2\sin^2(\frac{\pi}{p}qj))\Big]\\
&=-6+\frac{2}{p}\sum_{j=1}^{p-1}\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\Big[6+2\sin^2(\frac{\pi}{p}j)+2\sin^2(\frac{\pi}{p}qj)\Big]\\
&\phantom{==}+\frac{2}{p}\sum_{j=1}^{p-1}\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\Big[-4\sin^2(\frac{\pi}{p}j)\sin^2(\frac{\pi}{p}qj))\Big],
\end{split}\end{aligned}$$ which simplifies to $$\begin{aligned}
\begin{split}
\label{rhsf}
N(q,p)&=
-6+\frac{1}{p}\sum_{j=1}^{p-1}\Big[12\cot(\frac{\pi}{p}j)\cot(\frac{\pi}{p}qj)\Big]+\frac{1}{p}\sum_{j=1}^{p-1}\Big[2\sin(\frac{2\pi}{p}j)\cot(\frac{\pi}{p}qj)\Big]\\
&\phantom{==}+\frac{1}{p}\sum_{j=1}^{p-1}\Big[4\sin(\frac{\pi}{p}qj)\cos(\frac{\pi}{p}qj)\cot(\frac{\pi}{p}j)\Big]\\
&\phantom{==}-\frac{1}{p}\sum_{j=1}^{p-1}\Big[8\sin(\frac{\pi}{p}j)\cos(\frac{\pi}{p}j)\sin(\frac{\pi}{p}qj)\cos(\frac{\pi}{p}qj)\Big].
\end{split}\end{aligned}$$ The fifth term on the right hand side of sums to zero because $$\begin{aligned}
\begin{split}
\frac{-8}{p}\sum_{j=1}^{p-1}\sin(\frac{\pi}{p}j)\cos(\frac{\pi}{p}j)&\sin(\frac{\pi}{p}qj)\cos(\frac{\pi}{p}qj)=\frac{-4}{p}\sum_{j=1}^{p-1}\sin(\frac{2\pi}{p}j)\sin(\frac{2\pi}{p}qj)\\
&=\frac{-2}{p}\sum_{j=1}^{p-1}\Big[cos(\frac{2\pi}{p}(1-q)j)-cos(\frac{2\pi}{p}(1+q)j)\Big]
=0.
\end{split}\end{aligned}$$ Using Proposition \[fsp\], the fourth term on the right hand side of is $$\begin{aligned}
\frac{4}{p} \sum_{j=1}^{p-1} \sin(\frac{\pi}{p}qj)\cos(\frac{\pi}{p}qj)\cot(\frac{\pi}{p}j)
&=\frac{2}{p} \sum_{j=1}^{p-1} \sin(2 \frac{\pi}{p}qj)\cot(\frac{\pi}{p}j)
= -4 \bigg(\bigg( \frac{q}{p}\bigg)\bigg),\end{aligned}$$ and the third term on the right hand side of is $$\begin{aligned}
\frac{2}{p}\sum_{j=1}^{p-1}\sin(\frac{2\pi}{p}j)\cot(\frac{\pi}{p}qj)&=\frac{2}{p}\sum_{j=1}^{p-1}\sin(\frac{2\pi}{p}qq^{-1;p}j)\cot(\frac{\pi}{p}qj)\\
&=\frac{2}{p}\sum_{r=1}^{p-1}\sin(\frac{2\pi}{p}q^{-1;p}r)\cot(\frac{\pi}{p}r)
=-4\bigg(\bigg(\frac{q^{-1;p}}{p}\bigg)\bigg),\end{aligned}$$ where $r=jq^{-1;p}$, and this finishes the proof.
Since the formulas for $N(q,p)$ given in Theorem $\ref{N-nonexceptional}$ are the same in all cases except for when $q = p-1$, we make the following definition:
\[d-exc\] [*[ A singularity is said to be [*[exceptional]{}*]{} if it results from a $(p-1,p)$-action. Otherwise, it is called [*[non-exceptional]{}*]{}. ]{}*]{}
Explicit formula for $N(q,p)$ {#explicit}
=============================
We begin this section by proving reciprocity formulas for the individual summands of $N(q,p)$. Then, we use these relations to prove reciprocity formulas for $N(q,p)$, which will later be used to compute $N(q,p)$ explicitly. Since we have already computed $N(1,p)$, for the simplicity of presentation, we will assume that $q>1$ for the following. To simplify notation we let $A(q,p) = 48 s(q,p)$, where $s(q,p)$ is the Dedekind sum defined in .
\[reci\]Writing $p = eq - a$, the following reciprocity relations are satisfied:
1. $A(q,p)+A(p,q)=-12+4e-4\frac{a}{q}+4\frac{q}{p}+4\frac{1}{pq}$,
2. $-4\Big(\Big( \frac {q^{-1;p}}{p}\Big)\Big)-4\Big(\Big( \frac {p^{-1;q}}{q}\Big)\Big)=-\frac{4}{pq}$,
3. $-4\Big(\Big( \frac {q}{p}\Big)\Big)-4\Big(\Big( \frac {p}{q}\Big)\Big)=-4\frac{q}{p}+4\frac{a}{q}$.
By the reciprocity formula for Dedekind sums [@Rademacher2], we have that $$\begin{aligned}
\begin{split}
A(q,p)+A(p,q)&=-12+4\Big(\frac{p}{q}+\frac{q}{p}+\frac{1}{pq}\Big)\\
&=-12+4\Big(e-\frac{a}{q}+\frac{q}{p}+\frac{1}{pq}\Big)=-12+4e-4\frac{a}{q}+4\frac{q}{p}+4\frac{1}{pq}.
\end{split}\end{aligned}$$ Next, we have that $$\begin{aligned}
\begin{split}
-4\bigg(\bigg( \frac {q^{-1;p}}{p}\bigg)\bigg)&-4\bigg(\bigg( \frac {p^{-1;q}}{q}\bigg)\bigg)=\Big(-4\frac{q^{-1;p}}{p}+2\Big)+\Big(4\frac{a^{-1;q}}{q}-2\Big)\\
&=-4\frac{q^{-1;p}}{p}+4\frac{a^{-1;q}}{q}=4\frac{-qq^{-1;p}+a^{-1;q}p}{pq}.
\end{split}\end{aligned}$$ Next, using that $q^{-1;p}q=1+ a^{-1;q} p$ (see Proposition \[nt\]), we have $$\begin{aligned}
\begin{split}
-4\bigg(\bigg( \frac {q^{-1;p}}{p}\bigg)\bigg)&-4\bigg(\bigg( \frac {p^{-1;q}}{q}\bigg)\bigg)=4\frac{-qq^{-1;p}+a^{-1;q}p}{pq}\\
&=4\frac{-(1+\alpha p)+a^{-1;q}p}{pq}=-\frac{4}{pq}.
\end{split}
\end{aligned}$$ Finally, we have $$\begin{aligned}
\begin{split}
-4\bigg(\bigg( \frac {q}{p}\bigg)\bigg)-4\bigg(\bigg( \frac {p}{q}\bigg)\bigg)&=\Big(-4\frac{q}{p}+2\Big)+\Big(4\frac{a}{q}-2\Big)=-4\frac{q}{p}+4\frac{a}{q}.
\end{split}\end{aligned}$$
Next, we will prove useful reciprocity formulas for $N(q,p)$. Denote $$\begin{aligned}
\begin{split}
&R^+(q,p)=N(q,p)+N(p,q)\\
&R^-(q,p)=N(-q,p)+N(-p,q).
\end{split}\end{aligned}$$
Writing $p=eq-a$ with $0<a<q$, we have the following formulas: \[plus,minus\] $$\begin{aligned}
R^+(q,p)=
\begin{cases}
-4 &\text{ when } q=1 \text{ and } p=2\\
-14 &\text{ when $1<q=p-1$}\\
4p-14 &\text{ when $1=q<p-1$}\\
4e-22 &\text{ when } p=eq-1\\
4e-24 &\text{ when $2\leq a \leq q-1$},
\end{cases}\end{aligned}$$ and $$\begin{aligned}
R^-(q,p) =
\begin{cases}
-4 &\text{ when } q=1 \text{ and } p=2\\
-6 &\text{ when $1<q=p-1$}\\
-4p+4 &\text{ when $1=q<p-1$}\\
-4e+2 &\text{ when $p=eq-(q-1)$ and $1<q<p-1$}\\
-4e &\text{ when $1\leq a \leq q-2$ and $2<q$}.
\end{cases}\end{aligned}$$
The first three formulas for both $R^+(q,p)$ and $R^-(q,p)$ are easily computable from the cases where $N(q,p)$ is easy to compute. Denote by $C_{(\alpha,\beta)}$ the constant term in $N(\alpha,\beta)$, so $$\begin{aligned}
\label{cab}
C_{(\alpha,\beta)}=
\begin{cases}
-6 &\text{ for a non-exceptional singularity}\\
-4 &\text{ for an exceptional singularity}.
\end{cases}\end{aligned}$$ For the case when $p=eq-a$, where $1\leq a <q-1$, we have that $$\begin{aligned}
R^+(q,p)=N(q,p)+N(p,q)&=\bigg[C_{(q,p)}+A(q,p)-4\bigg(\bigg(\frac{q^{-1;p}}{p}\bigg)\bigg) -4\bigg(\bigg(\frac{q}{p}\bigg)\bigg)\bigg]\\
&\phantom{==}+\bigg[C_{(p,q)}+A(p,q)-4\bigg(\bigg(\frac{p^{-1;q}}{q}\bigg)\bigg)-4\bigg(\bigg(\frac{p}{q}\bigg)\bigg)\bigg].\end{aligned}$$ Then, by Proposition $\ref{reci}$, we see that $$\begin{aligned}
R^+(q,p)&=C_{(q,p)}+C_{(p,q)}+\Big[-12+4e-4\frac{a}{q}+4\frac{q}{p}+\frac{4}{pq}\Big]+\Big[-\frac{4}{pq}-4\frac{q}{p}+4\frac{a}{q}\Big]\\
&=4e+C_{(q,p)}+C_{(p,q)}-12,\end{aligned}$$ which proves the reciprocity formulas in each respective case. The proof for $R^-(q,p)$ is similar and is omitted.
We next use the above reciprocity relations to recursively compute an explicit formula for $N(q,p)$:
For $q$ and $p$ and relatively prime, we have \[N=\] $$\begin{aligned}
\label{nqpform}
N(q,p) =
\begin{cases}
\displaystyle \sum_{i=1}^{k}4e_i-12k-2 &\text{ when $q\not\equiv (p-1)$ mod $p$} \\
\displaystyle \sum_{i=1}^{k}4e_i-12k=-4p+4 &\text{ when $q\equiv (p-1)$ mod $p$},
\end{cases}\end{aligned}$$ where $k$ and $e_i$, $1 \leq i \leq k$, were defined above in the modified Euclidean algorithm .
We have already proved the second case in $\eqref{etc}$, and we will now prove the first case, so we only need consider $q\not\equiv (p-1) \text{ mod } p$. Since our formulas only depend upon $q$ mod $p$, we can assume that $1\leq q<p-1$. We begin by using Proposition \[plus,minus\] to compute $N(q,p)$ as follows: $$\begin{aligned}
N(q,p)&=R^+(q,p)-N(p,q)\\
&=R^+(q,p)-N(e_1q-a_1,q)\\
&=R^+(q,p)-N(-a_1,q)\\
&=R^+(q,p)-N(-a_1,q)-N(-q,a_1)+N(-q,a_1)\\
&=R^+(q,p)-R^-(a_1,q)+N(a_2,a_1)+N(a_1,a_2)-N(a_1,a_2)\\
&=R^+(q,p)-R^-(a_1,q)+R^+(a_2,a_1)-N(-a_3,a_2).\end{aligned}$$ Continuing this iteratively, we arrive at the formula $$\begin{aligned}
N(q,p)
&=\sum_{i=1}^{r+1}4e_i-24\Big\lceil \frac{r+1}{2} \Big\rceil \\
&\phantom{==}+\Big[(-1)^{r+1}R^{(-1)^{r+1}}(a_{r+1},a_r)+(-1)^{r+2}N((-1)^{r+2}a_{r+2},a_{r+1})\Big],\end{aligned}$$ where $a_r=e_{r+2}a_{r+1}-1$ or $a_r=e_{r+2}a_{r+1}-(a_{r+1}-1)$. It is only necessary to consider the four following cases:
1. When $r+2$ is even and $a_{r+2}=1$:\
$N(q,p)=\displaystyle \sum_{i=1}^{r+3}4e_i-12(r+2)-14$.
2. When $r+2$ is odd and $a_{r+2}=1$:\
$N(q,p)= \displaystyle \sum_{i=1}^{r+3}4e_i-12(r+1)-26$.
3. When $r+2$ is even and $a_{r+2}=a_{r+1}-1$:\
$N(q,p)= \displaystyle \sum_{i=1}^{r+2}4e_i-4a_{r+1}-12(r+2)+2$.
4. When $r+2$ is odd and $a_{r+2}=a_{r+1}-1$:\
$N(q,p)= \displaystyle \sum_{i=1}^{r+2}4e_i-4a_{r+1}-12(r+1)-10$.
The formulas for $N(q,p)$ in each case are a direct consequence of formula $\eqref{etc}$ and Proposition $\ref{plus,minus}$. In case (1) and case (2), $k=r+3$. So written in terms of $k$ we have $$\begin{aligned}
N(q,p) = \sum_{i=1}^{k}4e_i-12k-2,\end{aligned}$$ for both cases. Now, in case (3), $k=(a_{r+1}-1)+(r+2)$ and $e_i=2$ for $i\geq r+3$. Therefore we can check that $$\begin{aligned}
\begin{split}
\sum_{i=1}^{k}4e_i-12k-2&=\bigg[\sum_{i=1}^{r+2}4e_i-12(r+2)\bigg]+\bigg[\sum_{i=r+3}^{k}4e_i-12(a_{r+1}-1)-2\bigg]\\
&=\sum_{i=1}^{r+2}4e_i-12(r+2)-4a_{r+1}+2=N(q,p).
\end{split}\end{aligned}$$ Finally, in case (4), $k=(a_{r+1}-1)+(r+2)$ and $e_i=2$ for $i\geq r+3$, and the result holds similarly.
Theorem \[mainthm\] is then a trivial consequence of Theorem \[N=\] and .
[*Ashikaga and Ishizaka prove a recursive formula for the Dedekind sum in [@AshiIshi Theorem 1.1], which is equivalent to Theorem \[N=\]. However, our proof is more elementary and relies only on the reciprocity law for Dedekind sums. We will also need to use Proposition \[plus,minus\] below in Section \[wpssec\].* ]{}
Index on Calderbank-Singer spaces {#CSindex}
==================================
In this section, we prove the results regarding the Calderbank-Singer metrics. Let $k$ and $k'$ be the lengths of the modified Euclidean algorithm for $(q,p)$ and $(p-q,p)$ respectively.
It follows from that the compactified Calderbank-Singer space $(\hat{X}, \hat{g})$ satisfies $\tau_{top}(\hat{X})=-k$ and $\chi_{top}(\hat{X})=k+2$, so for a $(p-q,p)$-action when $q\neq 1$, the index is $$\begin{aligned}
Ind(\hat{X}, \hat{g})&=\frac{1}{2}(15\chi_{top}+29\tau_{top})+N(q,p)
=[-7k+15]+\Big[\sum_{i=1}^{k'}4e_i'-12k'-2\Big].\end{aligned}$$ We next use a 4-dimensional $(q,p)$-football, denoted by $S^4_{(q,p)}$, to relate $k$ and $k'$. This is defined using the $\Gamma_{(p,q)}$ action, acting as rotations around $x_5$-axis: $$\begin{aligned}
S^4_{(q,p)}=S^4/\Gamma_{(q,p)}.\end{aligned}$$ This quotient is an orbifold with two singular points, one of $(q,p)$-type, and the other of $(-q,p)$-type. Since $\chi_{top}(S^4_{(q,p)})=2$ and $\tau_{top}(S^4_{(q,p)})=0$, the index of on $S^4_{(q,p)}$ with the round metric $g_S$ is $$\begin{aligned}
Ind(S^4_{(q,p)},g_S)=
3 \text{ for $1<q<p-1$}.\end{aligned}$$ Using the formula $$\begin{aligned}
Ind(S^4_{(q,p)},g_S)=\frac{1}{2}(15\chi_{top}+29\tau_{top})+N(q,p)+N(-q,p),\end{aligned}$$ and Theorem \[N=\], we have $$\begin{aligned}
\begin{split}
-12&=N(q,p)+N(-q,p)=N(q,p)+N(p-q,p)\\
&=\Big [\sum_{i=1}^{k}4e_i-12k-2\Big]+\Big[\sum_{i=1}^{k'}4e_i'-12k'-2\Big],
\end{split}\end{aligned}$$ which yields the formula $$\begin{aligned}
k'=\frac{1}{12}\Big(8+\sum_{i=1}^{k}4e_i+\sum_{i=1}^{k'}4e_i'-12k\Big).\end{aligned}$$ Then, substituting this for $k$ in $Ind(\hat{X}, \hat{g})$ gives $$\begin{aligned}
\begin{split}
Ind(\hat{X}, \hat{g})&=[-7k+15]+\Big[\sum_{i=1}^{k'}4e_i'-\Big(8+\sum_{i=1}^{k}4e_i+\sum_{i=1}^{k'}4e_i'-12k
\Big)-2\Big]\\
&=5k+5-\sum_{i=1}^{k}4e_i.
\end{split}\end{aligned}$$
Next, when $q=1$, we have $k =1$, so the index is $$\begin{aligned}
Ind(\hat{X}, \hat{g}) =[-7k+15]+[-4p+4] =-4p+12.\end{aligned}$$
Calderbank-Singer showed that their toric metrics come in families of dimension $k-1$. It was proved by Dominic Wright that the moduli space of toric anti-self-dual metrics on the orbifolds are of dimension exactly $k -1$ [@Wright Corollary 1.1]. So as long as we show the moduli space is strictly larger than $k-1$, there must be non-toric deformations.
The $(1,2)$ case is the Eguchi-Hanson metric which has no deformations. For $q=1$ and $p > 2$, the $(1,p)$ type Calderbank-Singer spaces are exactly the LeBrun negative mass metrics on $\mathcal{O}(-p)$ found in [@LeBrunnegative]. It was shown in [@HondaOn] for $p = 3$, the moduli space of these metrics is of dimension $1$ so the result is true since $1 > 0 = k - 1$. For $p \geq 4$, by [@ViaclovskyIndex Theorem 1.9], the moduli space has dimension at least $4p - 12 > 0$ (in fact the dimension is exactly $4p-12$, see [@HondaOn Theorem 1.1]). So the result holds for $q = 1$ and $p \geq 3$. We also mention that [@HondaOn Theorem 1.1] determines exactly the identity component of the automorphism groups of the deformations.
Next, assume that $q = p-1$. In this case, the metrics are hyperkähler, and correspond to toric multi-Eguchi-Hason metrics. In this case, the moduli space of all hyperkähler metrics is known to be exactly of dimension $3(k-1)$.
Next, we assume that $1 < q < p-1$. As mentioned in the Introduction, from [@LeBrunMaskit Theorem 4.2], we know that $\dim(H^2(\hat{X},\hat{g}))=0$. Also, $\dim(H^0) = 2,$ since the metrics are toric and $q >1$. Therefore $$\begin{aligned}
\dim(H^1)=-Ind(\hat{X}, \hat{g}) +\dim(H^0) = -Ind(\hat{X}, \hat{g}) +2.\end{aligned}$$ When $q \neq 1$, we have that $$\begin{aligned}
-Ind =-5k-5+\sum_{i=1}^{k}4e_i.\end{aligned}$$ Since $e_i \geq 2 $ for all $i$ and since $q < p-1$, then $e_j \geq 3$ for some $j$, $1 \leq j \leq k$. Therefore $$\begin{aligned}
\dim(H^1)\geq 3k + 1.\end{aligned}$$ The actual moduli space is locally isomorphic to $H^1 / H^0$, so it has dimension at least $3k - 1 > 3(k-1)$.
Index on weighted projective spaces {#wpssec}
===================================
In this section we will study the index of the complex at the Bochner-Kähler metrics of Bryant [*[with reversed orientation to make them anti-self-dual]{}*]{}. This reversal of orientation makes the orbifold points have orientation-reversing conjugate actions as follows:
1. Around \[1,0,0\] there is a $(-q^{-1;r}p,r)$-action.
2. Around \[0,1,0\] there is a $(-p^{-1;q}r,q)$-action.
3. Around \[0,0,1\] there is a $(-r^{-1;p}q,p)$-action.
In the next subsection, we will present some elementary number theoretic propositions that we will use throughout our computations. After that, we will prove crucial reciprocity laws for sawtooth functions relating $r, q$ and $p$ and then employ these to prove our main formula for the index. Finally, we use this formula to prove Theorem \[wpsthm\].
Elementary number theoretic preliminaries
-----------------------------------------
Recall that for two relatively prime positive integers $1<\alpha < \beta$, that we denote $\alpha$’s inverse modulo $\beta$ by $\alpha^{-1;\beta}$, and $\beta$’s inverse modulo $\alpha$ by $\beta^{-1;\alpha}$. Since $\alpha < \beta$ we can write $$\begin{aligned}
\beta=e\alpha-a,\end{aligned}$$ where $e$ and $a$ are positive integers with $a<\alpha$. Then we have the following proposition:
\[nt\] We have the following identities:
1. $\beta^{-1;\alpha}=\alpha-a^{-1;\alpha}$
2. $\alpha \alpha^{-1;\beta}=1+a^{-1;\alpha}\beta$.
To prove the first identity, recall that $\beta=e\alpha-a$ so $$\begin{aligned}
\beta (\alpha-a^{-1;\alpha})&=(e\alpha-a)(\alpha-a^{-1;\alpha})
=e\alpha^2-e\alpha a^{-1;\alpha}-a\alpha+aa^{-1;\alpha}\equiv 1 \mod \alpha.\end{aligned}$$ This proves the first identity because $\alpha-a^{-1;\alpha}<\alpha$ and the multiplicative inverses are unique.
To prove second identity we first write $$\begin{aligned}
\alpha\alpha^{-1;\beta}=1+X\beta.\end{aligned}$$ Since $1<\alpha$ we know $X$ must be a positive integer. We can then solve for $$\begin{aligned}
\beta=\frac{\alpha \alpha^{-1;\beta}-1}{X}.\end{aligned}$$ Therefore $\beta=\frac{\alpha \alpha^{-1;\beta}-1}{X}=e\alpha-a$, so $$\begin{aligned}
\alpha\alpha^{-1;\beta}-1=e\alpha X-aX,\end{aligned}$$ from which we see that $$\begin{aligned}
aX=\alpha(eX-\alpha^{-1;\beta})+1,\end{aligned}$$ so $aX\equiv \text{1 mod $\alpha$}$. This proves the second identity because $X=\frac{\alpha \alpha^{-1;\beta}-1}{\beta}<\alpha$ and multiplicative inverses are unique.
For convenience, we define the fractional part of $x$ by $$\begin{aligned}
\{ x \} = x - \lfloor x \rfloor.\end{aligned}$$ We will use the following proposition extensively in the next section, the proof is elementary:
\[saw\] For any real $\alpha$ and $\beta$, both non-integral, $$\begin{aligned}
((\alpha+\beta))=
\begin{cases}
((\alpha))+((\beta))+\frac{1}{2} &\text{ when $\{\alpha\}+\{\beta\}<1$}\\
((\alpha))+((\beta))-\frac{1}{2} &\text{ when $\{\alpha\}+\{\beta\}>1$}\\
0 &\text{ when $\{\alpha\}+\{\beta\}=1$}.
\end{cases}\end{aligned}$$
Reciprocity formulas for sawtooth functions
-------------------------------------------
Let $r<q<p$ and write: $$\begin{aligned}
\begin{split}
&p=e_{pr}r-a_{pr}\\
&p=e_{pq}q-a_{pq}\\
&q=e_{qr}r-a_{qr}.
\end{split}\end{aligned}$$ We have the following identities from Proposition $\ref{nt}$ (1): $$\begin{aligned}
\begin{split}
&p^{-1;r}=r-a^{-1;r}_{pr}\\
&p^{-1;q}=q-a^{-1;q}_{pq}\\
&q^{-1;r}=r-a^{-1;r}_{qr}
\end{split}\end{aligned}$$ and from Proposition $\ref{nt}$ (2): $$\begin{aligned}
\begin{split}
&rr^{-1;p}=1+a^{-1;r}_{pr}p\\
&rr^{-1;q}=1+a^{-1;r}_{qr}q\\
&qq^{-1;p}=1+a^{-1;q}_{pq}p.
\end{split}\end{aligned}$$ We now use these identities to prove reciprocity laws for the sawtooth function. These reciprocity laws will be broken up into two theorems where the first is independent of $r+q$ in relation to $p$ and the second is dependent.
\[2\] We have the following reciprocity relations:
1. $\Big(\Big(\frac{qp^{-1;r}}{r}\Big)\Big)+ \Big(\Big(\frac{qr^{-1;p}}{p}\Big)\Big)=\frac{q}{pr}$\
2. $\Big(\Big(\frac{rp^{-1;q}}{q}\Big)\Big)+ \Big(\Big(\frac{rq^{-1;p}}{p}\Big)\Big)=\frac{r}{pq}$.
Consider the first reciprocity relation. We have $$\begin{aligned}
\begin{split}
\bigg(\bigg(\frac{qp^{-1;r}}{r}\bigg)\bigg)+ \bigg(\bigg(\frac{qr^{-1;p}}{p}\bigg)\bigg)&=\bigg(\bigg(\frac{q(r-a^{-1;r}_{pr})}{r}\bigg)\bigg)+\bigg(\bigg(\frac{rr^{-1;p}q}{pr}\bigg)\bigg)\\
&=\bigg(\bigg(q-\frac{qa^{-1;r}_{pr}}{r}\bigg)\bigg)+\bigg(\bigg(\frac{q}{pr}+\frac{qa^{-1;r}_{pr}}{r}\bigg)\bigg)\\
&=-\bigg(\bigg(\frac{qa^{-1;r}_{pr}}{r}\bigg)\bigg)+\bigg(\bigg(\frac{q}{pr}+\frac{qa^{-1;r}_{pr}}{r}\bigg)\bigg)
\end{split}\end{aligned}$$ Now, we can write $\frac{qa^{-1;r}_{pr}}{r}=X+\frac{C}{r}$ where $0<X$ and $0<C<r$ are positive integers so that $$\begin{aligned}
\label{1}-\bigg(\bigg(\frac{qa^{-1;r}_{pr}}{r}\bigg)\bigg)+\bigg(\bigg(\frac{q}{pr}+\frac{qa^{-1;r}_{pr}}{r}\bigg)\bigg)&=
-\bigg(\bigg(\frac{C}{r}\bigg)\bigg)+\bigg(\bigg(\frac{q}{pr}+\frac{C}{r}\bigg)\bigg).\end{aligned}$$ Since $0<C<r$ we know that $$\begin{aligned}
\frac{q}{pr}+\frac{C}{r}\leq \frac{q}{pr}+\frac{r-1}{r}=\frac{q}{pr}+\frac{pr-p}{pr}<1,\end{aligned}$$ because $q<p$, which implies that $$\begin{aligned}
\bigg\{\frac{p}{qr}\bigg\}+\bigg\{\frac{C}{r}\bigg\}<1.\end{aligned}$$ Therefore, by Proposition $\ref{saw}$, we can separate the second sawtooth function to get $$\begin{aligned}
\bigg(\bigg(\frac{q}{pr}+\frac{C}{r}\bigg)\bigg)=\bigg(\bigg(\frac{q}{pr}\bigg)\bigg)+\bigg(\bigg(\frac{C}{r}\bigg)\bigg)+\frac{1}{2}.\end{aligned}$$ Putting this back into $\eqref{1}$ we see that $$\begin{aligned}
\begin{split}
\bigg(\bigg(\frac{qp^{-1;r}}{r}\bigg)\bigg)&+ \bigg(\bigg(\frac{qr^{-1;p}}{p}\bigg)\bigg)=-\bigg(\bigg(\frac{C}{r}\bigg)\bigg)+\bigg(\bigg(\frac{q}{pr}+\frac{C}{r}\bigg)\bigg)\\
&=-\bigg(\bigg(\frac{C}{r}\bigg)\bigg)+\bigg(\bigg(\frac{q}{pr}\bigg)\bigg)+\bigg(\bigg(\frac{C}{r}\bigg)\bigg)+\frac{1}{2}=\bigg(\bigg(\frac{q}{pr}\bigg)\bigg)+\frac{1}{2}\\
&=\frac{q}{pr}-\Big\lfloor \frac{q}{pr}\Big \rfloor=\frac{q}{pr}.
\end{split}\end{aligned}$$ The proof of the second reciprocity relation exactly follows the proof of the first.
The next theorem gives a similar reciprocity relation, but it is dependent upon $r+q$ in relation to $p$.
\[3\] For $1<r<q<p$ we have the following reciprocity relation: $$\begin{aligned}
\bigg(\bigg(\frac{q^{-1;r}p}{r}\bigg)\bigg)+\bigg(\bigg(\frac{r^{-1;q}p}{q}\bigg)\bigg)=
\begin{cases}
\frac{p}{qr} &\text{when $r+q>p$}\\
\frac{p}{qr}-1 &\text{when $r+q=p$}\\
\frac{p}{qr}-\lfloor \frac{p}{qr}\rfloor &\text{when $r+q<p$}\\
&\text{and $\Big\{\frac{p}{qr}\Big\}<\Big\{\frac{q^{-1;r}p}{r}\Big\}$}\\
\frac{p}{qr}-\lfloor \frac{p}{qr}\rfloor-1 &\text{when $r+q<p$}\\
&\text{and $\Big\{\frac{p}{qr}\Big\}>\Big\{\frac{q^{-1;r}p}{r}\Big\}$}.
\end{cases}\end{aligned}$$
We begin in a similar way to the proof of Theorem \[2\]: $$\begin{aligned}
\begin{split}
\bigg(\bigg(\frac{q^{-1;r}p}{r}\bigg)\bigg)+\bigg(\bigg(\frac{r^{-1;q}p}{q}\bigg)\bigg)&=\bigg(\bigg(\frac{p(r-a^{-1;r}_{qr})}{r}\bigg)\bigg)+\bigg(\bigg(\frac{(1+a^{-1;r}_{qr}q)p}{qr}\bigg)\bigg)\\
&=-\bigg(\bigg(\frac{a^{-1;r}_{qr}p}{r}\bigg)\bigg)+\bigg(\bigg(\frac{p}{qr}+\frac{a^{-1;r}_{qr}p}{r}\bigg)\bigg).
\end{split}\end{aligned}$$ Now, we can write $\frac{a^{-1;r}_{qr}p}{r}=X+\frac{C}{r}$ where $0<X$ and $0<C<r$ are positive integers so that $$\begin{aligned}
-\bigg(\bigg(\frac{a^{-1;r}_{qr}p}{r}\bigg)\bigg)+\bigg(\bigg(\frac{p}{qr}+\frac{a^{-1;r}_{qr}p}{r}\bigg)\bigg)&=-\bigg(\bigg(\frac{C}{r}\bigg)\bigg)+\bigg(\bigg(\frac{p}{qr}+\frac{C}{r}\bigg)\bigg).\end{aligned}$$ The same argument that we used in the previous proof to split up the second sawtooth function will no longer work because $p>q$, which could allow $\frac{p}{qr}+\frac{C}{r}>1$. Lets consider the first case of this reciprocity relation when $r+q>p$. In this case we know that $p<rq$ so $p<2q$ since $1<r<q$. Now, we will show that $C\leq r-2$ and use this to prove the first case. Write $p$ as $$\begin{aligned}
p=k_{pr}r+m_{pr},\end{aligned}$$ where $k_{pr}=e_{pr}-1$ and $m_{pr}=r-a_{pr}$ are positive integers. We know that $\frac{C}{r}$ is going to be the fractional part of $\frac{pa^{-1;r}_{qr}}{r}$ which equals the fractional part of $\frac{m_{pr}a^{-1;r}_{qr}}{r}$. If this equals $\frac{r-1}{r}$ then $m_{pr}a^{-1;r}_{pr}\equiv -1 (r)$ and therefore $m_{pr}=r-a_{qr}$, because multiplicative inverses are unique, which implies that $a_{pr}=a_{qr}$ (denote this value by $A$). So we have that $$\begin{aligned}
e_{pr}r-A=p<q+r=(e_{qr}r-A)+r=(e_{qr}+q)r-A,\end{aligned}$$ which is a contradiction because $e_{pr}\geq e_{qr}+1$. Therefore $\frac{C}{r}\leq \frac{r-2}{r}$, so $$\begin{aligned}
\frac{p}{qr}+\frac{C}{r}\leq \frac{p}{qr}+\frac {r-2}{r}=\frac{p}{qr}+\frac{qr-2q}{qr}<1,\end{aligned}$$ which implies that $$\begin{aligned}
\bigg\{\frac{p}{qr}\bigg\}+\bigg\{\frac{C}{r}\bigg\}<1.\end{aligned}$$ Therefore, by Proposition $\ref{saw}$, we can separate the second sawtooth function to get $$\begin{aligned}
\begin{split}
-\bigg(\bigg(\frac{C}{r}\bigg)\bigg)+\bigg(\bigg(\frac{p}{qr}+\frac{C}{r}\bigg)\bigg)&=-\bigg(\bigg(\frac{C}{r}\bigg)\bigg)+\bigg(\bigg(\frac{p}{qr}\bigg)\bigg)+\bigg(\bigg(\frac{C}{r}\bigg)\bigg)+\frac{1}{2}\\
&=\bigg(\bigg(\frac{p}{qr}\bigg)\bigg)+\frac{1}{2}=\frac{p}{qr}-\Big\lfloor \frac{p}{qr}\Big \rfloor=\frac{p}{qr},
\end{split}\end{aligned}$$ which proves the first case.
Now, consider the second case when $r+q=p$. In this case we see that $$\begin{aligned}
\begin{split}
\bigg(\bigg(\frac{p}{qr}+\frac{pa^{-1;r}_{qr}}{r}\bigg)\bigg)&=\bigg(\bigg(\frac{p}{qr}+\frac{(r+q)a^{-1;r}_{qr}}{r}\bigg)\bigg)\\
&=\bigg(\bigg(\frac{p}{qr}-\frac{a_{qr}a^{-1;r}_{qr}}{r}\bigg)\bigg)=\bigg(\bigg(\frac{p}{qr}-\frac{1}{r}\bigg)\bigg)\\
&=\Big(\frac{p}{qr}-\frac{1}{r}\Big)-\Big\lfloor \frac{p}{qr}-\frac{1}{r} \Big\rfloor -\frac{1}{2}
=\frac{1}{q}-\frac{1}{2}.
\end{split}\end{aligned}$$ Then, we compare this to $$\begin{aligned}
\begin{split}
\bigg(\bigg(&\frac{p}{qr}\bigg)\bigg)+\bigg(\bigg(\frac{pa^{-1;r}_{qr}}{r}\bigg)\bigg)+\frac{1}{2}=\bigg(\bigg(\frac{p}{qr}\bigg)\bigg)+\bigg(\bigg(\frac{(r+q)a^{-1;r}_{qr}}{r}\bigg)\bigg)+\frac{1}{2}\\
&=\bigg(\bigg(\frac{p}{qr}\bigg)\bigg)+\bigg(\bigg(\frac{-a_{qr}a^{-1;r}_{qr}}{r}\bigg)\bigg)+\frac{1}{2}=\bigg(\bigg(\frac{p}{qr}\bigg)\bigg)-\bigg(\bigg(\frac{1}{r}\bigg)\bigg)+\frac{1}{2}\\
&=\Big(\frac{p}{qr}-\Big\lfloor \frac{p}{qr}\Big \rfloor-\frac{1}{2}\Big)-\Big(\frac{1}{r}-\Big\lfloor \frac{1}{r}\Big \rfloor -\frac{1}{2}\Big)+\frac{1}{2}\\
&=\frac{p}{qr}-\frac{1}{r}+\frac{1}{2}=\frac{1}{q}+\frac{1}{2}
=\bigg(\bigg(\frac{p}{qr}+\frac{pa^{-1;r}_{qr}}{r}\bigg)\bigg)+1.
\end{split}\end{aligned}$$ Therefore we see that $$\begin{aligned}
\begin{split}
-\bigg(\bigg(\frac{a^{-1;r}_{qr}p}{r}\bigg)\bigg)&+\bigg(\bigg(\frac{p}{qr}+\frac{a^{-1;r}_{qr}p}{r}\bigg)\bigg)\\
&=-\bigg(\bigg(\frac{a^{-1;r}_{qr}p}{r}\bigg)\bigg)+\bigg(\bigg(\frac{p}{qr}\bigg)\bigg)+\bigg(\bigg(\frac{pa^{-1;r}_{qr}}{r}\bigg)\bigg)-\frac{1}{2}\\
&=\bigg(\bigg(\frac{p}{qr}\bigg)\bigg)-\frac{1}{2}=\frac{p}{qr}-\Big\lfloor \frac{p}{qr} \Big\rfloor -1=\frac{p}{qr}-1,
\end{split}\end{aligned}$$ which proves the second case.
To prove the third and fourth cases we begin once again by using that $$\begin{aligned}
\bigg(\bigg(\frac{q^{-1;r}p}{r}\bigg)\bigg)+\bigg(\bigg(\frac{r^{-1;q}p}{q}\bigg)\bigg)=\bigg(\bigg(\frac{q^{-1;r}p}{r}\bigg)\bigg)+\bigg(\bigg(\frac{p}{qr}-\frac{q^{-1;r}p}{r}\bigg)\bigg).\end{aligned}$$ Notice that $$\begin{aligned}
\Big\{\frac{p}{qr}\Big\}+\Big\{\frac{-q^{-1;r}p}{r}\Big\}=\Big\{\frac{p}{qr}\Big\}+1-\Big\{\frac{q^{-1;r}p}{r}\Big\},\end{aligned}$$ which is never equal to one because $r$, $q$ and $p$ are relatively prime. Now, the rest of the proof of the third and fourth cases follows directly from Proposition $\ref{saw}$.
$\Gamma$-Index for weighted projective spaces
---------------------------------------------
First, recall Definition $\ref{d-exc}$: singularities resulting from a $(p-1,p)$-action are said to be [*[exceptional]{}*]{} and otherwise they are called [*[non-exceptional]{}*]{}. Consider the case when $1<r<q<p$ so that there are three singularities. Before giving theorems concerning the index, we will first examine what type singularities, non-exceptional or exceptional, are admitted around each orbifold point in the cases when $r+q>p$, $r+q=p$ and $r+q<p$.
\[gr\] When $r+q>p$ all three singularites are non-exceptional. When $r+q=p$ we have that
1. The singularity at $[1,0,0]$ is always exceptional.
2. The singularity at $[0,1,0]$ is always exceptional.
3. The singularity at $[0,0,1]$ is non-exceptional and comes from a $(1,p)$-action.
When $r+q<p$ we have that
1. The singularity at $[1,0,0]$ is exceptional if and only if $p\equiv q\text{ mod $r$}$.
2. The singularity at $[0,1,0]$ is exceptional if and only if $p\equiv r\text{ mod $q$}$.
3. The singularity at $[0,0,1]$ is always non-exceptional.
At $[1,0,0]$ the $(-q^{-1;r}p,r)$-action is equivalent to a $(-a^{-1;r}_{qr}a_{pr},r)$-action, and this is equivalent to a $(r-1,r)$-action if and only if $a_{pr}=a_{qr}$. If $ r + q > p$, suppose that $a_{pr}=a_{qr}$, then $$\begin{aligned}
p=e_{pr}r-a_{qr}, \text{ and } q=e_{qr}r-a_{qr},\end{aligned}$$ so $p<q+r=(e_{qr}+1)r-a_{qr}$, which is a contradiction because $e_{pr}\geq e_{qr}+1$. If $r+q=p$ we have that $$\begin{aligned}
p=q+r=(e_{qr}+1)r-a_{qr},\end{aligned}$$ so we see that $a_{pr}=a_{qr}$ since $a_{qr}<r$. If $ r + q < p$, then this happens if and only if $p\equiv q\text{ mod $r$}$.
At $[0,1,0]$, by Remark \[actrem\], the $(-p^{-1;q}r,q)$-action is equivalent to a $(-r^{-1;q}p,q)$-action. This is equivalent to a $(r^{-1;q}a_{pq},q)$-action, which is equivalent to a $(q-1,q)$-action if and only if $a_{pq}r^{-1;q}\equiv -1 \mod q$, which would imply that $a_{pq}=q-r$. If $ r + q > p$, suppose that $a_{pq}=q-r$, then $$\begin{aligned}
p=2q-a_{pq}=2q-(q-r) =q+r,\end{aligned}$$ which is a contradiction because $r+q>p$. If $r+q=p$, we have that $$\begin{aligned}
p=2q-(q-r),\end{aligned}$$ so we see that $a_{pq}=q-r$. If $ r + q < p$ then this happens if and only if $p\equiv r \mod q$.
At $[0,0,1]$ the $(-r^{-1;p}q,p)$-action is equivalent to a $(p-1,p)$-action if and only if $r^{-1;p}q\equiv 1 \mod p$. If $ r + q > p$, this condition would imply that $q=r$, which is a contradiction. If $r+q=p$ then the $(-r^{-1;p}q,p)$-action is obviously equivalent to a $(1,p)$-action since $q = p - r$. If $ r + q < p$ then $r^{-1;p}q\equiv 1 \mod p$ occurs if and only if $q=r$, but $q>r$ so this can never happen.
In the case $r + q < p$, we can add the following:
\[lee\] When $r+q<p$ and the singularities at $[1,0,0]$ and $[0,1,0]$ are both exceptional, we have that $p=Xqr+r+q$ for some integer $X$, and $$\begin{aligned}
\bigg\{\frac{p}{qr} \bigg\}>\bigg\{\frac{q^{-1;r}p}{r}\bigg\}.\end{aligned}$$
Since the singularities around $[1,0,0]$ and $[0,1,0]$ are both exceptional, from Proposition $\ref{gr}$ we know that $$\begin{aligned}
p\equiv q\text{ mod } r, \text{ and } p\equiv r \text{ mod } q.\end{aligned}$$ Therefore, we can write $$\begin{aligned}
p=Y_1q+r=Y_2r+q,\end{aligned}$$ and solve for $$\begin{aligned}
r=\frac{Y_1-1}{Y_2-1}q,\end{aligned}$$ which implies that $qX=Y_2-1$ for some $X$ in $\mathbb{Z}$, since $q$ and $r$ are relatively prime. Then solving for $Y_2=qX+1$ we see that $$\begin{aligned}
p=(qX+1)r+q =Xqr+r+q.\end{aligned}$$ Now, since $p=Xqr+r+q$ we see that $a_{pr}=a_{qr}$. Therefore $$\begin{aligned}
\begin{split}
\bigg\{\frac{p}{qr} \bigg\}- \bigg\{\frac{a^{-1;r}_{qr}a_{pr}}{r} \bigg\}&=\bigg\{\frac{Xqr+r+q}{qr} \bigg\}- \bigg\{\frac{a^{-1;r}_{qr}a_{qr}}{r} \bigg\}\\
&=\bigg\{\frac{1}{q}+\frac{1}{r} \bigg\}- \bigg\{\frac{1}{r} \bigg\}=\frac{1}{q}>0.
\end{split}\end{aligned}$$
The following is the main result of this section, which is the same as Theorem \[introthm\] upon identifying the integer $\epsilon$ with the number of exceptional singularities:
\[thm\] Let $g$ be the canonical Bochner-Kähler metric with reversed orientation on $\overline{\mathbb{CP}}^2_{(r,q,p)}$, and assume that $1<r<q<p$. If $r+q\geq p$ then $$\begin{aligned}
Ind(\overline{\mathbb{CP}}^2_{(r,q,p)},g)=2.\end{aligned}$$ If $r+q<p$ then $$\begin{aligned}
Ind(\overline{\mathbb{CP}}^2_{(r,q,p)},g)=
\begin{cases}
2+2\epsilon -4\lfloor \frac{p}{qr} \rfloor &\text{ when $\{ \frac{p}{qr}\}<\{\frac{q^{-1;r}p}{r}\}$}\\
-2+2\epsilon-4\lfloor \frac{p}{qr} \rfloor &\text{ when $\{ \frac{p}{qr}\}>\{\frac{q^{-1;r}p}{r}\}$},
\end{cases}\end{aligned}$$ where $\epsilon$ is the number of exceptional singularities, either $0$, $1$, or $2$.
Note that from Proposition $\ref{lee}$ the only instance when two exceptional singularities can occur is in the second case, thus there are really only five distinct cases. All of these cases do in fact occur, see Table \[casestable\].
$(r,q,p)$ $\epsilon$ $\{ \frac{p}{qr}\} - \{\frac{q^{-1;r}p}{r}\}$
------------ ------------ ----------------------------------------------- --
$(3,7,11)$ 0 $< 0$
$(3,7,41)$ 0 $> 0$
$(3,7,25)$ 1 $< 0$
$(3,7,13)$ 1 $> 0$
$(3,7,31)$ 2 $> 0$
: Cases in Theorem \[thm\][]{data-label="casestable"}
Since $1<r<q<p$, there are three singularities. Furthermore, $\chi_{top} = 3$ and $\tau_{top} = -1$ (see [@Dimca Appendix B]), so the $\Gamma$-index is $$\begin{aligned}
\begin{split}
Ind&=
8 + N(-q^{-1;r}p,r)+N(-p^{-1;q}r,q)+N(-r^{-1;p}q,p)\\
= 8 &+ \bigg[C_{(-q^{-1;r}p,r)}+A(-q^{-1;r}p,r)-4\bigg(\bigg( \frac{-q^{-1;r}p}{r}\bigg)\bigg)-4\bigg(\bigg(\frac{-p^{-1;r}q}{r}\bigg)\bigg)\bigg]\\
&+\bigg[C_{(-r^{-1;q}p,q)}+A(-p^{-1;q}r,q)-4\bigg(\bigg( \frac{-r^{-1;q}p}{q}\bigg)\bigg)-4\bigg(\bigg(\frac{-p^{-1;q}r}{q}\bigg)\bigg)\bigg]\\
&+\bigg[C_{(-r^{-1;p}q,p)}+A(-r^{-1;p}q,p)-4\bigg(\bigg( \frac{-r^{-1;p}q}{p}\bigg)\bigg)-4\bigg(\bigg(\frac{-q^{-1;p}r}{p}\bigg)\bigg)\bigg],
\end{split}\end{aligned}$$ recalling that $C_{(\alpha, \beta)}$ was defined above in . Then, using Rademacher’s triple reciprocity for Dedekind sums [@Rademacher1] $$\begin{aligned}
s(q^{-1;r}p,r)+s(p^{-1;q}r,q)+s(r^{-1;p}q,p) = -\frac{1}{4}+\frac{1}{12}\bigg(\frac{r}{pq}+\frac{q}{pr}+\frac{p}{qr}\bigg),\end{aligned}$$ we see that $$\begin{aligned}
\begin{split}
Ind&=8+[C_{(-q^{-1;r}p,r)}+C_{(-r^{-1;q}p,q)}+C_{(-r^{-1;p}q,p)}]\\
&\phantom{==}+48\bigg[\frac{1}{4}-\frac{1}{12}\bigg(\frac{r}{pq}+\frac{q}{pr}+\frac{p}{qr}\bigg)\bigg]\\
&\phantom{==}+4\bigg[\bigg(\bigg( \frac{q^{-1;r}p}{r}\bigg)\bigg)+\bigg(\bigg(\frac{p^{-1;r}q}{r}\bigg)\bigg)+\bigg(\bigg( \frac{r^{-1;q}p}{q}\bigg)\bigg)\bigg]\\
&\phantom{==}+4\bigg[\bigg(\bigg(\frac{p^{-1;q}r}{q}\bigg)\bigg)+\bigg(\bigg( \frac{r^{-1;p}q}{p}\bigg)\bigg)+\bigg(\bigg(\frac{q^{-1;p}r}{p}\bigg)\bigg)\bigg].
\end{split}\end{aligned}$$ Now, using our reciprocity laws for sawtooth functions, Theorems $\ref{2}$ and $\ref{3}$, and the restrictions on the types of singularities admitted, Proposition $\ref{gr}$, we complete the proof for each case.
When $r+q>p$: $$\begin{aligned}
Ind&=8+[-18]+48\bigg[\frac{1}{4}-\frac{1}{12}\bigg(\frac{r}{pq}+\frac{q}{pr}+\frac{p}{qr}\bigg)\bigg] +4\bigg[\frac{r}{pq}+\frac{q}{pr}+\frac{p}{qr}\bigg] =2.\end{aligned}$$
When $r+q=p$: $$\begin{aligned}
Ind&=8+[-14]+48\bigg[\frac{1}{4}-\frac{1}{12}\bigg(\frac{r}{pq}+\frac{q}{pr}+\frac{p}{qr}\bigg)\bigg] +4\bigg[\frac{r}{pq}+\frac{q}{pr}+\frac{p}{qr}-1\bigg] =2.\end{aligned}$$
When $r+q<p$ and $\{ \frac{p}{qr}\}<\{\frac{q^{-1;r}p}{r}\}$: $$\begin{aligned}
Ind&=8+[C_{(-q^{-1;r}p,r)}+C_{(-r^{-1;q}p,q)}+C_{(-r^{-1;p}q,p)}]\\
&\phantom{==}+48\bigg[\frac{1}{4}-\frac{1}{12}\bigg(\frac{r}{pq}+\frac{q}{pr}+\frac{p}{qr}\bigg)\bigg]+4\bigg[\frac{r}{pq}+\frac{q}{pr}+\frac{p}{qr}-\Big\lfloor \frac{p}{qr}\Big\rfloor\bigg]\\
&=20+[C_{(-q^{-1;r}p,r)}+C_{(-r^{-1;q}p,q)}+C_{(-r^{-1;p}q,p)}]-4\Big\lfloor \frac{p}{qr}\Big\rfloor\\
&=2+2\epsilon-4\Big\lfloor \frac{p}{qr}\Big\rfloor.\end{aligned}$$
When $r+q<p$ and $\{ \frac{p}{qr}\}>\{\frac{q^{-1;r}p}{r}\}$: $$\begin{aligned}
Ind&=8+[C_{(-q^{-1;r}p,r)}+C_{(-r^{-1;q}p,q)}+C_{(-r^{-1;p}q,p)}]\\
&\phantom{==}+48\bigg[\frac{1}{4}-\frac{1}{12}\bigg(\frac{r}{pq}+\frac{q}{pr}+\frac{p}{qr}\bigg)\bigg]+4\bigg[\frac{r}{pq}+\frac{q}{pr}+\frac{p}{qr}-1-\Big\lfloor \frac{p}{qr}\Big\rfloor\bigg]\\
&=16+[C_{(-q^{-1;r}p,r)}+C_{(-r^{-1;q}p,q)}+C_{(-r^{-1;p}q,p)}]-4\Big\lfloor \frac{p}{qr}\Big\rfloor\\
&=-2+2\epsilon-4\Big\lfloor \frac{p}{qr}\Big\rfloor.\end{aligned}$$ This completes the proof.
We also state the following theorem, which gives the index in the cases when there are strictly less than three singularities.
Let $g$ be the canonical Bochner-Kähler metric with reversed orientation on $\overline{\mathbb{CP}}^2_{(r,q,p)}$. When $1=r<q<p$ there are two singularities and $$\begin{aligned}
Ind(\overline{\mathbb{CP}}^2_{(1,q,p)},g)=
\begin{cases}
2 &\text{ when $q=p-1$}\\
-4\lfloor \frac{p}{q} \rfloor+6 &\text{ when $p=eq-(q-1)$ and $q\neq p-1$}\\
-4\lfloor \frac{p}{q} \rfloor+4 &\text{ when $1\leq a_{pq} \leq q-2$ and $q>2$}.
\end{cases}\end{aligned}$$ When $1=r=q<p$ there is one singularity and $$\begin{aligned}
Ind(\overline{\mathbb{CP}}^2_{(1,1,p)},g)=-4p+12.\end{aligned}$$
We have that $$\begin{aligned}
\frac{1}{2}(15\chi_{top}+20\tau_{top})=8,\end{aligned}$$ Since $1=r<q<p$ we know that $p>2$. The first case follows from the reciprocity formula for $R^-(q,p)$ in Proposition $\ref{plus,minus}$. The second case follows from $N(-1,p)=-4p+4$ in $\eqref{etc}$.
Proof of Theorem \[wpsthm\]
---------------------------
We first present a general result about $H^2(M,g)$ on certain self-dual Kähler orbifolds:
\[sdk\] Let $(M,g)$ be a compact self-dual Kähler orbifold and assume that the set $M^{>0} = \{ p \in M, R(p) > 0 \}$ is non-empty. With the reversed orientation to make $g$ anti-self-dual, we have $H^2(M,g) = 0$.
As mentioned in the Introduction, the metric $\tilde{g} = R^{-2} g$ is an Einstein metric, which is complete on components of $M^*$. If $Z \in S^2_0(\Lambda^2_+(T^*M))$ satisfies $\mathcal{D}_g^* Z = 0$, where $ \mathcal{D}^*_{g}$ is the adjoint of $\mathcal{D}_{g}$, then from conformal invariance $\mathcal{D}_{\tilde g}^* Z = 0$ when $Z$ is viewed as a $(1,3)$ tensor. We compute $$\begin{aligned}
|Z|_{\tilde{g}}^2 = \tilde{g}^{ip} \tilde{g}^{jq} Z_{ijk}^{\phantom{ijk}l}
Z_{pql}^{\phantom{pql}k}
= R^4 {g}^{ip} {g}^{jq} Z_{ijk}^{\phantom{ijk}l} Z_{pql}^{\phantom{pql}k} = R^4 |Z|_{g}^2,\end{aligned}$$ so we have $$\begin{aligned}
\label{zeqn}
|Z|_{\tilde{g}} = R^2 |Z|_{g}.\end{aligned}$$ Let $M^*_1$ denote any non-trivial component of $M^*$. Since the metric $\tilde{g}$ is Einstein on $M^*_1$, from [@Itoh Proposition 5.1], we have $$\begin{aligned}
\label{dds}
\mathcal{D}_{\tilde{g}} \mathcal{D}^*_{\tilde{g}} Z = \frac{1}{24}
( 3 \nabla^*_{\tilde{g}} \nabla_{\tilde{g}}
+ 2 R_{\tilde{g}})(2 \nabla^*_{\tilde{g}} \nabla_{\tilde{g}} + R_{\tilde{g}}) Z,\end{aligned}$$ where $R_{\tilde{g}}$ is the (constant) scalar curvature of the Einstein metric $\tilde{g}$ on $M^*_1$. If $M^{> 0} = M$, then the maximum principle immediately implies that $Z = 0$. Otherwise, there is an nontrivial open component of $M^*$, which we again call $M^*_1$. The metric $\tilde{g}$ is a complete Einstein metric on $M^*_1$, and shows that $|Z|_{\tilde{g}} (x) = o(1)$ as $r \rightarrow 0$, where $r$ is the distance to the zero set of the scalar curvature. Viewed on the complete manifold $(M^*_1, \tilde{g})$, $Z$ is then a decaying solution at infinity of $\eqref{dds}$. Since $R_{\tilde{g}}$ is a constant, a standard separation of variables argument (see for example [@Donn]) implies that $Z$ must decay faster than the inverse of any polynomial in the $\tilde{g}$ metric (it morever has exponential decay). Equivalently, $|Z|_g = O(r^k)$ as $r \rightarrow 0$ for any $k > 0$. This implies that $Z$ has a zero of infinite order along the zero set of the scalar curvature. The unique continuation principle for elliptic operators (see [@Aron]) then implies that $Z$ is identically zero.
As a corollary, we obtain
\[h2wps\] If $g$ is the canonical Bochner-Kähler metric with reversed orientation on $\overline{\mathbb{CP}}^2_{(r,q,p)}$, then $H^2(M,g) = 0$.
From [@DavidGauduchon Equation (2.32)], the set $M^{> 0}$ is non-empty. So this follows immediately from Proposition \[sdk\].
From Corollary \[h2wps\], $H^2(M,g) = 0$, so the actual moduli space is locally isomorphic to $H^1/ H^0$. Depending upon the action of $H^0$, the moduli space could therefore be of dimension $\dim(H^1)$, $\dim(H^1) - 1$, or $\dim(H^1) - 2$. The result then follows immediately from the determination of $H^1(M,g)$ in Theorem \[thm\].
Final remarks
-------------
We end with a non-rigorous remark on the number-theoretic condition appearing in Theorem \[thm\]. Figure \[plot\] contains a plot of the function $$\begin{aligned}
H(r,q,p(j)) = \Big\{ \frac{p}{qr} \Big\} - \Big\{\frac{q^{-1;r}p}{r}\Big\}\end{aligned}$$ for $r = 3$ and $q = 7$, where the horizontal axis indexes the $j$th prime. The plot begins at the fifth prime, $11$, and ends with the $100$th prime $541$. This, along with other empirical examples, indicates that the cases $H > 0$ and $H < 0$ occur with the approximately the same frequency.
![$H(3,7,p(j))$[]{data-label="plot"}](Plot.pdf)
[^1]: Research partially supported by NSF Grants DMS-0804042 and DMS-1105187
| {
"pile_set_name": "ArXiv"
} | ArXiv |
0.4cm
.25in .25in .25in
\[section\] \[theorem\][Theorem]{}
\[theorem\][Corollary]{}
\[theorem\][Lemma]{}
\[theorem\][Proposition]{}
\[section\]
[**Lexicographic Semigroupoids**]{}
[S. C. Power]{}
*Department of Mathematics and Statistics\
Lancaster University\
England LA1 4YF*
**Abstract**
[The natural lexicographic semigroupoids associated with Cantor product spaces indexed by countable linear orders are classified. Applications are given to the classification of triangular operator algebras which are direct limits of upper triangular matrix algebras. ]{}
Consider a Cantor space which is presented explicitly as an infinite product of finite topological spaces. The product presentation provides an equivalence relation $R$ consisting of the pairs $(x,y)$ of points $x$ and $y$ which disagree in at most finitely many coordinates. This equivalence relation supports a natural locally compact totally disconnected topology which makes $R$ a principal groupoid. It is well-known that in the case of countable products such topological equivalence relations are classified by the generalised integer obtained from the formal product of the cardinalities of the component spaces. Furthermore, this classification is closely related to the classification of C\*-algebras that are infinite tensor products of matrix algebras, the so-called UHF C\*-algebras. See, for example, Renault [@ren] and Power [@scp-book].
In the present paper we consider antisymmetric topological binary relations which are the lexicographic topological subrelations arising from infinite products indexed by general countable linear orderings. These natural semigroupoids are classified and their automorphism groups determined. This and related results enable us to give applications to the classification of triangular operator algebras which are themselves lexicographic products in an algebraic sense.
The binary relations may also be viewed as the (semigroupoid) lexicographic products of total orderings on finite sets, and in fact our methods are applicable to lexicographic products of connected antisymmetric finite partial orders. Although applications to approximately finite operator algebras gives our primary motivation it seems clear that lexicographic subrelations are interesting in their own right.
In the preliminary section we recall how the generalised integer associated with the presentation of the Cantor space gives a complete invariant for the associated approximately finite groupoid. In section 2 we classify the lexicographic products in the case of indexing by a countable dense order. It is interesting that the proof here is much more elementary than the case of indexing by $\ZZ$ which is taken up in section 3. Indeed the classification in the former case is less subtle and suggests, a posteriori, that the component coordinates must be accessible in purely order-topological terms. The proof in section 3 follows the order-topological methods in Power [@scp-outer] where automorphisms of the associated triangular algebras were studied. These algebras - the so-called alternation algebras - have been considered by a number of authors, namely Poon [@poo], Hopenwasser and Power [@hop-scp], and most recently, as part of a wider analysis, by Donsig and Hopenwasser [@don-hop]. Also in section 3 we identify the automorphism group of a lexicographic product over $\ZZ$. In section 4 we obtain the classification for the case of general countable linear orders and in the last section we give applications to operator algebras.
Preliminaries
==============
Let $\mu$ be a positive integer-valued function defined on a countable linear ordering $\Omega$. For notational convenience we denote the discrete finite topological space $\{1, \dots, n\}$ by \[$n$\]. Associate with $\mu$ the Cantor space
$$X_{\mu} = \prod_{w \in \Omega} [r_w]$$
where $r_w = \mu (w)$. Write $\tilde{R}_{\mu}$ for the equivalence relation described in the introduction. Let $R_{\mu}$ be the antisymmetric subrelation of points $(x, y) \in \tilde{R}_{\mu}$, for which $x$ preceeds $y$ in the lexicographic ordering. Thus $(y,x) \in
R_{\mu}$ if and only if $x = (x_w)$ and $y = (y_w)$, the coordinates $(x_w)$ and $ (y_w)$ are equal except in at most a finite number of coordinates, and $y_w < x_w$ for the first index $w$ where $x$ and $y$ differ.
The basic open-closed sets for the Cantor space $X_{\mu}$ are provided by the cylinder sets arising from the specification of a finite number of coordinates. The topology on $\tilde{R}_{\mu}$ arises in the following similar way. Let $F \subseteq \Omega$ be a finite subset, let $x^{\prime}_w, y^{\prime}_w$ be specified coordinates for $w \in F$, and let
$$E = \{(x,y) : x_w = y_w \hbox{ for } w \notin F, x_w =
x^{\prime}_w \hbox{ and } y_w = y^{\prime}_w \hbox{ for } w \in F\}.$$
The totality of these sets gives a base for the groupoid topology on $\tilde{R}_\mu$ and the subrelation $R_\mu$ carries the relative topology.
Notice that if $\pi_l, \pi_r : \tilde{R}_\mu \rightarrow X_\mu$ are the natural coordinate projections then $\pi_l$ and $\pi_r$ are continuous and are one to one when restricted to a basic open-closed set $E$. General open-closed sets with this property are called [*$G$-sets*]{} and these are important in the following two ways. Firstly they provide a class of sets which are invariant under [*groupoid isomorphism*]{}, that is, a binary relation isomorphism that is also a homeomorphism. Secondly they are used in the formulation of invariant measures on the underlying space. As a consequence groupoid isomorphisms conserve the invariant measures.
An [*invariant measure*]{} for a principal groupoid $\tilde{R}$ is a Borel measure $\lambda$ for the underlying topological space such that $\lambda
(\pi_l (G)) = \lambda (\pi_r (G))$ for every $G$-set $G$. It is a simple matter to check that $\tilde{R}_\mu$ possesses a unique invariant probability measure, namely the infinite product of the probability measures $\lambda_\omega$ on $[r_\omega]$ which assign equal mass $r_\omega^{-1}$ to each singleton set. At this stage we can deduce that if $\tilde{R}_\mu$ is isomorphic to the topological equivalence relation $\tilde{R_\nu}$, associated with $\nu : \Lambda \to \IN$, then the generalised integers for $\mu$ and $\nu$ agree. Indeed, the hypothesised isomorphism is a bijection $\alpha: X_\mu \rightarrow X_\nu$ such that the map $\alpha^{(2)} : \tilde{R}_\mu \to \tilde{R}_\nu$ is a bijection and a topological isomorphism. Since the invariant probability measures for $\tilde{R}_\mu$ and $\tilde{R}_\nu$ are unique they must correspond under $\alpha$, and from this it follows that they have the same range on open-closed sets. Thus the set of rationals $k/r$, with $k \in \ZZ$ and $r = r_{w_1}r_{w_2}\dots r_{w_n}$ for some $n \in \IN$, coincides with the corresponding set of rationals for $\nu$. Equivalently
$$\prod_{w \in \Omega}\mu (w) \ = \
\prod_{w \in \Lambda}\nu (w)$$
as generalised integers.
In the antisymmetric context we shall use the argument above in a local way (with various restrictions of $R_\mu$ in place of $\tilde{R}_\mu$) to obtain local comparisons of the data for two given lexicographic binary relations.
For $x$ in $X_\mu$ the (one-sided) [*orbit*]{} of $x$ is the set ${{\cal O}}(x)$ of points $y$ with $(y,x) \in R_\mu$, and the [*closed orbit*]{} of $x$ is $\overline{{{\cal O}}(x)}$ the closure of this set. Note that if $x =
(x_w)$ with $x_w = 1$ for all $w \leq w_1$ and $x_w = \mu (w)$ for all $w > w_1$, for some $w_1$ in $\Omega$, then $\overline{{{\cal O}}(x)}$ is the set of points $y = (y_w)$ with $y_w =1$ for $w \leq w_1$. This will be indicated by setting $y_{w_1} = 1$ and writing
$$\overline{{{\cal O}}(x)} = (..1..{|y_{w_1}|}..*..).$$
A similar such shorthand is used in the next section to indicate basic open-closed sets in $X_\mu$. Note that for the particular point $x$ above the restriction $R_\mu
\mid \overline{{{\cal O}}(x)}$ is isomorphic to a lexicographic ordering.
Countable dense orders
=======================
In this section we classify the topological binary relations $R_\mu$ in the cases when $\Omega$ is a dense linear ordering and $\mu : \Omega
\rightarrow \IN$ satisfies $\mu (w) \geq 2$ for all $w$ in $\Omega$. There are only four such linear orderings and these correspond to the presence and absence of maximal and minimal elements.
[**Theorem 2**]{} *Let $\mu : \Omega \rightarrow \IN
\backslash \{1\}$ and $\nu : \Omega^\prime \rightarrow \IN \backslash \{1\}$ be functions on the countable dense linear orderings $\Omega$ and $\Omega^\prime$. Then the lexicographic semigroupoids $R_\mu$ and $R_\nu$ are isomorphic if and only if there is an order bijection $\pi :
\Omega \rightarrow \Omega^\prime$ such that $\nu (\pi (w)) = \mu (w)$ for all $w$ in $\Omega$.*
Assume first that $\Omega$ and $\Omega^\prime$ have no minimal elements.
Observe that a point $x = (x_w)$ in $X_\mu$ has a closed orbit $\overline{{{\cal O}}(x)}$ which is a proper subset of $X_\mu$ if and only if there exists an index of $t$ in $\Omega$ such that $x_w =1$ for all $w \leq t$. Indeed, assume that this does not hold and consider an arbitrary point $y$ in $X_\mu$. For an index $s \in \Omega$, with $x_s > 1$, let $z^s = (z_w)$ where $z_w =
y_w$ for $w < s, z_s = x_s -1$, and $z_w = x_w$ for $w > s$. If $A$ is a basic open neighbourhood of $y$ then, from the assumption, it follows that there exists an element $z^s$ in $A$. Since $(z^s, x) \in R_\mu$ the observation follows.
Divide the set of points with proper closed orbit into two types. A point $X$ is of type 1 if there exists a first index $w_1$ such that $x_{w_1} \neq 1$, and is of type 2 otherwise. We claim that $x$ is type 1 if and only if $\overline{{{\cal O}}(x)}$ contains a closed orbit $\overline{{{\cal O}}(y)}$ which is a relatively open-closed proper subset. This property identifies the indexing in order topological terms and is the basis of the proof.
If $x$ is a type 2 point then there is an order interval decompositions $\Omega = \Omega_1 +
\Omega_2$, where $\Omega_2$ has no first element, such that $y$ belongs to the closed orbit of $x$ if and only if $y = (y_w)$ with $y_w =1$ for $w \in
\Omega_1$. To see this let $\Omega_1$ be the largest order ideal in $\Omega$ of indices $w$ with $x_w = 1$. Thus $\Omega = \Omega_1 +
\Omega_2$, where $\Omega_2$ has no first element and for any index $u$ in $\Omega_2$ there is a smaller index $v$ with $x_v > 1$. In particular $\overline{{{\cal O}}(x)}$ contains the set
$$(..1..\mid x_v-1\mid ..*..).$$
The union of all such sets is dense in the set
$$\overline{{{\cal O}}(x)} = (..1..\mid..*..)$$
associated with the decomposition of $\Omega$. It follows that $\overline{{{\cal O}}(x)} $ is precisely this latter set, as required.
A relatively open-closed subset of $\overline{{{\cal O}}(x)}$ contains a basic relatively open-closed neighbourhood of the form
$$(..1..|..*..|z_{w_1}|..*..|z_{w_2}|..*..,\dots ,..*..|z_{w_n}|..*..)$$
with $w_1 < w_2 < \dots < w_n$. Such a set contains points $z =
(z_w)$ with $z_t = 2$ for all $t$ in $\Omega_2$ with $t < w_1$. Since $\overline{{{\cal O}}(z}) = \overline{{{\cal O}}(x)}$ for such a point it follows that $\overline{{{\cal O}}(x)}$ cannot contain, properly, a relatively open-closed closed orbit.
We now identify the integers $\mu(w)$ in order-topological terms.
Let $E \subseteq X_\mu$ be an intersection of closed orbits of type 2 points which is not itself a closed orbit of a type 2 point. Then, in view of the description above of the closed orbits of type 2 points, E has the form
$$E^{(q)} = \bigcup^{\mu(q)}_{i=1} (..1..|i|..*..),$$
for some $q$ in $\Omega$. From this set and the relation $R_\mu$ we can discover $\mu(q)$ in order-topological terms as follows. There is a unique $R_\mu$-invariant probability measure, $\lambda$ say, on the set $E^{(q)}$. The sets
$$E^{(q)}_k = \bigcup^k_{i=1} (..1..|i|..*..)$$
are the only closed orbits contained in $E^{(q)}$ with positive $\lambda $ measure, and the reciprocal of the measure of the smallest such set is $\mu(q)$.
Suppose now that $\alpha^{(2)} : R_\mu \rightarrow R_\nu$ is a semigroupoid isomorphism implemented by $\alpha : X_\mu \rightarrow X_\nu$. In particular ${{\alpha}}$ is a homeomorphism. If $p \in \Omega^\prime$ then write $E^{(p)}$ for the sets in $X_\nu$ that are analogous to the subsets $E^{(q)}$. Then from the paragraph above it follows that ${{\alpha}}(E^{(q)}) = E^{\pi(q)}$ for some element $\pi(q)$ in $\Omega^{\prime}$. The map $\pi$ is an order isomorphism and from the characterisation of the numbers $\mu(q)$ above it follows that $\nu(\pi(q)) = \mu(q)$ for all $q$ in $\Omega$.
Suppose finally that $\Omega$ has a minimal element $w_0$. Then all points of $X_\mu$ except the unique maximal point $x_{max} = (\mu(w))$ have proper closed orbits. Also, the converse is true. Note that there exist open-closed orbits and a smallest open-closed orbit, namely
$$(1|..*..).$$
Since this set is determined in order-topological terms we can restrict considerations to this subset of $X_\mu$ and deduce the theorem in this case from the one already considered.
The Case $\Omega = \omega^* + \omega$
======================================
A classification is now given for the lexicographic semigroupoids $R_\mu$ in the case where, in the standard notation, $\Omega$ has order type $ \omega^* + \omega$. This order type is quite a bit more subtle than that of the dense orders in that semigroupoid isomorphisms may have to be effected by homeomorphisms defined in terms of a recoding.
Throughout this section assume that $\Omega = \ZZ \backslash \{0\}$ and that $\mu,\ \nu$ are maps from $\Omega$ to $\ZZ_+$ with
$$\mu (k) = r_k,\ \mu (-k) = s_k,\ \nu (k) = t_k, \ \nu (-k) = u_k,$$
for $k = 1, 2, \dots$. Associate with $\mu$ the pair $(\underline{r},
\underline{s})$ of generalised integers
$$\underline{r} = r_1 r_2 \dots,\ \
\underline{s} = s_1 s_2 \dots.$$
Define an equivalence relation $\sim$ on pairs of generalised (or finite) integers by $(\underline{r},
\underline{s}) \sim (\underline{t}, \underline{u})$ if and only if $\underline{r}\underline{s} = \underline{t}\underline{u}$ and there exist coprime natural numbers $a, b$ such that $b \underline{r} =
a \underline{t}$ and $a \underline{s} = b \underline{u}$.
[**Theorem 3**]{} *Let $\mu, \nu : \Omega \rightarrow \IN$, with associated pairs of generalised integers (possibly finite) $(\underline{r},
\underline{s})$ and $ (\underline{t}, \underline{u})$ respectively. Then the lexicographic semigroupoids $R_\mu$ and $R_\nu$ are isomorphic if and only if $(\underline{r}, \underline{s}) \sim (\underline{t},
\underline{u})$.*
We prove the necessity of the condition for isomorphism. The sufficiency direction is relatively straightforward and is left to the reader. (See also [@hop-scp] and [@poo].)
Let $\alpha : X_\mu \rightarrow X_\nu$ be a bijection such that $\alpha^{(2)} : R_\mu \rightarrow R_\nu$ is a topological isomorphism. The inversion map $\theta : X_\mu \times X_\mu \rightarrow X_\mu \times
X_\mu$ given by $\theta ((x, x^\prime)) = (x^\prime , x)$ is an automorphism of the equivalence relation $\tilde{R}_\mu$, and $\tilde{R}_\mu = R_\mu \cup \theta (R_\mu)$. It follows that $\alpha^{(2)}$ maps $\tilde{R}_\mu$ homeomorphically onto $\tilde{R}_\nu$. In particular, by the discussion in section 1, it follows that the generalised integer for $\tilde{R}_\mu$ coincides with that for $\tilde{R}_\nu$. That is, $\underline{r}\ \underline{s} =
\underline{t} \ \underline{u}$.
Let $X_{\mu, 0} \subseteq X_\mu$ be the set of points $x = (x_k)$ with $x_k =1$ for sufficiently small $k$. If $X_{\mu ,0}$ is a proper subset of $X_\mu$ then it is precisely the set of points in $X_\mu$ with proper closed orbits, and so $\alpha (X_{\mu , 0}) = X_{\nu ,0}$. The set $X_{\mu, 0} $ contains the special points for which, in addition, $x_k = r_k$ for all sufficiently large $k$. These special points can be characterised order topologically. Indeed they are precisely the points $x$ for which there exists a point $x^+$ whose closed orbit $ \overline{{{\cal O}}(x^+)}$ is the union of $\{x+\}$ and $ \overline{{{\cal O}}(x)}$. These so-called gap points are discussed in [@hop-scp] and [@scp-book]. Thus, if $x_* = (\dots, 1, 1, \hat{r}_1, r_2, \dots)$, where the symbol $\ \hat{}$ indicates the coordinate position $k =1$, then $\alpha (x_*)$ may be written as
$$(\dots, 1, 1, w_{-j +1}, w_{-j}, \dots , w_{j-1}, u_j, u_{j+1}, \dots)$$
for some positive integer $j$. We have
$$\overline{{{\cal O}}(x_*)} = (\dots 1, 1, \hat{*}, *, \dots),$$ and $$\overline{{{\cal O}}(\alpha (x_*))} = \{(\dots 1, w^\prime, y_j, y_{j+1},
\dots): w^\prime \in [1_j , w]\}$$
where $y_k \leq u_k$ for $k \geq j$, and where $w^\prime$ is any word of length $2j -2$ which preceeds, or is equal to, the word $w = (w_{-j
+1}, \dots , w_{j-1})$. That is, $w^\prime$ belongs to the lexicographic order interval $[1_j, w]$ where $1_j$ is the word with $2j-2$ coordinates all equal to 1. Let $n$ be the number of words in this order interval and note that the restricted topological equivalence relation $\tilde{R}_\nu \mid \overline{{{\cal O}}(\alpha (x_*))}$ has generalised integer $nt_j t_{j+1} \dots$. Since $\alpha$ induces an isomorphism between this relation and the topological equivalence relation $\tilde{R}_\mu \mid \overline{{{\cal O}}(x_*)}$ then, once again, by the discussion in section 1, we have $nt_j t_{j+1} \dots = r_1 r_2 \dots$, and hence $m \underline{r} = n \underline{t}$ where $m = t_1 t_2 \dots
t_{j-1}$. Although $\underline{r} \underline{s} = \underline{t}
\underline{u}$ we cannot yet conclude that $(\underline{r},
\underline{s}) \sim (\underline{t}, \underline{u})$.
Define $d_\mu : X_{\mu, 0} \rightarrow \IR$ by
$$d_\mu (x_k) = \sum^\infty_{k=1} \frac{x_k -1}{r_1 r_2 \dots r_k} +
\sum^\infty_{k=1} (x_{-k} -1) s_0 s_1 \dots s_{k-1},$$
where $s_0 =1$, and define $d_\nu : X_{\nu, 0} \rightarrow \IR$ similarly. If $x \in \overline{{{\cal O}}(x_*)}$ then we can interpret $d_\mu (x)$ as the measure of $\overline{{{\cal O}}(x)}$ with respect to the unique normalised $R_\mu-$invariant Borel measure on $\overline{{{\cal O}}(x_*)}$ Call this measure $\lambda_\mu$, and note that there is a unique $\tilde{R}_\mu$-invariant extension to $X_{\mu,0}$ which we also denote by $\lambda_\mu$.
Because of the uniqueness of normalised invariant measures it follows that $\lambda_\nu \circ \alpha = c \lambda_\mu$ and $d_\nu (\alpha (x))
= cd_\mu (x)$ for some positive constant $c$. Since $d_\mu (x_*) =1$ we have $c = d_\nu (\alpha (x_*))$. To see that $c = n/m$ note that there are precisely $m$ points in the product $[t_1] \times \dots \times [t_{j-1}]$. Thus, from the definition of $n$ and the product measure $\lambda_\nu$ we see that $\lambda_\nu(\overline{{{\cal O}}(\alpha (x_*))}) = n/m$.
We now use the connection $d_\nu (\alpha (x)) = cd_\nu (x)$ to show that $n \underline{s} = m \underline{u}$. We can assume that $ \underline{s}$ and $ \underline{u}$ are not finite.
Let $n/m = a/b$ where $(a, b) =1$ and suppose, by way of contradiction, that $a \underline{s}$ does not divide $b \underline{u}$. Then, there exists a prime number $p$ and a positive integer $g$ such that
$$p^g \mid a\underline{s},\ \ p^{g-1} \mid b\underline{u}, \ \
\hbox{ and } p^g \not{\mid}
\ b\underline{u}.$$
Choose $l$ large enough so that $as_1\dots s_l = k_1p^g$ for some integer $k_1$. Then
$$k_1 \frac{b}{a} \frac{u_1 u_2 \dots u_l}{s_1 s_2 \dots s_l} = b
\frac{u_1^\prime u^\prime_2 \dots u^\prime_l}{p}$$
and
$$(p, bu^\prime_1 u^\prime_2 \dots u^\prime_l u_{l+1} \dots u_v) =1$$
for all $v > l$, where $u^\prime_1, \dots u_l^\prime$ are factors of $u_1, \dots u_l$ respectively. Note that by increasing $l$, if necessary, and compensating with a multiple of $k_1$, we can arrange that the product $s = s_1 s_2 \dots s_l$ satisfies $p^{-1} <1 -s^{-1}$. These numerical relations lead to the contradiction that $\alpha (E) =
X_\nu$ where $E$ is the proper open-closed subset
$$E = \bigcup_{w \in W} (\dots * \dots, w, \dots * \dots)$$
where $W$ is the set of words $(s_l^\prime , \dots , s^\prime_1)$ which are strictly less than $(s_l, s_{l-1}, \dots, s_1)$ in the lexicographic ordering.
To see this let $y$ be an arbitrary point of $X_\nu$ and let $F_v (y)$ be the closed set
$$F_v (y) = \{y^\prime \in X_\nu : y^\prime = (y^\prime_k), y^\prime_k = y_k
\hbox{ for } k \geq -v\}$$
for $v = 1, 2, \dots$. The range of $d_\nu$ on $F_v (y) \cap X_{\nu,
0}$ is an arithmetic progression, namely,
$$d_\nu (F_v (y) \cap X_{\nu, 0}) = \{ku_1 u_2 \dots u_v + \xi : k \in
\ZZ_+\}$$
where $\xi = d_\nu (y^+)$, and where $y^+$ agress with $y$ in coordinates indexed by $\ZZ_+$ and is equal to 1 in the remaining coordinates.
On the other hand we have
$$d_\mu (E \cap X_{\nu , 0}) = \bigcup^\infty_{k=1} [ks, ks + (s-1)]$$
and so
$$d_\nu (\alpha (E) \cap X_{\mu , 0}) = \bigcup^\infty_{k=1} [cks, cks +
c(s-1)].$$
In particular, since $c = a/b$, the set $d_\nu (F_v (y) \cap X_{\nu, 0})$ meets $d_\nu
(\alpha (E) \cap X_{\nu, 0})$ if and only if the sets $B_v$ and $A$ meet, where
$$A = \bigcup^\infty_{k=1} [k, k+ (1-s^{-1})]$$
and
$$B_v = \{k \frac{b}{a} \frac{u_1 u_2 \dots u_v}{s} + \frac{\xi}{s} : k \in
\ZZ_+\}.$$
By our earlier remarks, and our choice of $l$, if $v > l$ then $B_v$ contains the set
$$B^\prime_v = \{k \frac{b u^\prime_1 u^\prime_2 \dots u^\prime_l
u_{l+1} \dots u_v}{p} + \frac{\xi}{s} : k \in \ZZ_+\}.$$
By the coprimality of $p$ and $b u^\prime_1 u^\prime_2 \dots u_v$ this set contains certain positive integral translates of the points
$$\frac{u}{p} + \frac{\xi}{s} , \hbox{ for } i = 1, \dots, p.$$
Since $p^{-1} < (1 -s^{-1})$ it follows that $B_v^\prime$ meets the set $A$. We have thus shown that
$$d_\nu (F_v (y) \cap X_{\nu, 0}) \cap d_\nu (\alpha (E) \cap X_{\nu, 0})
\neq \emptyset$$
and hence, by the openness of $\alpha (E)$, that
$$F_v (y) \cap \alpha (E) \neq \emptyset, \hbox{ for } v = 1, 2, \dots .$$
Since $y$ is the unique point in the intersection of the sets $F_v
(y)$, it follows that $y \in \alpha (E)$ and hence that $\alpha (E) = X_\nu$, the desired contradiction.
We have shown that $a \underline{s}$ divides $b \underline{u}$. Since $\alpha^{-1}$ is also an automorphism we conclude that $b \underline{u}$ divides $a \underline{s}$ and hence that $b \underline{u} =a
\underline{s}$, and $(\underline{r}, \underline{s}) \sim (\underline{t},
\underline{u})$.
[**Remarks 1.**]{} Note that from the proof above it follows that the restrictions $R_\mu \mid X_{\mu, 0}$ and $R_\nu \mid X_{\nu ,0}$ are isomorphic topological binary relatations if and only if $R_\mu$ and $R_\nu$ are isomorphic topological binary relations. We use this fact in the next section.
[**2.**]{} As we mention later, there is a close association between approximately finite topological binary relations and approximately finite triangular operator algebras. The algebras associated with the $\ZZ$-ordered relations above correspond to the so-called alternation algebras considered by Hopenwasser and Power [@hop-scp] and by Poon [@poo]. Their classification is also given, as part of a more general study, by Donsig and Hopenwasser [@don-hop]. The argument we give above follows very closely the method of [@scp-outer], which was restricted to the case $\mu = \nu$. (There is an inadequacy in the arithmetic progression argument of [@scp-outer] which is corrected in the somewhat more general argument above.)
[**3.**]{} In principle it should be possible to reformulate the arguments of Poon [@poo] and Hopenwasser and Donsig [@don-hop] in terms of binary relations to give alternative proofs of Theorem 3. The arguments we have given are also suited to other situations and in particular to the generalised alternation algebras associated with Markov chains and subshifts. (See [@scp-book].) We intend to report more fully on this elsewhere. However, let us note the following example from [@scp-book].
Let $R_\mu$ be the lexicographic topological binary relation for $\Omega = \ZZ$ with the function $\mu(w) = 2$ for all $w$. Let $R_1$ (resp. $R_2$) be the topological subrelation defined on the symbol subspace $X_1$ (resp. $X_2$) of $X_\mu$ for which the pair $00$ (resp. $11$) is forbidden. Then, in contrast to their generated equivalence relations, $R_1$ and $R_2$ are not isomorphic. Similarly the one-sided subrelations are not isomorphic.
[**Automorphisms**]{}
Fix $\mu : \Omega \rightarrow \IN$ as above with lexicographic semigroupoid $R_\mu$ and generalised integer pair $(\underline{r},\underline{s})$. Let $d$ be the number of primes $p$ such that both $\underline{r}$ and $\underline{s}$ are divisible by $p^\infty$.
[**Theorem 4**]{} [@scp-outer] *The semigroupoid automorphism group Aut$(R_\mu)$ is isomorphic to the restricted direct product $\ZZ^d$.*
An element $\alpha^{(2)} \in Aut (R_\mu)$ is a topological isomorphism induced by a bijection $\alpha : X_\mu \rightarrow X_\mu$. By the argument in the proof of Theorem 3, specialized to the case $\mu = \nu$, we have
$$d_\mu (\alpha (x)) = c d_\mu (x)$$
for $x \in X_{\mu , 0}$, where $c = a/b$ with $(a, b) = 1$, $a
\underline{r} = b \underline{r}$, and $b \underline{s} = a
\underline{s}$. Furthermore, if $c =1$ then $\alpha$ is trivial. These conditions imply that $a^\infty$ and $b^\infty$ divide $\underline{r}$ and $\underline{s}$. It follows that $c$ has the form $p_1^{e_1} p_2^{e_2} \dots p_k^{e_k}$ where $e_i \in \ZZ$ and each $p_i$ divides $\underline{r}$ and $\underline{s}$ with infinite multiplicity. The mapping $\alpha
\rightarrow c$ gives the desired isomorphism.
Suppose that $p$ is a prime such that $p^\infty$ divides $\underline{r}$ and $\underline{s}$. We note one way in which the order-preserving homeomorphism $\alpha$ corresponding to $p^{-1}$ may be identified.
Write $\underline{r} = pr_1 pr_2 pr_3 \dots, $ and $\underline{s} = s_1 p s_2 p
s_3 \dots$ and obtain the identification
$$\begin{array}{llc}
X_\mu & = & ( \dots [s_2] \times [p] \times [s_1]) \times ([p] \times [r_1]
\times [p] \times \dots)\\
& = & X_\mu^- \times X_\mu^+.
\end{array}$$
This is naturally isomorphic to
$$\begin{array}{llc}
X_\lambda & = & ( \dots [p] \times [s_1] \times [p]) \times ([r_1]
\times [p] \times [r_2] \times \dots )\\
& = & X_\lambda^1 \times X_\lambda^+
\end{array}$$
by a map $\beta = \beta^- \times \beta^+$ which respects the factors and which induces the natural semigroupoid isomorphism $R_\mu \rightarrow
R_\mu$. Let $\gamma : X_\lambda \rightarrow X_\mu$ be the right shift homeomorphism. Then $\gamma \circ \beta$ is an automorphism of $X_\mu$ and its associated constant is $p^{-1}$.
**Countable Linear Orderings**
==============================
Let $\Omega$ be a countable linear ordering and define an equivalence relation $\approx$ on $\Omega$ such that $w \approx v$ if the order intervals $[w, v]$ and $[v, w]$ are finite. Then the set $\Omega /
\approx$ of equivalence classes is linearly ordered and each equivalence class $< x>$ is itself a linearly ordered set which is isomorphic to a finite set or to one of $\ZZ_+, \ZZ_{-}$ and $\ZZ$. Let $\mu : \Omega \to \{2,3,\dots\}$. Then to each class $<x>$ we can associate a pair $p_\mu(<x>) = (\underline{r}, \underline{s})$, as in section 3, consisting of finite or generalized integers.
[**Theorem 5**]{}. *Let $\Omega,\ \Lambda$ be countable linear orderings with maps $\mu : \Omega \rightarrow \{2, 3, \dots \}, \nu : \Lambda \rightarrow
\{2, 3, \dots \}$. Then the lexicographic semigroupoids $R_\mu$ and $R_\nu$ are isomorphic if and only if there is an order preserving bijection $\pi : \Omega / \approx \ \rightarrow \ \Lambda / \approx$ such that $p_\mu (\pi (<w>)) \sim p_\mu (<w>)$ for all classes $<w>$ in $\Omega / \approx$.*
As in the proof of Theorem 2 declare a point $x = (x_w)$ in $X_\mu$ to be a type 1 point if there exists a first coordinate $w$ for which $x_w \neq 1$. Once again, as in the proof of Theorem 2, these points are identifiable in order-topological terms as those for which the closed (half) orbit $\overline{{{\cal O}}(x)}$ properly contains relatively open-closed closed orbits. If $w_2 > w_1$ then say that $w_1$ and $w_2$ are [*finitely equivalent*]{} if there exist type 1 points $x, y$ with first non-unit coordinates $x_{w_1}$ and $y_{w_2}$ respectively such that $\overline{{{\cal O}}(y)}$ has positive measure with respect to the unique invariant probability Borel measure on $\overline{{{\cal O}}(x)}$.
Note that $w_1$ and $w_2$ are finitely equivalent if and only if $w_1 \approx w_2$. In view of this it follows that if $\alpha^{(2)} :
R_\mu \rightarrow R_\nu$ is a topological isomorphism induced by the homeomorphism $\alpha$ then $\alpha$ induces a map $\pi : \Omega /
\approx \ \rightarrow \ \Lambda / \approx$. Indeed, if $x$ is a type 1 point associated with $w$ in $\Omega$ then $\alpha(x)$ is a type one point associated with $u$ in $\Lambda$ and we may define $\pi(<w>) = <u>$. Since the equivalence relation $\ \approx$ coincides with finite equivalence this is a well-defined bijection.
Fix a class $<w>, $ for some $w \in \Omega$, and define the set
$$X_\mu (<w>) = \bigcup_{x \in T(w)} \overline{{{\cal O}}(x)}$$
where the union is taken over the set $T(w)$ of type 1 points $x$ associated with the class $<w>$. Then $\alpha$ restricts to a homeomorphism
$$\beta : X_\mu (<w>) \rightarrow X_\nu (\pi (<w>)).$$
By considering the restriction of $\alpha^{(2)}$ to the set
$$R_\mu \cap
(X_\mu (<w>) \times X_\mu (<w>))$$
we shall show that the lexicographic semigroupoid $R(<w>, \mu)$ associated with the linear order $<w>$ and the function $\mu$ is isomorphic to $R (\pi (<w>),\nu)$. In view of the result in section 3 this will complete the proof.
To this end define the equivalence relation $E_\mu$ on $X_\mu (<w>)$ as the set of pairs $(x,y)$ for which $x_u = y_u$ for all $u$ in $<w>$, and similarly define $E_\nu$ on $X_\nu (\pi (<w>))$. Thus, the set of equivalence classes $X_\mu (<w>) /E_\mu$ is isomorphic to the set
$$X_{<w>} = \prod_{d \in <w>} [\mu (d)].$$
Furthermore $R_\mu$ and $E_\mu$ induce a binary relation, $R_\mu /
E_\mu$ say, on the set $X_\mu (<w>)/E_\mu. $ That is $ (x,y) \in R_\mu / E_\mu$ if and only if there exist $x^\prime, y^\prime$ in $X_\mu (<w>)$ with $(x^\prime , y^\prime) \in R_\mu, x^\prime \in x$ and $y^\prime \in y$. Also, under the natural identification above $R_\mu / E_\mu$ is the lexicographic semigroupoid $R (<w>, \mu )$ on $X_{<w>}$.
In view of these identifications it will be enough to show that the equivalence relation $E_\mu$ can be defined in an order-topological fashion and that the restriction of $\alpha$ to $X_\mu (<w>) $ induces a semigroupoid isomorphism from $R (<w>, \mu)$ to $R ( \pi (<w>) ,
\mu)$.
In section 2 we saw that, up to a constant multiplier, the set $X_\mu
(<w>) $ carries a unique $R_\mu$-invariant measure, $\lambda_\mu$ say. If $x, y \in X_\mu (<w>) $ and $\lambda_\mu (\overline{{{\cal O}}(x)}) =
\lambda_\mu (\overline{{{\cal O}}(y)})$ then there are two possibilities; either $x_u = y_u$ for all $u$ in $<w>$, or $x_u$ and $y_u$ correspond to “rational points” in the sense that the symmetric difference $\overline{{{\cal O}}(x)} \Delta \overline{{{\cal O}}(y)}$ is a singleton, namely $\{x\}$ or $\{y\}$. It follows that $E_\mu$ can be defined purely in order-topological terms and hence that $\alpha$ maps the $E_\mu$-equivalence classes to $E_\nu$-equivalence classes.
Since the given map $\alpha : X_\mu \rightarrow X_\nu$ is continuous and since the basic open-closed sets of $X_\mu$ and $X_\mu$ generate the topology of $X_\mu$ and $X_\mu$, respectively, it follows that the induced map
$$\alpha_i : X_\mu (<w>) / E_\mu \rightarrow X_\nu (\pi (<w>)) / E_\nu$$
is bicontinuous. Indeed $\alpha$ maps an $E_\mu$-saturated basic open-closed set to an $E_\nu$-saturated open-closed set and this is necessarily a finite union of $E_\nu$-saturated basic open-closed sets. Similarly, it follows that $\alpha_i^{(2)}$ induces a semigroupoid isomorphism from $R(<w>, \mu)$ to $R(\pi (<w>), \mu)$.
The arguments given above and in the previous section are also effective in the setting of infinite lexicographic products of partially ordered sets. This is illustrated in the following theorem in the case $\Omega = \IQ$ for which the proof in section 2 is applicable with little change.
For each rational $q \in \IQ$ let $\le_q$ be a connected partial ordering on the finite set $\{1,\dots,\mu(q)\}$. Then the product space $X_\mu$ carries a natural semigroupoid
$$R =
R(\{ \le_q : q \in \IQ\})$$
which is the subset of $R_\mu$ associated with the given partial orderings. That is, $(x,y) \in R$ if and only if $(x,y) \in \tilde{R}_\mu$ and $x_{w_0} \le_q y_{w_0}$ for the smallest index $w_0$ such that $x_w \ne y_w.$
[**Theorem 6**]{} *The semigroupoids $R(\{ \le_q : q \in \IQ\})$ and $R(\{ \prec_q : q \in \IQ\})$ are isomorphic if and only if there is an order bijection $\pi : \IQ \to \IQ$ such that the partial orderings $\le_q$ and $\prec_{\pi(q)}$ are isomorphic for all $q \in \IQ$.*
**Applications to Operator Algebras**
=====================================
The operator algebra $T_n$ is the subalgebra of the complex matrix algebra $ M_n$ consisting of upper triangular complex matrices and endowed with the usual operator norm. These algebras $A$ are [*triangular*]{} in the sense that $A \cap A^*$ is a maximal abelian self-adjoint subalgebra. Recently there has been considerable interest in classifying the many diverse families of triangular operator algebras arising as direct limits of these finite-dimensional algebras and their direct sums. (See [@scp-book].)
The following construction is given in [@scp-lex1].
Let $\Omega, \mu$ be as above, with $n_w = \mu (w)$. Let $F \subseteq \Omega$ be a finite subset, say $w_1 < w_2 < \dots < w_k$, and let $w_t < w < w_{t+1}$, for some $t$. Set $G = F \cup \{w\}$, $n_F = n_{w_1} n_{w_2}\dots n_{w_k}$, and $n_G =
n_\omega n_F$. Define a unital algebra injection $\phi_{F,G} : T_{n_F} \to
T_{n_G}$ as follows. View $T_{n_F}$ as the (maximal triangular) subalgebra of $M_{n_{w_1}} \otimes \dots \otimes M_{n_{w_k}}$ which is spanned by the matrix units
$$e_{{{\bf i}},{{\bf j}}} = e_{i_1, j_1} \otimes \dots \otimes e_{i_k,j_k}$$
where the multi-index ${{\bf i} } = (i_1, \dots, i_k)$ precedes ${{\bf j}} =
(j_1, \dots,
j_k)$ in the lexicographic ordering. Thus either ${{\bf i}} = {{\bf j}}$ or the first $i_p$ differing from $j_p$ is strictly less than $j_p$. Similarly identify $T_{n_G}$ for the ordered subset $G$ and set $\phi_{F,G}$ to be the linear extension of the correspondence
$$e_{{{\bf i}},{{\bf j}}} \to \sum^{n_\omega}_{s=1} e_{i_1, j_1} \otimes \dots
\otimes
e_{i_t, j_t} \otimes e_{s,s} \otimes e_{i_{t+1},j_{t+1}} \otimes \dots \otimes
e_{i_{k},j_{k}}$$
In a similar way, or by composing maps of the above type, define $\phi_{F,G}$ for $F \subseteq G$, general finite subsets. These maps are isometric and so determine the Banach algebra
$$A(\Omega, \nu) = {\lim_{\to}}_{F \in {\cal F}} T_{n_F}$$
where the direct limit is taken over the directed set ${\cal F}$ of finite subsets of $\Omega$. Each $\phi_{F,G}$ has an extension to a C\*-algebra injection from $M_{n_F}$ to $M_{n_G}$ and so it follows that we may view $A(\Omega, \nu)$ as a closed unital subalgebra of the UHF C\*-algebra $B(\Omega, \nu) = {\displaystyle \lim_\to M_{n_F}}$.
[**Theorem 7**]{} *The following statements are equivalent.*
\(i) $A(\Omega,\mu)$ and $A(\Lambda,\nu)$ are isometrically isomorphic Banach algebras.
\(ii) $R_\mu$ and $R_\nu$ are isomorphic lexicographic semigroupoids.
\(iii) There is an order preserving bijection $\pi : \Omega / \approx \ \rightarrow \ \Lambda / \approx$ such that $$p_\mu (\pi (<w>)) \sim p_\mu (<w>)$$ for all classes $<w>$ in $\Omega / \approx$.
The semigroupoids are readily identifiable with the topological fundamental relations of the operator algebras. (See, for example, the discussions of [@hop-scp] and [@scp-book].) The equivalence of (i) and (ii) is now immediate from Theorem 7.5 of [@scp-book] and so Theorem 5 completes the equivalences.
An immediate corollary of the last theorem is that there are uncountably many triangular algebras $A$ of the form $A(\Omega,\mu)$ with $C^*(A)$ equal to the $2^\infty$ UHF C\*-algebra. On the other hand there are only three such (infinite-dimensional) algebras of the form $A(\ZZ,\mu)$, namely the pure refinement algebra, the standard limit algebra and the alternation algebra with invariant $(2^\infty,2^\infty)$.
We now explain how the algebras above can also be interpreted in term of a lexicographic product operation at the algebraic level, as described in [@scp-lex1].
Let $A$ be an operator algebra admitting a subdiagonal decomposition in the sense that
$$A \ \ = \ \ A \cap A^* \ +\ A^0$$
where $A \cap A^*$ is a maximal abelian subalgebra of $A$ , and $A^0$ is the kernel of a contractive homomorphism $A \to A \cap A^*$. In particular, this holds if $A$ is a regular triangular subalgebra of an AF C\*-algebra ([@scp-book] or, more generally, if $A$ is a subdiagonal algebra in the sense of Arveson [@arv]. If $A$ and $B$ are triangular operator algebras admitting such decompositions then define their [*lexicographic product*]{} $A \star B$ to be the closed subalgebra of the injective tensor product $C^*(A) {\otimes}C^*(B)$ given by
$$A \star B = (A \cap A^*) {\otimes}B \ +\ A^0 {\otimes}C^*(B).$$
One can verify that the inclusions $\phi_{F,G}$ defined above coincide with the natural inclusions
$$T_{n_{w_1}}\star \dots \star T_{n_{w_k}}
\ \ \to \ \ T_{n_{w_1}} \star \dots
T_{n_{w_t}}\star T_{n_{w}}\star T_{n_{w_{t+1}}}\star \dots \star T_{n_{w_k}}.$$
In fact the lexicographic product is an associative operation and the algebras $A(\Omega,\mu)$ can be viewed (unambiguously) as infinite lexicographic products of upper triangular matrix algebras over the ordering $\Omega$. The following theorem is a corollary of Theorem 6.
[**Theorem 8**]{} *Let $G_q$ and $H_q$ for $q$ in $\IQ$ be connected transitive antisymmetric digraphs with triangular digraph algebras $A(G_q)$ and $A(H_q)$. Then the lexicographic products*
$$\prod_{q \in \ \IQ} \star A(G_q) \ \ \ \ \mbox{ and }\ \ \ \
\prod_{q \in \ \IQ} \star A(H_q)$$
are isometrically isomorphic triangular operator algebras if and only if there is an order bijection $\pi$ such that the digraphs $G_q$ and $H_{\pi(q)}$ are isomorphic for all rationals $q$.
[8]{}
W. B. Arveson, Analyticity in operator algebras, Amer. J. Math., 89 (1967), 578–642.
A.P. Donsig, Semisimple triangular AF algebras, J. Functional Anal., 111 (1993), 323-349.
A.P. Donsig and A. Hopenwasser, Order preservation in limit algebras, preprint, 1993.
A. Hopenwasser and S.C. Power, Classification of limits of triangular matrix algebras, Proc. Edinburgh Math. Soc., 36 (1992), 107-121.
P.S. Muhly and B. Solel, Subalgebras of groupoid C\*-algebras, J. für die Reine und Ange. Math. 402 (1989), 41–75.
Y.T. Poon, A complete isomorphism invariant for a class of triangular UHF algebras, preprint 1990, to appear in J. Operator Th..
J. Renault, A groupoid approach to C\*-algebras, Lecture Notes in Math. No. 793, Springer Verlag, Berlin-Heidelberg-New York 1980.
S.C. Power, Limit algebras: an introduction to subalgebras of C\*-algebras, Pitman Research Notes in Mathematics vol 278, Longman Scientific and Technical, England, New York, 1992.
S.C. Power, On the outer automorphism groups of triangular alternation limit algebras, preprint 1990, J. Functional Anal., 113 (1993), 462-471.
S.C. Power, Infinite lexicographic products of triangular algebras, [*Bull. London Math. Soc.*]{} to appear.
J. G. Rosenstein, Linear Orderings, Academic Press, London, New York, 1982.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this paper, we study the Fu-Yau equation on compact Hermitian manifolds and prove the existence of solutions of equation on astheno-Kähler manifolds. We also prove the uniqueness of solutions of Fu-Yau equation when the slope parameter $\alpha$ is negative.'
address:
- 'Institute of Mathematics, Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing 100190, P. R. China'
- 'School of Mathematical Sciences, University of Science and Technology of China, Hefei 230026, P. R. China'
- 'School of Mathematical Sciences, Peking University, Yiheyuan Road 5, Beijing 100871, P. R. China '
author:
- Jianchun Chu
- Liding Huang
- Xiaohua Zhu
title: 'The Fu-Yau equation on compact astheno-Kähler manifolds'
---
\[section\] \[theorem\][Lemma]{} \[theorem\][Corollary]{} \[theorem\][Proposition]{} \[section\] \[theor\] \[theorem\][Remark]{}
Introduction
============
Let $(M,\omega)$ be an $n$-dimensional compact Kähler manifold. As a reduced generalized Strominger system in higher dimensions, Fu and Yau introduced the following fully nonlinear equation for $\varphi$ [@FuY08], $$\label{Fu-Yau equation}
\begin{split}
{\sqrt{-1} \partial \overline{\partial}}(e^{{\varphi}}\omega & -\alpha e^{-{\varphi}}\rho)\wedge\omega^{n-2} \\
& +n\alpha{\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge{\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge\omega^{n-2}+\mu\frac{\omega^{n}}{n!}=0,
\end{split}$$ where $\alpha$ is a non-zero constant called the slope parameter, $\rho$ is a real smooth $(1,1)$ form, $\mu$ is a smooth function. For ${\varphi}$, we impose the elliptic condition $$\label{Elliptic condition}
\tilde{\omega}=e^{{\varphi}}\omega+\alpha e^{-{\varphi}}\rho+2n\alpha{\sqrt{-1} \partial \overline{\partial}}{\varphi}\in \Gamma_{2}(M)$$ and the normalization condition $$\label{Normalization condition}
\|e^{-{\varphi}}\|_{L^{1}}=A,$$ where $$\Gamma_{2}(M)=\{ \alpha\in A^{1,1}(M)~|~ \frac{\alpha^{1}\wedge\omega^{n-1}}{\omega^{n}}>0, \frac{\alpha^{2}\wedge\omega^{n-2}}{\omega^{n}}>0 \}$$ and $A^{1,1}(M)$ is the space of smooth real (1,1) forms on $M$.
When $n=2$, (\[Fu-Yau equation\]) is equivalent to the Strominger system on a toric fibration over a $K3$ surface constructed by Goldstein and Prokushki [@GP04], which was solved by Fu and Yau for $\alpha>0$ and $\alpha<0$ in [@FuY08] and [@FuY07], respectively. (\[Fu-Yau equation\]) is usually called Fu-Yau equation (cf. [@PPZ16b; @GM16]).
In case of $\alpha<0$, Phong, Picard and Zhang [@PPZ16b] recently proved the existence of solutions of (\[Fu-Yau equation\]) with the condition (\[Normalization condition\]) is replaced by $$\|e^{{\varphi}}\|_{L^{1}} = \frac{1}{A} \gg 1.$$ In [@CHZ18], we prove that there exists constant $A_{0}$ depending only on $\alpha$, $\rho$, $\mu$ and $(M,\omega)$ such that for any $A{\leqslant}A_{0}$ and any $\alpha\neq0$, (\[Fu-Yau equation\]) has a solution satisfying (\[Elliptic condition\]) and (\[Normalization condition\]). [^1] Our result is new and different from that of [@PPZ16b] which deals with only the case that $\alpha<0$.
Since the Strominger system comes from non-Kähler geometry [@Str86], it is natural to consider (\[Fu-Yau equation\]) on Hermitian manifolds. In the Kähler case, all the proofs in [@FuY08; @FuY07; @PPZ16b; @CHZ18] relied heavily on the Kähler condition $d\omega=0$. It seems to be very difficult to solve (\[Fu-Yau equation\]) on general Hermitian manifolds. In this paper, we focus on a class of Hermitian manifolds which satisfies the astheno-Kähler condition $$\label{Astheno-Kahler condition}
{\sqrt{-1} \partial \overline{\partial}}\omega^{n-2}=0.$$ Astheno-Kähler manifold was first introduced in the paper of Jost-Yau [@JY93], where they extended Siu’s rigidity results in Kähler manifolds to astheno-Kähler manifolds [@SY80]. Such manifolds have many naturally properties as Kähler manifolds. For example, every holomorphic $1$-form on a compact astheno-Kähler manifold is closed [@JY93 Lemma 6].
There are many examples of astheno-Kähler manifolds, see [@LY87; @LYZ94; @MT01; @Ma09; @FT11; @FGV16; @LU17]. For example, the product of a complex curve with a Kähler metric and a complex surface with a non-Kähler Gauduchon metric satisfies (\[Astheno-Kahler condition\]).
The purpose of this paper is to generalize the main result in [@CHZ18] to astheno-Kähler manifolds. Namely, we prove
\[Existence and Uniqueness Theorem\] Let $(M,\omega)$ be an $n$-dimensional compact astheno-Kähler manifold. Then there exist constants $A_{0}$, $C_{0}$, $\delta_{0}$, $M_{0}$ and $D_{0}$ depending only on $\alpha$, $\rho$, $\mu$ and $(M,\omega)$ such that for any $A{\leqslant}A_{0}$, there exists a unique solution ${\varphi}$ of (\[Fu-Yau equation\]) satisfying (\[Elliptic condition\]), (\[Normalization condition\]) and $$\label{Restrictions}
e^{-{\varphi}} {\leqslant}\delta_{0},~~|{\partial}{\overline{\partial}}{\varphi}|_{g} {\leqslant}D,~~D_{0}{\leqslant}D \text{~and~} A {\leqslant}\frac{1}{C_{0}M_{0}D}.$$
Since in the non-Kähler case the constant is not a trivial solution in the continuity method when $t=0$ in [@CHZ18], we introduce a new continuous path to solve (\[Fu-Yau equation\]) as follows ($t\in [0,1]$), $$\label{Fu-Yau equation t}
\begin{split}
{\sqrt{-1} \partial \overline{\partial}}(e^{{\varphi}}\omega & -t\alpha e^{-{\varphi}}\rho)\wedge\omega^{n-2}+n\alpha{\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge{\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge\omega^{n-2} \\
& +n\alpha(t-1){\sqrt{-1} \partial \overline{\partial}}h\wedge{\sqrt{-1} \partial \overline{\partial}}h\wedge\omega^{n-2}+t\mu\frac{\omega^{n}}{n!} = 0,
\end{split}$$ where $h$ is a smooth function. Clearly, (\[Fu-Yau equation t\]) is equivalent to (\[Fu-Yau equation\]) when $t=1.$ We will show that there is $h$ such that (\[Fu-Yau equation t\]) can be solved when $t=0$ (cf. Lemma \[Existence lemma\]).
In the proof of openness for the solvable set of $t$, the astheno-Kähler condition (\[Astheno-Kahler condition\]) will play an important role. (\[Astheno-Kahler condition\]) guarantees that the adjoint of linearized operator $L$ has no zero order terms (cf. (\[Definition of L\^\*\])), then the strong maximum principle can be applied.
(\[Astheno-Kahler condition\]) will also be used for the $C^{0}$-estimate (cf. Lemma \[Zero order estimate lemma\]). In fact, instead of $L^{1}$-integral of $\varphi$ in [@CHZ18], we first estimate a $L^{k_{0}}$-integral for some $k_{0}\ll1$, then apply the Moser iteration to derive the $C^0$-estimate. The $C^1, C^2$-estimates for solutions of (\[Fu-Yau equation t\]) can be obtained by the argument in [@CHZ18]. For the reader’s convenience, we give a sketch of the proofs in Section 3. Actually, the argument there are valid for solutions of (\[Fu-Yau equation\]) on any Hermitian manifolds $(M,\omega)$.
In the next part of this paper, we improve Theorem \[Existence and Uniqueness Theorem\] without the restriction condition (\[Restrictions\]) in case of $\alpha<0$. In fact, we prove the following uniqueness of Fu-Yau equation.
\[Uniqueness theorem alpha negative\] Let $\alpha<0$ and $(M,\omega)$ be an $n$-dimensional compact astheno-Kähler manifold. There exists a constant $A_{0}$ depending only on $\alpha$, $\rho$, $\mu$ and $(M,\omega)$ such that for any $A{\leqslant}A_{0}$, (\[Fu-Yau equation\]) has a unique smooth solution satisfying (\[Elliptic condition\]) and the $L^{n}$-normalization condition $$\label{L^n normalization condition}
\|e^{-{\varphi}}\|_{L^{n}}=A.$$
Furthermore, investigating the structure of the Fu-Yau equation, we obtain the monotonicity property of solutions.
\[Monotonicity theorem\] Let $\alpha<0$ and $(M,\omega)$ be an $n$-dimensional compact Kähler manifold. Suppose that ${\varphi}$ and ${\tilde{{\varphi}}}$ are solutions of (\[Fu-Yau equation\]) satisfying (\[Elliptic condition\]). If $\|e^{-{\varphi}}\|_{L^{n}}=A$, $\|e^{-{\tilde{{\varphi}}}}\|_{L^{n}}={\tilde{A}}$ and $A<{\tilde{A}}{\leqslant}A_{0}$, then we have ${\varphi}>{\tilde{{\varphi}}}$ on $M$, where $A_{0}$ is a constant depending only on $\alpha$, $\rho$, $\mu$ and $(M,\omega)$.
In addition, if $\textrm{tr}_{\omega}\rho{\geqslant}0$, both Theorem \[Uniqueness theorem alpha negative\] and \[Monotonicity theorem\] are still true when $L^{n}$ normalization condition (\[L\^n normalization condition\]) is replaced by a weaker condition $\|e^{-{\varphi}}\|_{L^{1}}=A$ (see Remark \[Special case remark\]). In particular, Theorem \[Uniqueness theorem alpha negative\] is an improvement of main results in [@FuY07; @PPZ16b] in Kähler case.
Roughly speaking, Theorem \[Uniqueness theorem alpha negative\] and \[Monotonicity theorem\] are consequences of a priori estimates for $\varphi$. Compared to the proof of Theorem \[Existence and Uniqueness Theorem\], we need to derive a strong $C^0, C^1, C^2$ estimates without (\[Restrictions\]). In order to use the blow-up argument for $C^1, C^2$ estimates, we establish an estimate $$\label{c1-c2-estimate}
\sup_{M}|{\partial}{\overline{\partial}}{\varphi}|_{g} {\leqslant}C_A(1+\sup_{M}|{\partial}{\varphi}|_{g}^{2}),$$ where $C_{A}$ is a constant depending only on $A$, $\alpha$, $\rho$, $\mu$ and $(M,\omega)$. Such a kind of estimate (\[c1-c2-estimate\]) was widely studied in Monge-Ampère equations and $\sigma_k$ Hessian equations (cf. [@WY09; @HMW10; @DK12; @TW17; @PPZ16b; @Sze15; @STW17]). In our case, we adopt an auxiliary function involving the largest eigenvalue $\lambda_{1}$ of ${\tilde{\omega}}$ with respect to $\omega$. This advantage gives us enough good third order terms to deal with the bad terms when we use the maximal principle as in [@CHZ17]. Also the sign of $\alpha$ plays a crucial role. We note that (\[Fu-Yau equation\]) is not degenerate when $\alpha<0$ (cf. (\[Fu-Yau equation 2-nd Hessian type\])).
As we know, (\[Fu-Yau equation\]) can be rewritten as a $\sigma_{2}$-type equation on a Hermitian manifold with function $F$ at the right hand including the gradient term of solution (cf. (\[Hessian type of Fu-Yau equation\]), (\[Fu-Yau equation 2-nd Hessian type\])). In [@CHZ17], we generalized $\sigma_{2}$-equation to an almost Hermitian manifold and obtained a $C^2$-estimate for the solutions, which depends only on the gradient of solutions and background data. It is interesting to studying the $C^2$-estimate for solutions of $\sigma_{k}$-type equation in space of $\Gamma_k$ $(k{\geqslant}2)$ of $k$-convex functions (cf. [@GRW15; @Sze15; @PPZ15; @PPZ16a], etc.). But it seems nontrivial to generalize the method for $\sigma_{2}$-equation to $\sigma_{k}$-equation even on Kähler manifolds if $F$ involves the gradient term of solution.
The organization of paper is as follows. In Section 2 and Section 3, we give the $C^0$-estimate, and $C^1, C^2$-estimates for solutions of (\[Fu-Yau equation\]) under the condition in Theorem \[Existence and Uniqueness Theorem\], respectively. Theorem \[Existence and Uniqueness Theorem\] is proved in Section 4. In Section 5, we improve the $C^0$-estimate in Section 2 in case of $\alpha <0$. In Section 6, we give another method to get strong $C^1, C^2$-estimates in case of $\alpha <0$. Theorem \[Uniqueness theorem alpha negative\] and Theorem \[Monotonicity theorem\] will be proved in Section 7, 8, respectively.
Zero order estimate (I)
=======================
In this section, we use the Moser iteration to do $C^{0}$-estimate for solutions ${\varphi}$ of (\[Fu-Yau equation\]). First, we prove a lemma for $L^{2}$-estimate of gradient ${\partial}{\varphi}$.
\[Zero order estimate lemma\] Let ${\varphi}$ be a smooth solution of (\[Fu-Yau equation\]) satisfying (\[Elliptic condition\]). Let $f(t)$ be a smooth function in $\mathbb R^1$ such that $f'{\geqslant}0$. Then we have $$\begin{aligned}
\label{l2-gradient}
& \int_{M}f'({\varphi})\sqrt{-1}{\partial}{\varphi}\wedge{\overline{\partial}}{\varphi}\wedge(e^{{\varphi}}\omega+\alpha e^{-{\varphi}}\rho)\wedge\omega^{n-2} \notag\\
&{\leqslant}-2\int_{M}f'({\varphi})\sqrt{-1}{\partial}{\varphi}\wedge(e^{{\varphi}}{\overline{\partial}}\omega-\alpha e^{-{\varphi}}{\overline{\partial}}\rho)\wedge\omega^{n-2}\notag \\
& -2\int_{M}f'({\varphi}){\overline{\partial}}{\varphi}\wedge(e^{{\varphi}}\omega-\alpha e^{-{\varphi}}\rho)\wedge\sqrt{-1}{\partial}\omega^{n-2}
+2\int_{M}f({\varphi})\mu\frac{\omega^{n}}{n!}.\end{aligned}$$
Since ${\tilde{\omega}}\in\Gamma_{2}(M)$ and $f'({\varphi}){\geqslant}0$, it is clear that $$\int_{M}f'({\varphi})\sqrt{-1}{\partial}{\varphi}\wedge{\overline{\partial}}{\varphi}\wedge{\tilde{\omega}}\wedge\omega^{n-2} {\geqslant}0.$$ By the Stokes’ formula , it follows $$\begin{split}
& \int_{M}f'({\varphi})\sqrt{-1}{\partial}{\varphi}\wedge{\overline{\partial}}{\varphi}\wedge(e^{{\varphi}}\omega+\alpha e^{-{\varphi}}\rho)\wedge\omega^{n-2} \\
& {\geqslant}-2n\alpha\int_{M}\sqrt{-1}{\partial}f({\varphi})\wedge{\overline{\partial}}{\varphi}\wedge{\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge\omega^{n-2} \\
&= -2n\alpha\int_{M}f({\varphi}){\overline{\partial}}{\varphi}\wedge{\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge\sqrt{-1}{\partial}\omega^{n-2} \\
&\ {} \ +2n\alpha\int_{M}f({\varphi}){\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge{\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge\omega^{n-2}.
\end{split}$$ On the other hand, by (\[Fu-Yau equation\]), we have $$\begin{split}
& 2n\alpha\int_{M}f({\varphi}){\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge{\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge\omega^{n-2} \\
= {} & -2\int_{M}f({\varphi}){\sqrt{-1} \partial \overline{\partial}}(e^{{\varphi}}\omega-\alpha e^{-{\varphi}}\rho)\wedge\omega^{n-2}-2\int_{M}f({\varphi})\mu\frac{\omega^{n}}{n!}.
\end{split}$$ Thus $$\label{Zero order estimate equ 1}
\begin{split}
& \int_{M}f'({\varphi})\sqrt{-1}{\partial}{\varphi}\wedge{\overline{\partial}}{\varphi}\wedge(e^{{\varphi}}\omega+\alpha e^{-{\varphi}}\rho)\wedge\omega^{n-2} \\
{\geqslant}{} & -2n\alpha\int_{M}f({\varphi}){\overline{\partial}}{\varphi}\wedge{\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge\sqrt{-1}{\partial}\omega^{n-2} \\
& -2\int_{M}f({\varphi}){\sqrt{-1} \partial \overline{\partial}}(e^{{\varphi}}\omega-\alpha e^{-{\varphi}}\rho)\wedge\omega^{n-2}-2\int_{M}f({\varphi})\mu\frac{\omega^{n}}{n!}.
\end{split}$$
For the first term on the right hand of the inequality (\[Zero order estimate equ 1\]), we have $$\label{Zero order estimate equ 2}
\begin{split}
& -2n\alpha\int_{M}f({\varphi}){\overline{\partial}}{\varphi}\wedge{\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge\sqrt{-1}{\partial}\omega^{n-2} \\
= {} & 2n\alpha\int_{M}{\overline{\partial}}(f({\varphi}){\overline{\partial}}{\varphi})\wedge\sqrt{-1}{\partial}{\varphi}\wedge\sqrt{-1}{\partial}\omega^{n-2} \\
& +2n\alpha\int_{M}f({\varphi}){\overline{\partial}}{\varphi}\wedge\sqrt{-1}{\partial}{\varphi}\wedge\sqrt{-1}~{\overline{\partial}}{\partial}\omega^{n-2} \\
= {} & 2n\alpha\int_{M}f({\varphi})\sqrt{-1}{\partial}{\varphi}\wedge{\overline{\partial}}{\varphi}\wedge{\sqrt{-1} \partial \overline{\partial}}\omega^{n-2}.
\end{split}$$ For the second term of (\[Zero order estimate equ 1\]), we compute $$\begin{split}
& -2\int_{M}f({\varphi}){\sqrt{-1} \partial \overline{\partial}}(e^{{\varphi}}\omega-\alpha e^{-{\varphi}}\rho)\wedge\omega^{n-2} \\
= {} & 2\int_{M}\sqrt{-1}{\partial}f({\varphi})\wedge{\overline{\partial}}(e^{{\varphi}}\omega-\alpha e^{-{\varphi}}\rho)\wedge\omega^{n-2} \\
& -2\int_{M}f({\varphi}){\overline{\partial}}(e^{{\varphi}}\omega-\alpha e^{-{\varphi}}\rho)\wedge\sqrt{-1}{\partial}\omega^{n-2}.
\end{split}$$ and then, $$\label{Zero order estimate equ 3}
\begin{split}
& -2\int_{M}f({\varphi}){\sqrt{-1} \partial \overline{\partial}}(e^{{\varphi}}\omega-\alpha e^{-{\varphi}}\rho)\wedge\omega^{n-2} \\
= {} & 2\int_{M}f'({\varphi})\sqrt{-1}{\partial}{\varphi}\wedge{\overline{\partial}}{\varphi}\wedge(e^{{\varphi}}\omega+\alpha e^{-{\varphi}}\rho)\wedge\omega^{n-2} \\
& +2\int_{M}f'({\varphi})\sqrt{-1}{\partial}{\varphi}\wedge(e^{{\varphi}}{\overline{\partial}}\omega-\alpha e^{-{\varphi}}{\overline{\partial}}\rho)\wedge\omega^{n-2} \\
& +2\int_{M}{\overline{\partial}}f({\varphi})\wedge(e^{{\varphi}}\omega-\alpha e^{-{\varphi}}\rho)\wedge\sqrt{-1}{\partial}\omega^{n-2} \\
& -2\int_{M}f({\varphi})(e^{{\varphi}}\omega-\alpha e^{-{\varphi}}\rho)\wedge{\sqrt{-1} \partial \overline{\partial}}\omega^{n-2}.
\end{split}$$ Thus substituting (\[Zero order estimate equ 2\]) and (\[Zero order estimate equ 3\]) into (\[Zero order estimate equ 1\]), we see that $$\begin{split}
& \int_{M}f'({\varphi})\sqrt{-1}{\partial}{\varphi}\wedge{\overline{\partial}}{\varphi}\wedge(e^{{\varphi}}\omega+\alpha e^{-{\varphi}}\rho)\wedge\omega^{n-2} \\
{\leqslant}{} & -2\int_{M}f'({\varphi})\sqrt{-1}{\partial}{\varphi}\wedge(e^{{\varphi}}{\overline{\partial}}\omega-\alpha e^{-{\varphi}}{\overline{\partial}}\rho)\wedge\omega^{n-2} \\
& -2\int_{M}f'({\varphi}){\overline{\partial}}{\varphi}\wedge(e^{{\varphi}}\omega-\alpha e^{-{\varphi}}\rho)\wedge\sqrt{-1}{\partial}\omega^{n-2} \\
& +2\int_{M}f({\varphi})(e^{{\varphi}}\omega-\alpha e^{-{\varphi}}\rho)\wedge{\sqrt{-1} \partial \overline{\partial}}\omega^{n-2} \\
& -2n\alpha\int_{M}f({\varphi})\sqrt{-1}{\partial}{\varphi}\wedge{\overline{\partial}}{\varphi}\wedge{\sqrt{-1} \partial \overline{\partial}}\omega^{n-2}+2\int_{M}f({\varphi})\mu\frac{\omega^{n}}{n!}.
\end{split}$$ Note that ${\sqrt{-1} \partial \overline{\partial}}\omega^{n-2}=0$. Hence, we get (\[l2-gradient\]).
By Lemma \[Zero order estimate lemma\], we prove the following $C^0$-estimate.
\[Zero order estimate\] Let ${\varphi}$ be a smooth solution of (\[Fu-Yau equation\]) satisfying (\[Elliptic condition\]) and (\[Normalization condition\]). There exist constants $\delta_{0}$, $A_{0}$ and $M_{0}$ depending only on $\alpha$, $\rho$, $\mu$ and $(M,\omega)$ such that if $$\label{two-condition}
e^{-{\varphi}}{\leqslant}\delta_{0} ~\text{~and~}~ \|e^{-{\varphi}}\|_{L^{1}} = A {\leqslant}A_{0},$$ then $$\frac{A}{M_{0}} {\leqslant}e^{-{\varphi}} {\leqslant}M_{0}A.$$
First, we estimate the positive infimum of $e^{{\varphi}}$. At the expense of decreasing $\delta_{0}$, we assume that $$\label{Infimum estimate equ 1}
e^{{\varphi}}\omega+\alpha e^{-{\varphi}}\rho {\geqslant}\frac{1}{2}e^{{\varphi}}\omega.$$ Then by taking $f({\varphi})=-e^{-(k+1){\varphi}}$ ($k{\geqslant}1$) in Lemma \[Zero order estimate lemma\], we have $$\begin{split}
\int_{M}e^{-k{\varphi}}|{\partial}{\varphi}|_{g}^{2}\omega^{n}
{\leqslant}C\int_{M}e^{-k{\varphi}}|{\partial}{\varphi}|_{g}\omega^{n}+C\int_{M}e^{-(k+1){\varphi}}\omega^{n}.
\end{split}$$ By the Cauchy-Schwarz inequality, it follows that $$\label{e-phi}
\int_{M}e^{-k{\varphi}}|{\partial}{\varphi}|_{g}^{2}\omega^{n} {\leqslant}C\int_{M}e^{-k{\varphi}}\omega^{n}.$$ Hence, by the above relation together with the Sobolev inequality, one can use the Moser iteration to derive $$\label{Infimum estimate equ 2}
\|e^{-{\varphi}}\|_{L^{\infty}} {\leqslant}C\|e^{-{\varphi}}\|_{L^{1}} = CA.$$
Next we estimate the supremum of $e^{{\varphi}}$. As in the proof of (\[e-phi\]), by taking $$f({\varphi}) = \frac{1}{k-1}e^{(k-1){\varphi}}$$ in Lemma \[Zero order estimate lemma\], we can also get $$\label{Supremum estimate equ 1}
\begin{split}
\int_{M}e^{k{\varphi}}|{\partial}{\varphi}|_{g}^{2}\omega^{n} {\leqslant}C\left(1+\frac{1}{|k-1|}\right)\int_{M}e^{k{\varphi}}\omega^{n}.
\end{split}$$
\[Claim\] There exists a positive constant $k_{0}\ll1$ depending only on $\alpha$, $\rho$, $\mu$ and $(M,\omega)$ such that $$\label{L^k0 estimate}
\|e^{{\varphi}}\|_{L^{k_{0}}} {\leqslant}\frac{C}{A}.$$
By (\[Supremum estimate equ 1\]), we use the Moser iteration to obtain $$\|e^{{\varphi}}\|_{L^{\infty}} {\leqslant}C_{0} \|e^{{\varphi}}\|_{L^2}.$$ By (\[L\^k0 estimate\]), it follows $$\|e^{{\varphi}}\|_{L^{\infty}} {\leqslant}C_{0}^{\frac{2}{k_{0}}} \|e^{{\varphi}}\|_{L^{k_{0}}} {\leqslant}\frac{C}{A}.$$ Thus, the proof of Proposition \[Zero order estimate\] is complete.
It remains to prove Claim \[Claim\]. Without loss of generality, we assume that ${\mathrm{Vol}}(M,\omega)=1$. We define $$U = \{x\in M~|~e^{-{\varphi}(x)}{\geqslant}\frac{A}{2} \}.$$ Then by (\[Infimum estimate equ 2\]), we have $$\begin{split}
A = {} & \int_{M}e^{-{\varphi}}\omega^{n} \\
= {} & \int_{U}e^{-{\varphi}}\omega^{n}+\int_{M\setminus U}e^{-{\varphi}}\omega^{n} \\
{\leqslant}{} & e^{-\inf_{M}{\varphi}}{\mathrm{Vol}}(U)+\frac{A}{2}(1-{\mathrm{Vol}}(U)) \\
{\leqslant}{} & \left(C-\frac{1}{2}\right)A{\mathrm{Vol}}(U)+\frac{A}{2},
\end{split}$$ which implies $$\label{Supremum estimate equ 2}
{\mathrm{Vol}}(U) {\geqslant}\frac{1}{C_{0}}.$$ On the other hand, by the Poincaré inequality, we have $$\int_{M}e^{k_{0}{\varphi}}\omega^{n}-\left(\int_{M}e^{\frac{k_{0}{\varphi}}{2}}\omega^{n}\right)^{2}
{\leqslant}C\int_{M}|{\partial}e^{\frac{k_{0}{\varphi}}{2}}|_{g}^{2}\omega^{n}
{\leqslant}Ck_{0}^{2}\int_{M}e^{k_{0}{\varphi}}\omega^{n}.$$ It then follows that $$\label{Supremum estimate equ 3}
\int_{M}e^{k_{0}{\varphi}}\omega^{n}
{\leqslant}\frac{1}{1-C_{0}k_{0}^{2}}\left(\int_{M}e^{\frac{k_{0}{\varphi}}{2}}\omega^{n}\right)^{2}.$$ Combining this with the Cauchy-Schwarz inequality, we obtain $$\begin{split}
& \left(\int_{M}e^{\frac{k_{0}{\varphi}}{2}}\omega^{n}\right)^{2} \\
{\leqslant}{} & (1+C_{0})\left(\int_{U}e^{\frac{k_{0}{\varphi}}{2}}\omega^{n}\right)^{2}
+\left(1+\frac{1}{C_{0}}\right)\left(\int_{M\setminus U}e^{\frac{k_{0}{\varphi}}{2}}\omega^{n}\right)^{2}\\
{\leqslant}{} & \frac{(1+C_{0})2^{k_{0}}}{A^{k_{0}}}({\mathrm{Vol}}(U))^{2}+\left(1+\frac{1}{C_{0}}\right)(1-{\mathrm{Vol}}(U))^{2}\int_{M}e^{k_{0}{\varphi}}\omega^{n}\\
{\leqslant}{} & \frac{(1+C_{0})2^{k_{0}}}{A^{k_{0}}}
+\left(1-\frac{1}{C_{0}^{2}}\right)\frac{1}{1-C_{0}k_{0}^{2}}\left(\int_{M}e^{\frac{k_{0}{\varphi}}{2}}\omega^{n}\right)^{2}.
\end{split}$$ By choosing $k_{0}\ll 1$, we see that $$\left(\int_{M}e^{\frac{k_{0}{\varphi}}{2}}\omega^{n}\right)^{2} {\leqslant}\frac{C}{A^{k_{0}}}.$$ Thus, we get from (\[Supremum estimate equ 3\]), $$\int_{M}e^{k_{0}{\varphi}}\omega^{n} {\leqslant}\frac{C}{A^{k_{0}}}.$$ Claim \[Claim\] is proved.
First and second order estimates (I)
====================================
In this section, we give a sketch of proofs of $C^1, C^2$ estimates of $\varphi$. As in [@CHZ18], the basic idea is to rewrite (\[Fu-Yau equation\]) as a $\sigma_2$-type equation, $$\label{Hessian type of Fu-Yau equation}
\sigma_{2}({\tilde{\omega}}) = \frac{n(n-1)}{2}\left(e^{2{\varphi}}-4\alpha e^\varphi|{\partial}{\varphi}|_{g}^{2}\right)+\frac{n(n-1)}{2}f,$$ where $$\label{Definition of f}
\begin{split}
& \ f\omega^{n} \\
= & \ 2\alpha\rho\wedge\omega^{n-1}+\alpha^{2}e^{-2{\varphi}}\rho^{2}\wedge\omega^{n-2}-4n\alpha\mu\frac{\omega^{n}}{n!} \\
+ & \ 4n\alpha^{2}e^{-{\varphi}}\sqrt{-1}\left({\partial}{\varphi}\wedge{\overline{\partial}}{\varphi}\wedge\rho-{\partial}{\varphi}\wedge{\overline{\partial}}\rho
-{\partial}\rho\wedge{\overline{\partial}}{\varphi}+{\partial}{\overline{\partial}}\rho\right)\wedge\omega^{n-2} \\[1mm]
+ & \ 4n\alpha e^{{\varphi}}\sqrt{-1}\left({\partial}\omega\wedge{\overline{\partial}}{\varphi}+{\partial}{\varphi}\wedge{\overline{\partial}}\omega+{\partial}{\overline{\partial}}\omega\right)\wedge\omega^{n-2}.
\end{split}$$ As in [@CHZ18], we define $\hat{\omega}=e^{-{\varphi}}{\tilde{\omega}}$. Then (\[Hessian type of Fu-Yau equation\]) becomes $$\label{New Hessian type of Fu-Yau equation}
\sigma_{2}(\hat{\omega}) = \frac{n(n-1)}{2}\left(1-4\alpha e^{-{\varphi}}|{\partial}{\varphi}|_{g}^{2}\right)+\frac{n(n-1)}{2}e^{-2{\varphi}}f.$$ Since $\omega$ is not Kähler, the function $f$ is more complicated than one in [@CHZ18]. Precisely, more terms involving $e^{{\varphi}}$ and ${\partial}{\varphi}$ appears. However, for the right hand side of (\[New Hessian type of Fu-Yau equation\]), the leading term is still $-2n(n-1)\alpha e^{-{\varphi}}|{\partial}{\varphi}|_{g}^{2}$. Thus we will obtain a similar inequality as in Kähler case when we differentiate (\[New Hessian type of Fu-Yau equation\]). This is why we can prove an analogy of [@CHZ18 Propsition 3.1, 4.1] as follows.
\[First and second order estimate\] Let ${\varphi}$ be a smooth solution of (\[Fu-Yau equation\]) satisfying (\[Elliptic condition\]) and $\frac{1}{M_{0}A}{\leqslant}e^{-{\varphi}}{\leqslant}M_{0}A$ for some uniform constant $M_0$. There exist uniform constants $D_{0}$ and $C_{0}$ such that if $$\label{condition of first and second order estimate}
|{\partial}{\overline{\partial}}{\varphi}|_{g} {\leqslant}D, ~~D_{0}{\leqslant}D \text{~and~} A{\leqslant}A_{D}:=\frac{1}{C_{0}M_{0}D},$$ then $$|{\partial}{\varphi}|_{g}^{2}{\leqslant}M_{1} \text{~and~} |{\partial}{\overline{\partial}}{\varphi}|_{g} {\leqslant}\frac{D}{2}.$$
i). $C^1$-estimate. As in [@CHZ18], we consider the following auxiliary function, $$Q = \log|{\partial}{\varphi}|_{g}^{2}+\frac{{\varphi}}{B},$$ where $B$ is a uniform constant to be determined. Let $x_{0}$ be the maximum point of $Q$ and $\{e_{i}\}_{i=1}^{n}$ be a local unitary frame in a neighbourhood of $x_{0}$ such that, at $x_{0}$, $$\label{tilde gij}
\tilde{g}_{i{\overline}{j}}
= \delta_{i{\overline}{j}}\tilde{g}_{i{\overline}{i}}
= \delta_{i{\overline}{j}}(e^{{\varphi}}+\alpha e^{-{\varphi}}\rho_{i{\overline{i}}}+2n\alpha {\varphi}_{i{\overline{i}}}).$$ We use the following notations $$F^{i{\overline{j}}} = \frac{{\partial}\sigma_{2}(\hat{\omega})}{{\partial}\hat{g}_{i{\overline{j}}}}
\text{~and~}
F^{i{\overline{j}},k{\overline{l}}} = \frac{{\partial}^{2}\sigma_{2}(\hat{\omega})}{{\partial}\hat{g}_{i{\overline{j}}}{\partial}\hat{g}_{k{\overline{l}}}},$$ where $\hat{\omega}=e^{-{\varphi}}{\tilde{\omega}}$. By (\[condition of first and second order estimate\]), we know that $$\label{Bound of Fij}
\left|F^{i{\overline{i}}}-(n-1)\right| {\leqslant}\frac{1}{100}.$$ Then by a direct calculation, we have $$\label{Maximum principle gradient 1}
\begin{split}
F^{i{\overline{j}}}e_{i}{\overline{e}}_{j}(|{\partial}{\varphi}|_{g}^{2})
{\geqslant}{} & \frac{4}{5}\sum_{i,j}(|e_{i}{\overline{e}}_{j}({\varphi})|^{2}+|e_{i}e_{j}({\varphi})|^{2})-C|{\partial}{\varphi}|_{g}^{2} \\
& +\sum_{k}F^{i{\overline{i}}}\left(e_{k}({\varphi}_{i{\overline{i}}}){\varphi}_{{\overline{k}}}+{\overline{e}}_{k}({\varphi}_{i{\overline{i}}}){\varphi}_{k}\right).
\end{split}$$
Next, we deal with the terms involving three derivatives of ${\varphi}$ in . Differentiating along $e_{k}$ at $x_{0}$, we have $$\label{first differentiate of equation}
\begin{split}
2n\alpha F^{i{\overline{i}}}e_{k}({\varphi}_{i{\overline{i}}})
= {} & 2\alpha e^{-{\varphi}}{\varphi}_{k}F^{i{\overline{i}}}\rho_{i{\overline{i}}}-\alpha e^{-{\varphi}}F^{i{\overline{i}}}e_{k}(\rho_{i{\overline{i}}})+2n\alpha{\varphi}_{k}F^{i{\overline{i}}}{\varphi}_{i{\overline{i}}} \\[1mm]
& -2\alpha n(n-1)(|{\partial}{\varphi}|_{g}^{2})_{k}+2\alpha n(n-1)|{\partial}{\varphi}|_{g}^{2}{\varphi}_{k} \\
& -n(n-1)e^{-{\varphi}}f{\varphi}_{k}+\frac{n(n-1)}{2}e^{-{\varphi}}f_{k},
\end{split}$$ which implies $$\label{Gradient estimate equation 7}
\begin{split}
& \sum_{k}F^{i{\overline{i}}}\left(e_{k}({\varphi}_{i{\overline{i}}}){\varphi}_{{\overline{k}}}+{\overline{e}}_{k}({\varphi}_{i{\overline{i}}}){\varphi}_{k}\right) \\[1mm]
{\geqslant}{} & -Ce^{-{\varphi}}|{\partial}{\varphi}|_{g}^{2}-Ce^{-{\varphi}}|{\partial}{\varphi}|_{g}+2|{\partial}{\varphi}|_{g}^{2}F^{i{\overline{i}}}{\varphi}_{i{\overline{i}}} \\[1mm]
& +2(n-1)|{\partial}{\varphi}|_{g}^{4}
-2(n-1)\textrm{Re}\left(\sum_{k}(|{\partial}{\varphi}|_{g}^{2})_{k}{\varphi}_{{\overline{k}}}\right) \\
& -\frac{n-1}{\alpha}e^{-{\varphi}}|{\partial}{\varphi}|_{g}^{2}f
+\frac{n-1}{2\alpha}e^{-{\varphi}}\textrm{Re}\left(\sum_{k}f_{k}{\varphi}_{{\overline{k}}}\right).
\end{split}$$ For the third and fourth term of (\[Gradient estimate equation 7\]), by the argument of [@CHZ18 (3.14)], we obtain $$\label{Gradient estimate equation 10}
\begin{split}
& 2|{\partial}{\varphi}|_{g}^{2}F^{i{\overline{i}}}{\varphi}_{i{\overline{i}}}+2(n-1)|{\partial}{\varphi}|_{g}^{4} \\[2mm]
{\geqslant}{} & -\frac{1}{10}\sum_{i,j}|e_{i}{\overline{e}}_{j}({\varphi})|^{2}-\left(Ce^{-2{\varphi}}+\frac{1}{B}\right)|{\partial}{\varphi}|_{g}^{4}-C|{\partial}{\varphi}|_{g}^{2}.
\end{split}$$ For the last two terms of (\[Gradient estimate equation 7\]). By the similar calculation of [@CHZ18 (3.10)] and the expression of $f$ (\[Definition of f\]), at $x_{0}$, we get $$\label{the first order of f}
\begin{split}
& -\frac{n-1}{\alpha}e^{-{\varphi}}|{\partial}{\varphi}|_{g}^{2}f+\frac{n-1}{2\alpha}e^{-{\varphi}}\textrm{Re}\left(\sum_{k}f_{k}{\varphi}_{{\overline{k}}}\right) \\[1mm]
{\geqslant}{} & -C\left(e^{-2{\varphi}}|{\partial}{\varphi}|_{g}+e^{-2{\varphi}}+|{\partial}{\varphi}|_{g}\right)\sum_{i,j}(|e_{i}{\overline{e}}_{j}({\varphi})|+|e_{i}e_{j}({\varphi})|) \\
& -Ce^{-{\varphi}}|{\partial}{\varphi}|_{g}^{4}-Ce^{-{\varphi}}|{\partial}{\varphi}|_{g}^{3}-Ce^{-{\varphi}}|{\partial}{\varphi}|_{g}^{2}-Ce^{-{\varphi}}|{\partial}{\varphi}|_{g} \\[2mm]
& -C|{\partial}{\varphi}|^{3}_{g}-C|{\partial}{\varphi}|^{2}_{g} \\[1mm]
{\geqslant}{} & -\frac{1}{10}\sum_{i,j}(|e_{i}{\overline{e}}_{j}({\varphi})|^{2}+|e_{i}e_{j}({\varphi})|^{2})-\left(Ce^{-{\varphi}}+\frac{1}{2B}\right)|{\partial}{\varphi}|_{g}^{4} \\
& -CB^{3},
\end{split}$$ where we used the Cauchy-Schwarz inequality in the last inequality. Substituting (\[Gradient estimate equation 10\]) and (\[the first order of f\]) into (\[Gradient estimate equation 7\]), we derive $$\label{third term of gradient-lapalce}
\begin{split}
& \sum_{k}F^{i{\overline{i}}}\left(e_{k}({\varphi}_{i{\overline{i}}}){\varphi}_{{\overline{k}}}+{\overline{e}}_{k}({\varphi}_{i{\overline{i}}}){\varphi}_{k}\right) \\
{\geqslant}{} & -\frac{1}{5}\sum_{i,j}(|e_{i}{\overline{e}}_{j}({\varphi})|^{2}+|e_{i}e_{j}({\varphi})|^{2})
-2(n-1){\rm Re}\left(\sum_{k}(|{\partial}{\varphi}|_{g}^{2})_{k}{\varphi}_{{\overline{k}}}\right) \\
& -\left(Ce^{-{\varphi}}+\frac{3}{2B}\right)|{\partial}{\varphi}|_{g}^{4}-C|{\partial}{\varphi}|_{g}^{2}-CB^{3}.
\end{split}$$ Hence, substituting this into , we see that $$\label{Maximum principle gradient}
\begin{split}
& F^{i{\overline{j}}}e_{i}{\overline{e}}_{j}(|{\partial}{\varphi}|_{g}^{2}) \\
{\geqslant}{} & \frac{3}{5}\sum_{i,j}(|e_{i}{\overline{e}}_{j}({\varphi})|^{2}+|e_{i}e_{j}({\varphi})|^{2})
-2(n-1)\textrm{Re}\left(\sum_{k}(|{\partial}{\varphi}|_{g}^{2})_{k}{\varphi}_{{\overline{k}}}\right) \\
& -\left(Ce^{-{\varphi}}+\frac{3}{2B}\right)|{\partial}{\varphi}|_{g}^{4}-C|{\partial}{\varphi}|_{g}^{2}-CB^{3}.
\end{split}$$ By the maximum principle, at $x_{0}$, we obtain $$\label{Gradient estimate equation 3}
\begin{split}
0 {\geqslant}{} & F^{i{\overline{j}}}e_{i}{\overline{e}}_{j}(Q) \\
{\geqslant}{} & \frac{1}{2|{\partial}{\varphi}|_{g}^{2}}\sum_{i,j}(|e_{i}{\overline{e}}_{j}({\varphi})|^{2}+|e_{i}e_{j}({\varphi})|^{2})
-\frac{2(n-1)\textrm{Re}\left(\sum_{k}(|{\partial}{\varphi}|_{g}^{2})_{k}{\varphi}_{{\overline{k}}}\right)}{|{\partial}{\varphi}|_{g}^{2}} \\
& -\frac{F^{i{\overline{i}}}|e_{i}(|{\partial}{\varphi}|_{g}^{2})|^{2}}{|{\partial}{\varphi}|_{g}^{4}}
-\left(Ce^{-{\varphi}}+\frac{1}{B}\right)|{\partial}{\varphi}|_{g}^{2}-C+\frac{1}{B}F^{i{\overline{i}}}e_{i}{\overline{e}}_{i}({\varphi}).
\end{split}$$ On the other hand, by the fact $dQ=0$ and the Cauchy-Schwarz inequality, we get (cf. [@CHZ18 (3.17)-(3.19)]) $$\begin{split}
& -\frac{2(n-1)\textrm{Re}\left(\sum_{k}(|{\partial}{\varphi}|_{g}^{2})_{k}{\varphi}_{{\overline{k}}}\right)}{|{\partial}{\varphi}|_{g}^{2}}
-\frac{F^{i{\overline{i}}}|e_{i}(|{\partial}{\varphi}|_{g}^{2})|^{2}}{|{\partial}{\varphi}|_{g}^{4}}
+\frac{1}{B}F^{i{\overline{i}}}e_{i}{\overline{e}}_{i}({\varphi}) \\
{\geqslant}{} & -\frac{1}{4|{\partial}{\varphi}|_{g}^{2}}\sum_{i,j}|e_{i}{\overline{e}}_{j}({\varphi})|^{2}+
\frac{2(n-1)}{B}|{\partial}{\varphi}|_{g}^{2}-\frac{C}{B^{2}}|{\partial}{\varphi}|_{g}^{2}.
\end{split}$$ Substituting this into (\[Gradient estimate equation 3\]), we see that $$\label{gradient lalpace}
\begin{split}
0 {\geqslant}{} & \frac{1}{4|{\partial}{\varphi}|_{g}^{2}}\sum_{i,j}\left(|e_{i}{\overline{e}}_{j}({\varphi})|^{2}+|e_{i}e_{j}({\varphi})|^{2}\right)-C_{0}B^{3} \\
& +\left(\frac{4n-7}{2B}-\frac{C_{0}}{B^{2}}-C_{0}e^{-{\varphi}}\right)|{\partial}{\varphi}|_{g}^{2}.
\end{split}$$ Since $A\ll1$, we may assume $$C_{0}e^{-{\varphi}} {\leqslant}\frac{1}{32C_{0}}.$$ By choosing $B = 4C_{0}$ in (\[gradient lalpace\]), we see that $$|{\partial}{\varphi}|_{g}^{2}(x_{0}) {\leqslant}2^{11}C_{0}^{5}.$$ Note that $\frac{1}{M_{0}A}{\leqslant}e^{-{\varphi}}{\leqslant}M_{0}A$. Hence, we obtain $$\label{the first estimate}
\max_{M}|{\partial}{\varphi}|_{g}^{2} {\leqslant}e^{\frac{1}{B}(\sup_{M}{\varphi}-\inf_{M}{\varphi})}|{\partial}{\varphi}|_{g}^{2}(x_{0}) {\leqslant}C.$$
ii). $C^2$-estimate. The proof is almost as same as [@CHZ18 Propsition 4.1]. We consider the following auxiliary function, $$Q = |{\partial}{\overline{\partial}}{\varphi}|_{g}^{2} + B|{\partial}{\varphi}|_{g}^{2},$$ where $B$ is a uniform constant to be determined. Let $x_{0}$ be the maximum point of $Q$ and $\{e_{i}\}_{i=1}^{n}$ be the local unitary frame such that ${\tilde{g}}(x_{0})$ is diagonal. By direct calculation, we have $$\label{Second order estimate equation 1}
\begin{split}
F^{i{\overline{i}}}e_{i}{\overline{e}}_{i}(|{\partial}{\overline{\partial}}{\varphi}|_{g}^{2})
= {} & 2 \sum_{k,l}F^{i{\overline{i}}}e_{i}{\overline{e}}_{i}({\varphi}_{k{\overline{l}}}){\varphi}_{l{\overline{k}}}+2\sum_{k,l}F^{i{\overline{i}}}e_{i}({\varphi}_{k{\overline{l}}}){\overline{e}}_{i}({\varphi}_{l{\overline{k}}}) \\
{\geqslant}{} & -2|{\partial}{\overline{\partial}}{\varphi}|_{g}\sum_{k,l}|F^{i{\overline{i}}}e_{i}{\overline{e}}_{i}({\varphi}_{k{\overline{l}}})|+\frac{1}{2}\sum_{i,j,p}|e_{p}e_{i}{\overline{e}}_{j}({\varphi})|^{2} \\
& -C\sum_{i,j}(|e_{i}{\overline{e}}_{j}({\varphi})|^{2}+|e_{i}e_{j}({\varphi})|^{2})-C.
\end{split}$$ To deal with the fourth order terms $ \sum_{k,l}F^{i{\overline{i}}}e_{i}{\overline{e}}_{i}({\varphi}_{k{\overline{l}}}){\varphi}_{l{\overline{k}}}$ in (\[Second order estimate equation 1\]), we differentiate (\[New Hessian type of Fu-Yau equation\]) twice along $e_{k}$ and ${\overline{e}}_{l}$, we get $$\begin{split}
&F^{i{\overline{j}},p{\overline{q}}}e_{k}(e^{-{\varphi}}{\tilde{g}}_{i{\overline{j}}}){\overline{e}}_{l} (e^{-{\varphi}}{\tilde{g}}_{p{\overline{q}}})
+F^{i{\overline{j}}}e_{k}{\overline{e}}_{l}(e^{-{\varphi}}{\tilde{g}}_{i{\overline{j}}}) \\
&= -2n(n-1)\alpha e_{k}{\overline{e}}_{l}(e^{-{\varphi}}|{\partial}{\varphi}|_{g}^{2})+\frac{n(n-1)}{2}e_{k}{\overline{e}}_{l}(e^{-2{\varphi}}f).
\end{split}$$ By the similar argument of [@CHZ18 Lemma 4.2] and the expression of $f$ (\[Definition of f\]), we obtain $$\begin{split}
|F^{i{\overline{i}}}e_{i}{\overline{e}}_{i}({\varphi}_{k{\overline{l}}})|
{\leqslant}{} & 8n|\alpha| e^{-{\varphi}}\sum_{i,j,p}|e_{p}e_{i}{\overline{e}}_{j}({\varphi})|^{2}+C\sum_{i,j,p}|e_{p}e_{i}{\overline{e}}_{j}({\varphi})| \\
& +C\sum_{i,j}(|e_{i}{\overline{e}}_{j}({\varphi})|^{2}+|e_{i}e_{j}({\varphi})|^{2})+C.
\end{split}$$ Substituting this into (\[Second order estimate equation 1\]) and using , we obtain $$\begin{split}
F^{i{\overline{i}}}e_{i}{\overline{e}}_{i}(|{\partial}{\overline{\partial}}{\varphi}|_{g}^{2})
{\geqslant}-C_{0}(D+1)\sum_{i,j}(|e_{i}{\overline{e}}_{j}({\varphi})|^{2}+|e_{i}{e}_{j}({\varphi})|^{2})-C_{0}.
\end{split}$$ On the other hand, by and $C^{1}$-estimate, we have $$F^{i{\overline{i}}}e_{i}{\overline{e}}_{i}(|{\partial}{\varphi}|_{g}^{2})
{\geqslant}\frac{1}{2}\sum_{i,j}(|e_{i}{\overline{e}}_{j}({\varphi})|^{2}+|e_{i}{e}_{j}({\varphi})|^{2})-C_{1}.$$ Hence, by the maximum principle, at $x_{0}$, we get $$\begin{split}
0 {\geqslant}{} & F^{i{\overline{i}}}e_{i}{\overline{e}}_{i}(Q) \\[1mm]
= {} & F^{i{\overline{i}}}e_{i}{\overline{e}}_{i}(|{\partial}{\overline{\partial}}{\varphi}|_{g}^{2})+BF^{i{\overline{i}}}e_{i}{\overline{e}}_{i}(|{\partial}{\varphi}|_{g}^{2}) \\
{\geqslant}{} & \left(\frac{B}{2}-C_{0}D-C_{0}\right)\sum_{i,j}(|e_{i}{\overline{e}}_{j}({\varphi})|^{2}+|e_{i}{e}_{j}({\varphi})|^{2})-C_{0}-C_{1}B.
\end{split}$$ Choose $B=8C_{0}D+8C_{0}$. It follows that $$|{\partial}{\overline{\partial}}{\varphi}|_{g}^{2}(x_{0}) {\leqslant}C.$$ Therefore, by , at the expense of increasing $D_{0}$, we obtain $$\max_{M}|{\partial}{\overline{\partial}}{\varphi}|_{g}^{2} {\leqslant}|{\partial}{\overline{\partial}}{\varphi}|_{g}^{2}(x_{0})+BC {\leqslant}CD {\leqslant}\frac{D^{2}}{4}.$$
Proof of Theorem \[Existence and Uniqueness Theorem\] {#Proof of existence theorem}
=====================================================
In this section, we solve (\[Fu-Yau equation t\]) when any $t\in [0,1]$. The following lemma shows the existence of solutions when $t=0$.
\[Existence lemma\] Let $(M,\omega)$ be an $n$-dimensional compact astheno-Kähler manifold. Then there exists a function $h\in C^{\infty}(M)$, unique up to addition of a constant, such that $$\label{h-solution}
{\sqrt{-1} \partial \overline{\partial}}(e^{h}\omega)\wedge\omega^{n-2} = 0.$$
First we prove the existence. We define an elliptic operator ${\tilde{L}}$ by $$({\tilde{L}}u)\omega^{n} = {\sqrt{-1} \partial \overline{\partial}}(u\omega)\wedge\omega^{n-2}.$$ Let ${\tilde{L}}^{*}$ be a $L^{2}$-adjoint operator of ${\tilde{L}}$. Then by Stokes’ formula and the condition ${\sqrt{-1} \partial \overline{\partial}}\omega^{n-2}=0$, we see that $$({\tilde{L}}^{*}v)\omega^{n}
= {\sqrt{-1} \partial \overline{\partial}}v\wedge\omega^{n-1}+\sqrt{-1}{\partial}v\wedge\omega\wedge{\overline{\partial}}\omega^{n-2}-{\overline{\partial}}v\wedge\omega\wedge\sqrt{-1}{\partial}\omega^{n-2}.$$ It follows that ${\rm Ker}{\tilde{L}}^{*}=\{\text{constants}\}$ and $\textrm{Ind}({\tilde{L}})=\textrm{Ind}({\tilde{L}}^{*})=0$. Thus $$\textrm{dim}(\textrm{Ker}{\tilde{L}}) = 1.$$ Denote the generator of $\textrm{Ker}{\tilde{L}}$ by $v_{0}$. Then $v_{0}$ does not change the sign, and we may assume that $v_{0}{\geqslant}0$. By the strong maximum principle, we know that $v_{0}>0$. Hence, $h=\log v_{0}$ satisfies (\[h-solution\]).
For the uniqueness, by $\textrm{dim}(\textrm{Ker}{\tilde{L}})=1$, we see that the only solution of (\[h-solution\]) is $h+c$, where $c$ is a constant.
Choose the function $h$ in (\[Fu-Yau equation t\]) as a solution of (\[h-solution\]). We consider solution $\varphi=\varphi_t$ of (\[Fu-Yau equation t\]) which satisfies the elliptic condition $$\label{Elliptic condition t}
e^{{\varphi}}\omega+t\alpha e^{-{\varphi}}\rho+2n\alpha{\sqrt{-1} \partial \overline{\partial}}{\varphi}\in \Gamma_{2}(M)$$ and the normalization condition $$\label{Normalization condition t}
\|e^{-{\varphi}}\|_{L^{1}}=A.$$ By taking ${\varphi}_{0}=h+\log\|e^{-h}\|_{L^{1}}-\log A$, it is easy to see that $\|e^{-{\varphi}_{0}}\|_{L^{1}}=A$ and ${\varphi}_{0}$ is a solution of (\[Fu-Yau equation t\]) when $t=0$. Moreover, since $A\ll 1$, $$e^{{\varphi}_{0}}\omega+2n\alpha{\sqrt{-1} \partial \overline{\partial}}{\varphi}_{0}=\frac{\|e^{-h}\|_{L^{1}}}{A}e^{h}\omega+2n\alpha{\sqrt{-1} \partial \overline{\partial}}h > 0.$$ Thus ${\varphi}_{0}$ satisfies (\[Elliptic condition t\]) and (\[Normalization condition t\]).
For a fixed $\beta\in (0,1)$, we define $$\begin{split}
B & = \{ {\varphi}\in C^{2,\beta}(M) ~|~ \|e^{-{\varphi}}\|_{L^{1}} = A \},\\[2mm]
B_{1} & = \{ ({\varphi},t)\in B\times[0,1] ~|~ \text{${\varphi}$ satisfies (\ref{Elliptic condition t})} \},\\
B_{2} & = \{ u\in C^{\beta}(M) ~|~ \int_{M}u\omega^{n}=0 \}.
\end{split}$$ Then $B_{1}$ is an open subset of $B\times[0,1]$. Since $\int_{M}\mu\omega^{n}=0$, we introduce a map $\Phi:B_{1}\rightarrow B_{2}$, $$\begin{split}
\Phi({\varphi},t)\omega^{n} = {} &{\sqrt{-1} \partial \overline{\partial}}(e^{{\varphi}}\omega-t\alpha e^{-{\varphi}}\rho)\wedge\omega^{n-2} \\[1mm]
& +n\alpha{\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge{\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge\omega^{n-2} \\
& +2n\alpha(t-1){\sqrt{-1} \partial \overline{\partial}}h\wedge{\sqrt{-1} \partial \overline{\partial}}h\wedge\omega^{n-2}+t\mu\frac{\omega^{n}}{n!}.
\end{split}$$ Set $$I=\{ t\in [0,1] ~|~ \text{there exists $({\varphi},t)\in B_{1}$ such that $\Phi({\varphi},t)=0$} \}.$$ Then the existence of solutions of (\[Fu-Yau equation\]) is reduced to proving that $I$ is both open and closed.
**Openness.** Suppose that $\hat{t}\in I$ and there exists $(\hat{{\varphi}},\hat{t})\in B_{1}$ such that $\Phi(\hat{{\varphi}},\hat{t})=0$. Let $$L: \{ u\in C^{2,\beta}(M) ~|~ \int_{M}ue^{-\hat{{\varphi}}}\omega^{n}=0 \}
\rightarrow \{ v\in C^{\beta}(M) ~|~ \int_{M}v\omega^{n}=0 \}$$ be a linearized operator of $\Phi$ at $\hat{{\varphi}}$. Then $$\begin{split}
(Lu)\omega^{n} = {} & {\sqrt{-1} \partial \overline{\partial}}(ue^{\hat{{\varphi}}}\omega+\hat{t}\alpha ue^{-\hat{{\varphi}}}\rho)\wedge\omega^{n-2} \\
& +2n\alpha{\sqrt{-1} \partial \overline{\partial}}\hat{{\varphi}}\wedge{\sqrt{-1} \partial \overline{\partial}}u\wedge\omega^{n-2}.
\end{split}$$ By the implicit function theorem, it suffices to prove that $L$ is injective and surjective.
Let $L^{*}$ be a $L^{2}$-adjoint operaor of $L$. By the fact ${\sqrt{-1} \partial \overline{\partial}}\omega^{n-2}=0$ and Stokes’ formula, it follows that $$\label{Definition of L^*}
\begin{split}
L^{*}(v)\omega^{n}
= {} & {\sqrt{-1} \partial \overline{\partial}}v\wedge\left(e^{\hat{{\varphi}}}\omega+\hat{t}\alpha e^{-\hat{{\varphi}}}\rho+2n\alpha{\sqrt{-1} \partial \overline{\partial}}\hat{{\varphi}}\right)\wedge\omega^{n-2} \\
& +\sqrt{-1}{\partial}v\wedge\left(e^{\hat{{\varphi}}}\omega+\hat{t}\alpha e^{-\hat{{\varphi}}}\rho+2n\alpha{\sqrt{-1} \partial \overline{\partial}}\hat{{\varphi}}\right)\wedge{\overline{\partial}}\omega^{n-2} \\
& -{\overline{\partial}}v\wedge\left(e^{\hat{{\varphi}}}\omega+\hat{t}\alpha e^{-\hat{{\varphi}}}\rho+2n\alpha{\sqrt{-1} \partial \overline{\partial}}\hat{{\varphi}}\right)\wedge\sqrt{-1}{\partial}\omega^{n-2}.
\end{split}$$ Thus $L^{*}$ has no zero order terms. By the strong maximum principle, we see that $$\label{Kernel of L*}
\textrm{Ker}L^{*} \subset \{ \text{constants} \}.$$ As a consequence, $\textrm{Ker}L \subset \{ cu_{0} ~|~ c\in\mathbf{R} \}$ for some smooth function $u_0$ by $\textrm{Ind}(L)=0$. On the other hand, again by the strong maximum principle, we may assume that $u_{0}>0$. Thus $$u_{0} \notin \{ u\in C^{2,\beta}(M) ~|~ \int_{M}ue^{-\hat{{\varphi}}}\omega^{n}=0 \},$$ which implies $\textrm{Ker}L=0$, and so $L$ is injective.
Next, for any $$w \in \{ v\in C^{\beta}(M) ~|~ \int_{M}v\omega^{n}=0 \},$$ by the Fredholm alternative and regularity theory of elliptic equations, there exists a function ${\tilde{u}}\in C^{2,\beta}(M)$ such that $L{\tilde{u}}=w$. It then follows that $$L({\tilde{u}}+c_{0}u_{0}) = w \text{~and~} {\tilde{u}}+c_{0}u_{0} \in \{ u\in C^{2,\beta}(M) ~|~ \int_{M}ue^{-\hat{{\varphi}}}\omega^{n}=0 \},$$ where $$c_{0} = -\frac{\int_{M}{\tilde{u}}e^{-\hat{{\varphi}}}\omega^{n}}{\int_{M}u_{0}e^{-\hat{{\varphi}}}\omega^{n}}.$$ This implies that $L$ is surjective.
**Closeness.** First we prove the $C^{0}$-estimate along (\[Fu-Yau equation t\]). Recalling ${\varphi}_{0}=h+\log\|e^{-h}\|_{L^{1}}-\log A$ and $A\ll1$, we have $\sup_{M}e^{-{\varphi}_{0}}{\leqslant}M_{0}A$. We claim
$$\sup_{M}e^{-{\varphi}_{t}}{\leqslant}2M_{0}A, ~\forall~ t\in[0,1].$$
If the claim is false, there exists ${\tilde{t}}\in(0,1)$ such that $$\label{Claim 2 equation 1}
\sup_{M}e^{-{\varphi}_{{\tilde{t}}}} = 2M_{0}A.$$ Since $A\ll1$, we assume that $2M_{0}A{\leqslant}\delta_{0}$, where $\delta_{0}$ is the constant in Proposition \[Zero order estimate\]. Then, applying Proposition \[Zero order estimate\] while $\rho$ and $\mu$ are replaced by $$t\rho \text{~and~} \frac{n\alpha(t-1)n!{\sqrt{-1} \partial \overline{\partial}}h\wedge{\sqrt{-1} \partial \overline{\partial}}h\wedge\omega^{n-2}}{\omega^{n}}+t\mu,$$ we obtain $\sup_{M}e^{-{\varphi}_{{\tilde{t}}}}{\leqslant}M_{0}A$, which contradicts with (\[Claim 2 equation 1\]). Thus the claim is true.
By Claim 2, we see that Proposition \[Zero order estimate\] and Proposition \[First and second order estimate\] hold for ${\varphi}_{t}$. As a consequence, we get the $C^{2}$-estimate for $\varphi_t$ along (\[Fu-Yau equation t\]). Then combining the $C^{2,\alpha}$-estimate (cf. [@TWWY15 Theorem 1.1]) and the bootstrapping argument, we complete the proof of closeness (for more details, we refer the reader to [@CHZ18 Section 5.2]).
**Uniqueness.** The uniqueness of solutions of (\[Fu-Yau equation\]) can be proved by a similar argument of [@CHZ18 Section 5.3] (also see the proof of Theorem \[Uniqueness theorem alpha negative\] in Section 7 below). It suffices to prove that $\varphi_0$ of (\[Fu-Yau equation t\]) is unique when $t=0$ by the estimates in Proposition \[Zero order estimate\] and Proposition \[First and second order estimate\]. But the latter is guarantted by Remark \[unique-2-h\] in Section 7.
Zero order estimate (II)
========================
In this section, we improve Proposition \[Zero order estimate\] in the case of $\alpha<0$. The key point is to drop the condition $e^{-{\varphi}}{\leqslant}\delta_{0}$. We begin with the following lemma.
\[Effective zero order estimate lemma\] Let $\alpha<0$ and ${\varphi}$ be a smooth solution of (\[Fu-Yau equation\]) satisfying (\[Elliptic condition\]) and (\[L\^n normalization condition\]). There exists constant $A_{0}$ depending only on $\alpha$, $\rho$, $\mu$ and $(M,\omega)$ such that if any $A{\leqslant}A_{0}$, then $$\label{c0-2}
e^{-{\varphi}} {\leqslant}C_{0}.$$
The elliptic condition ${\tilde{\omega}}\in\Gamma_{2}(M)$ implies that $$\int_{M}e^{-k{\varphi}}{\tilde{\omega}}\wedge\omega^{n-1} {\geqslant}0.$$ Namely, $$0 {\leqslant}\int_{M}e^{-k{\varphi}}(e^{{\varphi}}\omega+\alpha e^{-{\varphi}}\rho+2n\alpha{\sqrt{-1} \partial \overline{\partial}}{\varphi})\wedge\omega^{n-1} .$$ By the Stokes’ formula, for $k>1$, it follows that $$\label{Zero order estimate equation 2}
\begin{split}
0 {\leqslant}{} & \int_{M}e^{-(k-1){\varphi}}\omega^{n}-|\alpha|\int_{M}e^{-(k+1){\varphi}}\rho\wedge\omega^{n-1} \\
& +2n\alpha\int_{M}e^{-k{\varphi}}{\overline{\partial}}{\varphi}\wedge\sqrt{-1}{\partial}\omega^{n-1} \\
& -2n|\alpha|k\int_{M}e^{-k{\varphi}}\sqrt{-1}{\partial}{\varphi}\wedge{\overline{\partial}}{\varphi}\wedge\omega^{n-1} \\
{\leqslant}{} & C\int_{M}\left(e^{-(k-1){\varphi}}+e^{-(k+1){\varphi}}\right)\omega^{n} \\
& -\frac{2n\alpha}{k}\int_{M}e^{-k{\varphi}}{\sqrt{-1} \partial \overline{\partial}}\omega^{n-1} \\
& -2n|\alpha|k\int_{M}e^{-k{\varphi}}\sqrt{-1}{\partial}{\varphi}\wedge{\overline{\partial}}{\varphi}\wedge\omega^{n-1}.
\end{split}$$ Thus by the Sobolev inequality, we obtain $$\label{Zero order estimate equation 1}
\begin{split}
~ & \left(\int_{M}e^{-k\beta{\varphi}}\omega^{n}\right)^{\frac{1}{\beta}}\\
{\leqslant}& C_{1}k\int_{M}e^{-(k-1){\varphi}}\omega^{n}+C_{1}k\int_{M}e^{-(k+1){\varphi}}\omega^{n}+C\int_{M}e^{-k{\varphi}}\omega^{n},
\end{split}$$ where $\beta=\frac{n}{n-1}$.
Next, we prove that $\|e^{-{\varphi}}\|_{L^{n+1}}{\leqslant}C$ by (\[Zero order estimate equation 1\]). In fact, by taking $k=n$ in (\[Zero order estimate equation 1\]) and the Hölder inequality, we see that $$\begin{split}
\|e^{-{\varphi}}\|_{L^{n\beta}}^{n}
{\leqslant}{} & C_{1}n\|e^{-{\varphi}}\|_{L^{n}}^{n-1}+C_{1}n\|e^{-{\varphi}}\|_{L^{n}}\|e^{-{\varphi}}\|_{L^{n\beta}}^{n}+C\|e^{-{\varphi}}\|_{L^{n}} \\
{\leqslant}{} & CA^{n-1}+C_{1}nA\|e^{-{\varphi}}\|_{L^{n\beta}}^{n}+CA.
\end{split}$$ Note that $C_{1}nA{\leqslant}\frac{1}{2}$. Then $$\|e^{-{\varphi}}\|_{L^{n\beta}} {\leqslant}C.$$ Thus we get $$\label{ln-integral}
\|e^{-{\varphi}}\|_{L^{n+1}} {\leqslant}\|e^{-{\varphi}}\|_{L^{n\beta}} {\leqslant}C.$$
Finally, we use the iteration to obtain (\[c0-2\]). Let $H=e^{-{\varphi}}+1$. It suffices to prove that $\|H\|_{L^{\infty}}{\leqslant}C$. By (\[Zero order estimate equation 1\]), it is easy to see that $$\|H\|_{L^{k\beta}} {\leqslant}(Ck)^{\frac{1}{k}}\|H\|_{L^{k+1}}^{\frac{k+1}{k}}.$$ For $j{\geqslant}0$, we define $$p_{j} = \beta^{j}+\frac{\beta}{\beta-1}, ~a_{j} = (Cp_{j}-p_{j})^{\frac{1}{p_{j}-1}} \text{~and~} b_{j} = \frac{p_{j}}{p_{j}-1}.$$ It then follows that $$\|H\|_{L^{p_{j+1}}} {\leqslant}a_{j}\|H\|_{L^{p_{j}}}^{b_{j}},$$ which implies $$\label{iteration-h}
\|H\|_{L^{p_{j+1}}} {\leqslant}a_{j}a_{j-1}^{b_{j}}\cdots a_{0}^{b_{j}\cdots b_{1}}\|H\|_{L^{p_{0}}}^{b_{j}\cdots b_{0}}.$$ Note that $\prod_{i=1}^{\infty}a_{i}<\infty$ and $\prod_{i=1}^{\infty}b_{i}<\infty$. Thus as $j\rightarrow\infty$, we obtain from (\[iteration-h\]), $$\|H\|_{L^{\infty}} {\leqslant}C\|H\|_{L^{n+1}}^{C}{\leqslant}C'.$$ Here we used (\[ln-integral\]).
Now, we apply Lemma \[Effective zero order estimate lemma\] and Lemma \[Zero order estimate lemma\] to improve Proposition \[Zero order estimate\] as follows.
\[Stronger zero order estimate\] Let $\alpha<0$ and ${\varphi}$ be a smooth solution of (\[Fu-Yau equation\]) satisfying (\[Elliptic condition\]) and (\[L\^n normalization condition\]). There exist constants $A_{0}$ and $M_{0}$ depending only on $\alpha$, $\rho$, $\mu$ and $(M,\omega)$ such that if $A{\leqslant}A_{0}$, then $$\frac{A}{M_{0}} {\leqslant}e^{-{\varphi}} {\leqslant}M_{0}A.$$
By taking $f({\varphi})=-e^{-k{\varphi}}$ ($k{\geqslant}2$) in Lemma \[Zero order estimate lemma\], we have $$\begin{split}
& k\int_{M}e^{-k{\varphi}}(e^{{\varphi}}\omega+\alpha e^{-{\varphi}}\rho)\wedge\sqrt{-1}{\partial}{\varphi}\wedge{\overline{\partial}}{\varphi}\wedge\omega^{n-2} \\
{\leqslant}{} & -2k\int_{M}e^{-k{\varphi}}\sqrt{-1}{\partial}{\varphi}\wedge(e^{{\varphi}}{\overline{\partial}}\omega-\alpha e^{-{\varphi}}{\overline{\partial}}\rho)\wedge\omega^{n-2} \\
& -2k\int_{M}e^{-k{\varphi}}{\overline{\partial}}{\varphi}\wedge(e^{{\varphi}}\omega-\alpha e^{-{\varphi}}\rho)\wedge\sqrt{-1}{\partial}\omega^{n-2}
-2\int_{M}e^{-k{\varphi}}\mu\frac{\omega^{n}}{n!}.
\end{split}$$ It follows that $$\int_{M}e^{-(k-1){\varphi}}|{\partial}{\varphi}|_{g}^{2}\omega^{n}
{\leqslant}C\int_{M}\left(e^{-(k+1){\varphi}}|{\partial}{\varphi}|_{g}+e^{-k{\varphi}}\right)\omega^{n}.$$ Combining this with Lemma \[Effective zero order estimate lemma\] and the Cauchy-Shwarz inequality, we get $$\begin{split}
\int_{M}e^{-(k-1){\varphi}}|{\partial}{\varphi}|_{g}^{2}\omega^{n}
{\leqslant}{} & C\int_{M}\left(e^{-(k+3){\varphi}}+e^{-k{\varphi}}\right)\omega^{n} \\
{\leqslant}{} & C(C_{0}^{4}+C_{0})\int_{M}e^{-(k-1){\varphi}}\omega^{n}.
\end{split}$$ Thus by the Moser iteration, we derive $$\|e^{-{\varphi}}\|_{L^{\infty}} {\leqslant}C\|e^{-{\varphi}}\|_{L^{1}} {\leqslant}CA.$$ Note that $A\ll 1$. Hence $e^{-{\varphi}}\ll 1$. Now we can apply Proposition \[Zero order estimate\] to obtain $$\frac{A}{M_{0}} {\leqslant}e^{-{\varphi}} {\leqslant}M_{0}A.$$
First and second order estimates (II)
=====================================
In this section, we provide another proof to derive a prior $C^1,C^2$ estimates for $\varphi$ of (\[Fu-Yau equation\]) in case of $\alpha<0$, but without the restriction condition (\[Restrictions\]). For convenience, we say a constant $C$ is uniform if it depends only on $\alpha$, $\rho$, $\mu$ and $(M,\omega)$, and we use $C_{A}$ to denote a uniform constant depending on $A$. The main goal in this section is to prove the following proposition.
\[Effective second order estimate\] Let $\alpha<0$ and ${\varphi}$ be a smooth solution of (\[Fu-Yau equation\]) on a Hermitian manifold $(M,\omega)$, which satisfies (\[Elliptic condition\]) and (\[L\^n normalization condition\]). Then $$\sup_{M}|{\partial}{\overline{\partial}}{\varphi}|_{g} {\leqslant}C_{A}\sup_{M}|{\partial}{\varphi}|_{g}^{2}+C_{A},$$ where $C_{A}$ is a uniform constant depending on $A$.
For simplicity, we write (\[Fu-Yau equation t\]) as $$\label{Fu-Yau equation 2-nd Hessian type}
\sigma_{2}({\tilde{\omega}}) = F,$$ where $$F= \frac{n(n-1)}{2}\left(e^{2{\varphi}}+4|\alpha|e^\varphi|{\partial}{\varphi}|_{g}^{2}\right)+\frac{n(n-1)}{2}f$$ and $f$ is defined by (\[Definition of f\]). Let $\lambda_{1}{\geqslant}\lambda_{2}{\geqslant}\cdots{\geqslant}\lambda_{n}$ be the eigenvalues of ${\tilde{\omega}}$ with respect to $\omega$. Since ${\tilde{\omega}}\in\Gamma_{2}(M)$, by Proposition \[Stronger zero order estimate\], it is clear that $$\label{ddbar and lambda}
|{\partial}{\overline{\partial}}{\varphi}|_{g} {\leqslant}C_{A}\lambda_{1}+C_{A}.$$
In Hermitian case, more troublesome terms will appear when we commute the covariant derivatives (cf. (\[Commutation formulas\]) below). To deal with these bad terms, we consider the following auxiliary function as in [@CHZ17], $$Q = \log\lambda_{1}+h(|{\partial}{\varphi}|_{g}^{2})+e^{B{\varphi}},$$ where $B$ is a constant to be determined later, $$h(t) = -\frac{1}{2}\log(2K-t) \text{~and~} K = \sup_{M}|{\partial}{\varphi}|_{g}^{2}+1.$$ By directly calculation, we have $$\label{h'}
\frac{1}{4K}{\leqslant}h'{\leqslant}\frac{1}{2K} \text{~and~} h''=2(h')^{2}.$$
Let $x_{0}$ be the maximum point of $Q$. Around $x_{0}$, we choose holomorphic coordinate $(z^{1},z^{2},\cdots,z^{n})$ such that at $x_{0}$, $$g_{i{\overline{j}}} = \delta_{ij} \text{~and~}
{\tilde{g}}_{i{\overline{j}}} = \delta_{ij}\lambda_{i}.$$ To prove Proposition \[Effective second order estimate\], by (\[ddbar and lambda\]), it suffices to prove $\lambda_{1}{\leqslant}C_{A}K$. Without loss of generality, we assume that $\lambda_{1}\gg C_{A}K$. Moreover, We may suppose that $Q$ is smooth at $x_{0}$. Otherwise, we just need to apply a perturbation argument (cf. [@Sze15; @STW17; @CTW16]).
In the following calculation, we use the covariant derivatives with respect to the Chern connection $(\nabla, T^{\mathbb C}M)$ induced by $\omega$. Let us recall the commutation formulas for covariant derivatives: $$\label{Commutation formulas}
\begin{split}
{\varphi}_{i{\overline{j}}k} {} & = {\varphi}_{ki{\overline{j}}}-T_{ki}^{p}{\varphi}_{p{\overline{j}}}-R_{i{\overline{j}}k}^{\ \ \ \ p}{\varphi}_{p}, \\
{\varphi}_{i{\overline{j}}k{\overline{l}}} = {\varphi}_{k{\overline{l}}i{\overline{j}}}-T_{ki}^{p}{\varphi}_{p{\overline{l}}{\overline{j}}} {} & -{\overline{T_{lj}^{q}}}{\varphi}_{k{\overline{q}}i}
+{\varphi}_{p{\overline{j}}}R_{k{\overline{l}}i}^{\ \ \ \ p}-{\varphi}_{p{\overline{l}}}R_{i{\overline{j}}k}^{\ \ \ \ p}
-T_{ik}^{p}{\overline{T_{lj}^{q}}}{\varphi}_{p{\overline{q}}},
\end{split}$$ where $T_{ij}^{k}$ and $R_{i{\overline{j}}k}^{\ \ \ \ l}$ are components of torsion tensor and curvature tensor induced by the Chern connection. Let $$G^{i{\overline{j}}} = \frac{{\partial}\sigma_{2}^{\frac{1}{2}}({\tilde{\omega}})}{{\partial}{\tilde{g}}_{i{\overline{j}}}},~
G^{i{\overline{j}},k{\overline{l}}} = \frac{{\partial}^{2}\sigma_{2}^{\frac{1}{2}}({\tilde{\omega}})}{{\partial}{\tilde{g}}_{i{\overline{j}}}{\partial}{\tilde{g}}_{k{\overline{l}}}}.$$ The following lemmas are devoted to deriving lower bounds of $G^{k{\overline{k}}}{\varphi}_{k{\overline{k}}}$, $G^{k{\overline{k}}}(|{\partial}{\varphi}|_{g}^{2})_{k{\overline{k}}}$ and $G^{k{\overline{k}}}(\lambda_{1})_{k{\overline{k}}}$.
\[Lemma 1\] At $x_{0}$, we have $$G^{k{\overline{k}}}{\varphi}_{k{\overline{k}}} {\geqslant}\sum_{k}G^{k{\overline{k}}}-CF^{\frac{1}{2}}$$ and $$\begin{split}
G^{k{\overline{k}}}(|{\partial}{\varphi}|_{g}^{2})_{k{\overline{k}}}
{\geqslant}{} & \frac{1}{2}\sum_{i}G^{k{\overline{k}}}\left(|{\varphi}_{ik}|^{2}+|{\varphi}_{i{\overline{k}}}|^{2}\right)
-C_{A}K^{\frac{3}{2}} \\
& -C_{A}K^{\frac{1}{2}}\sum_{i,k}\left(|{\varphi}_{ik}|+|{\varphi}_{i{\overline{k}}}|\right)-C\sum_{k}G^{k{\overline{k}}}.
\end{split}$$
By the definition of ${\tilde{\omega}}$, we have $$\begin{split}
2n\alpha G^{k{\overline{k}}}{\varphi}_{k{\overline{k}}}
= {} & G^{k{\overline{k}}}\left({\tilde{g}}_{k{\overline{k}}}-e^{{\varphi}}g_{k{\overline{k}}}-\alpha e^{-{\varphi}}\rho_{k{\overline{k}}}\right) \\[2mm]
= {} & \sigma_{2}^{\frac{1}{2}}({\tilde{\omega}})-e^{{\varphi}}\sum_{k}G^{k{\overline{k}}}-\alpha e^{-{\varphi}}G^{k{\overline{k}}}\rho_{k{\overline{k}}} \\
= {} & F^{\frac{1}{2}}-e^{{\varphi}}\sum_{k}G^{k{\overline{k}}}-\alpha e^{-{\varphi}}G^{k{\overline{k}}}\rho_{k{\overline{k}}}.
\end{split}$$ Note that $A\ll1$ and $\alpha<0$. Then by Proposition \[Stronger zero order estimate\], we get $$G^{k{\overline{k}}}{\varphi}_{k{\overline{k}}} {\geqslant}\sum_{k}G^{k{\overline{k}}}-CF^{\frac{1}{2}}.$$
To prove the second inequality in the lemma, we compute $$\label{Lemma 1 equation 1}
\begin{split}
& G^{k{\overline{k}}}(|{\partial}{\varphi}|_{g}^{2})_{k{\overline{k}}} \\
= {} & \sum_{i}G^{k{\overline{k}}}\left(|{\varphi}_{ik}|^{2}+|{\varphi}_{i{\overline{k}}}|^{2}\right)
+2\textrm{Re}\left(\sum_{i}G^{k{\overline{k}}}{\varphi}_{i}{\varphi}_{{\overline{i}}k{\overline{k}}}\right) \\
{\geqslant}{} & \sum_{i}G^{k{\overline{k}}}\left(|{\varphi}_{ik}|^{2}+|{\varphi}_{i{\overline{k}}}|^{2}\right)
+2\textrm{Re}\left(\sum_{i}G^{k{\overline{k}}}{\varphi}_{i}{\varphi}_{k{\overline{k}}{\overline{i}}}\right) \\[1mm]
& -C\sum_{i}G^{k{\overline{k}}}\left(|{\varphi}_{ik}|+|{\varphi}_{i{\overline{k}}}|\right) \\
{\geqslant}{} & \frac{1}{2}\sum_{i}G^{k{\overline{k}}}\left(|{\varphi}_{ik}|^{2}+|{\varphi}_{i{\overline{k}}}|^{2}\right)
+2\textrm{Re}\left(\sum_{i}G^{k{\overline{k}}}{\varphi}_{i}{\varphi}_{k{\overline{k}}{\overline{i}}}\right)
-C\sum_{k}G^{k{\overline{k}}}.
\end{split}$$ On the other hand, by differentiating (\[Fu-Yau equation 2-nd Hessian type\]) along $\nabla_{{\overline{i}}}$, we see that $$2\textrm{Re}\left(\sum_{i}G^{k{\overline{k}}}{\varphi}_{i}{\varphi}_{k{\overline{k}}{\overline{i}}}\right)
= 2\textrm{Re}\left(\sum_{i}{\varphi}_{i}(F^{\frac{1}{2}})_{{\overline{i}}}\right)
= \frac{1}{F^{\frac{1}{2}}}\textrm{Re}\left(\sum_{i}{\varphi}_{i}F_{{\overline{i}}}\right).$$ Since $$F {\geqslant}\frac{1}{C}(e^{2{\varphi}}+e^{{\varphi}}|{\partial}{\varphi}|_{g}^{2})$$ and $$|F_{{\overline{i}}}| {\leqslant}Ce^{2{\varphi}}|{\partial}{\varphi}|_{g}+Ce^{{\varphi}}|{\partial}{\varphi}|_{g}^{3}
+Ce^{{\varphi}}(|{\partial}{\varphi}|_{g}+1)\sum_{i,k}\left(|{\varphi}_{ik}|+|{\varphi}_{i{\overline{k}}}|\right)+C,$$ we get $$\begin{split}
2\textrm{Re}\left(\sum_{i}G^{k{\overline{k}}}{\varphi}_{i}{\varphi}_{{\overline{i}}k{\overline{k}}}\right)
{\geqslant}{} & -Ce^{\frac{3}{2}{\varphi}}|{\partial}{\varphi}|_{g}-Ce^{\frac{1}{2}{\varphi}}|{\partial}{\varphi}|_{g}^{3} \\
& +Ce^{\frac{1}{2}{\varphi}}(|{\partial}{\varphi}|_{g}+1)\sum_{i,k}\left(|{\varphi}_{ik}|+|{\varphi}_{i{\overline{k}}}|\right)-C.
\end{split}$$ Hence, substituting this into (\[Lemma 1 equation 1\]), we obtain $$\begin{split}
G^{k{\overline{k}}}(|{\partial}{\varphi}|_{g}^{2})_{k{\overline{k}}}
{\geqslant}{} & \frac{1}{2}\sum_{i}G^{k{\overline{k}}}\left(|{\varphi}_{ik}|^{2}+|{\varphi}_{i{\overline{k}}}|^{2}\right)
-Ce^{\frac{3}{2}{\varphi}}|{\partial}{\varphi}|_{g}-Ce^{\frac{1}{2}{\varphi}}|{\partial}{\varphi}|_{g}^{3} \\[1mm]
& -Ce^{\frac{1}{2}{\varphi}}(|{\partial}{\varphi}|_{g}+1)\sum_{i,k}\left(|{\varphi}_{ik}|+|{\varphi}_{i{\overline{k}}}|\right)-C\sum_{k}G^{k{\overline{k}}}-C \\
{\geqslant}{} & \frac{1}{2}\sum_{i}G^{k{\overline{k}}}\left(|{\varphi}_{ik}|^{2}+|{\varphi}_{i{\overline{k}}}|^{2}\right)
-C_{A}K^{\frac{3}{2}} \\[1mm]
& -C_{A}K^{\frac{1}{2}}\sum_{i,k}\left(|{\varphi}_{ik}|+|{\varphi}_{i{\overline{k}}}|\right)-C\sum_{k}G^{k{\overline{k}}}.
\end{split}$$
\[Lemma 2\] At $x_{0}$, we have $$\begin{split}
G^{k{\overline{k}}}(\lambda_{1})_{k{\overline{k}}}
{\geqslant}{} & -G^{i{\overline{j}},k{\overline{l}}}\nabla_{1}{\tilde{g}}_{i{\overline{j}}}\nabla_{{\overline{1}}}{\tilde{g}}_{p{\overline{q}}}
-C\sum_{k}G^{k{\overline{k}}}|{\varphi}_{1{\overline{1}}k}|-C_{A}\sum_{k}|{\varphi}_{k1{\overline{1}}}| \\
& -C_{A}(\lambda_{1}+K)\sum_{i}G^{i{\overline{i}}}-C_{A}K^{-\frac{1}{2}}\sum_{i,k}|{\varphi}_{i{\overline{k}}}|^{2}
-C_{A}K^{\frac{3}{2}}.
\end{split}$$
By the formulas for the derivatives of $\lambda_{1}$ (cf. [@Sze15; @STW17; @CTW16]), we have $$\label{Lemma 2 equation 1}
\begin{split}
G^{k{\overline{k}}}(\lambda_{1})_{k{\overline{k}}}
&= G^{k{\overline{k}}}\nabla_{{\overline{k}}}\nabla_{k}{\tilde{g}}_{1{\overline{1}}}
+2\sum_{j>1}G^{k{\overline{k}}}\frac{|\nabla_{k}{\tilde{g}}_{1{\overline{j}}}|^{2}}{\lambda_{1}-\lambda_{j}} \\
& {\geqslant}2n\alpha G^{k{\overline{k}}}{\varphi}_{1{\overline{1}}k{\overline{k}}}
+\frac{2}{n\lambda_{1}}\sum_{j>1}G^{k{\overline{k}}}|\nabla_{k}{\tilde{g}}_{i{\overline{j}}}|^{2}
-Ce^{{\varphi}}(\lambda_{1}+K)\sum_{i}G^{i{\overline{i}}} \\
&{\geqslant}2n\alpha G^{k{\overline{k}}}{\varphi}_{1{\overline{1}}k{\overline{k}}}
+\frac{2}{n\lambda_{1}}\sum_{j>1}G^{k{\overline{k}}}|{\varphi}_{j{\overline{k}}{\overline{1}}}|^{2}
-Ce^{{\varphi}}(\lambda_{1}+K)\sum_{i}G^{i{\overline{i}}} \\
\end{split}$$ On the other hand, by differentiating (\[Fu-Yau equation 2-nd Hessian type\]) along $\nabla_{{\overline{1}}}\nabla_{1}$, we have $$G^{k{\overline{k}}}\nabla_{{\overline{1}}}\nabla_{1}(e^{{\varphi}}g_{k{\overline{k}}}+\alpha e^{-{\varphi}}\rho_{k{\overline{k}}}+2n\alpha{\varphi}_{k{\overline{k}}})
= (F^{\frac{1}{2}})_{1{\overline{1}}}-G^{i{\overline{j}},k{\overline{l}}}\nabla_{1}{\tilde{g}}_{i{\overline{j}}}\nabla_{{\overline{1}}}{\tilde{g}}_{p{\overline{q}}}.$$ Then by the commutation formula (\[Commutation formulas\]), we see that $$\begin{split}
2n\alpha G^{k{\overline{k}}}{\varphi}_{1{\overline{1}}k{\overline{k}}}
{\geqslant}{} & -C\sum_{i,k}G^{k{\overline{k}}}|{\varphi}_{i{\overline{k}}{\overline{1}}}|-Ce^{{\varphi}}(\lambda_{1}+K)\sum_{i}G^{i{\overline{i}}} \\
& +(F^{\frac{1}{2}})_{1{\overline{1}}}-G^{i{\overline{j}},k{\overline{l}}}\nabla_{1}{\tilde{g}}_{i{\overline{j}}}\nabla_{{\overline{1}}}{\tilde{g}}_{p{\overline{q}}}.
\end{split}$$ Hence, substituting this into (\[Lemma 2 equation 1\]) and using the Cauchy-Schwarz inequality, we obtain $$\label{Lemma 2 equation 4}
\begin{split}
G^{k{\overline{k}}}(\lambda_{1})_{k{\overline{k}}}
{\geqslant}{} & -C\sum_{k}G^{k{\overline{k}}}|{\varphi}_{1{\overline{1}}k}|-Ce^{{\varphi}}(\lambda_{1}+K)\sum_{i}G^{i{\overline{i}}} \\
& +(F^{\frac{1}{2}})_{1{\overline{1}}}-G^{i{\overline{j}},k{\overline{l}}}\nabla_{1}{\tilde{g}}_{i{\overline{j}}}\nabla_{{\overline{1}}}{\tilde{g}}_{p{\overline{q}}}.
\end{split}$$
Next, we deal with the term $(F^{\frac{1}{2}})_{1{\overline{1}}}$. Clearly, $$\label{Lemma 2 equation 2}
(F^{\frac{1}{2}})_{1{\overline{1}}} = \frac{1}{2F^{\frac{1}{2}}}\left(F_{1{\overline{1}}}-\frac{|F_{1}|^{2}}{2F}\right).$$ For the first term of (\[Lemma 2 equation 2\]), we compute $$\begin{split}
F_{1{\overline{1}}} = {} & 2n(n-1)|\alpha|(e^{{\varphi}}|{\partial}{\varphi}|_{g}^{2})_{1{\overline{1}}}+\frac{n(n-1)}{2}(e^{2{\varphi}})_{1{\overline{1}}}+\frac{n(n-1)}{2}f_{1{\overline{1}}} \\[1mm]
{\geqslant}{} & 2n(n-1)|\alpha|e^{{\varphi}}\sum_{k}\left(|{\varphi}_{k1}|^{2}+|{\varphi}_{k{\overline{1}}}|^{2}\right)-Ce^{{\varphi}}|{\partial}{\varphi}|_{g}\sum_{k}|{\varphi}_{k1{\overline{1}}}| \\
& -Ce^{{\varphi}}|{\partial}{\varphi}|_{g}^{2}|{\varphi}_{1{\overline{1}}}|-Ce^{{\varphi}}|{\partial}{\varphi}|_{g}^{4}
-Ce^{{\varphi}}|{\partial}{\varphi}|_{g}^{2}\sum_{k}\left(|{\varphi}_{k1}|+|{\varphi}_{k{\overline{1}}}|\right) \\
& -Ce^{2{\varphi}}|{\varphi}_{1{\overline{1}}}|+\frac{n(n-1)}{2}f_{1{\overline{1}}} \\
{\geqslant}{} & \frac{3}{2}n(n-1)|\alpha|e^{{\varphi}}\sum_{k}\left(|{\varphi}_{k1}|^{2}+|{\varphi}_{k{\overline{1}}}|^{2}\right)-Ce^{{\varphi}}(|{\partial}{\varphi}|_{g}+1)\sum_{k}|{\varphi}_{k1{\overline{1}}}| \\
& -Ce^{{\varphi}}|{\partial}{\varphi}|_{g}^{4}-Ce^{3{\varphi}}.
\end{split}$$ For the second term of (\[Lemma 2 equation 2\]), we have $$\begin{split}
F_{1} = {} & 2n(n-1)|\alpha|e^{{\varphi}}\nabla_{1}|{\partial}{\varphi}|_{g}^{2}+2n(n-1)|\alpha|e^{{\varphi}}|{\partial}{\varphi}|_{g}^{2}{\varphi}_{1} \\
& +n(n-1)e^{2{\varphi}}{\varphi}_{1}+\frac{n(n-1)}{2}f_{1},
\end{split}$$ which implies $$\begin{split}
|F_{1}|^{2} {\leqslant}{} & \left(4+\frac{1}{200}\right)n^{2}(n-1)^{2}\alpha^{2}e^{2{\varphi}}|{\partial}{\varphi}|_{g}^{2}\sum_{k}|{\varphi}_{k1}|^{2} \\
& +Ce^{2{\varphi}}|{\partial}{\varphi}|_{g}^{2}\sum_{k}|{\varphi}_{k{\overline{1}}}|^{2}
+Ce^{2{\varphi}}|{\partial}{\varphi}|_{g}^{6}+Ce^{4{\varphi}}|{\partial}{\varphi}|_{g}^{2}+C|f_{1}|^{2} \\
{\leqslant}{} & \left(4+\frac{1}{100}\right)n^{2}(n-1)^{2}\alpha^{2}e^{2{\varphi}}|{\partial}{\varphi}|_{g}^{2}\sum_{k}|{\varphi}_{k1}|^{2}
+Ce^{2{\varphi}}\sum_{k}|{\varphi}_{k1}|^{2} \\
& +Ce^{2{\varphi}}(|{\partial}{\varphi}|_{g}^{2}+1)\sum_{k}|{\varphi}_{k{\overline{1}}}|^{2}
+Ce^{2{\varphi}}|{\partial}{\varphi}|_{g}^{6}+Ce^{4{\varphi}}|{\partial}{\varphi}|_{g}^{2}+C. \\
\end{split}$$ On the other hand, by the definition of $F$, we have $$2F {\geqslant}\frac{99}{100}n(n-1)\left(e^{2{\varphi}}+4|\alpha|e^{{\varphi}}|{\partial}{\varphi}|_{g}^{2}\right).$$ Thus, substituting these estimates into (\[Lemma 2 equation 2\]), we get $$\label{Lemma 2 equation 3}
\begin{split}
& (F^{\frac{1}{2}})_{1{\overline{1}}} \\
{\geqslant}{} & -\frac{C}{F^{\frac{1}{2}}}\left(e^{{\varphi}}(|{\partial}{\varphi}|_{g}+1)\sum_{k}|{\varphi}_{k1{\overline{1}}}|+e^{{\varphi}}\sum_{k,l}|{\varphi}_{k{\overline{l}}}|^{2}
+e^{{\varphi}}|{\partial}{\varphi}|_{g}^{4}+e^{3{\varphi}}\right) \\
{\geqslant}{} & -Ce^{\frac{1}{2}{\varphi}}\sum_{k}|{\varphi}_{k1{\overline{1}}}|-Ce^{{\varphi}}F^{-\frac{1}{2}}\sum_{k,l}|{\varphi}_{k{\overline{l}}}|^{2}
-Ce^{\frac{1}{2}{\varphi}}|{\partial}{\varphi}|_{g}^{3}-Ce^{2{\varphi}}.
\end{split}$$ We note that a similar estimate of (\[Lemma 2 equation 3\]) was also appeared in [@PPZ16b].
Combining (\[Lemma 2 equation 4\]) and (\[Lemma 2 equation 3\]), we finally prove that $$\begin{split}
& \ G^{k{\overline{k}}}(\lambda_{1})_{k{\overline{k}}} \\[1mm]
{\geqslant}{} & -G^{i{\overline{j}},k{\overline{l}}}\nabla_{1}{\tilde{g}}_{i{\overline{j}}}\nabla_{{\overline{1}}}{\tilde{g}}_{p{\overline{q}}}
-C\sum_{k}G^{k{\overline{k}}}|{\varphi}_{1{\overline{1}}k}|-Ce^{\frac{1}{2}{\varphi}}\sum_{k}|{\varphi}_{k1{\overline{1}}}| \\
& -Ce^{{\varphi}}(\lambda_{1}+K)\sum_{i}G^{i{\overline{i}}}-Ce^{{\varphi}}F^{-\frac{1}{2}}\sum_{i,k}|{\varphi}_{i{\overline{k}}}|^{2}
-Ce^{\frac{1}{2}{\varphi}}|{\partial}{\varphi}|_{g}^{3}-Ce^{2{\varphi}} \\
{\geqslant}{} & -G^{i{\overline{j}},k{\overline{l}}}\nabla_{1}{\tilde{g}}_{i{\overline{j}}}\nabla_{{\overline{1}}}{\tilde{g}}_{p{\overline{q}}}
-C\sum_{k}G^{k{\overline{k}}}|{\varphi}_{1{\overline{1}}k}|-C_{A}\sum_{k}|{\varphi}_{k1{\overline{1}}}| \\
& -C_{A}(\lambda_{1}+K)\sum_{i}G^{i{\overline{i}}}-C_{A}K^{-\frac{1}{2}}\sum_{i,k}|{\varphi}_{i{\overline{k}}}|^{2}
-C_{A}K^{\frac{3}{2}}.
\end{split}$$
Using the above lemmas, we prove the following lower bound of $G^{k{\overline{k}}}Q_{k{\overline{k}}}$ at $x_{0}$.
At $x_{0}$, for any ${\varepsilon}\in(0,1)$, we have $$\label{Lower bound}
\begin{split}
0 {\geqslant}G^{k{\overline{k}}}Q_{k{\overline{k}}}
{\geqslant}{} & -\frac{G^{i{\overline{j}},k{\overline{l}}}\nabla_{1}{\tilde{g}}_{i{\overline{j}}}\nabla_{{\overline{1}}}{\tilde{g}}_{p{\overline{q}}}}{\lambda_{1}}
-(1+{\varepsilon})\frac{G^{k{\overline{k}}}|\nabla_{k}{\tilde{g}}_{1{\overline{1}}}|^{2}}{\lambda_{1}^{2}} \\
& +\frac{h'}{4}\sum_{i}G^{k{\overline{k}}}\left(|{\varphi}_{ik}|^{2}+|{\varphi}_{i{\overline{k}}}|^{2}\right)
+h''G^{k{\overline{k}}}|\nabla_{k}|{\partial}{\varphi}|_{g}^{2}|^{2} \\
& +B^{2}e^{B{\varphi}}G^{k{\overline{k}}}|{\varphi}_{k}|^{2}+\left(\frac{B}{2}e^{B{\varphi}}-\frac{C_{A}}{{\varepsilon}}\right)\sum_{k}G^{k{\overline{k}}}.
\end{split}$$
By , Lemma \[Lemma 1\] and Lemma \[Lemma 2\], we see that $$\label{Lower bound equation 1}
\begin{split}
& G^{k{\overline{k}}}Q_{k{\overline{k}}}(x_{0})\\
{\geqslant}{} & -\frac{G^{i{\overline{j}},k{\overline{l}}}\nabla_{1}{\tilde{g}}_{i{\overline{j}}}\nabla_{{\overline{1}}}{\tilde{g}}_{p{\overline{q}}}}{\lambda_{1}}
-\frac{G^{k{\overline{k}}}|\nabla_{k}{\tilde{g}}_{1{\overline{1}}}|^{2}}{\lambda_{1}^{2}}
-\frac{C\sum_{k}G^{k{\overline{k}}}|{\varphi}_{1{\overline{1}}k}|}{\lambda_{1}} \\
& -\frac{C_{A}\sum_{k}|{\varphi}_{k1{\overline{1}}}|}{\lambda_{1}}
+\frac{h'}{2}\sum_{i}G^{k{\overline{k}}}\left(|{\varphi}_{ik}|^{2}+|{\varphi}_{i{\overline{k}}}|^{2}\right)
-C_{A}K^{\frac{3}{2}}h'\\
& +h''G^{k{\overline{k}}}|\nabla_{k}|{\partial}{\varphi}|_{g}^{2}|^{2}
-C_{A}K^{\frac{1}{2}}h'\sum_{i,k}\left(|{\varphi}_{ik}|+|{\varphi}_{i{\overline{k}}}|\right)
-C_{A}K^{\frac{1}{2}} \\
& +B^{2}e^{B{\varphi}}G^{k{\overline{k}}}|{\varphi}_{k}|^{2}
+\left(Be^{B{\varphi}}-C_{A}\right)\sum_{k}G^{k{\overline{k}}}-CBe^{B{\varphi}}F^{\frac{1}{2}}.
\end{split}$$ On the other hand, by the definition of ${\tilde{\omega}}$, we have $$\label{the first order derivative of tildeomega}
2n\alpha{\varphi}_{1{\overline{1}}k} = \nabla_{k}{\tilde{g}}_{1{\overline{1}}}-e^{{\varphi}}{\varphi}_{k}-\alpha(e^{-{\varphi}}\rho_{1{\overline{1}}})_{k},$$ which implies $$\label{Lower bound equation 2}
\begin{split}
-\frac{C\sum_{k}G^{k{\overline{k}}}|{\varphi}_{1{\overline{1}}k}|}{\lambda_{1}}
{\geqslant}{} & -\frac{C}{\lambda_{1}}\sum_{k}G^{k{\overline{k}}}(|\nabla_{k}{\tilde{g}}_{1{\overline{1}}}|+e^{{\varphi}}K) \\
{\geqslant}{} & -{\varepsilon}\frac{G^{k{\overline{k}}}|\nabla_{k}{\tilde{g}}_{1{\overline{1}}}|^{2}}{\lambda_{1}^{2}}
-\frac{C_{A}}{{\varepsilon}}\sum_{k}G^{k{\overline{k}}},
\end{split}$$ where ${\varepsilon}\in(0,1)$. Note that $\nabla_{k}Q(x_{0})=0$. Then $$\label{dQ=0}
\frac{\nabla_{k}{\tilde{g}}_{1{\overline{1}}}}{\lambda_{1}} = -h'\nabla_{k}|{\partial}{\varphi}|_{g}^{2}-Be^{B{\varphi}}{\varphi}_{k}.$$ Thus combining this with (\[the first order derivative of tildeomega\]), it follows that $$-\frac{C_{A}\sum_{k}|{\varphi}_{1{\overline{1}}k}|}{\lambda_{1}}
{\geqslant}-C_{A}K^{\frac{1}{2}}h'\sum_{i,k}\left(|{\varphi}_{ik}|+|{\varphi}_{i{\overline{k}}}|\right)-C_{A}K^{\frac{1}{2}}Be^{B{\varphi}}.$$ Hence, substituting the above estimates into (\[Lower bound equation 1\]), we get $$\label{Lower bound equation 3}
\begin{split}
0 {\geqslant}{} & -\frac{G^{i{\overline{j}},k{\overline{l}}}\nabla_{1}{\tilde{g}}_{i{\overline{j}}}\nabla_{{\overline{1}}}{\tilde{g}}_{p{\overline{q}}}}{\lambda_{1}}
-(1+{\varepsilon})\frac{G^{k{\overline{k}}}|\nabla_{k}{\tilde{g}}_{1{\overline{1}}}|^{2}}{\lambda_{1}^{2}}
+h''G^{k{\overline{k}}}|\nabla_{k}|{\partial}{\varphi}|_{g}^{2}|^{2}\\
& +\frac{h'}{2}\sum_{i}G^{k{\overline{k}}}\left(|{\varphi}_{ik}|^{2}+|{\varphi}_{i{\overline{k}}}|^{2}\right)
-C_{A}K^{\frac{1}{2}}h'\sum_{i,k}\left(|{\varphi}_{ik}|+|{\varphi}_{i{\overline{k}}}|\right) \\
& +B^{2}e^{B{\varphi}}G^{k{\overline{k}}}|{\varphi}_{k}|^{2}+\left(Be^{B{\varphi}}-\frac{C_{A}}{{\varepsilon}}\right)\sum_{k}G^{k{\overline{k}}} \\
& -C_{A}K^{\frac{3}{2}}h'-C_{A}K^{\frac{1}{2}}
-CBe^{B{\varphi}}F^{\frac{1}{2}}-C_{A}K^{\frac{1}{2}}Be^{B{\varphi}}.
\end{split}$$ By the definitions of $G^{k{\overline{k}}}$, it is clear that $$\sum_{k}G^{k{\overline{k}}} =\frac{n-1}{2F^{\frac{1}{2}}}\lambda_{1}+(n-1)G^{1{\overline}{1}}{\geqslant}\frac{n-1}{2F^{\frac{1}{2}}}\lambda_{1}.$$ Combining this with and $G^{k{\overline}{k}}{\geqslant}\frac{F^{\frac{1}{2}}}{n\lambda_{1}}$, it follows that $$\begin{split}
C_{A}K^{\frac{1}{2}}h'\sum_{i,k}\left(|{\varphi}_{ik}|+|{\varphi}_{i{\overline{k}}}|\right)
{\leqslant}{} & \frac{h'}{10}\frac{F^{\frac{1}{2}}}{n\lambda_{1}}\sum_{i,k}\left(|{\varphi}_{ik}|^{2}+|{\varphi}_{i{\overline{k}}}|^{2}\right)
+C_{A}Kh'\frac{\lambda_{1}}{F^{\frac{1}{2}}} \\
{\leqslant}{} & \frac{h'}{10}\sum_{i}G^{k{\overline{k}}}\left(|{\varphi}_{ik}|^{2}+|{\varphi}_{i{\overline{k}}}|^{2}\right)+C_{A}\sum_{k}G^{k{\overline{k}}}.
\end{split}$$ By the definition of $F$ and $\lambda_{1}\gg C_{A}K$, we have $$C_{A}K^{\frac{3}{2}}h'+C_{A}K^{\frac{1}{2}}+CBe^{B{\varphi}}F^{\frac{1}{2}}+C_{A}K^{\frac{1}{2}}Be^{B{\varphi}}
{\leqslant}C_{A}K^{\frac{1}{2}}Be^{B{\varphi}}$$ and $$C_{A}K^{\frac{1}{2}}Be^{B{\varphi}}
{\leqslant}\frac{n-1}{4F^{\frac{1}{2}}}\lambda_{1}Be^{B{\varphi}}
{\leqslant}\frac{B}{2}e^{B{\varphi}}\sum_{k}G^{k{\overline{k}}}.$$ Substituting these estimates into (\[Lower bound equation 3\]), we obtain (\[Lower bound\]) immediately.
By (\[dQ=0\]) and , we have $$\label{GkkDQ=0}
\begin{split}
(1+{\varepsilon})\frac{G^{k{\overline{k}}}|\nabla_{k} {\tilde{g}}_{1{\overline}{1}}|^{2}}{\lambda_{1}^{2}}
= {} & (1+{\varepsilon})G^{k{\overline{k}}}|h'\nabla_{k}|{\partial}{\varphi}|^{2}_{g}+Be^{B{\varphi}}{\varphi}_{k}|^{2} \\
{\leqslant}{} & 2(h')^{2}G^{k{\overline}{k}}|\nabla_{k}|{\partial}{\varphi}|^{2}_{g}|^{2}+CB^{2}e^{2B{\varphi}}G^{k{\overline{k}}}|{\varphi}_{k}|^{2} \\
{\leqslant}{} & h''G^{k{\overline}{k}}|\nabla_{k}|{\partial}{\varphi}|_{g}^{2}|^{2}+CB^{2}e^{2B{\varphi}}G^{k{\overline{k}}}|{\varphi}_{k}|^{2}.
\end{split}$$ Since $G^{i{\overline{j}},k{\overline{l}}}\nabla_{1}{\tilde{g}}_{i{\overline{j}}}\nabla_{{\overline{1}}}{\tilde{g}}_{p{\overline{q}}}{\leqslant}0$, by (\[Lower bound\]) (taking ${\varepsilon}=\frac{1}{2}$), we get $$\label{Inequality for i>1}
\begin{split}
0 {\geqslant}{} & \frac{h'}{8}\sum_{i}G^{k{\overline{k}}}\left(|{\varphi}_{ik}|^{2}+|{\varphi}_{i{\overline{k}}}|^{2}\right)-C_{A}KB^{2}e^{2B{\varphi}}\sum_{k}G^{k{\overline{k}}} \\
{\geqslant}{} & \frac{1}{16K}\sum_{k}G^{k{\overline}{k}}|{\varphi}_{k{\overline}{k}}|^{2}-C_{A}B^{2}e^{2B{\varphi}}K\sum_{k}G^{k{\overline{k}}}.
\end{split}$$ On other hand hand, for the $\sigma_2$ function, we have (cf. [@LT94 Theorem 1])), $$G^{i{\overline}{i}}{\geqslant}\frac{1}{C}\sum_{k}G^{k{\overline}{k}}, ~\forall ~i{\geqslant}2.$$ Substituting this into (\[Inequality for i>1\]), it follows that $$0 {\geqslant}\frac{1}{16K}G^{i{\overline{i}}}|{\varphi}_{i{\overline{i}}}|^{2}-C_{A}B^{2}e^{2B{\varphi}}G^{i{\overline{i}}}.$$ Hence, we obtain $$\label{bounded of lambdak}
\lambda_{i} {\leqslant}C_{A}|{\varphi}_{i{\overline{i}}}|+C_{A} {\leqslant}C_{A,B}K, ~\forall i{\geqslant}2,$$ where $C_{A,B}$ is a uniform constant depending on $A$ and $B$.
By and for $k=1$, we have $$\label{Lower bound 1}
\begin{split}
0
{\geqslant}{} & -\frac{G^{i{\overline{j}},k{\overline{l}}}\nabla_{1}{\tilde{g}}_{i{\overline{j}}}\nabla_{{\overline{1}}}{\tilde{g}}_{k{\overline{l}}}}{\lambda_{1}}
-(1+{\varepsilon})\sum_{k{\geqslant}2}\frac{G^{k{\overline{k}}}|\nabla_{k}{\tilde{g}}_{1{\overline{1}}}|^{2}}{\lambda_{1}^{2}} \\
& +\frac{h'}{8}\sum_{i}G^{k{\overline{k}}}\left(|{\varphi}_{ik}|^{2}+|{\varphi}_{i{\overline{k}}}|^{2}\right)
+\sum_{k{\geqslant}2}h''G^{k{\overline{k}}}|\nabla_{k}|{\partial}{\varphi}|_{g}^{2}|^{2} \\
& +B^{2}e^{B{\varphi}}G^{k{\overline{k}}}|{\varphi}_{k}|^{2}+\left(\frac{B}{2}e^{B{\varphi}}-\frac{C_{A}}{{\varepsilon}}\right)\sum_{k}G^{k{\overline{k}}}
-CKB^2e^{2B{\varphi}}G^{1{\overline}{1}}.
\end{split}$$ We need to deal with bad third order term $$\label{Bad third order term}
\begin{split}
& (1+{\varepsilon})\sum_{k{\geqslant}2}\frac{G^{k{\overline{k}}}|\nabla_{k}{\tilde{g}}_{1{\overline{1}}}|^{2}}{\lambda_{1}^{2}} \\
= {} & (1-2{\varepsilon})\sum_{k{\geqslant}2}\frac{G^{k{\overline{k}}}|\nabla_{k}{\tilde{g}}_{1{\overline{1}}}|^{2}}{\lambda_{1}^{2}}
+3{\varepsilon}\sum_{k{\geqslant}2}\frac{G^{k{\overline{k}}}|\nabla_{k}{\tilde{g}}_{1{\overline{1}}}|^{2}}{\lambda_{1}^{2}}
\end{split}$$ For the first term of (\[Bad third order term\]), we use to see that $$\label{third 1}
\begin{split}
& (1-2{\varepsilon})\sum_{k{\geqslant}2}\frac{G^{k{\overline{k}}}|\nabla_{k}{\tilde{g}}_{1{\overline{1}}}|^{2}}{\lambda_{1}^{2}} \\
{\leqslant}{} & (1-{\varepsilon})\sum_{k{\geqslant}2}\frac{G^{k{\overline{k}}}|\nabla_{1}{\tilde{g}}_{k{\overline{1}}}|^{2}}{\lambda_{1}^{2}}+\frac{C}{{\varepsilon}}\sum_{k}G^{k{\overline{k}}} \\
{\leqslant}{} & (1-{\varepsilon})\sum_{k{\geqslant}2}\frac{\lambda_{1}+C_{A,B}K}{\lambda_{1}^{2}}|\nabla_{1}{\tilde{g}}_{k{\overline{1}}}|^{2}+\frac{C}{{\varepsilon}}\sum_{k}G^{k{\overline{k}}} \\
{\leqslant}{} & -\sum_{k{\geqslant}2}\frac{G^{1{\overline{k}},k{\overline{1}}}|\nabla_{1}{\tilde{g}}_{k{\overline{1}}}|^{2}}{\lambda_{1}}+\frac{C}{{\varepsilon}}\sum_{k}G^{k{\overline{k}}},
\end{split}$$ as long as $\lambda_{1}{\geqslant}\frac{C_{A,B}K}{{\varepsilon}}$. For the second term of (\[Bad third order term\]), we use (\[dQ=0\]) to get $$\label{third 2}
\begin{split}
& 3{\varepsilon}\sum_{k{\geqslant}2}\frac{G^{k{\overline{k}}}|\nabla_{k}{\tilde{g}}_{1{\overline{1}}}|^{2}}{\lambda_{1}^{2}}\\
{\leqslant}{} & 6{\varepsilon}(h')^{2}\sum_{k{\geqslant}2}G^{k{\overline{k}}}|\nabla_{k}|{\partial}{\varphi}|_{g}^{2}|^{2}+6{\varepsilon}B^{2}e^{2B{\varphi}}\sum_{k{\geqslant}2}G^{k{\overline}{k}}|{\varphi}_{k}|^{2}.
\end{split}$$ Thus substituting , and into , we obtain $$\begin{split}
0 {\geqslant}{} & \frac{h'}{8}\sum_{i}G^{k{\overline{k}}}\left(|{\varphi}_{ik}|^{2}+|{\varphi}_{i{\overline{k}}}|^{2}\right)
+\left(B^{2}e^{B{\varphi}}-6{\varepsilon}B^{2}e^{2B{\varphi}}\right)G^{k{\overline{k}}}|{\varphi}_{k}|^{2} \\
& +\left(\frac{B}{2}e^{B{\varphi}}-\frac{C_{A}}{{\varepsilon}}\right)\sum_{k}G^{k{\overline{k}}}-C_{A}KB^2e^{2B{\varphi}}G^{1{\overline}{1}}.
\end{split}$$ Choose $B=12C_{A}+1$ and ${\varepsilon}=\frac{e^{-B{\varphi}({x_{0}})}}{6}$, so that $$B^{2}e^{B{\varphi}}-6{\varepsilon}B^{2}e^{2B{\varphi}}=0 \text{~and~} \left(\frac{B}{2}e^{B{\varphi}}-\frac{C_{A}}{{\varepsilon}}\right)\sum_{k}G^{k{\overline{k}}} {\geqslant}0.$$ Hence, $$\frac{h'}{8}G^{1{\overline{1}}}|{\varphi}_{1{\overline{1}}}|^{2}{\leqslant}C_{A}KB^2e^{2B{\varphi}}G^{1{\overline}{1}},$$ which implies $\lambda_{1}{\leqslant}C_{A}K$. We complete the proof.
As a corollary of Proposition \[Effective second order estimate\], we obtain the following estimate.
\[A priori estimates\] Let $\alpha<0$ and ${\varphi}$ be a smooth solution of (\[Fu-Yau equation t\]) satisfying (\[Elliptic condition t\]) and (\[L\^n normalization condition\]). Then there exists a uniform constant $A_{0}$ such that if $A{\leqslant}A_{0}$, then we have the following estimate $$\|{\varphi}\|_{C^{k}} {\leqslant}C_{A,k},$$ where $C_{A,k}$ depends only on $A$, $k$, $\alpha$, $\rho$, $\mu$ and $(M,\omega)$.
By Proposition \[Effective second order estimate\] while $\rho$ and $\mu$ are replaced by $$t\rho \text{~and~} \frac{n\alpha(t-1)n!{\sqrt{-1} \partial \overline{\partial}}h\wedge{\sqrt{-1} \partial \overline{\partial}}h\wedge\omega^{n-2}}{\omega^{n}}+t\mu,$$ we use the blow-up argument to derive (cf. [@DK12; @Sze15]), $$\sup_{M}|{\partial}{\overline{\partial}}{\varphi}|_{g} {\leqslant}C_{A}.$$ By the $C^{2,\alpha}$-estimate (cf. [@TWWY15 Theorem 1.1]), it follows $$\|\varphi\|_{C^{2,\alpha}} {\leqslant}C_{A}'.$$ Hence, by the bootstrapping argument, we complete the proof.
Proof of Theorem \[Uniqueness theorem alpha negative\]
======================================================
In this section, we give the proof of Theorem \[Uniqueness theorem alpha negative\]. First we prove the uniqueness of solutions of (\[Fu-Yau equation t\]) in case of $\alpha<0$ when $t=0$.
\[Uniqueness t=0\] When $t=0$, (\[Fu-Yau equation t\]) has a unique solution satisfying (\[Elliptic condition t\]) and (\[L\^n normalization condition\]) $${\varphi}_{0} = h+\ln\|e^{-h}\|_{L^{n}}-\ln A,$$ where $h$ is the function in Lemma \[Existence lemma\].
It suffices to prove that ${\varphi}-h$ is constant. For convenience, we define $${\tilde{{\varphi}}} = {\varphi}-h \text{~and~} \omega_{h} = e^{h}\omega.$$ Then, when $t=0$, (\[Fu-Yau equation t\]) and (\[Elliptic condition t\]) can be expressed as $$\begin{split}
{\sqrt{-1} \partial \overline{\partial}}(e^{{\tilde{{\varphi}}}}\omega_{h})\wedge\omega^{n-2} & +n\alpha{\sqrt{-1} \partial \overline{\partial}}{\tilde{{\varphi}}}\wedge{\sqrt{-1} \partial \overline{\partial}}{\tilde{{\varphi}}}\wedge\omega^{n-2} \\
& +2n\alpha{\sqrt{-1} \partial \overline{\partial}}{\tilde{{\varphi}}}\wedge{\sqrt{-1} \partial \overline{\partial}}h\wedge\omega^{n-2} = 0
\end{split}$$ and $$e^{{\tilde{{\varphi}}}}\omega_{h}+2n\alpha{\sqrt{-1} \partial \overline{\partial}}({\tilde{{\varphi}}}+h) \in \Gamma_{2}(M).$$
By the similar calculation of Lemma \[Zero order estimate lemma\], we have $$\label{Uniqueness t=0 equation 1}
\begin{split}
& \int_{M}e^{2{\tilde{{\varphi}}}}\sqrt{-1}{\partial}{\tilde{{\varphi}}}\wedge{\overline{\partial}}{\tilde{{\varphi}}}\wedge\omega_{h}\wedge\omega^{n-2} \\
{\geqslant}{} & -2n\alpha\int_{M}\sqrt{-1}{\partial}e^{{\tilde{{\varphi}}}}\wedge{\overline{\partial}}{\tilde{{\varphi}}}\wedge{\sqrt{-1} \partial \overline{\partial}}({\tilde{{\varphi}}}+h)\wedge\omega^{n-2} \\
= {} & -2n\alpha\int_{M}e^{{\tilde{{\varphi}}}}\wedge{\overline{\partial}}{\tilde{{\varphi}}}\wedge{\sqrt{-1} \partial \overline{\partial}}({\tilde{{\varphi}}}+h)\wedge\sqrt{-1}{\partial}\omega^{n-2} \\
& +2n\alpha\int_{M}e^{{\tilde{{\varphi}}}}\wedge{\sqrt{-1} \partial \overline{\partial}}{\tilde{{\varphi}}}\wedge{\sqrt{-1} \partial \overline{\partial}}({\tilde{{\varphi}}}+h)\wedge\omega^{n-2} \\
= {} & -2n\alpha\int_{M}e^{{\tilde{{\varphi}}}}{\sqrt{-1} \partial \overline{\partial}}({\tilde{{\varphi}}}+h)\wedge{\sqrt{-1} \partial \overline{\partial}}\omega^{n-2} \\
& -2\int_{M}e^{{\tilde{{\varphi}}}}{\sqrt{-1} \partial \overline{\partial}}(e^{{\tilde{{\varphi}}}}\omega_{h})\wedge\omega^{n-2} \\
& -2n\alpha\int_{M}e^{{\tilde{{\varphi}}}}{\sqrt{-1} \partial \overline{\partial}}{\tilde{{\varphi}}}\wedge{\sqrt{-1} \partial \overline{\partial}}h\wedge\omega^{n-2}.
\end{split}$$ Since ${\sqrt{-1} \partial \overline{\partial}}\omega^{n-2}=0$, the first term of (\[Uniqueness t=0 equation 1\]) vanishes. For the second term, we compute $$\begin{split}
& -2\int_{M}e^{{\tilde{{\varphi}}}}{\sqrt{-1} \partial \overline{\partial}}(e^{{\tilde{{\varphi}}}}\omega_{h})\wedge\omega^{n-2} \\
= {} & 2\int_{M}e^{{\tilde{{\varphi}}}}\sqrt{-1}{\partial}{\tilde{{\varphi}}}\wedge{\overline{\partial}}(e^{{\tilde{{\varphi}}}}\omega_{h})\wedge\omega^{n-2}
-2\int_{M}e^{{\tilde{{\varphi}}}}{\overline{\partial}}(e^{{\tilde{{\varphi}}}}\omega_{h})\wedge\sqrt{-1}{\partial}\omega^{n-2} \\
= {} & 2\int_{M}(e^{2{\tilde{{\varphi}}}}\sqrt{-1}{\partial}{\tilde{{\varphi}}}\wedge{\overline{\partial}}{\tilde{{\varphi}}}\wedge\omega_{h}
+\frac{1}{2}\sqrt{-1}{\partial}e^{2{\tilde{{\varphi}}}}\wedge{\overline{\partial}}\omega_{h})\wedge\omega^{n-2} \\
& -\int_{M}{\overline{\partial}}e^{2{\tilde{{\varphi}}}}\wedge\omega_{h}\wedge\sqrt{-1}{\partial}\omega^{n-2}
-2\int_{M}e^{2{\tilde{{\varphi}}}}{\overline{\partial}}\omega_{h}\wedge\sqrt{-1}{\partial}\omega^{n-2} \\
= {} &\int_{M}e^{2{\tilde{{\varphi}}}}\left(-\partial({\overline}{\partial}\omega_{h}\wedge\omega^{n-1})
+{\overline}{\partial}(\omega_{h}\wedge{\partial}\omega^{n-1})-2{\overline}{\partial}\omega_{h}\wedge\partial\omega^{n-1}\right)\\
&+2\int_{M}e^{2{\tilde{{\varphi}}}}\sqrt{-1}{\partial}{\tilde{{\varphi}}}\wedge{\overline{\partial}}{\tilde{{\varphi}}}\wedge\omega_{h}\\
= {} & 2\int_{M}e^{2{\tilde{{\varphi}}}}\sqrt{-1}{\partial}{\tilde{{\varphi}}}\wedge{\overline{\partial}}{\tilde{{\varphi}}}\wedge\omega_{h}\wedge\omega^{n-2},
\end{split}$$ where we used the relations in the last equality, $${\sqrt{-1} \partial \overline{\partial}}\omega^{n-2}=0~{\rm and}~ {\sqrt{-1} \partial \overline{\partial}}(e^{h}\omega)\wedge\omega^{n-2}=0.$$ For the third term of (\[Uniqueness t=0 equation 1\]), we see that $$\begin{split}
& -2n\alpha\int_{M}e^{{\tilde{{\varphi}}}}{\sqrt{-1} \partial \overline{\partial}}{\tilde{{\varphi}}}\wedge{\sqrt{-1} \partial \overline{\partial}}h\wedge\omega^{n-2} \\
= {} & 2n\alpha\int_{M}e^{{\tilde{{\varphi}}}}\sqrt{-1}{\partial}{\tilde{{\varphi}}}\wedge{\overline{\partial}}{\tilde{{\varphi}}}\wedge{\sqrt{-1} \partial \overline{\partial}}h\wedge\omega^{n-2} \\
& -2n\alpha\int_{M}{\overline{\partial}}e^{{\tilde{{\varphi}}}}\wedge{\sqrt{-1} \partial \overline{\partial}}h\wedge\sqrt{-1}{\partial}\omega^{n-2} \\
= {} & 2n\alpha\int_{M}e^{{\tilde{{\varphi}}}}\sqrt{-1}{\partial}{\tilde{{\varphi}}}\wedge{\overline{\partial}}{\tilde{{\varphi}}}\wedge{\sqrt{-1} \partial \overline{\partial}}h\wedge\omega^{n-2}.
\end{split}$$ Substituting the above estimates into (\[Uniqueness t=0 equation 1\]), we get the inequality $$\label{Uniqueness t=0 equation 2}
\begin{split}
& \int_{M}e^{2{\tilde{{\varphi}}}}\sqrt{-1}{\partial}{\tilde{{\varphi}}}\wedge{\overline{\partial}}{\tilde{{\varphi}}}\wedge\omega_{h}\wedge\omega^{n-2} \\
{\leqslant}{} & -2n\alpha\int_{M}e^{{\tilde{{\varphi}}}}\sqrt{-1}{\partial}{\tilde{{\varphi}}}\wedge{\overline{\partial}}{\tilde{{\varphi}}}\wedge{\sqrt{-1} \partial \overline{\partial}}h\wedge\omega^{n-2} \\
{\leqslant}{} & C\int_{M}e^{{\tilde{{\varphi}}}}\sqrt{-1}{\partial}{\tilde{{\varphi}}}\wedge{\overline{\partial}}{\tilde{{\varphi}}}\wedge\omega_{h}\wedge\omega^{n-2}.
\end{split}$$ On the other hand, by Proposition \[Stronger zero order estimate\], we have $$\label{tilde-phi}
e^{{\tilde{{\varphi}}}} = e^{{\varphi}-h} {\geqslant}\frac{1}{CA}.$$ Combining this with (\[Uniqueness t=0 equation 2\]) and $A\ll1$, we prove $$\int_{M}e^{2{\tilde{{\varphi}}}}\sqrt{-1}{\partial}{\tilde{{\varphi}}}\wedge{\overline{\partial}}{\tilde{{\varphi}}}\wedge\omega_{h}\wedge\omega^{n-2} = 0,$$ which implies ${\partial}{\tilde{{\varphi}}}=0$. Therefore, ${\tilde{{\varphi}}}$ is constant.
\[unique-2-h\] Lemma \[Uniqueness t=0\] is also true from the above proof if (\[L\^n normalization condition\]) is replaced by the condition (\[two-condition\]), and the solution of (\[Fu-Yau equation t\]) when $t=0$ is given by ${\varphi}_{0} = h+\ln\|e^{-h}\|_{L^{1}}-\ln A$. The reason is that (\[tilde-phi\]) holds by Proposition \[Zero order estimate\]. In this case, the lemma holds for any $\alpha\neq0$.
Now we are in a position to prove Theorem \[Uniqueness theorem alpha negative\].
The existence is proved in Theorem \[Existence and Uniqueness Theorem\]. It suffices to prove the uniqueness. Assume that we have two solutions ${\varphi}$ and ${\varphi}'$. Then as in [@CHZ18 Section 5.3], we use the continuity method to solve (\[Fu-Yau equation t\]) from $t=1$ to $0$. By Theorem \[A priori estimates\], there are two families of solutions $\{{\varphi}_{t}\}$ and $\{{\varphi}_{t}'\}$ of (\[Fu-Yau equation t\]) satisfying (\[Elliptic condition t\]) and $L^{n}$-normalization conditions. We also have ${\varphi}_{1}={\varphi}$ and ${\varphi}_{1}'={\varphi}'$. By Lemma \[Uniqueness t=0\], we see that $${\varphi}_{0} = {\varphi}_{0}' = -\ln A.$$ Let $$J = \{ t\in [0,1] ~|~ {\varphi}_{t}={\varphi}_{t}'\}.$$ Clearly, $J$ is closed. Applying the implicit function theorem, we see that $J$ is open. Then $J=[0,1]$ and so $${\varphi}= {\varphi}_{1} = {\varphi}_{1}' = {\varphi}'.$$ This completes the proof of uniqueness.
Proof of Theorem \[Monotonicity theorem\]
=========================================
In this section, we give the proof of Theorem \[Monotonicity theorem\]. When $A$ is sufficiently small, Theorem \[Uniqueness theorem alpha negative\] implies that there exists a unique solution of (\[Fu-Yau equation t\]) satisfying (\[Elliptic condition t\]) and (\[L\^n normalization condition\]). For convenience, we denote it by ${\varphi}_{t,A}$.
\[Smooth lemma\] ${\varphi}_{t,A}$ is smooth with respect to $t$ and $A$.
For $\beta\in(0,1)$, we define the sets $$\begin{split}
{\tilde{B}} & = C^{2,\beta}(M)\times[0,1]\times[0,1],\\[2mm]
{\tilde{B}}_{1} & = \{ ({\varphi},t,A)\in {\tilde{B}} ~|~ \text{${\varphi}$ satisfies (\ref{Elliptic condition t})} \}, \\
{\tilde{B}}_{2} & = \textbf{R}\times\{ u\in C^{\beta}(M) ~|~ \int_{M}u\omega^{n}=0 \},
\end{split}$$ and map $\Phi: {\tilde{B}}_{1}\rightarrow {\tilde{B}}_{2}$ $$\Phi({\varphi},t,A) = \left(\Phi_{1}({\varphi},A),\Phi_{2}({\varphi},t)\right),$$ where $$\Phi_{1}({\varphi},A) = \int_{M}e^{-n{\varphi}}\omega^{n}-A^{n}$$ and $$\begin{split}
\Phi_{2}({\varphi},t) = {} & {\sqrt{-1} \partial \overline{\partial}}(e^{{\varphi}}\omega-t\alpha e^{-{\varphi}}\rho)\wedge\omega^{n-2} \\[2mm]
& +n\alpha{\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge{\sqrt{-1} \partial \overline{\partial}}{\varphi}\wedge\omega^{n-2} \\
& +n\alpha(t-1){\sqrt{-1} \partial \overline{\partial}}h\wedge{\sqrt{-1} \partial \overline{\partial}}h\wedge\omega^{n-2}+t\mu\frac{\omega^{n}}{n!}.
\end{split}$$
By the same argument of Section \[Proof of existence theorem\], we have $$(D_{{\varphi}}\Phi)_{(\hat{{\varphi}},\hat{t},\hat{A})}(u) = \left(-\int_{M}ne^{-n\hat{{\varphi}}}u\omega^{n},Lu\right),$$ where $$\label{Definition of L}
\begin{split}
(Lu)\omega^{n} = {} & {\sqrt{-1} \partial \overline{\partial}}(ue^{\hat{{\varphi}}}\omega+\hat{t}\alpha ue^{-\hat{{\varphi}}}\rho)\wedge\omega^{n-2} \\
& +2n\alpha{\sqrt{-1} \partial \overline{\partial}}\hat{{\varphi}}\wedge{\sqrt{-1} \partial \overline{\partial}}u\wedge\omega^{n-2}.
\end{split}$$ and $$\label{Kernel of L}
\textrm{Ker}L = \{ cu_{0} ~|~ c\in\mathbf{R} \},$$ where $u_{0}$ is a positive function. Similarly, we see that $(D_{{\varphi}}\Phi)_{(\hat{{\varphi}},\hat{t},\hat{A})}$ is invertible. Using the implicit function theorem, near $(\hat{{\varphi}},\hat{t},\hat{A})$, there exists a smooth map $F(t,A)$ such that $\Phi(F(t,A),t,A)=(0,0)$, where ${\varphi}_{\hat{t},\hat{A}}$ (for convenience, we denote it by $\hat{{\varphi}}$) satisfies $$\Phi(\hat{{\varphi}},\hat{t},\hat{A}) = (0,0).$$ This implies that $F(t,A)$ is the solution of (\[Fu-Yau equation t\]) satisfying the elliptic and $L^{n}$-normalization conditions. Thanks to Theorem \[Uniqueness theorem alpha negative\], we have ${\varphi}_{t,A}=F(t,A)$. Hence, ${\varphi}_{t,A}$ is smooth at $(\hat{t},\hat{A})$. Since $(\hat{t},\hat{A})$ is arbitrary, we complete the proof.
By the definition of ${\varphi}_{t,A}$, it suffices that prove ${\varphi}_{1,A}>{\varphi}_{1,{\tilde{A}}}$. Define ${\varphi}(s)={\varphi}_{1,A^{s}{\tilde{A}}^{1-s}}$. Then it follows that $$\begin{split}
{\sqrt{-1} \partial \overline{\partial}}(e^{{\varphi}(s)}\omega & -t\alpha e^{-{\varphi}(s)}\rho)\wedge\omega^{n-2}+n\alpha{\sqrt{-1} \partial \overline{\partial}}{\varphi}(s)\wedge{\sqrt{-1} \partial \overline{\partial}}{\varphi}(s)\wedge\omega^{n-2} \\
& +n\alpha(t-1){\sqrt{-1} \partial \overline{\partial}}h\wedge{\sqrt{-1} \partial \overline{\partial}}h\wedge\omega^{n-2}+t\mu\frac{\omega^{n}}{n!} = 0,
\end{split}$$ and $$\int_{M}e^{-n{\varphi}(s)}\omega^{n} = A^{ns}{\tilde{A}}^{n(1-s)}.$$ By Lemma \[Smooth lemma\], we can differentiate the above two equations with respect to $s$, respectively, and we obtain $$\frac{{\partial}{\varphi}(s)}{{\partial}s} \in \textrm{Ker}L \text{~and~} \int_{M}e^{-n{\varphi}(s)}\frac{{\partial}{\varphi}(s)}{{\partial}s}\omega^{n} > 0.$$ Recalling (\[Kernel of L\]), we see that $\frac{{\partial}{\varphi}(s)}{{\partial}s} > 0$, which implies $${\varphi}_{1,A}-{\varphi}_{1,{\tilde{A}}} = \int_{0}^{1}\frac{{\partial}{\varphi}(s)}{{\partial}s}ds > 0.$$ Theorem \[Monotonicity theorem\] is proved.
\[Special case remark\] When $\textrm{tr}_{\omega}\rho{\geqslant}0$, using (\[Zero order estimate equation 2\]) and Sobolev inequality, we obtain $$\left(\int_{M}e^{-k\beta{\varphi}}\omega^{n}\right)^{\frac{1}{\beta}}
{\leqslant}Ck\int_{M}e^{-(k-1){\varphi}}\omega^{n},$$ which implies $\|e^{-{\varphi}}\|_{L^{\infty}}{\leqslant}C$. By the similar argument in Section 5, we get the analogous estimate of Theorem \[A priori estimates\] under the normalization $\|e^{-{\varphi}}\|_{L^{1}}=A$. This estimate is enough for the proofs of Theorem \[Uniqueness theorem alpha negative\] and \[Monotonicity theorem\].
[99]{}
*J. Chu, L. Huang* and *X. Zhu,* [The $2$-nd Hessian type equation on almost Hermitian manifolds,]{} preprint, arXiv: 1707.04072.
*J. Chu, L. Huang* and *X. Zhu,* [The Fu-Yau equation in higher dimensions,]{} preprint, arXiv: 1801.09351.
*J. Chu, V. Tosatti* and *B. Weinkove,* [The Monge-Ampère equation for non-integrable almost complex structures,]{} to appear in J. Eur. Math. Soc. (JEMS)
*S. Dinew* and *S. Ko[ł]{}odziej,* [Liouville and Calabi-Yau type theorems for complex Hessian equations,]{} Amer. J. Math. **139** (2017), no. 2, 403–415.
*A. Fino, G. Grantcharov* and *L. Vezzoni,* [Astheno-Kähler and balanced structures on fibrations,]{} preprint, arXiv: 1608.06743.
*A. Fino* and *A. Tomassini,* [On astheno-Kähler metrics,]{} J. London Math. Soc. **83** (2011), 290–308.
*J.-X. Fu* and *S.-T. Yau,* [A Monge-Ampère-type equation motivated by string theory,]{} Comm. Anal. Geom. **15** (2007), no. 1, 29–75.
*J.-X. Fu* and *S.-T. Yau,* [The theory of superstring with flux on non-Kähler manifolds and the complex Monge-Ampère equation,]{} J. Differential Geom. **78** (2008), no. 3, 369–428.
*M. Garcia-Fernandez,* [Lectures on the Strominger system,]{} preprint, arXiv: 1609.02615.
*E. Goldstein* and *S. Prokushkin*, [Geometric model for complex non-Kähler manifolds with $SU(3)$ structure,]{} Comm. Math. Phys. **251** (2004), no. 1, 65–78.
*P. Guan, C. Ren* and *Z. Wang,* [Global $C^{2}$-estimates for convex solutions of curvature equations,]{} Comm. Pure Appl. Math. **68** (2015), no. 8, 1287–1325.
*Z. Hou, X.-N. Ma* and *D. Wu,* [A second order estimate for complex Hessian equations on a compact Kähler manifold,]{} Math. Res. Lett. **17** (2010), no. 3, 547–561.
*J. Jost* and *S.-T. Yau,* [A nonlinear elliptic system for maps from Hermitian to Riemannian manifolds and rigidity theorems in Hermitian geometry]{}, Acta Math. **170** (1993), no. 2, 221–254; Correction, Acta Math. **173** (1994), no. 2, 307.
*M. Lin* and *N. S. Trudinger,* [On some inequalities for elementary symmetric functions,]{} Bull. Aust. Math. Soc. **50** (1994), 317–326.
*A. Latorre* and *L. Ugarte,* [On non-Kähler compact complex manifolds with balanced and astheno-Kähler metrics,]{} Com. Ren. Acad. Sci. Math. **355** (2017), 90–93.
*J. Li* and *S.-T. Yau,* [Hermitian-Yang-Mills connection on non-Kähler manifolds,]{} Mathematical aspects of string theory (San Diego, Calif., 1986), 560–573, Adv. Ser. Math. Phys., 1, World Sci. Publishing, Singapore, 1987.
*J. Li, S.-T. Yau* and *F. Zheng,* [On projectively flat Hermitian manifolds,]{} Comm. Anal. Geom. **2** (1994), 103–109.
*K. Matsuo,* [Astheno-Kähler structures on Calabi-Eckmann manifolds]{}, Colloq. Math. **115** (2009), no. 1, 33–39.
*K. Matsuo* and *T. Takahashi,* [On compact astheno-Kähler manifolds]{}, Colloq. Math. **89** (2001), no. 2, 213–221.
*D. H. Phong, S. Picard* and *X. Zhang,* [On estimates for the Fu-Yau generalization of a Strominger system,]{} to appear in J. Reine Angew. Math.
*D. H. Phong, S. Picard* and *X. Zhang,* [A second order estimate for general complex Hessian equations,]{} Anal. PDE **9** (2016), no. 7, 1693–1709.
*D. H. Phong, S. Picard* and *X. Zhang,* [The Fu-Yau equation with negative slope parameter,]{} Invent. Math. **209** (2017), no. 2, 541–576.
*D. H. Phong, S. Picard* and *X. Zhang,* [Fu-Yau Hessian equations,]{} preprint, arXiv: 1801.09842.
*Y. T. Siu*, [The complex analyticity of harmonic maps and the strong rigidity of compact Kähler manifolds]{}. Ann. of Math. **112** (1980), 73–111.
*A. Strominger,* [Superstrings with torsion,]{} Nuclear Phys. B **274** (1986), no. 2, 253–284.
*G. Székelyhidi,* [Fully non-linear elliptic equations on compact Hermitian manifolds,]{} to appear in J. Differential Geom.
*G. Székelyhidi, V. Tosatti* and *B. Weinkove,* [Gauduchon metrics with prescribed volume form]{}, Acta Math. **219** (2017), no. 1, 181–211.
*V. Tosatti, Y. Wang, B. Weinkove* and *X. Yang,* [$C^{2,\alpha}$ estimate for nonlinear elliptic equations in complex and almost complex geometry,]{} Calc. Var. Partial Differential Equations **54** (2015), no. 1, 431–453.
*V. Tosatti* and *B. Weinkove* [The Monge-Amp¨¨re equation for $(n-1)$-plurisubharmonic functions on a compact Kähler manifold]{}, J. Amer. Math. Soc. **30** (2017), no. 2, 311–346.
*M. Warren* and *Y. Yuan,* [Hessian estimates for the sigma-2 equation in dimension 3]{}, Comm. Pure Appl. Math. **62** (2009), no. 3, 305–321.
[^1]: Phong, Pacard and Zhang posted a paper [@PPZ18] with a similar result after we posted the paper in arXiv.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this work, we propose a subspace-based algorithm for DOA estimation which iteratively reduces the disturbance factors of the estimated data covariance matrix and incorporates prior knowledge which is gradually obtained on line. An analysis of the MSE of the reshaped data covariance matrix is carried out along with comparisons between computational complexities of the proposed and existing algorithms. Simulations focusing on closely-spaced sources, where they are uncorrelated and correlated, illustrate the improvements achieved.'
author:
- |
Silvio F. B. Pinto $^1$ and Rodrigo C. de Lamare $^{1,2}$\
Center for Telecommunications Studies (CETUC)\
$^1$ Pontifical Catholic University of Rio de Janeiro, RJ, Brazil.\
$^2$ Department of Electronics, University of York, UK\
Emails: [email protected], [email protected]
title: |
Multi-Step Knowledge-Aided Iterative ESPRIT\
for Direction Finding\
---
Introduction
============
In array signal processing, direction-of-arrival (DOA) estimation is a key task in a broad range of important applications including radar and sonar systems, wireless communications and seismology [@Vantrees]. Traditional high-resolution methods for DOA estimation such as the multiple signal classification (MUSIC) method [@schimdt], the root-MUSIC algorithm [@Barabell], the estimation of signal parameters via rotational invariance techniques (ESPRIT) [@Roy] and subspace techniques ,[@Steinwandt; @Wang; @Qiu] exploit the eigenstructure of the input data matrix. These techniques may fail for reduced data sets or low signal-to-noise ratio (SNR) levels where the expected estimation error is not asymptotic to the Cramér-Rao bound (CRB) [@Thomas]. The accuracy of the estimates of the covariance matrix is of fundamental importance in parameter estimation. Low levels of SNR or short data records can result in significant divergences between the true and the sample data covariance matrices. In practice, only a modest number of data snapshots is available and when the number of snapshots is similar to the number of sensor array elements, the estimated and the true subspaces can differ significantly. Several approaches have been developed with the aim of enhancing the computation of the covariance matrix [@Carlson]-[@Qian] and for dealing with large sensor-array systems large [@mmimo; @wence; @Costa; @delamare_ieeproc; @TDS_clarke; @TDS_2; @switch_int; @switch_mc; @smce; @TongW; @jpais_iet; @TARMO; @badstbc; @baplnc; @keke1; @kekecl; @keke2; @wlbd; @Tomlinson; @dopeg_cl; @peg_bf_iswcs; @gqcpeg; @peg_bf_cl; @Harashima; @mbthpc; @zuthp; @rmbthp; @Hochwald; @BDVP],[@delamare_mber; @rontogiannis; @delamare_itic; @stspadf; @choi; @stbcccm; @FL11; @jio_mimo; @peng_twc; @spa; @spa2; @jio_mimo; @P.Li; @jingjing; @memd; @did; @bfidd; @mbdf; @bfidd; @mserrr; @shaowcl08].
Diagonal loading [@Carlson] and shrinkage [@Chen; @ruan1; @ruan2] techniques can enhance the estimate of the data covariance matrix by weighing and individually increasing its diagonal by a real constant. Nevertheless, the eigenvectors remain the same, which leads to unaltered estimates of the signal and noise projection matrices obtained from the enhanced covariance matrix. Additionally, an improvement of the estimates of the covariance matrix can be achieved by employing forward/backward averaging and spatial smoothing approaches [@Pillai; @Evans]. The former leads to twice the number of the original samples and its corresponding enhancement. The latter extracts the array covariance matrix as the average of all covariance matrices from its sub-arrays, resulting in a greater number of samples. Both techniques are employed in signal decorrelation. An approach to improve MUSIC dealing with the condition in which the number of snapshots and the sensor elements approach infinity was presented in [@Mestre]. Nevertheless, this technique is not that effective for reduced number of snapshots. Other approaches to deal with reduced data sets or low SNR levels [@Gershman; @Qian] consist of reiterating the procedure of adding pseudo-noise to the observations which results in new estimates of the covariance matrix. Then, the set of solutions is computed from previously stored DOA estimates. In [@Vorobyov2], two aspects resulting from the computation of DOAs for reduced data sets or low SNR levels have been studied using the root-MUSIC technique. The first aspect dealt with the probability of estimated signal roots taking a smaller magnitude than the estimated noise roots, which is an anomaly that leads to wrong choices of the closest roots to the unit circle. To mitigate this problem, different groups of roots are considered as potential solutions for the signal sources and the most likely one is selected [@Stoica]. The second aspect previously mentioned, shown in [@Johnson], refers to the fact that a reduced part of the true signal eigenvectors exists in the sample noise subspace (and vice-versa). Such coexistence has been expressed by a Frobenius norm of the related irregularity matrix and introduced its mathematical foundation. An iterative technique to enhance the efficacy of root-MUSIC by reducing this anomaly making use of the gradual reshaping of the sample data covariance matrix has been reported. Inspired by the work in [@Vorobyov2], we have developed an ESPRIT-based method known as Two-Step KAI-ESPRIT (TS-ESPRIT) [@Pinto], which combines that modifications of the sample data covariance matrix with the use of prior knowledge [@Guerci1]-[@Guerci2] about the covariance matrix of a set of impinging signals to enhance the estimation accuracy in the finite sample size region. In practice, this prior knowledge could be from the signals coming from known base stations or from static users in a system. TS-ESPRIT determines the value of a correction factor that reduces the undesirable terms in the estimation of the signal and noise subspaces in an iterative process, resulting in better estimates.
In this work [@Pinto2; @Pinto3], we present the Multi-Step KAI ESPRIT (MS-KAI-ESPRIT) approach that refines the covariance matrix of the input data via multiple steps of reduction of its undesirable terms. This work presents the MS-KAI-ESPRIT in further detail, an analysis of the mean squared error (MSE) of the data covariance matrix free of undesired terms (side effects), a more accurate study of the computational complexity and a comprehensive study of MS-KAI-ESPRIT and other competing techniques for scenarios with both uncorrelated and correlated signals. Unlike TS-ESPRIT, which makes use of only one iteration and available known DOAs, MS-KAI-ESPRIT employs multiple iterations and obtains prior knowledge on line. At each iteration of MS-KAI-ESPRIT, the initial Vandermonde matrix is updated by replacing an increasing number of steering vectors of initial estimates with their corresponding refined versions. In other words, at each iteration, the knowledge obtained on line is updated, allowing the direction finding algorithm to correct the sample covariance matrix estimate, which yields more accurate estimates.
In summary, this work has the following contributions:
- The proposed MS-KAI-ESPRIT technique.
- An MSE analysis of the covariance matrix obtained with the proposed MS-KAI-ESPRIT algorithm.
- A comprehensive performance study of MS-KAI-ESPRIT and competing techniques.
This paper is organized as follows. Section II describes the system model. Section III presents the proposed MS-KAI-ESPRIT algorithm. In section IV, an analytical study of the MSE of the data covariance matrix free of side-effects is carried out together with a study of the computational complexity of the proposed and competing algorithms. In Section V, we present and discuss the simulation results. Section VI concludes the paper.
System Model {#sysmodel}
=============
Let us assume that *P* narrowband signals from far-field sources impinge on a uniform linear array (ULA) of $M\ (M
> \textit{P})$ sensor elements from directions ${\boldsymbol \theta}=[\theta_{1},\theta_{2},\ldots, \theta_P]^T$. We also consider that the sensors are spaced from each other by a distance $ d\leq\frac{\lambda_{c}}{2}$, where $\lambda_{c}$ is the signal wavelength, and that without loss of generality, we have ${\frac{-\pi}{2}\leq\theta_{1}\leq\theta_{2}\leq\ldots
\leq\theta_P\leq \frac{\pi}{2}}$.
The $i$th data snapshot of the $M$-dimensional array output vector can be modeled as $$\bm x(i)=\bm A\,s(i)+\bm n(i),\qquad i=1,2,\ldots,N,
\label{model}$$ where $\bm s(i)=[s_{1}(i),\ldots,s_{P}(i)]^T
\in\mathbb{C}^{\mathit{P\times1}}$ represents the zero-mean source data vector, $\bm n(i) \in\mathbb{C}^{\mathit{M \times 1}}$ is the vector of white circular complex Gaussian noise with zero mean and variance $\sigma_n^2$, and $N$ denotes the number of available snapshots.
The Vandermonde matrix $\bm A(\bm \Theta)=[\bm
a(\theta_{1}),\ldots,\bm a(\theta_{P})] \in\mathbb
{C}^{\mathit{M\times P}}$, known as the array manifold, contains the array steering vectors $\bm a(\theta_j)$ corresponding to the $n$th source, which can be expressed as $$\bm a(\theta_n)=[1,e^{j2\pi\frac{d}{\lambda_{c}}
\sin\theta_n},\ldots,e^{j2\pi(M-1)\frac{d}{\lambda_{c}}\sin\theta_n}]^T,
\label{steer}$$ where $n=1,\ldots, P$. Using the fact that $\bm s(i)$ and $\bm n(i)$ are modeled as uncorrelated linearly independent variables, the $M\times M$ signal covariance matrix is calculated by $$\bm R=\mathbb E\left[\bm x(i)\bm x^H(i)
\right]=\bm A\,\bm R_{ss}\bm A^H+
\sigma_n^2\bm I_M,
\label{covariance}$$ where the superscript *H* and $\mathbb E[\cdot]$ in $\bm
R_{ss}=\mathbb E[\bm s(i)\bm s^H(i)]$ and in $\mathbb E[\bm n(i)\bm
n^H(i)]=\sigma_n^2\bm I_M^{}$ denote the Hermitian transposition and the expectation operator and $\bm I_M$ stands for the $M$-dimensional identity matrix. Since the true signal covariance matrix is unknown, it must be estimated and a widely-adopted approach is the sample average formula given by $$\bm {\hat{R}}=\frac{1}{N} \sum\limits^{N}_{i=1}\bm x(i)\bm x^H(i),
\label{covsample}$$ whose estimation accuracy is dependent on $N$.
Proposed MS-KAI-ESPRIT Algorithm
=================================
In this section, we present the proposed MS-KAI-ESPRIT algorithm and detail its main features. We start by expanding using as derived in [@Vorobyov2]: $$\begin{aligned}
\bm {\hat{R}}=\frac{1}{N} \sum\limits^{N}_{i=1}(\bm A\,s(i)+\bm
n(i))\:(\bm A\,s(i)+\bm n(i))^H \nonumber\\= \bm
A\left\lbrace\frac{1}{N} \sum\limits^{N}_{i=1}\bm s(i)\bm
s^H(i)\right\rbrace\bm A^H+\:\frac{1}{N} \sum\limits^{N}_{i=1}\bm
n(i)\bm n^H(i)\;+\nonumber\\\underbrace{\bm A\left\lbrace\frac{1}{N}
\sum\limits^{N}_{i=1}\bm s(i)\bm n^H(i)\right\rbrace\:
+\:\left\lbrace\frac{1}{N} \sum\limits^{N}_{i=1}\bm n(i)\bm
s^H(i)\right\rbrace\bm{A}^{H}}_{{"undesirable terms"}}
\label{expandedcovsample}\end{aligned}$$ The first two terms of in can be considered as estimates of the two summands of given in , which represent the signal and the noise components, respectively. The last two terms in are undesirable side effects, which can be seen as estimates for the correlation between the signal and the noise vectors. The system model under study is based on noise vectors which are zero-mean and also independent of the signal vectors. Thus, the signal and noise components are uncorrelated to each other. As a consequence, for a large enough number of samples $N$, the last two terms of tend to zero. Nevertheless, in practice the number of available samples can be limited. In such situations, the last two terms in may have significant values, which causes the deviation of the estimates of the signal and the noise subspaces from the true signal and noise subspaces.
The key point of the proposed MS-KAI-ESPRIT algorithm is to modify the sample data covariance matrix estimate at each iteration by gradually incorporating the knowledge provided by the newer Vandermonde matrices which progressively embody the refined estimates from the preceding iteration. Based on these updated Vandermonde matrices, refined estimates of the projection matrices of the signal and noise subspaces are calculated. These estimates of projection matrices associated with the initial sample covariance matrix estimate and the reliability factor are employed to reduce its side effects and allow the algorithm to choose the set of estimates that has the highest likelihood of being the set of the true DOAs. The modified covariance matrix is computed by computing a scaled version of the undesirable terms of $\bm {\hat{R}}$, as pointed out in .
The steps of the proposed algorithm are listed in Table \[Multi\_Step\_KAI\]. The algorithm starts by computing the sample data covariance matrix . Next, the DOAs are estimated using the ESPRIT algorithm. The superscript $(\cdot)^{(1)}$ refers to the estimation task performed in the first step. Now, a procedure consisting of $n=1:P$ iterations starts by forming the Vandermonde matrix using the DOA estimates. Then, the amplitudes of the sources are estimated such that the square norm of the differences between the observation vector and the vector containing estimates and the available known DOAs is minimized. This problem can be formulated [@Vorobyov2] as: $$\begin{aligned}
\hat{\bm{s}}(i)=\arg\min_{\substack{\bm
s}}\parallel\bm{x}(i)-\hat{\bm{A}}\mathbf{s}\parallel^2_2.
\label{minimization1}\end{aligned}$$ The minimization of is achieved using the least squares technique and the solution is described by $$\hat{\bm{s}}(i)=(\mathbf{\hat{A}}^{H}\:\mathbf{\hat{A}})^{-1}\:\mathbf{\hat{A}}\:\bm{x}(i)
\label{minimization2}$$ The noise component is then estimated as the difference between the estimated signal and the observations made by the array, as given by $$\begin{aligned}
\hat{\bm n}(i)=\bm x(i)\:-\: \hat{\bm A}\:\hat{\bm s}(i).
\label{noise_component}\end{aligned}$$ After estimating the signal and noise vectors, the third term in can be computed as: $$\begin{aligned}
\bm{V}&\triangleq \hat{\bm{A}}\left\lbrace\frac{1}{N}
\sum\limits^{N}_{i=1}\bm \hat{\mathbf{s}}(i)\bm
\hat{\mathbf{n}}^H(i)\right\rbrace\nonumber\\&=\hat{\bm{A}}\left\lbrace\frac{1}{N}
\sum\limits^{N}_{i=1}(\mathbf{\hat{A}}^{H}\:\mathbf{\hat{A}})^{-1}\mathbf{\hat{A}}^{H}\bm{x}(i)\right.\nonumber\\&\left.\times(\bm{x}^{H}(i)-\bm{x}^{H}(i)\hat{\mathbf{A}}(\hat{\mathbf{A}}^{H}\hat{\mathbf{A}})^{-1}\:\hat{\mathbf{A}}^{H})\right\rbrace\nonumber\\&=\mathbf{\hat{Q}}_{A}\left\lbrace\frac{1}{N}
\sum\limits^{N}_{i=1}
\bm{x}(i)\bm{x}^H(i)\:\left(\mathbf{I}_{M}\:-\:\hat{\mathbf{Q}}_{A}\right)
\right\rbrace\nonumber\\&=\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}^{\perp},
\label{terms_deducted}\end{aligned}$$ where $$\mathbf{\hat{Q}}_{A}\triangleq \mathbf{\hat{A}}\:(\mathbf{\hat{A}}^{H}\:\mathbf{\hat{A}})^{-1}\:\mathbf{\hat{A}}^{H}
\label{signal_projection}$$ is an estimate of the projection matrix of the signal subspace, and $$\mathbf{\hat{Q}}_{A}^{\perp}\triangleq\mathbf{I}_{M}\:-\:\mathbf{\hat{Q}}_{A}
\label{noise_projection}$$ is an estimate of the projection matrix of the noise subspace.
Next, as part of the procedure consisting of $n=1:P $ iterations, the modified data covariance matrix $\mathbf{\hat{R}}^{(n+1)}$ is obtained by computing a scaled version of the estimated terms from the initial sample data covariance matrix as given by $$\label{modified_data_covariance}
\mathbf{\hat{R}}^{(n+1)} = \mathbf{\hat{R}}\:-\:\mathrm{\mu}\:(\mathbf{V}^{(n)}\:+\:\mathbf{V}^{(n)H}),$$ where the superscript $(\cdot)^{(n)}$ refers to the $n^{th} $ iteration performed. The scaling or reliability factor increases from 0 to 1 incrementally, resulting in modified data covariance matrices. Each of them gives origin to new estimated DOAs also denoted by the superscript $(\cdot)^{(n+1)}$ by using the ESPRIT algorithm, as briefly described ahead.
In this work, the rank *P* is assumed to be known, which is an assumption frequently found in the literature. Alternatively, the rank *P* could be estimated by model-order selection schemes such as Akaike´s Information Theoretic Criterion (AIC) [@Schell] and the Minimum Descriptive Length (MDL) Criterion [@Rissanen].
In order to estimate the signal and the orthogonal subspaces from the data records, we may consider two approaches [@Vaccaro; @Haardt]: the direct data approach and the covariance approach. The direct data approach makes use of singular value decomposition(SVD) of the data matrix $\mathbf{X}$, composed of the $i$th data snapshot of the $M$-dimensional array data vector: $$\begin{aligned}
\bm X=&[\bm
{x}(1),\bm{x}(2),\ldots,\bm{x}(N)]\nonumber\\=&\bm A[\bm
{s}(1),\bm{s}(2),\ldots,\bm{s}(N)]+[\bm
{n}(1),\bm
{n}(2),\ldots,\bm{n}(N)]\nonumber\\=&\bm{A(\Theta)\;S}+\:\bm{N}\;\in\mathbb{C}^{\mathit{M \times N}}
\label{Data_matrix}\end{aligned}$$
Since the number of the sources is assumed known or can be estimated by AIC[@Schell] or MDL[@Rissanen] , as previously mentioned, we can write $\mathbf{X}$ as:
$$\begin{aligned}
\begin{array}{ccc}
\bm{X}
\end{array}& =\left[ \begin{array}{ccc}
\mathbf{\hat{U}}_{s} & \mathbf{\hat{U}}_{n} \\
\end{array} \right]\left[ \begin{array}{ccc}
\mathbf{\hat{\Gamma}}_{s} & 0 \\
0 & \mathbf{\hat{\Gamma}}_{n}
\end{array} \right]\left[ \begin{array}{ccc}
\mathbf{\hat{U}}_{s}^{H} \\
\mathbf{\hat{U}}_{n}^{H}
\end{array} \right],
\label{SVD}\end{aligned}$$
where the diagonal matrices $\mathbf{\hat{\Gamma}}_{s}$ and $\mathbf{\hat{\Gamma}}_{n}$ contain the $\mathit{P}$ largest singular values and the $\mathit{M-P}$ smallest singular values, respectively. The estimated signal subspace $\mathbf{\hat{U}}_{s}$ $\in\mathbb{C}^{\mathit{M \times P}}$ consists of the singular vectors corresponding to $\mathbf{\hat{\Gamma}}_{s}$ and the orthogonal subspace $\mathbf{\hat{U}}_{n}$ $\in\mathbb{C}^{\mathit{M
\times (M-P)}}$ is related to $\mathbf{\hat{\Gamma}}_{n}$. If the signal subspace is estimated a rank-*P* approximation of the SVD can be applied.
The covariance approach applies the eigenvalue decomposition (EVD) of the sample covariance matrix , which is related to the data matrix : $$\bm {\hat{R}}=\frac{1}{N} \sum\limits^{N}_{i=1}\bm x(i)\bm x^H(i)=\frac{1}{N}\bm {X}\bm {X}^H\;\in\mathbb{C}^{\mathit{M \times M}
\label{rel_datamatrix_covsample}}$$ Then, the EVD of can be carried out as follows: $$\begin{array}{ccc}
\mathbf{\hat{R}}
\end{array} =\left[ \begin{array}{ccc}
\mathbf{\hat{U}}_{s} & \mathbf{\hat{U}}_{n} \\
\end{array} \right]\left[ \begin{array}{ccc}
\mathbf{\hat{\Lambda}}_{s} & 0 \\
0 & \mathbf{\hat{\Lambda}}_{n}
\end{array} \right]\left[ \begin{array}{ccc}
\mathbf{\hat{U}}_{s}^{H} \\
\mathbf{\hat{U}}_{n}^{H}
\end{array} \right],
\label{Detalha_ESPRIT}$$ where the diagonal matrices $\mathbf{\hat{\Lambda}_{s}}$ and $\mathbf{\hat{\Lambda}_{n}}$ contain the *P* largest and the *M-P* smallest eigenvalues, respectively. The estimated signal subspace $\mathbf{\hat{U}}_{s}$ $\in$ $\mathbb{C}^{\mathit{M\times P}}$ corresponding to $\mathbf{\hat{\Gamma}_{s}}$ and the orthogonal subspace $\mathbf{\hat{U}}_{n}$ $\in$ $\mathbb{C}^{\mathit{M\times(M-P) }}$ complies with $\hat{\Gamma}_{n}$. If the signal subspace is estimated a rank-P approximation of the EVD can be applied. With infinite precision arithmetic, both SVD and EVD can be considered equivalent. However, as in practice, finite precision arithmetic is employed, ’squaring’ the data to obtain the Gramian $\bm {X}\bm {X}^H$ can result in round-off errors and overflow. These are potential problems to be aware when using the covariance approach. Now, we can briefly review ESPRIT. We start by forming a twofold subarray configuration, as each row of the array steering matrix $\bm A(\bm \Theta)$ corresponds to one sensor element of the antenna array. The subarrays are specified by two $\mathit{(s\times
M)}$-dimensional selection matrices $\mathbf{J_{1}}$ and $\mathbf{J_{2}}$ which choose $\mathit{s}$ elements of the $\mathit{M}$ existing sensors, respectively, where $\mathit{s}$ is in the range $\mathit{P\leq s < M}$. For maximum overlap, the matrix $\mathbf{J_{1}}$ selects the first $\mathit{s=M-1}$ elements and the matrix $\mathbf{J_{2}}$ selects the last $\mathit{s=M-1}$ rows of $\bm A(\bm \Theta)$.
Since the matrices $\mathbf{J_{1}}$ and $\mathbf{J_{2}}$ have now been computed, we can estimate the operator $ \mathbf{\Psi} $ by solving the approximation of the shift invariance equation given by $$\mathbf{J}_{1}\:\mathbf{\hat{U}}_{s}\:\mathbf{\Psi}\:\approx\:\mathbf{J}_{2}\:\mathbf{\hat{U}}_{s}.
\label{shift_invariance_equation}$$ where $\hat{U}_{s}$ is obtained in .
Using the least squares (LS) method, which yields $$\hat{\mathbf{\Psi}}=\arg\min_{\substack{\mathbf
{\Psi}}}\parallel\mathbf{J}_{2}\:\mathbf{\hat{U}}_{s}\:-\:\mathbf{J}_{1}\:\mathbf{\hat{U}}_{s}\:\mathbf{\Psi}\parallel_{F}\:=\:\left(
\mathbf{J}_{1}\:\hat{\mathbf{U}}_{s}\right)
^{\dagger}\:\mathbf{J}_{2}\:\hat{\mathbf{U}}_{s},$$ where $\parallel\cdot\parallel_{F}$ denotes the Frobenius norm and $\left( \cdot\right)^{\dagger}$ stands for the pseudo-inverse.
Lastly, the eigenvalues $\lambda_{i}$ of $\hat{\mathbf{\Psi}}$ contain the estimates of the spatial frequencies $\gamma_{i}$ computed as: $$\gamma_{i}\:=\:\arg\left(\lambda_{i} \right),
\label{spatial_frequencies}$$ so that the DOAs can be calculated as: $$\hat{\theta}_{i}\:=\:\arcsin\left(\frac{\gamma_{i}\:\lambda_{c}}{2\pi\:\mathit{d}} \right)
\label{doas_ESPRIT}$$ where for and $\mathrm{i=1,\cdots,P}$.
Then, a new Vandermonde matrix $\mathbf{\hat{B}}^{(n+1)}$ is formed by the steering vectors of those refined estimates of the DOAs. By using this updated matrix, it is possible to compute the refined estimates of the projection matrices of the signal and the noise subspaces.
Next, employing the refined estimates of the projection matrices, the initial sample data matrix, $\bm {\hat{R}}$, and the number of sensors and sources, the stochastic maximum likelihood objective function $\mathit{U^{(n+1)}(\mu)}$ [@Stoica] is computed for each value of at the $n^{th}$ iteration, as follows: $$\begin{split}
\mathit{U^{(n+1)}(\mu)} & =\mathrm{ln\:det}
\Big(\mathbf{\hat{Q}}_{B}^{(n+1)}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{B}^{(n+1)}
\\
& \quad + \dfrac{{\rm
Trace}\{\mathbf{\hat{Q}}_{B}^{\perp\:(n+1)}\:\mathbf{\hat{R}}\}}
{\mathrm{M-P}}\:\mathbf{\hat{Q}}_{B}^{\:(n+1)\perp} \Big).
\end{split}
\label{SML_objective_function}$$
The previous computation selects the set of unavailable DOA estimates that have a higher likelihood at each iteration. Then, the set of estimated DOAs corresponding to the optimum value of that minimizes also at each $n^{th}$ iteration is determined. Finally, the output of the proposed MS-KAI-ESPRIT algorithm is formed by the set of the estimates obtained at the $P^{th}$ iteration, as described in Table \[Multi\_Step\_KAI\].
\[Multi\_Step\_KAI\]
Analysis {#Analysis}
=========
In this section, we carry out an analysis of the MSE of the data covariance matrix free of side effects along with a study of the computational complexity of the proposed MS-KAI-ESPRIT and existing direction finding algorithms.
MSE Analysis {#MSE_analysis}
------------
In this subsection we show that at the first of the $P$ iterations, the MSE of the data covariance matrix free of side effects $\mathbf{\hat{R}}^{(n+1)}$ is less than or equal to the MSE of that of the original one $\mathbf{\hat{R}}$. This can be formulated as: $$\begin{aligned}
\mathrm{MSE}\left( \mathbf{\hat{R}}^{(n+1)}\right) \leq\mathrm{MSE}\left(\mathbf{\hat{R}}\right)
\label{problem_formulation1}\end{aligned}$$ or, alternatively, as $$\begin{aligned}
\mathrm{MSE}\left( \mathbf{\hat{R}}^{(n+1)}\right)- \mathrm{MSE}\left(\mathbf{\hat{R}}\right) \leq0
\label{problem_formulation2}\end{aligned}$$ The proof of this inequality is provided in the Appendix.
Computational Complexity Analysis {#computational_analysis}
---------------------------------
In this section, we evaluate the computational cost of the proposed MS-KAI-ESPRIT algorithm which is compared to the following classical subspace methods: ESPRIT [@Roy], MUSIC [@schimdt], Root-MUSIC [@Barabell], Conjugate Gradient (CG) [@Semira], Auxiliary Vector Filtering (AVF) [@Grover] and TS-ESPRIT [@Pinto]. The ESPRIT and MUSIC-based methods use the Singular Value Decomposition (SVD) of the sample covariance matrix . The computational complexity of MS-KAI-ESPRIT in terms of number of multiplications and additions is depicted in Table \[Comput\_Complexity1\], where $\mathrm{\tau}=\frac{1}{ \iota} +1$. The increment ${ \iota}$ is defined in Table \[Multi\_Step\_KAI\]. As can be seen, for this specific configuration used in the simulations \[simulations\] MS-KAI-ESPRIT shows a relatively high computational burden with $\mathcal{O}\mathit{(P\tau(3M^{3}+8MN^{2}))}$, where $\tau$ is typically an integer that ranges from $ 1$ to $20 $. It can be noticed that for the configuration used in the simulations $(P=4,
M=40, N=25)$ $ 3M^{3}$ and $8MN^{2}$ are comparable, resulting in two dominant terms. It can also be seen that the number of multiplications required by the proposed algorithm is more significant than the number of additions. For this reason, in Table \[Comput\_Complexity2\], we computed only the computational burden of the previously mentioned algorithms in terms of multiplications for the purpose of comparisons. In that table, $\Delta$ stands for the search step.
Next, we will evaluate the influence of the number of sensor elements on the number of multiplications based on the specific configuration described in Table \[sysmodel\]. Supposing $\mathrm{P=4}$ narrowband signals impinging a ULA of $\mathrm{M}$ sensor elements and $\mathrm{N=25} $ available snapshots, we obtain Fig. \[figura:Multiplications\_reviewer2\]. We can see the main trends in terms of computational cost measured in multiplications of the proposed and analyzed algorithms. By examining Fig. \[figura:Multiplications\_reviewer2\], it can be noticed that in the range $M=\left[ 20\ 70\right] $ sensors, the curves describing the exact number of multiplications in MS-KAI-ESPRIT and AVF tend to merge. For $M=40$, this ratio tends to $1$, i.e. the number of multiplications are almost equivalent.
[|l|l|p[6cm]{}|]{} &\
------------------------------------------------------------------------
& $P\:\mathrm{\tau [\frac{10}{3}M^{3}+M^{2}(3P+2)+M(\frac{5}{2}P^{2}+\frac{1}{2}P+8N^{2})}$\
MS-KAI&$\mathrm{+P^{2}(\frac{17}{2}P+\frac{1}{2})]}$\
-ESPRIT&\
(Proposed)& $\mathrm{+P\:[2M^{3}+M^{2}(P)+M(\frac{3}{2}P^{2}+\frac{1}{2}P)+P^{2}(\frac{P}{2}+\frac{3}{2}) ]}$\
& $\mathrm{+2M^{2}(P)+M(P^{2}-P+8N^{2})+P^{2}(8P-1)}$\
&\
------------------------------------------------------------------------
& $P\:\mathrm{\tau
[\frac{10}{3}M^{3}+M^{2}(3P-1)+M(\frac{5}{2}P^{2}-\frac{9}{2}P+8N^{2})}$\
&$\mathrm{+P(8P^{2}-2P-\frac{5}{2})]}$\
& $\mathrm{+P\:[2M^{3}+M^{2}(P-2)+M(\frac{3}{2}P^{2}-\frac{1}{2}P)-P(P+\frac{1}{2}) ]}$\
& $\mathrm{+2M^{2}(P)+M(P^{2}-4P+8N^{2})+P(8P^{2}-P-2)}$\
\[Comput\_Complexity1\]
[|l|l|p[6cm]{}|]{} Algorithm & Multiplications\
MUSIC [@schimdt] & $\mathrm{\frac{180}{\Delta}[M^{2}+M(2-P)-P]+8MN^{2}}$\
root-MUSIC[@Barabell] & $\mathrm{2M^{3}-M^{2}P+8MN^{2}}$\
AVF [@Grover] & $\mathrm{\frac{180}{\Delta}[M^{2}(3P+1)+M(4P-2)+P+2]}$\
&$\mathrm{+M^{2}N}$\
CG [@Semira] & $\mathrm{\frac{180}{\Delta}[M^{2}(P+1)+M(6P+2)+P+1]+M^{2}N}$\
ESPRIT[@Roy] & $\mathrm{2M^{2}P+M(P^{2}-2P+8N^{2})+8P^{3}-P^{2} }$\
& $\mathrm{\tau[3M^{3}+M^{2}(3P+2)+M(\frac{5}{2}P^{2}-\frac{3}{2}P+8N^{2})}$\
&$\mathrm{+P^{2}(\frac{17}{2}P+\frac{1}{2})+1]}$\
TS-ESPRIT [@Pinto]\*& $\mathrm{+[2M^{3}+M^{2}(3P)+M(\frac{5}{2}P^{2}-\frac{3}{2}P+8N^{2})}$\
& $\mathrm{+P^{2}(\frac{17}{2}P+\frac{1}{2})]}$\
\[Comput\_Complexity2\]
![Number of multiplications as powers of 10 versus number of sensors for $P=4$, $N=25$. []{data-label="figura:Multiplications_reviewer2"}](fig1.eps){width="8cm" height="6cm"}
Simulations
===========
In this section, we examine the performance of the proposed MS-KAI-ESPRIT in terms of probability of resolution and RMSE and compare them to the standard ESPRIT [@Roy], the Iterative ESPRIT (IESPRIT), which is also developed here by combining the approach in [@Vorobyov2] that exploits knowledge of the structure of the covariance matrix and its perturbation terms, the Conjugate Gradient (CG) [@Semira], the Root-MUSIC [@Barabell], and the MUSIC [@schimdt] algorithms. Despite TS-ESPRIT is based on the knowledge of available known DOAS and the proposed MS-KAI-ESPRIT does not have access to prior knowledge, TS-ESPRIT is plotted with the aim of illustrating the comparisons. For a fair comparison in terms of RMSE and probability of resolution of all studied algorithms, we suppose that we do not have prior knowledge, that is to say that although we have available known DOAs, we compute TS-ESPRIT as they were unavailable. We employ a ULA with *M=40* sensors, inter-element spacing $\Delta=\frac{\lambda_{c}}{2}$ and assume there are four uncorrelated complex Gaussian signals with equal power impinging on the array. The closely-spaced sources are separated by *$2.4^{o}$*, at $\mathrm
(10.2^{o},12.6^{o},15^{o},17.4^{o})$, and the number of available snapshots is *N*=25. For TS-ESPRIT, as previously mentioned, we presume a priori knowledge of the last true DOAS $\mathrm(15^{o},17.4^{o})$
In Fig. \[figura:DSP\_PR\_2ponto4deg\_40sens\_25snap\_100runs\], we show the probability of resolution versus SNR. We take into account the criterion [@Stoica3], in which two sources with DOA $\theta_{1}$ and $\theta_{2}$ are said to be resolved if their respective estimates $\hat{\theta}_{1}$ and $\hat{\theta}_{2}$ are such that both $\left|\hat{\theta}_{1} -\theta_{1}\right|$ and $\left|\hat{\theta}_{2} -\theta_{2}\right|$ are less than $\left|\theta_{1} -\theta_{2}\right|/2$. The proposed MS-KAI-ESPRIT algorithm outperforms IESPRIT developed here, based on [@Vorobyov2], and the standard ESPRIT [@Roy] in the range between $-6$ and $ 5dB $ and MUSIC [@schimdt] from $-6$ to $
8.5dB $. MS-KAI-ESPRIT also outperforms CG [@Semira] and Root-Music [@Barabell] throughout the whole range of values. The poor performance of the latter could be expected from the results for two closed signals obtained in [@Vorobyov2]. When compared to TS-ESPRIT, which as previously discussed, was supposed to have the best performance, the proposed MS-KAI-ESPRIT algorithm is outperformed by the former only in the range between $ -6 $ and $ -2dB $. From this last point to $ 20dB $ its performance is superior or equal to the other algorithms.
In Fig. \[figura:DSP\_RMSE\_CRB\_2ponto4deg\_40sens\_25snap\_100runs\], it is shown the RMSE in dB versus SNR, where the term CRB refers to the square root of the deterministic Cramér-Rao bound [@Stoica4]. The RMSE is defined as: $$\centering \mathrm{RMSE}
=\sqrt{\frac{1}{L\:P}\sum\limits^{L}_{l=1}\sum\limits^{P}_{p=1}\bm(\theta_p
-\bm \hat{\theta}_p(l))^{2}}, \label{RMSE_run}$$ where $L$ is the number of trials.
The results show the superior performance of MS-KAI-ESPRIT in the range between $-2.5$ and $5$ dB. From this last point to $20$ dB, MS-KAI-ESPRIT, IESPRIT, ESPRIT and TS-ESPRIT have similar performance. The only range in which MS-KAI-ESPRIT is outperformed lies in the range between $-6$ and $-2.5$ dB. From this last point to $20$ dB its performance is better or similar to the others.
![Probability of resolution versus SNR with $P=4$ uncorrelated sources, $M=40$, $N=25$, $L=100$ runs[]{data-label="figura:DSP_PR_2ponto4deg_40sens_25snap_100runs"}](fig2.eps){width="8cm" height="6cm"}
![RMSE and the square root of CRB versus SNR with $P=4$ uncorrelated sources, $M=40$, $N=25$, $L=100$ runs[]{data-label="figura:DSP_RMSE_CRB_2ponto4deg_40sens_25snap_100runs"}](fig3.eps){width="8cm" height="6cm"}
Now, we focus on the performance of MS-KAI-ESPRIT under more severe conditions, i.e., we analyze it in terms of RMSE when at least two of the four equal-powered Gaussian signals are strongly correlated, as shown in the following signal correlation matrix $\bm R_{ss}$ : $$\bm R_{ss}=\sigma_{s}^{2}\begin{bmatrix}
\label{signals_correlated_matrix}
1 & 0.9 & 0.6 & 0 \\
0.9 & 1 & 0.4 & 0.5 \\
0.6 & 0.4 & 1 & 0 \\
0 & 0.5 & 0 & 1
\end{bmatrix}.$$
The signal-to-noise ratio $\text(SNR)$ is defined as $SNR\triangleq 10\log_{10}\left(\frac{\sigma_{s}^{2}}{\sigma_{n}^{2}} \right)$.
![RMSE and the square root of CRB versus SNR with $P=4$ correlated sources, $M=40$, $N=25$, $L=250$ runs[]{data-label="figura:RMSE_dB_correlated_2ponto4deg_40sens_25snap_250runs"}](fig4.eps){width="8cm" height="6cm"}
In Fig. \[figura:RMSE\_dB\_correlated\_2ponto4deg\_40sens\_25snap\_250runs\], we can see the performance of the same algorithms plotted in Fig. \[figura:DSP\_RMSE\_CRB\_2ponto4deg\_40sens\_25snap\_100runs\] in terms of $\mathrm{RMSE(dB)}$ versus SNR computed after $\mathrm{250}$ runs, when the signal correlation matrix is given by . As can be seen, the superior performance of MS-KAI-ESPRIT occurs in the whole range between $4.0$ and $ 12$ dB , which can be considered a small but consistent gain. From $ 12$dB to $ 20$dB MS-KAI-ESPRIT, TS-ESPRIT, IESPRIT and ESPRIT have similar performance. The values for which MS-KAI-ESPRIT is outperformed are in the range between $ -6.0 $ and $4.0$dB. In Fig. \[RMSE\_dB\_corr\_letter\_steps\_full\_reviewer1\], we have provided further simulations to illustrate the performance of each iteration of MS-KAI ESPRIT in terms of RMSE. The resulting iterations can be compared to each other and to the the original ESPRIT, which corresponds to the first step of MS-KAI ESPRIT. For this purpose, we have considered the same scenario employed before, except for the number of the trials, which is $L=200$ runs for all simulations. In particular, we have considered the case of correlated sources. From Fig. \[RMSE\_dB\_corr\_letter\_steps\_reviewer1\], which is a magnified detail of Fig. \[RMSE\_dB\_corr\_letter\_steps\_full\_reviewer1\], it can be seen that the estimates become more accurate with the increase of iterations.
![RMSE for each iteration of MS-KAI ESPRIT,original ESPRIT and CRB versus SNR with $P=4$ correlated sources, $M=40$, $N=25$, $L=200$ runs[]{data-label="RMSE_dB_corr_letter_steps_full_reviewer1"}](fig5.eps){width="8cm" height="6cm"}
![RMSE for each iteration of MS-KAI ESPRIT,original ESPRIT and CRB versus SNR with $P=4$ correlated sources, $M=40$, $N=25$, $L=200$ runs -magnification []{data-label="RMSE_dB_corr_letter_steps_reviewer1"}](fig6.eps){width="8cm" height="6cm"}
Conclusions
===========
We have proposed the MS-KAI-ESPRIT algorithm which exploits the knowledge of source signals obtained on line and the structure of the covariance matrix and its perturbations. An analytical study of the MSE of this matrix free of side effects has shown that it is less or equal than the MSE of the original matrix, resulting in better performance of MS-KAI-ESPRIT especially in scenarios where limited number of samples are available. The proposed MS-KAI-ESPRIT algorithm can obtain significant gains in RMSE or probability of resolution performance over previously reported techniques, and has excellent potential for applications with short data records in large-scale antenna systems for wireless communications, radar and other large sensor arrays. The relatively high computational burden required, which is associated with extra matrix multiplications, the increment $\iota$ applied to reduce the undesirable side effects and the iterations needed to progressively incorporate the knowledge obtained on line as newer estimates can be justified for the superior performance achieved. Future work will consider approaches to reducing the computational cost.
[Appendix]{} {#appendix .unnumbered}
============
[Here, we prove the inequality described in Section \[MSE\_analysis\]. We start by expressing the MSE of the original data covariance matrix as:]{} $$\begin{aligned}
\mathrm{MSE}\left( \mathbf{\hat{R}}\right) =\mathbb E\left[\|\mathbf{\hat{R}}-
\mathbf{R}\|^2_F\right].
\label{problem_formulation3}\end{aligned}$$ where $\mathbf{R} $ is the true covariance matrix . Similarly, the MSE of the data covariance matrix free of side effects $\mathbf{\hat{R}}^{(n+1)}$ can be expressed for the first iteration $n=1$ by making use of , as follows $$\begin{aligned}
&\mathrm{MSE}\left( \mathbf{\hat{R}}^{(n+1)}\right)\big\rvert_{n=1}\nonumber\\&=\mathrm{MSE}\left( \mathbf{\hat{R}}^{(2)}\right)=\mathbb E\left[\|\mathbf{\hat{R}}^{(2)}-
\mathbf{R}\|^2_F\right]\nonumber\\
&=\mathbb E\left[\| \mathbf{\hat{R}}\:-\:\mathrm{\mu}\:(\mathbf{V}^{(1)}\:+\:\mathbf{V}^{(1)H})-
\mathbf{R}\|^2_F\right]\nonumber\\
&=\mathbb E\left[\|\left( \mathbf{\hat{R}}-
\mathbf{R}\right) \:-\:\mathrm{\mu}\:(\mathbf{V}^{(1)}\:+\:\mathbf{V}^{(1)H})\|^2_F\right]
\label{MSE_eq1}\end{aligned}$$ where for the sake of simplicity, from now on we omit the superscript $^{(1)}$, which refers to the first iteration. In order to expand the result in , we make use of the following proposition:
: The squared Frobenius norm of the difference between any two matrices $\mathbf{A}$ $\in\mathbb{C}^{\mathit{m \times m}}$ and $\mathbf{B}$ $\in\mathbb{C}^{\mathit{m \times m}}$ is given by $$\begin{aligned}
\|\mathbf{A}-
\mathbf{B}\|^2_F=\|\mathbf{A}\|^2_F +\|\mathbf{B}\|^2_F -\left( \operatorname{Tr}\mathbf{A}^{H}\mathbf{B} + \operatorname{Tr}\mathbf{A}\mathbf{B}^{H}\right)\end{aligned}$$ :\
The Frobenius norm of any $\mathbf{D}$ $\in$ $\mathbb{C}^{\mathit{m \times
m}}$ matrix is defined [@Vantrees] as $$\begin{aligned}
{
\|\mathbf{D}\|_F=
\left( \sum\limits_{i=1}^{m}\sum\limits_{j=1}^{m}\left| d_{ij}\right|^{2}\right )^\frac{1}{2}=\left[\operatorname{Tr}\left(\mathbf{ D}^{H}\mathbf{ D}\right) \right]^{\frac{1}{2}}}
\label{Proof_Lemma_1}\end{aligned}$$ We express $\mathbf{D} $ as a difference between two matrices $\mathbf{A}$ and $\mathbf{B}$, both also $\in\mathbb{C}^{\mathit{m \times m}}$. Making use of Lemma1 and the properties of the trace, we obtain [$$\begin{aligned}
&\|\mathbf{A}-
\mathbf{B}\|^2_F= \operatorname{Tr}\left[\left( \mathbf{A}-\mathbf{B}\right)^{H} \left( \mathbf{A}-\mathbf{B}\right)\right]\nonumber\\
&= \operatorname{Tr}\left[ \left( \mathbf{A}^{H}-\mathbf{B}^{H}\right) \left( \mathbf{A}-\mathbf{B}\right)\right]\nonumber\\
&= \operatorname{Tr}\left[\left( \mathbf{A}^{H}\mathbf{A}\right)
-\operatorname{Tr}\left( \mathbf{A}^{H}\mathbf{B}\right)- \operatorname{Tr}\left( \mathbf{B}^{H}\mathbf{A}\right)+\operatorname{Tr}\left( \mathbf{B}^{H}\mathbf{B}\right)\right]\nonumber\\
&=\|\mathbf{A}\|^2_F +\|\mathbf{B}\|^2_F -\left( \operatorname{Tr}\mathbf{A}^{H}\mathbf{B} + \operatorname{Tr}\mathbf{A}\mathbf{B}^{H}\right),
\end{aligned}$$]{} which is the desired result.
Now, assuming that the true $\mathbf{R}$ [@Haykin] and the data covariance matrices $\mathbf{\hat{R}} $ [@Haykin] are Hermitian and using combined with Lemma1, the cyclic [@Graybill] property of the trace and the linearity [@Karr] property of the expected value, we get $$\begin{aligned}
\mathrm{MSE}\left( \mathbf{\hat{R}}^{(2)}\right)& =\mathbb
E\left\lbrace \|\mathbf{\hat{R}}-
\mathbf{R}\|^2_F + \mathrm{\mu}^{2}\:\|\mathbf{V}\:+\:\mathbf{V}^{H}\|^2_F\right.\nonumber\\
&\left.-\operatorname{Tr}\left[ \left(\mathbf{\hat{R}}-
\mathbf{R}\right)^{H} \mu\left(\mathbf{V}\:+\:\mathbf{V}^{H} \right) \right]\right.\nonumber\\
& \left.-\operatorname{Tr}\left[\mu \left(\mathbf{V}+
\mathbf{V}^{H}\right)^{H} \left(\mathbf{\hat{R}}\:+\:\mathbf{R} \right)\right]\right\rbrace \nonumber\\
& = \mathbb E\left\lbrace \|\mathbf{\hat{R}}-
\mathbf{R}\|^2_F + \mathrm{\mu}^{2}\:\|\mathbf{V}\:+\:\mathbf{V}^{H}\|^2_F\right.\nonumber\\
&\left.-\mu\operatorname{Tr}\left[ \left(\mathbf{\hat{R}}-
\mathbf{R}\right)^{H} \left(\mathbf{V}\:+\:\mathbf{V}^{H} \right) \right]\right.\nonumber\\
& \left.-\mu\operatorname{Tr}\left[ \left(\mathbf{V}+
\mathbf{V}^{H}\right)^{H} \left(\mathbf{\hat{R}}\:+\:\mathbf{R} \right)\right]\right\rbrace \nonumber\\
&=\mathbb E\left\lbrace \|\mathbf{\hat{R}}-
\mathbf{R}\|^2_F + \mathrm{\mu}^{2}\:\|\mathbf{V}\:+\:\mathbf{V}^{H}\|^2_F\right.\nonumber\\
&\left.-\mu\operatorname{Tr}\left[ \left(\mathbf{\hat{R}}-
\mathbf{R}\right) \left(\mathbf{V}\:+\:\mathbf{V}^{H} \right) \right]\right.\nonumber\\
& \left.-\mu\operatorname{Tr}\left[ \left(\mathbf{V}^{H}+
\mathbf{V}\right)\left(\mathbf{\hat{R}}\:+\:\mathbf{R} \right)\right]\right\rbrace \nonumber\\
&=\mathbb E\left\lbrace \|\mathbf{\hat{R}}-
\mathbf{R}\|^2_F + \mathrm{\mu}^{2}\:\|\mathbf{V}\:+\:\mathbf{V}^{H}\|^2_F\right.\nonumber\\
&\left.-\mu\operatorname{Tr}\left[ \left(\mathbf{\hat{R}}-
\mathbf{R}\right) \left(\mathbf{V}\:+\:\mathbf{V}^{H} \right) \right]\right.\nonumber\\
& \left.-\mu\operatorname{Tr}\left[ \left(\mathbf{\hat{R}}\:+\:\mathbf{R} \right)\left(\mathbf{V}+
\mathbf{V}^{H}\right)\right]\right\rbrace\nonumber\\
&=\mathbb E\left\lbrace \|\mathbf{\hat{R}}-
\mathbf{R}\|^2_F\right\rbrace +\mathrm{\mu}^{2} \mathbb E\left\lbrace\|\mathbf{V}\:+\:\mathbf{V}^{H}\|^2_F\right\rbrace \mathbb \nonumber\\
&-2\mu\mathbb E\left\lbrace \operatorname{Tr}\left[ \left(\mathbf{\hat{R}}-
\mathbf{R}\right) \left(\mathbf{V}\:+\:\mathbf{V}^{H} \right) \right]\right\rbrace\nonumber\\
&=\mathrm{MSE}\left( \mathbf{\hat{R}}\right) +\mathrm{\mu}^{2} \mathbb E\left\lbrace\|\mathbf{V}\:+\:\mathbf{V}^{H}\|^2_F\right\rbrace \mathbb \nonumber\\
&-2\mu\mathbb E\left\lbrace \operatorname{Tr}\left[ \left(\mathbf{\hat{R}}-
\mathbf{R}\right) \left(\mathbf{V}\:+\:\mathbf{V}^{H} \right) \right]\right\rbrace
\label{MSE_eq2}
$$ By moving the first summand of to its first element, we obtain the intended expression for the difference between the $MSEs$ of the data covariance matrix free of perturbations and the original one, i.e.: $$\begin{aligned}
\mathrm{MSE}\left( \mathbf{\hat{R}}^{(n+1)}\right)\big\rvert_{n=1}-\mathrm{MSE}\left( \mathbf{\hat{R}}\right)=\mathrm{\mu}^{2} \mathbb E\left\lbrace\|\mathbf{V}+\mathbf{V}^{H}\|^2_F\right\rbrace \mathbb \nonumber\\
\quad -2\mu\mathbb E\left\lbrace \operatorname{Tr}\left[ \left(\mathbf{\hat{R}}-
\mathbf{R}\right) \left(\mathbf{V}\:+\:\mathbf{V}^{H} \right)
\right]\right\rbrace.
\label{Difference_MSEs}\end{aligned}$$
Now, we expand the expressions inside braces of the second member of individually. We start with the first summand $$\begin{aligned}
\|\mathbf{V}+
\mathbf{V}^{H}\|^2_F& =\|\mathbf{V}\|^2_F +\|\mathbf{V}^{H}\|^2_F +\operatorname{Tr}\left(\mathbf{V}^{H}\mathbf{V}^{H} \right)+\nonumber\\
& \quad \operatorname{Tr}\left(\mathbf{(V^{\mathit{H}})}^{H}\mathbf{V} \right)\nonumber\\
&=\|\mathbf{V}\|^2_F +\|\mathbf{V}^{H}\|^2_F
+\operatorname{Tr}\left(\mathbf{V}^{H}\mathbf{V}^{H} \right)+
\operatorname{Tr}\left(\mathbf{V}\mathbf{V} \right).
\label{By_products}\end{aligned}$$
The equation can be computed by using the projection matrices of the signal and the noise subspaces and the data covariance matrix by using , , the idempotence [@Vantrees] [@Graybill] of $\mathbf{\hat{Q}}_{A}$ and the cyclic property [@Graybill] of the trace. Starting with the computation of its fourth summand, we have $$\begin{aligned}
\operatorname{Tr}\left(\mathbf{V}\mathbf{V} \right)& = \operatorname{Tr}\left[\left( \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}^{\perp} \right)\left( \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}^{\perp} \right)\right] \nonumber\\
&= \operatorname{Tr}\left[ \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\left(\mathbf{I_{M}}-\mathbf{\hat{Q}}_{A} \right)\: \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\left(\mathbf{I_{M}}-\mathbf{\hat{Q}}_{A} \right) \right]\nonumber\\
&=\operatorname{Tr}\left[\left( \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}- \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\right)\right.\nonumber\\
&\left.\left( \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}- \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\right) \right]\nonumber\\
&= \operatorname{Tr}\left[\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}} -\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\right.\nonumber\\
&\left.-\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}} +\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\right] \nonumber\\
&=\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\right) -\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\right) \nonumber\\
&-\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\right)\nonumber\\ & +\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\right)\nonumber\\
&= \operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\right)-\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\right)\nonumber\\
&-\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\right)+\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\right)=0.
\label{Trace_VV}\end{aligned}$$ Taking into account that the data covariance matrix $ \mathbf{\hat{R}} $ and the estimate of the projection matrix of the noise subspace $ \mathbf{\hat{Q}}_{A}^{\perp} $ are Hermitian, we can evaluate the third summand of as follows: $$\begin{aligned}
&\operatorname{Tr}\left(\mathbf{V}^{H}\mathbf{V}^{H} \right)= \operatorname{Tr}\left[\left( \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}^{\perp} \right)^{H}\left( \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}^{\perp} \right)^{H}\right] \nonumber\\
&= \operatorname{Tr}\left\lbrace \left[ \left( \mathbf{\hat{Q}}_{A}^{\perp}\right)^{H} \:\mathbf{\hat{R}}^{H}\:\mathbf{\hat{Q}}_{A}^{H} \right] \left[ \left( \mathbf{\hat{Q}}_{A}^{\perp}\right)^{H} \:\mathbf{\hat{R}}^{H}\:\mathbf{\hat{Q}}_{A}^{H} \right] \right\rbrace \nonumber\\
&= \operatorname{Tr}\left\lbrace \left[ \mathbf{\hat{Q}}_{A}^{\perp} \:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A} \right] \left[ \mathbf{\hat{Q}}_{A}^{\perp} \:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A} \right] \right\rbrace \nonumber\\
&= \operatorname{Tr}\left\lbrace \left[ \mathbf{\left(I_{M}-\mathbf{\hat{Q}}_{A} \right)\mathbf{\hat{R}}\:\hat{Q}}_{A}\right] \left[ \:\mathbf{\left(I_{M}-\mathbf{\hat{Q}}_{A} \right)\mathbf{\hat{R}}\:\hat{Q}}_{A} \right]\right\rbrace \nonumber\\
&=\operatorname{Tr}\left\lbrace \left[ \mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}- \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\right]\left[ \mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}- \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\right] \right\rbrace
\nonumber\\
&=\operatorname{Tr}\left\lbrace \mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A} -\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\:\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\right.\nonumber\\
&\left.-\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A} +\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\right\rbrace \nonumber\\
&=\operatorname{Tr}\left(\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\right) -\operatorname{Tr}\left(\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\right) \nonumber\\
&-\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\right) +\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\right)\nonumber\\
&= \operatorname{Tr}\left(\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\right)-\operatorname{Tr}\left(\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\right)\nonumber\\
&-\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\right)+\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\right)=0.
\label{Trace_VHVH}\end{aligned}$$ By using , we can expand the first and the second summands of as follows: $$\begin{aligned}
&\|\mathbf{V}\|^2_F +\|\mathbf{V}^{H}\|^2_F =\operatorname{Tr}\left(\mathbf{V}^{H}\mathbf{V} \right)+\operatorname{Tr}\left(\left( \mathbf{V}^{H}\right)^{H} \mathbf{V}^{H} \right)\nonumber\\
&= \operatorname{Tr}\left(\mathbf{V}^{H}\mathbf{V} \right)+\operatorname{Tr}\left(\mathbf{V}\mathbf{V}^{H} \right)\nonumber\\
&=\operatorname{Tr}\left(\mathbf{V}\mathbf{V}^{H}
\right)+\operatorname{Tr}\left(\mathbf{V}\mathbf{V}^{H}
\right)=2\operatorname{Tr}\left(\mathbf{V}\mathbf{V}^{H} \right).
\label{Sq_Frob_norm_of_VmaisVH}\end{aligned}$$ Equation can be expressed in terms of the projection matrices of the signal and the noise subspaces and the data covariance, in a similar way as for the third and fourth summands of , as follows: $$\begin{aligned}
& 2\operatorname{Tr}\left(\mathbf{V}\mathbf{V}^{H} \right)=2\operatorname{Tr}\left[ \left( \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}^{\perp} \right) \left( \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}^{\perp} \right)^{H}\right]\nonumber\\
&=2\operatorname{Tr}\left\lbrace\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \left(\mathbf{ I_{M} }-\mathbf{\hat{Q}}_{A} \right) \left[ \mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \left(\mathbf{ I_{M} }-\mathbf{\hat{Q}}_{A} \right)\right]^{H} \right\rbrace \nonumber\\
&=2\operatorname{Tr}\left\lbrace\left(\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}- \mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\right)\left(\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}- \mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\right)^{H} \right\rbrace \nonumber\\
&=2\operatorname{Tr}\left\lbrace\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}- \mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\right.\nonumber\\
&\left.- \mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}+ \mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right\rbrace \nonumber\\
&=2\left\lbrace\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A} \right) -\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right) \right.\nonumber\\
&\left.-\operatorname{Tr}\left( \mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\right) +\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right) \right\rbrace \nonumber\\
&=2\left\lbrace\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}} \right) -\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right) \right.\nonumber\\
&\left.-\operatorname{Tr}\left( \mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\right) +\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right) \right\rbrace \nonumber\\
&=2\left\lbrace\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}} \right) -\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right)\right\rbrace
\label{Double_tr_V_mais_VH}\end{aligned}$$ From , , , and , we obtain the first summand of , as follows: $$\begin{aligned}
\mathrm{\mu}^{2} \mathbb {E}\left\lbrace\|\mathbf{V}+\mathbf{V}^{H}\|^2_F\right\rbrace& =\mathrm{2\mu}^{2} \mathbb{E}\left\lbrace\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}} \right) \right.\nonumber\\
&\left.-\operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right)\right\rbrace
\label{First_summ_Difference_MSEs_key}\end{aligned}$$ In order to finish the expansion of the expressions inside braces of the second member of , now we deal with its second summand, in which we make use of the cyclic property [@Graybill] of the trace and the idempotence [@Vantrees] [@Graybill] of $\mathbf{\hat{Q}}_{A}$.
$$\begin{aligned}
& \operatorname{Tr}\left[\left( \mathbf{\hat{R}}-\mathbf{R}\right) \left( \mathbf{V}+\mathbf{V}^{H}\right) \right] =\left\lbrace \operatorname{Tr}\left( \mathbf{\hat{R}}-\mathbf{R}\right)\right.\nonumber\\
&\left.\left[ \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}^{\perp} + \left( \mathbf{\hat{Q}}_{A}\:\mathbf{\hat{R}}\:\mathbf{\hat{Q}}_{A}^{\perp} \right)^{H}\right]\right\rbrace \nonumber\\
&=\operatorname{Tr}\left\lbrace\left( \mathbf{\hat{R}}-\mathbf{R}\right)\left[ \mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \left(\mathbf{ I_{M} }-\mathbf{\hat{Q}}_{A} \right)\right.\right.\nonumber\\
&\left.\left.+ \left( \mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \left(\mathbf{ I_{M} }-\mathbf{\hat{Q}}_{A} \right)\right)^{H} \right] \right\rbrace \nonumber\\
&=\operatorname{Tr}\left\lbrace\left( \mathbf{\hat{R}}-\mathbf{R}\right)\left[\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}-\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\right.\right.\nonumber\\
&\left.\left.+\left( \mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}-\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\right)^{H} \right] \right\rbrace \nonumber\\
&=\operatorname{Tr}\left\lbrace\left( \mathbf{\hat{R}}-\mathbf{R}\right)\left[\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}-\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}+\mathbf{\hat{R}\mathbf{\hat{Q}}_{A}}-\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A} \right] \right\rbrace \nonumber\\
&=\operatorname{Tr}\left\lbrace \mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}+\mathbf{\hat{R}}\mathbf{\hat{R}\mathbf{\hat{Q}}_{A}}-2\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A} \right.\nonumber\\
&\left.-\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}-\mathbf{R}\mathbf{\hat{R}\mathbf{\hat{Q}}_{A}}+2\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\right\rbrace \nonumber\\
&=\operatorname{Tr}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}+\operatorname{Tr}\mathbf{\hat{R}}\mathbf{\hat{R}\mathbf{\hat{Q}}_{A}}-2\operatorname{Tr}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A} \nonumber\\
&-\operatorname{Tr}\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}-\operatorname{Tr}\mathbf{R}\mathbf{\hat{R}\mathbf{\hat{Q}}_{A}}+2\operatorname{Tr}\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\nonumber\\
&=\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}}+\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}}-2\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \nonumber\\
&-\operatorname{Tr}\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}-\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{R}}+2\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\nonumber\\
&=2\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}}-2\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}-\operatorname{Tr}\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \nonumber\\
&-\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{R}}+2\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\nonumber\\
&=2\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}}-2\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}-\operatorname{Tr}\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \nonumber\\
&-\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{R}}+2\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}
\label{tr_Rsample_R_V_mais_VH}\end{aligned}$$
By using , we can straightforwardly write the second summand of the second member of in terms of the projection matrices of the signal and the noise subspaces and the data covariance matrix as follows: $$\begin{aligned}
&-2\mu\mathbb E\left\lbrace \operatorname{Tr}\left[ \left(\mathbf{\hat{R}}-
\mathbf{R}\right) \left(\mathbf{V}\:+\:\mathbf{V}^{H} \right) \right]\right\rbrace
\nonumber\\
&=-2\mu\mathbb E\left\lbrace 2\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}}-2\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}-\operatorname{Tr}\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\right. \nonumber\\
&\left.-\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{R}}+2\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right\rbrace\nonumber\\
&= -4\mu\mathbb E\left\lbrace \operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}}-\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\right\rbrace\nonumber\\
& -2\mu\left\lbrace-\operatorname{Tr}\mathbb E\left[ \mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\right]-\operatorname{Tr}\mathbb E\left[ \mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{R}} \right]\right. \nonumber\\
&\left.+2\operatorname{Tr}\mathbb E\left[ \mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right] \right\rbrace \nonumber\\
&= -4\mu\mathbb E\left\lbrace \operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}}-\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\right\rbrace\nonumber\\
& -2\mu\left\lbrace-\operatorname{Tr}\mathbb \mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A} \mathbb E\left[\mathbf{\hat{R}}\right]-\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbb E\left[\mathbf{\hat{R}} \right]\right. \nonumber\\
&\left.+2\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{Q}}_{A} \mathbb E\left[\mathbf{\hat{R}}\right] \right\rbrace
\label{Sec_summ_Difference_MSEs_key}\end{aligned}$$ Now, by using and , and assuming that $\mathbb
E\left[\mathbf{\hat{R}}\right] $ is an unbiased estimate of $\mathbf{\hat{R}}$, i.e., $\mathbb E\left[\mathbf{\hat{R}}\right]
=\mathbf{R}$, we can rewrite as follows: $$\begin{aligned}
&\mathrm{MSE}\left( \mathbf{\hat{R}}^{(n+1)}\right)\big\rvert_{n=1}-\mathrm{MSE}\left( \mathbf{\hat{R}}\right)=\mathrm{\mu}^{2} \mathbb E\left\lbrace\|\mathbf{V}+\mathbf{V}^{H}\|^2_F\right\rbrace \mathbb \nonumber\\
&-2\mu\mathbb E\left\lbrace \operatorname{Tr}\left[ \left(\mathbf{\hat{R}}-
\mathbf{R}\right) \left(\mathbf{V}\:+\:\mathbf{V}^{H} \right) \right]\right\rbrace \nonumber\\
&=\mathrm{2\mu}^{2} \mathbb{E}\left\lbrace\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}} -\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right\rbrace\nonumber\\
&-4\mu\mathbb E\left\lbrace \operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}}-\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\right\rbrace\nonumber\\
& -2\mu\left\lbrace-\operatorname{Tr}\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A} \mathbf{R}-\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{R}\right. \nonumber\\
&\left.+2\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{Q}}_{A} \mathbf{R} \right\rbrace\nonumber\\
&=\mathrm{2\mu}^{2} \mathbb{E}\left\lbrace\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}} -\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right\rbrace\nonumber\\
&-4\mu\mathbb E\left\lbrace \operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}}-\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\right\rbrace\nonumber\\
& -2\mu\left\lbrace-2\operatorname{Tr}\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A} \mathbf{R} +2\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{Q}}_{A} \mathbf{R} \right\rbrace\nonumber\\
&=\mathrm{2\mu}^{2} \mathbb{E}\left\lbrace\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}} -\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right\rbrace\nonumber\\
&-4\mu\mathbb E\left\lbrace \operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}}-\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\right\rbrace\nonumber\\
& -4\mu\left\lbrace\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A} \mathbf{R}\mathbf{R} -\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{Q}}_{A} \mathbf{R} \right\rbrace\nonumber\\
&=\left( \mathrm{2\mu}^{2}-\mathrm{4\mu}\right) \mathbb{E}\left\lbrace\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}} -\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right\rbrace\nonumber\\
& -4\mu\left\lbrace\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A} \mathbf{R}\mathbf{R} -\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{Q}}_{A} \mathbf{R} \right\rbrace
\label{Difference_MSEs_final}\end{aligned}$$ Next, we will discuss equation . For this purpose, we assume that the estimate of the projection matrix of the signal subspace $\mathbf{\hat{Q}}_{A}$ [@Vantrees], the true $\mathbf{R}$ [@Haykin] and the data covariance matrices $
\mathbf{\hat{R}} $ [@Haykin] are Hermitian. For the next steps we will make use of the following Theorem which is proved in [@Chang]:
: For two Hermitian matrices $\mathbf{A}$ and $\mathbf{B}$ of the same order, $$\begin{aligned}
\label{Theorem_1}
& \operatorname{Tr}\left(\mathbf{A}\mathbf{B} \right)^{2^{k}}\leq \operatorname{Tr}\left(\mathbf{A}^{2^{k}}\mathbf{B}^{2^{k}} \right),\end{aligned}$$ where k is in integer.\
By replacing $\mathbf{A}$ with $\mathbf{\hat{Q}}_{A}$ and $\mathbf{B}$ with $\mathbf{\hat{R}}$ in and also considering $ k=1 $ , we have $$\begin{aligned}
& \operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right)^{2}\leq \operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}^{2}\mathbf{\hat{R}}^{2} \right)\nonumber\\
&\therefore\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\leq \operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}}\nonumber\\
&\Rightarrow \operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}}-\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\geq0
\label{Inequality_2}\end{aligned}$$
Similarly, making $\mathbf{A}=\mathbf{\hat{Q}}_{A}$ and $\mathbf{B}=\mathbf{R}$ for $ k=1 $, we obtain $$\begin{aligned}
& \operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}\mathbf{R} \right)^{2}\leq \operatorname{Tr}\left(\mathbf{\hat{Q}}_{A}^{2}\mathbf{R}^{2} \right)\nonumber\\
&\therefore\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{R} \mathbf{\hat{Q}}_{A}\mathbf{R}\leq \operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{R}\nonumber\\&\Rightarrow \operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{R}-\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{R} \mathbf{\hat{Q}}_{A}\mathbf{R}\geq0
\label{Inequality_3}\end{aligned}$$ Next, we analyze the behavior of the expressions $-4\mu$ and $\left(\mathrm{2\mu}^{2}-\mathrm{4\mu}\right)$ based on the reliability factor $\mu $ $\in$ $[0\;1]$, as defined in . In order to illustrate the case being studied, we assume that both expressions are continuous functions as depicted in Fig. \[Behavior\_reliability\_factor4\].
![Behavior of $\left(\mathrm{2\mu}^{2}-\mathrm{4\mu}\right)$ and $-4\mu$ for $\mu$ $\in$ $[0\;1]$[]{data-label="Behavior_reliability_factor4"}](fig7.eps){width="8cm" height="6cm"}
It can be seen in it that in the range $ [0\;1] $ both expressions assume values $\mathrm{f}(\mu)\leq0 $, i.e.: $$\begin{aligned}
\mathrm{For}\:\mu\in[0\;1]: \left \{
\begin{aligned}
&\left(\mathrm{2\mu}^{2}-\mathrm{4\mu}\right)\leq0 \\
&-4\mu\leq0
\end{aligned} \right.
\label{Mu_conditions}\end{aligned}$$ Now, we can consider the traces which form the subtraction in as different random variables $ \mathit{y}\left(\omega \right) $ and $ \mathit{x}\left(\omega \right) $, i.e.: $$\begin{aligned}
\begin{aligned}
&\left.
\begin{aligned}
&\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}}=\mathit{y}\left(\omega \right)\\
&\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}=\mathit{x}\left(\omega \right)
\end{aligned}
\right \},\:\forall\: \omega\:\in\Omega.
\end{aligned}
\label{Monotonicity_conditions_traces}\end{aligned}$$ In addition, we can suppose that there is a random variable $
\mathit{z}\left(\omega \right) $ always greater than zero, i.e., $
\mathit{z}\left(\omega \right)\geq0 $, so that $$\begin{aligned}
\mathit{z}\left(\omega \right) =\mathit{y}\left(\omega \right)-\mathit{x}\left(\omega \right)\geq0,\;\forall\: \omega\:\in\Omega
\label{Monotonicity_conditions_rv}\end{aligned}$$ Taking the expectation of and applying its properties of linearity and monotonicity [@Karr; @JM], we obtain $$\begin{aligned}
&\mathbb{E}\left[ \mathit{z}\left(\omega \right)\right]
=\mathbb{E}\left[ \mathit{y}\left(\omega
\right)-\mathit{x}\left(\omega \right)\right] \geq0,
\label{Expect_Monotonicity_conditions_rv}\end{aligned}$$ which, by making use of , results in $$\begin{aligned}
&\mathbb{E}\left[\mathit{z}\left(\omega \right)\right] =\mathbb{E}\left[ \mathit{y}\left(\omega \right)-\mathit{x}\left(\omega \right)\right]\nonumber\\
&=\mathbb{E}\left\lbrace\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}} -\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right\rbrace \geq0
\label{Expect_Monotonicity_conditions_trace}\end{aligned}$$ Next, we can combine the inequalities with to compute the second member of , for $ \mu\in[0\;1]$.
For its first summand, we combine and , as follows: $$\begin{aligned}
\left\lbrace
\begin{aligned}
&\mathbb{E}\left\lbrace\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}} -\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right\rbrace \geq0\\
&\left(\mathrm{2\mu}^{2}-\mathrm{4\mu}\right)\leq0,\;\mu\in[0\;1],
\end{aligned}
\right.
\label{Proof_MSE_final_firstsummand_1}\end{aligned}$$ to obtain in a straightforward way $$\begin{aligned}
&\left(\mathrm{2\mu}^{2}-\mathrm{4\mu}\right)\mathbb{E}\left\lbrace\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}} -\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right\rbrace \leq0
\label{Proof_MSE_final_first summand_2}\end{aligned}$$ Similarly, we can compute its second member, by combining and , as described by $$\begin{aligned}
\left\lbrace
\begin{aligned}
&\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{R} -\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \geq0\\
&-\mathrm{4\mu}\leq0,\;\mu\in[0\;1],
\end{aligned}
\right.
\label{Proof_MSE_final_second_summand_1}\end{aligned}$$ to obtain also straightforwardly the expression given by $$\begin{aligned}
&-\mathrm{4\mu}\left\lbrace\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{R} -\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{Q}}_{A}\mathbf{R} \right\rbrace \leq0
\label{Proof_MSE_final_second_summand_2}\end{aligned}$$ By combining the inequalities and with , we have $$\begin{aligned}
&\mathrm{MSE}\left( \mathbf{\hat{R}}^{(n+1)}\right)\big\rvert_{n=1}-\mathrm{MSE}\left( \mathbf{\hat{R}}\right)\nonumber\\
&=\underbrace{\left( \mathrm{2\mu}^{2}-\mathrm{4\mu}\right) \mathbb{E}\left\lbrace\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{R}} -\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}}\mathbf{\hat{Q}}_{A}\mathbf{\hat{R}} \right\rbrace}_{\leq\;0}\nonumber\\
&\underbrace{-4\mu\left\lbrace\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{\hat{Q}}_{A} \mathbf{R}\mathbf{R} -\operatorname{Tr}\mathbf{\hat{Q}}_{A}\mathbf{R}\mathbf{\hat{Q}}_{A} \mathbf{R} \right\rbrace}_{\leq\;0}\\
& \therefore\;\mathrm{MSE}\left( \mathbf{\hat{R}}^{(n+1)}\right)\big\rvert_{n=1}-\mathrm{MSE}\left( \mathbf{\hat{R}}\right)\leq\:0
\label{Proof_final}\end{aligned}$$ which is the desired result.
[100]{}
H. L. Van Trees, , John Wiley &Sons, 2002.
R. Schmidt, “Multiple emitter location and signal parameter estimation” *IEEE Trans on Antennas and Propagation*, vol.34, No.3, Mar 1986, pp 276-280.
A. J. Barabell, “Improving the resolution performance of eigenstructure-based direction-finding algorithms,” in Proc. ICASSP, Boston, MA, Apr. 1983, pp. 336–339.
R. Roy and T. Kailath, “Estimation of signal parameters via rotational invariance techniques”, *IEEE Trans. Acoust., Speech., Signal Processing*, vol. 37, July 1989, pp 984-995.
L. L. Scharf and D. W. Tufts, “Rank reduction for modeling stationary signals," *IEEE Transactions on Acoustics, Speech and Signal Processing*, vol. ASSP-35, pp. 350-355, March 1987.
A. M. Haimovich and Y. Bar-Ness, “An eigenanalysis interference canceler," [*IEEE Trans. on Signal Processing*]{}, vol. 39, pp. 76-84, Jan. 1991.
D. A. Pados and S. N. Batalama “Joint space-time auxiliary vector filtering for DS/CDMA systems with antenna arrays” *IEEE Transactions on Communications*, vol. 47, no. 9, pp. 1406 - 1415, 1999.
J. S. Goldstein, I. S. Reed and L. L. Scharf “A multistage representation of the Wiener filter based on orthogonal projections” *IEEE Transactions on Information Theory*, vol. 44, no. 7, 1998.
Y. Hua, M. Nikpour and P. Stoica, “Optimal reduced rank estimation and filtering,” IEEE Transactions on Signal Processing, pp. 457-469, Vol. 49, No. 3, March 2001.
M. L. Honig and J. S. Goldstein, “Adaptive reduced-rank interference suppression based on the multistage Wiener filter," *IEEE Transactions on Communications*, vol. 50, no. 6, June 2002.
E. L. Santos and M. D. Zoltowski, “On Low Rank MVDR Beamforming using the Conjugate Gradient Algorithm", *Proc. IEEE International Conference on Acoustics, Speech and Signal Processing*, 2004.
Q. Haoli and S.N. Batalama, “Data record-based criteria for the selection of an auxiliary vector estimator of the MMSE/MVDR filter", *IEEE Transactions on Communications*, vol. 51, no. 10, Oct. 2003, pp. 1700 - 1708.
R. C. de Lamare and R. Sampaio-Neto, “Reduced-Rank Adaptive Filtering Based on Joint Iterative Optimization of Adaptive Filters", *IEEE Signal Processing Letters*, Vol. 14, no. 12, December 2007.
Z. Xu and M.K. Tsatsanis, “Blind adaptive algorithms for minimum variance CDMA receivers," *IEEE Trans. Communications*, vol. 49, No. 1, January 2001.
R. C. de Lamare and R. Sampaio-Neto, “Low-Complexity Variable Step-Size Mechanisms for Stochastic Gradient Algorithms in Minimum Variance CDMA Receivers", *IEEE Trans. Signal Processing*, vol. 54, pp. 2302 - 2317, June 2006.
C. Xu, G. Feng and K. S. Kwak, “A Modified Constrained Constant Modulus Approach to Blind Adaptive Multiuser Detection," *IEEE Trans. Communications*, vol. 49, No. 9, 2001.
Z. Xu and P. Liu, “Code-Constrained Blind Detection of CDMA Signals in Multipath Channels," *IEEE Sig. Proc. Letters*, vol. 9, No. 12, December 2002.
R. C. de Lamare and R. Sampaio Neto, “Blind Adaptive Code-Constrained Constant Modulus Algorithms for CDMA Interference Suppression in Multipath Channels”, *IEEE Communications Letters*, vol 9. no. 4, April, 2005.
L. Landau, R. C. de Lamare and M. Haardt, “Robust adaptive beamforming algorithms using the constrained constant modulus criterion," IET Signal Processing, vol.8, no.5, pp.447-457, July 2014.
R. C. de Lamare, “Adaptive Reduced-Rank LCMV Beamforming Algorithms Based on Joint Iterative Optimisation of Filters", *Electronics Letters*, vol. 44, no. 9, 2008.
R. C. de Lamare and R. Sampaio-Neto, “Adaptive Reduced-Rank Processing Based on Joint and Iterative Interpolation, Decimation and Filtering", *IEEE Transactions on Signal Processing*, vol. 57, no. 7, July 2009, pp. 2503 - 2514.
R. C. de Lamare and Raimundo Sampaio-Neto, “Reduced-rank Interference Suppression for DS-CDMA based on Interpolated FIR Filters", *IEEE Communications Letters*, vol. 9, no. 3, March 2005.
R. C. de Lamare and R. Sampaio-Neto, “Adaptive Reduced-Rank MMSE Filtering with Interpolated FIR Filters and Adaptive Interpolators", *IEEE Signal Processing Letters*, vol. 12, no. 3, March, 2005.
R. C. de Lamare and R. Sampaio-Neto, “Adaptive Interference Suppression for DS-CDMA Systems based on Interpolated FIR Filters with Adaptive Interpolators in Multipath Channels", *IEEE Trans. Vehicular Technology*, Vol. 56, no. 6, September 2007.
R. C. de Lamare, “Adaptive Reduced-Rank LCMV Beamforming Algorithms Based on Joint Iterative Optimisation of Filters," Electronics Letters, 2008.
R. C. de Lamare and R. Sampaio-Neto, “Reduced-rank adaptive filtering based on joint iterative optimization of adaptive filters", *IEEE Signal Process. Lett.*, vol. 14, no. 12, pp. 980-983, Dec. 2007.
R. C. de Lamare, M. Haardt, and R. Sampaio-Neto, “Blind Adaptive Constrained Reduced-Rank Parameter Estimation based on Constant Modulus Design for CDMA Interference Suppression", *IEEE Transactions on Signal Processing*, June 2008.
M. Yukawa, R. C. de Lamare and R. Sampaio-Neto, “Efficient Acoustic Echo Cancellation With Reduced-Rank Adaptive Filtering Based on Selective Decimation and Adaptive Interpolation," IEEE Transactions on Audio, Speech, and Language Processing, vol.16, no. 4, pp. 696-710, May 2008.
R. C. de Lamare and R. Sampaio-Neto, “Reduced-rank space-time adaptive interference suppression with joint iterative least squares algorithms for spread-spectrum systems," *IEEE Trans. Vehi. Technol.*, vol. 59, no. 3, pp. 1217-1228, Mar. 2010.
R. C. de Lamare and R. Sampaio-Neto, “Adaptive reduced-rank equalization algorithms based on alternating optimization design techniques for MIMO systems," *IEEE Trans. Vehi. Technol.*, vol. 60, no. 6, pp. 2482-2494, Jul. 2011.
R. C. de Lamare, L. Wang, and R. Fa, “Adaptive reduced-rank LCMV beamforming algorithms based on joint iterative optimization of filters: Design and analysis," Signal Processing, vol. 90, no. 2, pp. 640-652, Feb. 2010.
R. Fa, R. C. de Lamare, and L. Wang, “Reduced-Rank STAP Schemes for Airborne Radar Based on Switched Joint Interpolation, Decimation and Filtering Algorithm," *IEEE Transactions on Signal Processing*, vol.58, no.8, Aug. 2010, pp.4182-4194.
L. Wang and R. C. de Lamare, “Low-Complexity Adaptive Step Size Constrained Constant Modulus SG Algorithms for Blind Adaptive Beamforming”, *Signal Processing*, vol. 89, no. 12, December 2009, pp. 2503-2513.
L. Wang and R. C. de Lamare, “Adaptive Constrained Constant Modulus Algorithm Based on Auxiliary Vector Filtering for Beamforming," IEEE Transactions on Signal Processing, vol. 58, no. 10, pp. 5408-5413, Oct. 2010.
L. Wang, R. C. de Lamare, M. Yukawa, “Adaptive Reduced-Rank Constrained Constant Modulus Algorithms Based on Joint Iterative Optimization of Filters for Beamforming,” *IEEE Transactions on Signal Processing*, vol.58, no.6, June 2010, pp.2983-2997.
L. Wang, R. C. de Lamare and M. Yukawa, “Adaptive reduced-rank constrained constant modulus algorithms based on joint iterative optimization of filters for beamforming", IEEE Transactions on Signal Processing, vol.58, no. 6, pp. 2983-2997, June 2010.
L. Wang and R. C. de Lamare, “Adaptive constrained constant modulus algorithm based on auxiliary vector filtering for beamforming", IEEE Transactions on Signal Processing, vol. 58, no. 10, pp. 5408-5413, October 2010.
R. Fa and R. C. de Lamare, “Reduced-Rank STAP Algorithms using Joint Iterative Optimization of Filters," IEEE Transactions on Aerospace and Electronic Systems, vol.47, no.3, pp.1668-1684, July 2011.
Z. Yang, R. C. de Lamare and X. Li, “L1-Regularized STAP Algorithms With a Generalized Sidelobe Canceler Architecture for Airborne Radar," IEEE Transactions on Signal Processing, vol.60, no.2, pp.674-686, Feb. 2012.
Z. Yang, R. C. de Lamare and X. Li, “Sparsity-aware space–time adaptive processing algorithms with L1-norm regularisation for airborne radar", IET signal processing, vol. 6, no. 5, pp. 413-423, 2012.
Neto, F.G.A.; Nascimento, V.H.; Zakharov, Y.V.; de Lamare, R.C., “Adaptive re-weighting homotopy for sparse beamforming,” in Signal Processing Conference (EUSIPCO), 2014 Proceedings of the 22nd European , vol., no., pp.1287-1291, 1-5 Sept. 2014
Almeida Neto, F.G.; de Lamare, R.C.; Nascimento, V.H.; Zakharov, Y.V.,“Adaptive reweighting homotopy algorithms applied to beamforming," IEEE Transactions on Aerospace and Electronic Systems, vol.51, no.3, pp.1902-1915, July 2015.
L. Wang, R. C. de Lamare and M. Haardt, “Direction finding algorithms based on joint iterative subspace optimization," IEEE Transactions on Aerospace and Electronic Systems, vol.50, no.4, pp.2541-2553, October 2014.
S. D. Somasundaram, N. H. Parsons, P. Li and R. C. de Lamare, “Reduced-dimension robust capon beamforming using Krylov-subspace techniques," IEEE Transactions on Aerospace and Electronic Systems, vol.51, no.1, pp.270-289, January 2015.
S. Xu and R.C de Lamare, , *Distributed conjugate gradient strategies for distributed estimation over sensor networks*, Sensor Signal Processing for Defense SSPD, September 2012.
S. Xu, R. C. de Lamare, H. V. Poor, “Distributed Estimation Over Sensor Networks Based on Distributed Conjugate Gradient Strategies", IET Signal Processing, 2016 (to appear).
S. Xu, R. C. de Lamare and H. V. Poor, *Distributed Compressed Estimation Based on Compressive Sensing*, IEEE Signal Processing letters, vol. 22, no. 9, September 2014.
S. Xu, R. C. de Lamare and H. V. Poor, “Distributed reduced-rank estimation based on joint iterative optimization in sensor networks," in Proceedings of the 22nd European Signal Processing Conference (EUSIPCO), pp.2360-2364, 1-5, Sept. 2014
S. Xu, R. C. de Lamare and H. V. Poor, “Adaptive link selection strategies for distributed estimation in diffusion wireless networks," in Proc. IEEE International Conference onAcoustics, Speech and Signal Processing (ICASSP), , vol., no., pp.5402-5405, 26-31 May 2013.
S. Xu, R. C. de Lamare and H. V. Poor, “Dynamic topology adaptation for distributed estimation in smart grids," in Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP), 2013 IEEE 5th International Workshop on , vol., no., pp.420-423, 15-18 Dec. 2013.
S. Xu, R. C. de Lamare and H. V. Poor, “Adaptive Link Selection Algorithms for Distributed Estimation", EURASIP Journal on Advances in Signal Processing, 2015.
N. Song, R. C. de Lamare, M. Haardt, and M. Wolf, “Adaptive Widely Linear Reduced-Rank Interference Suppression based on the Multi-Stage Wiener Filter," IEEE Transactions on Signal Processing, vol. 60, no. 8, 2012.
N. Song, W. U. Alokozai, R. C. de Lamare and M. Haardt, “Adaptive Widely Linear Reduced-Rank Beamforming Based on Joint Iterative Optimization," IEEE Signal Processing Letters, vol.21, no.3, pp. 265-269, March 2014.
R.C. de Lamare, R. Sampaio-Neto and M. Haardt, “Blind Adaptive Constrained Constant-Modulus Reduced-Rank Interference Suppression Algorithms Based on Interpolation and Switched Decimation,” *IEEE Trans. on Signal Processing*, vol.59, no.2, pp.681-695, Feb. 2011.
Y. Cai, R. C. de Lamare, “Adaptive Linear Minimum BER Reduced-Rank Interference Suppression Algorithms Based on Joint and Iterative Optimization of Filters," IEEE Communications Letters, vol.17, no.4, pp.633-636, April 2013.
R. C. de Lamare and R. Sampaio-Neto, “Sparsity-Aware Adaptive Algorithms Based on Alternating Optimization and Shrinkage," IEEE Signal Processing Letters, vol.21, no.2, pp.225,229, Feb. 2014.
J. Steinwandt, R. C. de Lamare and M. Haardt, “Beamspace direction finding based on the conjugate gradient and the auxiliary vector filtering algorithms”, Signal Processing, vol. 93, no. 4, April 2013, pp. 641-651.
L. Wang, R. C. de Lamare and M. Haardt, “Direction finding algorithms based on joint iterative subspace optimization,” *IEEE Transactions on Aerospace and Electronic Systems*, vol. 50, no. 4, pp. 2541-2553, October 2014.
L. Qiu, Y. Cai, R. C. de Lamare and M. Zhao, “Reduced-Rank DOA Estimation Algorithms Based on Alternating Low-Rank Decomposition,” *IEEE Signal Processing Letters*, vol. 23, no. 5, pp. 565-569, May 2016.
J. Thomas, L. Scharf, and D. Tufts, “The probability of a subspace swap in the SVD,” IEEE Trans. Signal Process., vol. 43, no. 3, pp. 730–736, Mar. 1995.
B. D. Carlson, “Covariance matrix estimation errors and diagonal loading in adaptive arrays,” IEEE Trans. Aerosp. Electron. Syst., vol. 24, no. 4, pp. 397–401, Jul. 1988.
Y. Chen, A. Wiesel, Y. C. Eldar, and A. O. Hero, “Shrinkage algorithms for MMSE covariance estimation,” IEEE Trans. Signal Process., vol. 58, no. 10, pp. 5016–5028, Oct. 2010.
H. Ruan and R. C. de Lamare, “Robust Adaptive Beamforming Using a Low-Complexity Shrinkage-Based Mismatch Estimation Algorithm," IEEE Signal Processing Letters, vol. 21, no. 1, pp. 60-64, Jan. 2014.
H. Ruan and R. C. de Lamare, “Robust Adaptive Beamforming Based on Low-Rank and Cross-Correlation Techniques," IEEE Transactions on Signal Processing, vol. 64, no. 15, pp. 3919-3932, Aug.1, 1 2016.
S. U. Pillai and B. H. Known, “Forward/backward spatial smoothing techniques for coherent signal identification,” IEEE Trans. Acoust., Speech, Signal Process., vol. 37, no. 1, pp. 8–15, Jan. 1989.
J. E. Evans, J. R. Johnson, and D. F. Sun, Application of Advanced Signal Processing Techniques to Angle of Arrival Estimation in ATC Navigation and Surveillance Systems. Lexington, MA, USA: MIT Lincoln Lab., June 1982.
X. Mestre and M. A. Lagunas, “Modified subspace algorithms for DOA estimation with large arrays,” IEEE Trans. Signal Process., vol. 56, no. 2, pp. 598–614, Feb. 2008.
A. B. Gershman and J. F. Böhme, “Improved DOA estimation via pseudorandom resampling of spatial spectrum,” IEEE Signal Process. Lett., vol. 4, no. 2, pp. 54–57, Feb. 1997.
V. Vasylyshyn, “Removing the outliers in root-MUSIC via pseudonoise resampling and conventional beamformer,” Signal Process., vol. 93, no. 12, pp. 3423–3429, Dec. 2013.
C. Qian, L. Huang, and H. C. So, “Improved unitary root-MUSIC for DOA estimation based on pseudo-noise resampling,” IEEE Signal Process. Lett., vol. 21, no. 2, pp. 140–144, Feb. 2014.
R. C. de Lamare, “Massive MIMO Systems: Signal Processing Challenges and Future Trends", Radio Science Bulletin, December 2013.
W. Zhang, H. Ren, C. Pan, M. Chen, R. C. de Lamare, B. Du and J. Dai, “Large-Scale Antenna Systems With UL/DL Hardware Mismatch: Achievable Rates Analysis and Calibration", IEEE Trans. Commun., vol.63, no.4, pp. 1216-1229, April 2015.
M. Costa, “Writing on dirty paper,” *IEEE Trans. Inform. Theory*, vol. 29, no. 3, pp. 439-441, May 1983.
R. C. de Lamare and A. Alcaim, “Strategies to improve the performance of very low bit rate speech coders and application to a 1.2 kb/s codec” IEE Proceedings- Vision, image and signal processing, vol. 152, no. 1, February, 2005.
P. Clarke and R. C. de Lamare, “Joint Transmit Diversity Optimization and Relay Selection for Multi-Relay Cooperative MIMO Systems Using Discrete Stochastic Algorithms,” *IEEE Communications Letters*, vol.15, no.10, pp.1035-1037, October 2011.
P. Clarke and R. C. de Lamare, “Transmit Diversity and Relay Selection Algorithms for Multirelay Cooperative MIMO Systems” *IEEE Transactions on Vehicular Technology*, vol.61, no. 3, pp. 1084-1098, October 2011.
Y. Cai, R. C. de Lamare, and R. Fa, “Switched Interleaving Techniques with Limited Feedback for Interference Mitigation in DS-CDMA Systems," IEEE Transactions on Communications, vol.59, no.7, pp.1946-1956, July 2011.
Y. Cai, R. C. de Lamare, D. Le Ruyet, “Transmit Processing Techniques Based on Switched Interleaving and Limited Feedback for Interference Mitigation in Multiantenna MC-CDMA Systems," IEEE Transactions on Vehicular Technology, vol.60, no.4, pp.1559-1570, May 2011.
T. Wang, R. C. de Lamare, and P. D. Mitchell, “Low-Complexity Set-Membership Channel Estimation for Cooperative Wireless Sensor Networks," IEEE Transactions on Vehicular Technology, vol.60, no.6, pp.2594-2607, July 2011.
T. Wang, R. C. de Lamare and A. Schmeink, “Joint linear receiver design and power allocation using alternating optimization algorithms for wireless sensor networks,” *IEEE Trans. on Vehi. Tech.*, vol. 61, pp. 4129-4141, 2012.
R. C. de Lamare, “Joint iterative power allocation and linear interference suppression algorithms for cooperative DS-CDMA networks", IET Communications, vol. 6, no. 13 , 2012, pp. 1930-1942.
T. Peng, R. C. de Lamare and A. Schmeink, “Adaptive Distributed Space-Time Coding Based on Adjustable Code Matrices for Cooperative MIMO Relaying Systems”, *IEEE Transactions on Communications*, vol. 61, no. 7, July 2013.
T. Peng and R. C. de Lamare, “Adaptive Buffer-Aided Distributed Space-Time Coding for Cooperative Wireless Networks," IEEE Transactions on Communications, vol. 64, no. 5, pp. 1888-1900, May 2016.
J. Gu, R. C. de Lamare and M. Huemer, “Buffer-Aided Physical-Layer Network Coding with Optimal Linear Code Designs for Cooperative Networks," IEEE Transactions on Communications, 2018.
K. Zu, R. C. de Lamare, “Low-Complexity Lattice Reduction-Aided Regularized Block Diagonalization for MU-MIMO Systems”, IEEE. Communications Letters, Vol. 16, No. 6, June 2012, pp. 925-928.
K. Zu, R. C. de Lamare, “Low-Complexity Lattice Reduction-Aided Regularized Block Diagonalization for MU-MIMO Systems”, IEEE. Communications Letters, Vol. 16, No. 6, June 2012.
K. Zu, R. C. de Lamare and M. Haart, “Generalized design of low-complexity block diagonalization type precoding algorithms for multiuser MIMO systems", IEEE Trans. Communications, 2013.
M. Tomlinson, “New automatic equaliser employing modulo arithmetic,” *Electronic Letters*, vol. 7, Mar. 1971.
C. T. Healy and R. C. de Lamare, “Decoder-optimised progressive edge growth algorithms for the design of LDPC codes with low error floors", *IEEE Communications Letters*, vol. 16, no. 6, June 2012, pp. 889-892.
A. G. D. Uchoa, C. T. Healy, R. C. de Lamare, R. D. Souza, “LDPC codes based on progressive edge growth techniques for block fading channels", *Proc. 8th International Symposium on Wireless Communication Systems (ISWCS)*, 2011, pp. 392-396.
A. G. D. Uchoa, C. T. Healy, R. C. de Lamare, R. D. Souza, “Generalised Quasi-Cyclic LDPC codes based on progressive edge growth techniques for block fading channels", *Proc. International Symposium Wireless Communication Systems (ISWCS)*, 2012, pp. 974-978.
A. G. D. Uchoa, C. T. Healy, R. C. de Lamare, R. D. Souza, “Design of LDPC Codes Based on Progressive Edge Growth Techniques for Block Fading Channels", *IEEE Communications Letters*, vol. 15, no. 11, November 2011, pp. 1221-1223.
H. Harashima and H. Miyakawa, “Matched-transmission technique for channels with intersymbol interference,” *IEEE Trans. Commun.*, vol. 20, Aug. 1972.
K. Zu, R. C. de Lamare and M. Haardt, “Multi-branch tomlinson-harashima precoding for single-user MIMO systems," in Smart Antennas (WSA), 2012 International ITG Workshop on , vol., no., pp.36-40, 7-8 March 2012.
K. Zu, R. C. de Lamare and M. Haardt, “Multi-Branch Tomlinson-Harashima Precoding Design for MU-MIMO Systems: Theory and Algorithms," IEEE Transactions on Communications, vol.62, no.3, pp.939,951, March 2014.
L. Zhang, Y. Cai, R. C. de Lamare and M. Zhao, “Robust Multibranch Tomlinson–Harashima Precoding Design in Amplify-and-Forward MIMO Relay Systems," IEEE Transactions on Communications, vol.62, no.10, pp.3476,3490, Oct. 2014.
W. Zhang et al., “Widely Linear Precoding for Large-Scale MIMO with IQI: Algorithms and Performance Analysis," IEEE Transactions on Wireless Communications, vol. 16, no. 5, pp. 3298-3312, May 2017.
B. Hochwald, C. Peel and A. Swindlehurst, “A vector-perturbation technique for near capacity multiantenna multiuser communication - Part II: Perturbation,” *IEEE Trans. Commun.*, vol. 53, no. 3, Mar. 2005.
C. B. Chae, S. Shim and R. W. Heath, “Block diagonalized vector perturbation for multiuser MIMO systems,” *IEEE Trans. Wireless Commun.*, vol. 7, no. 11, pp. 4051 - 4057, Nov. 2008.
R. C. de Lamare, R. Sampaio-Neto, “Adaptive MBER decision feedback multiuser receivers in frequency selective fading channels", *IEEE Communications Letters*, vol. 7, no. 2, Feb. 2003, pp. 73 - 75.
A. Rontogiannis, V. Kekatos, and K. Berberidis,“ A Square-Root Adaptive V-BLAST Algorithm for Fast Time-Varying MIMO Channels,” *IEEE Signal Processing Letters*, Vol. 13, No. 5, pp. 265-268, May 2006.
R. C. de Lamare, R. Sampaio-Neto, A. Hjorungnes, “Joint iterative interference cancellation and parameter estimation for CDMA systems", *IEEE Communications Letters*, vol. 11, no. 12, December 2007, pp. 916 - 918.
Y. Cai and R. C. de Lamare, “Adaptive Space-Time Decision Feedback Detectors with Multiple Feedback Cancellation”, *IEEE Transactions on Vehicular Technology*, vol. 58, no. 8, October 2009, pp. 4129 - 4140.
J. W. Choi, A. C. Singer, J Lee, N. I. Cho, “Improved linear soft-input soft-output detection via soft feedback successive interference cancellation," *IEEE Trans. Commun.*, vol.58, no.3, pp.986-996, March 2010.
R. C. de Lamare and R. Sampaio-Neto, “Blind adaptive MIMO receivers for space-time block-coded DS-CDMA systems in multipath channels using the constant modulus criterion," IEEE Transactions on Communications, vol.58, no.1, pp.21-27, January 2010.
R. Fa, R. C. de Lamare, “Multi-Branch Successive Interference Cancellation for MIMO Spatial Multiplexing Systems", *IET Communications*, vol. 5, no. 4, pp. 484 - 494, March 2011.
R.C. de Lamare and R. Sampaio-Neto, “Adaptive reduced-rank equalization algorithms based on alternating optimization design techniques for MIMO systems," IEEE Trans. Veh. Technol., vol. 60, no. 6, pp. 2482-2494, July 2011.
P. Li, R. C. de Lamare and R. Fa, “Multiple Feedback Successive Interference Cancellation Detection for Multiuser MIMO Systems," *IEEE Transactions on Wireless Communications*, vol. 10, no. 8, pp. 2434 - 2439, August 2011.
R.C. de Lamare, R. Sampaio-Neto, “Minimum mean-squared error iterative successive parallel arbitrated decision feedback detectors for DS-CDMA systems," IEEE Trans. Commun., vol. 56, no. 5, May 2008, pp. 778-789.
R.C. de Lamare, R. Sampaio-Neto, “Minimum mean-squared error iterative successive parallel arbitrated decision feedback detectors for DS-CDMA systems," IEEE Trans. Commun., vol. 56, no. 5, May 2008.
R.C. de Lamare and R. Sampaio-Neto, “Adaptive reduced-rank equalization algorithms based on alternating optimization design techniques for MIMO systems," IEEE Trans. Veh. Technol., vol. 60, no. 6, pp. 2482-2494, July 2011.
P. Li, R. C. de Lamare and J. Liu, “Adaptive Decision Feedback Detection with Parallel Interference Cancellation and Constellation Constraints for Multiuser MIMO systems”, IET Communications, vol.7, 2012, pp. 538-547.
J. Liu, R. C. de Lamare, “Low-Latency Reweighted Belief Propagation Decoding for LDPC Codes," IEEE Communications Letters, vol. 16, no. 10, pp. 1660-1663, October 2012.
C. T. Healy and R. C. de Lamare, “Design of LDPC Codes Based on Multipath EMD Strategies for Progressive Edge Growth," IEEE Transactions on Communications, vol. 64, no. 8, pp. 3208-3219, Aug. 2016.
P. Li and R. C. de Lamare, Distributed Iterative Detection With Reduced Message Passing for Networked MIMO Cellular Systems, IEEE Transactions on Vehicular Technology, vol.63, no.6, pp. 2947-2954, July 2014.
A. G. D. Uchoa, C. T. Healy and R. C. de Lamare, “Iterative Detection and Decoding Algorithms For MIMO Systems in Block-Fading Channels Using LDPC Codes," IEEE Transactions on Vehicular Technology, 2015.
R. C. de Lamare, “Adaptive and Iterative Multi-Branch MMSE Decision Feedback Detection Algorithms for Multi-Antenna Systems”, *IEEE Trans. Wireless Commun.*, vol. 14, no. 10, October 2013.
A. G. D. Uchoa, C. T. Healy and R. C. de Lamare, “Iterative Detection and Decoding Algorithms for MIMO Systems in Block-Fading Channels Using LDPC Codes," IEEE Transactions on Vehicular Technology, vol. 65, no. 4, pp. 2735-2741, April 2016.
Y. Cai, R. C. de Lamare, B. Champagne, B. Qin and M. Zhao, “Adaptive Reduced-Rank Receive Processing Based on Minimum Symbol-Error-Rate Criterion for Large-Scale Multiple-Antenna Systems,” in IEEE Transactions on Communications, vol. 63, no. 11, pp. 4185-4201, Nov. 2015.
Z. Shao, R. C. de Lamare and L. T. N. Landau, “Iterative Detection and Decoding for Large-Scale Multiple-Antenna Systems with 1-Bit ADCs," IEEE Wireless Communications Letters, 2018.
M. Shaghaghi and S. A. Vorobyov, “Subspace leakage analysis and improved DOA estimation with small sample size”, IEEE Trans. Signal Process., vol. 63, no.12, pp 3251-3265, Jun.2015.
P. Stoica and A. Nehorai, “Performance study of conditional and unconditional direction-of-arrival estimation,” IEEE Trans. Acoust., Speech, Signal Process., vol. 38, no. 10, pp. 1783–1795, Oct. 1990.
B. A. Johnson, Y. I. Abramovich, and X. Mestre, “MUSIC, G-MUSIC, and maximum-likelihood performance breakdown,” IEEE Trans. Signal Process., vol. 56, no. 8, pp. 3944–3958, Aug. 2008.
S. F. B. Pinto, R. C. de Lamare, Two-Step Knowledge-aided Iterative ESPRIT Algorithm, Twenty First ITG Workshop on Smart Antennas, 15-17 March 2017, Berlin, Germany.
W. L. Melvin and J. R. Guerci, “Knowledge-aided signal processing: a new paradigm for radar and other advanced sensors,” IEEE Transactions on Aerospace and Electronic Systems, vol. 42, no. 3, pp. 983–996, July 2006.
W. L. Melvin and G. A. Showman, “An approach to knowledge-aided covariance estimation,” IEEE Transactions on Aerospace and Electronic Systems, vol. 42, no. 3, pp.1021–1042, July 2006.
J. S. Bergin, C. M. Teixeira, P. M. Techau, and J. R. Guerci, “Improved clutter mitigation performance using knowledge-aided space-time adaptive processing,” IEEE Transactions on Aerospace and Electronic Systems, vol. 42, no. 3, pp. 997–1009, July 2006.
J. Steinwandt, R. C. de Lamare and M. Haardt, “Knowledge-aided direction finding based on Unitary ESPRIT,” 2011 Conference Record of the Forty Fifth Asilomar Conference on Signals, Systems and Computers (ASILOMAR), Pacific Grove, CA, 2011, pp. 613-617.
P. Stoica, J. Li, X. Zhu, and J. R. Guerci, "On using a priori knowledge in space-time adaptive processing,” IEEE Transactions on Signal Processing, vol. 56, no. 6, pp. 2598-2602, June 2008.
G. Bouleux, P. Stoica, and R. Boyer, “An optimal prior knowledge-based DOA estimation method,” in 17th European Signal Processing Conference (EUSIPCO), Aug. 2009, pp. 869-873.
P. Stoica, J. Li, X. Zhu, and J. R. Guerci, “On using a priori knowledge in spacetime adaptive processing,” IEEE Transactions on Signal Processing, vol. 56, no. 6, pp. 2598–2602, June 2008.
S. F. B. Pinto, R. C. de Lamare, “Multi-Step Knowledge-Aided Iterative ESPRIT for Direction Finding", submitted to 22nd International Conference on Digital Signal Processing, 23-25 August 2017 (DSP 2017), London, United Kingdom.
S. F. B. Pinto and R. C. de Lamare, “Multi-Step Knowledge-Aided Iterative ESPRIT: Design and Analysis," IEEE Transactions on Aerospace and Electronic Systems, 2018.
Simon Haykin, Adaptive Filter Theory, fourth edition, 2003.
F.A. Graybill, Matrices with Applications in Statistics, Wadsworth Publishing Company, Inc., Second Edition, 1983.
Alan F. Karr, Probability, Springer-Verlag NY, Second Edition, 1993.
Da-Wei Chang, “A Matrix Trace Inequality for Products of Hermitian Matrices”, Journal of Mathematical Analysis and Applications 237, pp.721-725, 1999.
H.Semira, H.Belkacemi, S.Marcos, “High-resolution source localization algorithm based on the conjugate gradient”, EURASIP Journal on Advances in Signal Processing, 2007(2)(2007)1–9.
R.Grover, D.A.Pados, M.J. Medley, “Subspace direction finding with an auxiliary-vector basis”, IEEE Transactions on Signal Processing, 55 (2) (2007) pp.758–763.
J.P.A. Almeida, J.M.P. Fortes, W.A. Finamore, Probability, Random Variables and Stochastic Processes, PUC-RIO/Interciencia, 2008.
P.Stoica and A.B.Gershman, ”Maximum-likelihood DOA estimation by data-supported grid search”, IEEE Signal Processing Letters, vol. 6, no. 10, pp. 273- 275, Oct 1999.
P.Stoica and Arye Nehorai, "MUSIC, maximum Likelihood, and Cramer-Rao Bound”, IEEE Transactions on Acoustics, Speech and Signal Processing, vol. 37,no. 5, pp. 720- 741, May 1989.
J.C.Liberti Jr, Theodore S. rappaport, “Smart antennas for Wireless Communications: IS-95 and Third Generation CDMA Applications”, Chapter 9, pp 253-284, Prentice Hall, 1999.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We report on the usage of a linear optics phase gate for distinguishing all four Bell states simultaneously in a quantum teleportation and entanglement swapping protocol. This is demonstrated by full state tomography of the one and two qubit output states of the two protocols, yielding average state fidelities of about $0.83$ and $0.77$, respectively. In addition, the performance of the teleportation channel is characterised by quantum process tomography. The non-classical properties of the entanglement swapping output states are further confirmed by the violation of a CHSH-type Bell inequality of $2.14$ on average.'
address: |
$^{1}$Department für Physik, Ludwig-Maximilians-Universität, D-80797 München, Germany\
$^{2}$Max-Planck-Institut für Quantenoptik, D-85748 Garching, Germany\
$^{3}$Materials Department, University of Oxford, Oxford OX1 3PH, United Kingdom\
$^{4}$Fakultät für Physik, Universität Wien, A-1090 Wien, Austria
author:
- 'Christian Schmid$^{1,2}$, Nikolai Kiesel$^{1,2}$, Ulrich K. Weber$^{3}$, Rupert Ursin$^{4}$, Anton Zeilinger$^{4}$ and Harald Weinfurter$^{1,2}$'
title: Quantum teleportation and entanglement swapping with linear optics logic gates
---
Introduction
============
Quantum teleportation [@bennett93] and entanglement swapping [@zukowski93] are fundamental elements of quantum communication protocols and thus play an important role in a number of applications. Both processes rely on the projection of two qubits onto maximally entangled Bell states. As every qubit can be analysed only separately, this detection requires to map the four Bell states one-to-one onto four distinguishable, separable states. Such a disentangling operation can be realised by elementary two-qubit quantum gates, e.g., a controlled not (<span style="font-variant:small-caps;">cnot</span>) or controlled phase gate (<span style="font-variant:small-caps;">cphase</span>). Teleportation was demonstrated already with a number of different systems, where the gate operations can be achieved (e.g. [@nielsen98; @furusawa98; @riebe04; @barrett04]). However, while photons doubtlessly are the most proper quantum system for communication tasks, the implementation of two photon quantum gates is not straight forward as there is no photon-photon interaction with reasonable coupling strength[^1]. Beginning with the initial experiments [@bouwtele; @swap], two photon interference [@weinfurter94; @braunstein95] was employed to identify up to two of the four Bell states and, recently, a probabilistic identification of three Bell states using POVM-operators was demonstrated [@gisin05]. As introduced by Knill, Laflamme and Milburn (KLM) [@KLM01], all optical two qubit quantum logic can be achieved near deterministically using linear optics plus conditioned detection and ancillary qubits. The latter can be omitted when probabilistic gate operation is sufficient [@Ralph_PG; @HT_PG]. Several experiments already proved the feasibility of these approaches (see [@pan01; @pittman02; @San04] or [@obrien03], respectively). Recently a significant improvement with respect to reliability and stability of a linear optics logic gate was reported [@pgsimpleII; @pgsimple; @pgsimpleIII] which allows to employ such gates in multi-photon quantum communication protocols. Here we report on the implementation of quantum teleportation and entanglement swapping including probabilistic, complete Bell state analysis (BSA) accomplished by the use of a linear optics <span style="font-variant:small-caps;">cphase</span> gate.
Experiment
==========
The Bell State measurement {#sec:bsa}
--------------------------
Let us start with shortly sketching the functionality of the gate [@pgsimple]:\
The operation is defined by $$\label{eq:PG} \textsc{cphase}= \left\{
\begin{array}{ccr}
{\mbox{$\mid \! HH \, \rangle$}}& \rightarrow & {\mbox{$\mid \! HH \, \rangle$}}\phantom{,} \\ {\mbox{$\mid \! HV \, \rangle$}} & \rightarrow & {\mbox{$\mid \! HV \, \rangle$}}\phantom{,} \\
{\mbox{$\mid \! VH \, \rangle$}}& \rightarrow & {\mbox{$\mid \! VH \, \rangle$}}\phantom{,} \\
{\mbox{$\mid \! VV \, \rangle$}}& \rightarrow &-{\mbox{$\mid \! VV \, \rangle$}},
\end{array}\right.$$ where the logical 0 and 1 are represented by the linear horizontal (H) and vertical (V) polarisation states of a photon respectively. To obtain the $\pi$-phaseshift for the term only, the gate-input photons are overlapped on a beam splitter with polarisation-dependent splitting ratio (PDBS), where the transmission for horizontal polarisation $T_H=1$, and for vertical polarisation $T_V=1/3$. As horizontal polarisation is not affected, no interference can occur for ${\mbox{$\mid \! HH \, \rangle$}}$ and the state does not change. The same holds for ${\mbox{$\mid \! HV \, \rangle$}}$ and ${\mbox{$\mid \! VH \, \rangle$}}$ where the photons are distinguishable by polarisation and therefore do not interfere. Only if two vertically polarised photons are passing the gate two photon interference occurs. For this case we obtain a $\pi$-phase shift if the ratio of the amplitudes for both photons being reflected is greater than the one for both being transmitted. In order to obtain equal amplitudes for all four output states a transmission $T_V=1/3$ is required, together with two beam splitters with reversed splitting ratio ($T_H=1/3$, $T_V=1$) placed after each output of the PDBS (see [figure \[fig:setup\]]{}). The gate operation succeeds if one photon is detected in each of the two outputs of the gate, which occurs in 1/9 of all cases.
The <span style="font-variant:small-caps;">cphase</span> gate can be used to perform a complete Bell state projection measurement by mapping four Bell states onto four orthogonal product states. Considering as input, for example, the maximally entangled Bell state ${\mbox{$\mid \! \widetilde{\phi^+} \, \rangle$}} =
1/\sqrt{2}({\mbox{$\mid \! H+ \, \rangle$}}+{\mbox{$\mid \! V- \, \rangle$}})$, where $+\,(-)$ denotes ${+45 \,^{\circ}}\,
({-45 \,^{\circ}})$ linear polarisation, the gate will do the following operation: $$\begin{aligned}
{\mbox{$\mid \! \widetilde{\phi^+} \, \rangle$}} & = & \frac{1}{2}\;\bigl({\mbox{$\mid \! HH \, \rangle$}}+{\mbox{$\mid \! HV \, \rangle$}}+{\mbox{$\mid \! VH \, \rangle$}}-{\mbox{$\mid \! VV \, \rangle$}}\bigr) \nonumber \\
& \stackrel{{\tiny \textsc{cphase}}}{\rightleftharpoons} &
\frac{1}{2}\;\bigl({\mbox{$\mid \! HH \, \rangle$}}+{\mbox{$\mid \! HV \, \rangle$}}+{\mbox{$\mid \! VH \, \rangle$}}+{\mbox{$\mid \! VV \, \rangle$}}\bigr) \nonumber \\
& = & \frac{1}{2} \; \bigl({\mbox{$\mid \! H \, \rangle$}}+{\mbox{$\mid \! V \, \rangle$}}\bigr) \otimes
\bigl({\mbox{$\mid \! H \, \rangle$}}+{\mbox{$\mid \! V \, \rangle$}}\bigr)={\mbox{$\mid \! ++ \, \rangle$}}.\end{aligned}$$ This means, the gate transforms between the product state ${\mbox{$\mid \! ++ \, \rangle$}}$ and the maximally entangled Bell state ${\mbox{$\mid \! \widetilde{\phi^+} \, \rangle$}}$, (analogously we obtain for ${\mbox{$\mid \! \widetilde{\psi^+} \, \rangle$}} \rightleftharpoons {\mbox{$\mid \! +- \, \rangle$}}$, $
{\mbox{$\mid \! \widetilde{\phi^-} \, \rangle$}} \rightleftharpoons {\mbox{$\mid \! -+ \, \rangle$}}$ and ${\mbox{$\mid \! \widetilde{\psi^-} \, \rangle$}} \rightleftharpoons {\mbox{$\mid \! -- \, \rangle$}}$). Consequently, by detecting one of these four product states *behind* the phase gate, we know that the photons have been in the corresponding Bell state *before* the phase gate.
For quantum teleportation and entanglement swapping it is, in principle, not necessary to project onto all four Bell states to perform these protocols. The first realizations indeed used the projection onto a single Bell-state only, neglecting the other cases. This results in a success probability of $1/4$ [@bouwtele; @swap] . The best success probability achieved is $1/2$ and is known to be the theoretical limit when using linear optics without ancillary qubits [@weinfurter94; @braunstein95; @Lue99; @Vai99]. Even though we do not neglect any Bell state in our scheme, our success probability is limited by the efficiency of the gate operation, which is $1/9$, and therefore lower than in the other schemes. However, the beauty in the application of the <span style="font-variant:small-caps;">cphase</span> gate is the possibility to detect all four Bell states in a setup just as simple as the (single-state) Bell state projection of the initial demonstration of quantum teleportation [@bouwtele].
Moreover, for quantum teleportation and entanglement swapping one might even mimic the detection of all four Bell states with a setup that detects only a restricted number, by randomly switching between the detected set of states. We like to emphasize, however, that this approach relies only on a statistical mixture of all four Bell states. In contrast, in an <span style="font-variant:small-caps;">cphase</span> gate scheme, a coherent superposition of all Bell states is obtained. This is a fundamental difference and might be crucial for other tasks that rely on the detection of Bell states. Such a situation is present, e.g., in the quantum games scheme, where the referee relies on a quantum gate to analyze entangled states [@games]. As discussed before, the referee will, due to the limited success probability of BSA, have to discard several games. But by detecting coherently all four states he does not give the players any chance to cheat, as it might be possible for other BSA implementations.
Photon state preparation and detection
--------------------------------------
In the experiment, the input states for teleportation and entanglement swapping are generated with spontaneous parametric down conversion. A 2 mm thick BBO ($\beta$-Barium Borate) crystal is pumped by UV pulses with a central wavelength of 390 nm and an average power of 700 mW from a frequency-doubled mode-locked Ti:sapphire laser (pulse length 130 fs). After passing the crystal the beam is reflected back by a UV-mirror in a distance of about 3 cm to enable SPDC also into a second pair of beams. We use degenerate, non-collinear type-II phase matching to obtain pairs of orthogonally polarised photons at a wavelength of $\lambda \simeq 780$ nm in the forward and backward direction of the BBO crystal, respectively. The photons propagating along the characteristic intersection lines of the emission cones are coupled into single mode fibers defining the four spatial modes $a$, $b$, $c$, $d$. The spectral selection is done with narrow bandwidth interference filters F ($\Delta\lambda = 2$ nm in the <span style="font-variant:small-caps;">cphase</span> gate and $\Delta\lambda = 3$ nm in modes $a$ and $d$) before detection. For initial alignment of the spatial overlap at the partially polarising beam splitter in the <span style="font-variant:small-caps;">cphase</span> gate we use the two photons of one pair for higher count rates, whereas the temporal overlap can be aligned via Hong-Ou-Mandel interference of two independently created, heralded single photons in the forward and backward direction (see [figure \[fig:setup\]]{}). The polarisation states of the photonic qubits are analysed by half- and quarter wave-plates in combination with a polarising beam splitter cube and detected by avalanche photon diodes (APD). The setup is stable over several days with typical detection rates of 180 fourfold coincidence counts per hour. The coincidence count rates have to be corrected for different detector efficiencies in the polarisation analysis of modes $a$, $b$, $c$ and $d$, which are determined relative to each other. The errors on all quantities are deduced according to Poissonian counting statistics of the raw detection-events and the detection-efficiencies.
![\[fig:setup\]**Experimental setup for the quantum teleportation and the entanglement swapping experiment, respectively.** The three and four photon states are provided by two EPR-pairs originating from type II spontaneous parametric down conversion (SPDC) processes. UV pulses are used to pump a $\beta$-Barium Borate (BBO) crystal twice. If a photon pair is created in each of the two passages, teleportation or entanglement swapping can be performed. After passing the crystal, the beam is reflected back by a UV-mirror (M). Half- and quarter wave plates (HWP, QWP) in conjunction with polarising beam splitters (PBS) are used for the polarisation analysis (PA). The complete BSA is done by a controlled phase gate (CPHASE) consisting of three polarisation dependent beam splitters (PDBS$_1$ - PDBS$_3$). The photons are spectrally selected with interference filters.](fig1.eps){width="8.5cm"}
Teleportation
-------------
The goal of quantum teleportation is to transfer the most general polarisation state ${\mbox{$\mid \! \chi \, \rangle$}}_c=(\alpha {\mbox{$\mid \! H \, \rangle$}}_c+\beta
{\mbox{$\mid \! V \, \rangle$}}_c)$ with arbitrary amplitudes $\alpha$, $\beta$ of the photon in mode $c$ onto the photon in mode $a$. In order to do so, we need, firstly, a maximally entangled Bell state in modes $a$ and $b$, and, secondly, a complete Bell state projection measurement between the photons in mode $b$ and $c$. We obtain the Bell state by proper alignment of the photon pair originating from the forward downconversion. The photon which will carry the state ${\mbox{$\mid \! \chi \, \rangle$}}_c$ is provided by the backward emission of the down conversion which is operated as a heralded single photon source with the photon in mode $d$ initialising the trigger. The polarisation state ${\mbox{$\mid \! \chi \, \rangle$}}_c$ can be prepared by a polariser in front of the fiber coupler in mode $c$ and proper alignment of the fibre’s polarisation controller.
To demonstrate teleportation, we have prepared the states ${\mbox{$\mid \! H \, \rangle$}}$, ${\mbox{$\mid \! V \, \rangle$}}$, ${\mbox{$\mid \! + \, \rangle$}}$, ${\mbox{$\mid \! R \, \rangle$}}=1/\sqrt{2}\;({\mbox{$\mid \! H \, \rangle$}}+\rmi{\mbox{$\mid \! V \, \rangle$}})$ as input states and carried out a single qubit tomography [@qutomo] in the output mode $a$. From this we obtain the density matrix $\rho_{\mathrm{exp}}$ of the experimentally teleported states and can calculate the fidelities to the input states, $\mathcal{F}_H=0.93 \pm 0.02,\;
\mathcal{F}_V=0.75 \pm 0.05,\; \mathcal{F}_+=0.79 \pm 0.02,\;
\mathcal{F}_R=0.84 \pm 0.03$, with $$\mathcal{F}_k=\phantom{\langle}_c{\mbox{$\langle \, \chi_k \! \mid$}} \;\mathcal{U}_i
\rho_{\mathrm{exp}} \mathcal{U}^{\dagger}_i \; {\mbox{$\mid \! \chi_k \, \rangle$}}_c
\;\;,$$ where $k=H,V,+$, and $R$.
Depending on the outcome of the Bell projection measurement one has to apply one out of four unitary operations $($represented by the identity or one of the three Pauli matrices, respectively, $\mathcal{U}_i={\leavevmode\hbox{\small1\normalsize\kern-.33em1}},\; {\ensuremath{\sigma_x}},\; {\ensuremath{\sigma_z}},\; \mathrm{or}\; \rmi
{\ensuremath{\sigma_y}})$ in order to recover the original state in the teleported mode. Therefore, the fidelities are calculated after application of the unitary operation on the data and averaging over the four different results of the BSA.
As can be seen, the quality of the output states differs for the various input states. This can be understood by considering the influence of imperfect gate operation. For the experimental gate the main reason for deviation from ideal performance is caused by lack of interference at PDBS$_1$. From the considerations in section \[sec:bsa\] it can be easily seen that for perfectly distinguishable photons the probability to obtain a coincidence detection is enhanced by a factor of five for the input state ${\mbox{$\mid \! VV \, \rangle$}}$. This is because if the photons do not interfere, the probabilities rather than the amplitudes for both being reflected or both being transmitted add up to $(1/3)^2+(-2/3)^2=5/9$.
Taking that into account, it is obvious that the teleportation works best for the state ${\mbox{$\mid \! H \, \rangle$}}$, as in this instance no interference is required. Consequently, from this point of view, the output state for the input ${\mbox{$\mid \! V \, \rangle$}}$ is expected to be the worst. The states ${\mbox{$\mid \! + \, \rangle$}}$ and ${\mbox{$\mid \! R \, \rangle$}}$ should be teleported approximately at the same quality on average. However, for the state ${\mbox{$\mid \! + \, \rangle$}}$ the fidelity of the output state depends on the result of the measurement in the <span style="font-variant:small-caps;">cphase</span> gate. The measured fidelities exhibit roughly the expected behavior: The loss in quality for ${\mbox{$\mid \! H \, \rangle$}}$ is not caused by lack of interference but determined by impurity of the input states. For ${\mbox{$\mid \! R \, \rangle$}}$, ${\mbox{$\mid \! + \, \rangle$}}$ and ${\mbox{$\mid \! V \, \rangle$}}$ both effects are relevant. However, ${\mbox{$\mid \! V \, \rangle$}}$ is maximally impaired by the imperfect interference and exhibits indeed the lowest fidelity. Still, despite all imperfections it is important to note that the average fidelities are all well above the optimal classical limit of $2/3$.
The four chosen input states represent a tomographic set out of which we can evaluate a teleportation process tomography [@tomo]. From this tomography one obtains the matrix $\mathcal{M}_{\mathrm{exp}}$ representing the performance of the teleportation process (see [figure \[fig:teletomo\]]{}). In this representation an ideal teleportation ($\mathcal{M}_{\mathrm{theo}}$) corresponds to the identity operation. Thus the height of the $({\leavevmode\hbox{\small1\normalsize\kern-.33em1}}, {\leavevmode\hbox{\small1\normalsize\kern-.33em1}})$-entry of $\mathcal{M}_{\mathrm{exp}}$ directly gives the so called process fidelity [@profid]: $$F_p=\mathrm{Tr}[\mathcal{M}_{\mathrm{theo}}\mathcal{M}_{\mathrm{exp}}]$$ which is the overlap between the experimentally obtained and the theoretically expected matrix, and which is the measure for the quality of the implemented teleportation process. In our experiment we reached $F_p=0.75$. The limiting factor for the process fidelity is the fidelity of the state which is teleported worst. Following the previous discussions this is the state ${\mbox{$\mid \! V \, \rangle$}}$ for which the output state fidelity reaches an average value comparable to $F_p$.
Entanglement swapping
---------------------
The inherent quantum features of the teleportation process are best seen by performing entanglement swapping. In this quantum communication method two photons, which have never interacted in the past, become entangled by teleporting the state of one photon of an entangled pair onto one photon of another entangled pair. In the experiment described before, the teleportation of a polarised photon does not succeed always, e.g. due to experimental restrictions like limited detection efficiencies. Hence, it could be argued that the observed teleportation fidelities are a result of statistical averaging over many measurements. Such arguments can be directly refuted for entanglement swapping. Here, the teleported photon is part of an entangled pair, in that sense it is not polarised. Therefore, the outcome of a measurement on this photon only is completely random. If the observed teleportation results for individual one-photon output states were attributed to statistical averaging, the analogue experimental procedure would thus unavoidably lead to a random result for the correlation measurements on two-photon output states. In the following, however, it will be proven that indeed quantum correlations can be observed. This confirms the entanglement contained in the swapped photon pair and proves that teleportation succeeds for every single instance.
![\[fig:teletomo\] **Experimentally reconstructed teleportation process tomography matrix $\mathcal{M}_{\mathrm{exp}}$.** As the ideal teleportation process equals the ${\leavevmode\hbox{\small1\normalsize\kern-.33em1}}$-operation, the height of the $({\leavevmode\hbox{\small1\normalsize\kern-.33em1}},
{\leavevmode\hbox{\small1\normalsize\kern-.33em1}})$-entry of $\mathcal{M}_{\mathrm{exp}}$ serves directly as a measure for the fidelity of the experimentally achieved process. The smallness of the imaginary parts, which are all below 0.1, also confirms the quality of our teleportation procedure.](fig2.eps){width="8.5cm"}
To perform entanglement swapping we start with two entangled photon pairs, each in the state ${\mbox{$\mid \! \phi^+ \, \rangle$}}=1/\sqrt{2}({\mbox{$\mid \! HH \, \rangle$}}+{\mbox{$\mid \! VV \, \rangle$}})$ emitted by our down conversion source in the forward and in the backward direction, respectively. As before, we accomplish the Bell projection measurement between modes $b$ and $c$ by the use of the <span style="font-variant:small-caps;">cphase</span> gate. Consequently, by projecting photons from these two modes onto a Bell state, the photons from mode $a$ and $d$ will be left in a maximally entangled state. Which Bell state they form again depends on the result of the Bell state measurement in modes $b$ and $c$: $$\begin{aligned}
{\mbox{$\mid \! \Psi \, \rangle$}}_{abcd} ={\mbox{$\mid \! \phi^+ \, \rangle$}}_{ab} {\mbox{$\mid \! \phi^+ \, \rangle$}}_{cd}& =
&\frac{1}{2}\,\Bigl({\mbox{$\mid \! \widetilde{\phi^+} \, \rangle$}}_{ad}{\mbox{$\mid \! \widetilde{\phi^+} \, \rangle$}}_{bc}
+{\mbox{$\mid \! \widetilde{\psi^+} \, \rangle$}}_{ad}{\mbox{$\mid \! \widetilde{\psi^+} \, \rangle$}}_{bc} \nonumber\\
&+&{\mbox{$\mid \! \widetilde{\phi^-} \, \rangle$}}_{ad}{\mbox{$\mid \! \widetilde{\phi^-} \, \rangle$}}_{bc}
+{\mbox{$\mid \! \widetilde{\psi^-} \, \rangle$}}_{ad}{\mbox{$\mid \! \widetilde{\psi^-} \, \rangle$}}_{bc}\Bigr)\end{aligned}$$
![\[fig:swap\] **Experimentally reconstructed density matrix $\rho_{\mathrm{exp}}$ of the swapped states.** Different outcomes of the projection measurement in the <span style="font-variant:small-caps;">cphase</span> gate result in different swapped states. In all four cases, the four columns, which are significantly different from noise clearly signal the respective swapped, entangled state with high fidelity.](fig3.eps){width="8.5cm"}
In order to determine how close the experimentally obtained states are to the expected ones and whether they are indeed entangled we performed a two-qubit tomography for photons detected in modes $a$ and $d$ depending on the result in the BSA. From this we obtain the experimental density matrices $\rho_{\mathrm{exp}}$ (see [figure \[fig:swap\]]{}), out of which we are able to calculate the states’ fidelity $\mathcal{F}$, as well as their logarithmic negativity $\mathcal{N}$ [@neg]. The latter is, as an entanglement measure, zero for separable states and equal to one for maximally entangled states. As one can see from [table \[tab:swap\]]{} we get an entangled state for each of the four Bell state projections in the <span style="font-variant:small-caps;">cphase</span> gate with fidelities of up to $0.803$ relative to the corresponding expected Bell state and with an average of 0.773 for all simultaneously detected Bell states.
[ccc]{} Bell state observed & Fidelity $\mathcal{F}_{\mathrm{exp}}$ & Negativity $\mathcal{N}$\
${\mbox{$\mid \! \widetilde{\phi^+} \, \rangle$}}_{cb}$ & $0.777 \pm 0.031$ & $0.660 \pm 0.051$\
${\mbox{$\mid \! \widetilde{\psi^+} \, \rangle$}}_{cb}$ & $0.776 \pm 0.029$ & $0.666 \pm 0.048$\
${\mbox{$\mid \! \widetilde{\phi^-} \, \rangle$}}_{cb}$ & $0.736 \pm 0.031$ & $0.582 \pm 0.055$\
${\mbox{$\mid \! \widetilde{\psi^-} \, \rangle$}}_{cb}$ & $0.803 \pm 0.027$ & $0.720 \pm 0.042$\
Quantum teleportation enables efficient communication of quantum information between remote partners and thus is a core element of future long distance quantum networks. From that point of view entanglement swapping is particularly useful, provided one obtains a swapped state which is entangled strongly enough such as to exhibit non-local correlations. To check the non-classical properties of our swapped states we show that they violate a CHSH-type Bell inequality [@chsh]. Using the <span style="font-variant:small-caps;">cphase</span> gate this can be done at the same time for all four Bell states by measuring the correlation coefficient: $$\label{eq:chsh1}
{\mbox{$\mid S_\pm \mid$}}:= {\mbox{$\mid \pm\langle \widehat{A}, \widehat{D} \rangle \mp
\langle \widehat{A}, \widehat{d} \rangle + \langle \widehat{a},
\widehat{D} \rangle + \langle \widehat{a}, \widehat{d} \rangle \mid$}}$$ Herein $\langle \widehat{A}, \widehat{D} \rangle$, $\langle
\widehat{A}, \widehat{d} \rangle$, $\langle \widehat{a},
\widehat{D} \rangle$ and $\langle \widehat{a}, \widehat{d}
\rangle$ are the expectation values of four local operators which correspond to a polarisation measurement under four sets of angles; ${0 \,^{\circ}}\,\mathrm{for}\,\widehat{a}$, ${-
22.5 \,^{\circ}}\,\mathrm{for}\,\widehat{D}$, ${-
45 \,^{\circ}}\,\mathrm{for}\,\widehat{A}$ and ${-67.5 \,^{\circ}}\,\mathrm{for}\,\widehat{d}$, respectively. $\widehat{A}$, $\widehat{a}$ are acting on qubits in mode $a$ and $\widehat{D}$, $\widehat{d}$ on qubits in mode $d$.
For local hidden variable models ${\mbox{$\mid S_{\pm} \mid$}}$ is bounded from above by $2$. In our experiment we were able to violate this limit for each of the four Bell states (${\mbox{$\mid \! \widetilde{\phi^+} \, \rangle$}}_{ad}: S_{+}=-2.20\pm 0.17$, ${\mbox{$\mid \! \widetilde{\psi^+} \, \rangle$}}_{ad}: S_{+}=2.13\pm 0.15$, ${\mbox{$\mid \! \widetilde{\phi^-} \, \rangle$}}_{ad}: S_{-}=2.12\pm 0.16$, ${\mbox{$\mid \! \widetilde{\psi^-} \, \rangle$}}_{ad}: S_{-}=-2.12\pm 0.18$). Due to the limited measurement time for each of the four cases the error is relatively high compared to the actual violation. However, the average value of $2.14 \pm 0.08$ confirms the violation of the Bell inequality.
Discussion and conclusion
=========================
To summarise, we have performed complete BSA in a teleportation and entanglement swapping experiment by applying a probabilistic, linear optics <span style="font-variant:small-caps;">cphase</span> gate for photons. The teleported polarisation states showed fidelities clearly above the classical bound. The quality of the implemented teleportation and the fact that we were able to achieve an efficient quantum channel was confirmed by reconstruction of the quantum process matrix. Running the entanglement swapping protocol yields high fidelities and states which are entangled strong enough to violate a Bell inequality. So, a universal two photon gate based on linear optics was successfully applied for the first time in quantum communication protocols.
Our experiment is a further demonstration that linear optics gates are no longer feasible just in principle but have reached a level of functionality and simplicity which allows their implementation in quantum information applications. The combination with recently developed active feed-forward techniques [@pre06] will additionally open up new vistas for linear optics quantum computation.
This work was supported by the DFG-Cluster of Excellence MAP and the European Commission through the EU Project RamboQ and QAP.\
R. U. and A. Z. acknowledge the support by QCCM.
References {#references .unnumbered}
==========
[10]{} Bennett C H, Brassard G, Crépeau C, Jozsa R, Peres A and Wootters W K 1993 Teleporting an unknown quantum state via dual classical and Einstein-Podolsky-Rosen channels. [*Phys. Rev. Lett.*]{} [**70**]{} 1895–-1899.
Żukowski M, Zeilinger A, Horne M A and Ekert A K 1993 "Event-ready-detectors” Bell experiment via entanglement swapping. [*Phys. Rev. Lett.*]{} [**71**]{} 4287–4290.
Nielsen M A, Knill E and LaFlamme R 1998 Complete quantum teleportation using nuclear magnetic resonance. [*Nature*]{} [**396**]{} 52–55.
Furusawa A, Sørensen J, Braunstein S, Fuchs C, Kimble J and Polzik E 1998 Unconditional quantum teleportation. [*Science*]{} [**282**]{} 706–709.
Riebe M [*et al.*]{} 2004 Deterministic quantum teleportation with atoms. [*Nature*]{} [**429**]{}, 734–-737.
Barrett M D [*et al.*]{} 2004 Deterministic quantum teleportation of atomic qubits. [*Nature*]{} [**429**]{} 737–-739.
Kim Y-H, Kulik S P and Shih Y 2001 Quantum Teleportation of a Polarization State with a Complete Bell State Measurement. [*Phys. Rev. Lett.*]{} [**86**]{}, 1370–1373.
Schuck C, Huber G, Kurtsiefer C and Weinfurter H 2006 Complete Deterministic Linear Optics Bell State Analysis. [*Phys. Rev. Lett.*]{} [**96**]{} 190501.
Bouwmeester D, Pan J W, Mattle K, Eibl M, Weinfurter H and Zeilinger A 1997 Experimental quantum teleportation. [*Nature*]{} [**390**]{} 575–-579.
Pan J W, Bouwmeester D, Weinfurter H and Zeilinger A 1998 Experimental entanglement swapping: entangling photons that never interacted. [*Phys. Rev. Lett.*]{} [**80**]{} 3891–3894.
Weinfurter H 1994 Experimental Bell-state analysis. [*Europhys. Lett.*]{} [**25**]{} 559.
Braunstein S L and Mann A 1995 Measurement of the Bell operator and quantum teleportation. [*Phys. Rev.*]{} A [**51**]{} R1727–R1730.
van Houwelingen J A W, Brunner N, Beverators A, Zbinden H and Gisin N 2006 Quantum teleportation with a three-Bell-state analyzer. [*Phys. Rev. Lett.*]{} [**96**]{} 130502.
Knill E, Laflamme R and Milburn G J 2001 A scheme for efficient quantum computation with linear optics. [*Nature*]{} [**409**]{} 46-–52.
Ralph T C, Langford N K, Bell T B and White A G 2002 Linear optical controlled-NOT gate in the coincidence basis. [*Phys. Lett.*]{} A [**65**]{} 062324.
Hofmann H F and Takeuchi S 2002 Quantum phase gate for photonic qubits using only beam splitters and postselection. [*Phys. Rev.*]{} A [**66**]{} 024308.
Gasparoni S, Pan J W, Walther P, Rudolph T and Zeilinger A 2004 Realization of a photonic controlled-NOT gate sufficient for quantum computation. [*Phys. Rev. Lett.*]{} [**93**]{} 020504.
Pittman T B, Jacobs B C and Franson J D 2002 Demonstration of nondeterministic quantum logic operations using linear optical elements. [*Phys. Rev. Lett*]{} [**88**]{} 257902.
Sanaka K, Jennewein T, Pan J-W, Resch K and Zeilinger A 2004 Experimental nonlinear sign-shift for linear optics quantum computation. [*Phys. Rev. Lett.*]{} [**92**]{} 017902.
O’Brien J L, Pryde G J, White A G, Ralph T C and Branning D 2003 Demonstration of an all-optical quantum controlled-not gate. [*Nature*]{} [**426**]{} 264–-267.
Langford N K, Weinhold T J, Prevedel R, Gilchrist A, O’Brien J L, Pryde G J and White A G 2005 Demonstration of a simple entangling optical gate and its use in Bell-state analysis. [*Phys. Rev. Lett.*]{} [**95**]{} 210504.
Kiesel N, Schmid C, Weber U, Ursin R and Weinfurter H 2005 Linear optics controlled-phase gate made simple. [*Phys. Rev. Lett.*]{} [**95**]{} 210505.
Okamoto R, Hofmann H F, Takeuchi S and Sasaki K 2005 Demonstration of an optical quantum controlled-NOT gate without path interference. [*Phys. Rev. Lett.*]{} [**95**]{} 210506.
Lütkenhaus N, Calsamiglia J and Suominen K-A 1999 Bell measurements for teleportation. [*Phys. Rev.*]{} A [**59**]{} 3295–3300.
Vaidman L and Yoran N 1999 Methods for reliable teleportation. [*Phys. Rev.*]{} A [**59**]{} 116–125.
Eisert J, Wilkens M, and Lewenstein M 1999 Quantum Games and Quantum Strategies. [*Phys. Rev. Lett.*]{} [**83**]{} 3077.
James D, Kwiat P, Munro W and White A 2001 Measurement of qubits. [*Phys. Rev.*]{} A [**64**]{} 052312.
Chuang I L and Nielsen M A 1997 Prescription for experimental determination of the dynamics of a quantum black box. [*J. Mod. Opt.*]{} [**44**]{} 2455–-2467.
Gilchrist A, Langford N K and Nielsen M A 2004 Distance measures to compare real and ideal quantum processes. quant-ph/0408063.
Vidal G and Werner R F 2002 Computable measure of entanglement. [*Phys. Rev.*]{} A [**65**]{} 032314.
Clauser J F, Horne M A, Shimony A and Holt R A 1969 Proposed experiment to test local hidden-variable theories. [*Phys. Rev. Lett.*]{} [**23**]{} 880–-884.
Prevedel R, Walther P, Tiefenbacher F, Böhi P, Kaltenbaek R, Jennewein T and Zeilinger A 2006 High-speed linear optics quantum computing using active feed-forward. [*Nature.*]{} [**445**]{} 65–69.
[^1]: Quantum teleportation of a photon polarization state with complete Bell state analysis was once demonstrated using non-linear effects, though with vanishingly small probability [@shih01]. Deterministic schemes relying on entanglement in additional degrees of freedom are not suited for teleportation [@Schuck].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The self-enrichment of massive star clusters by p-processed elements is shown to increase significantly with increasing gas density as a result of enhanced star formation rates and stellar scatterings compared to the lifetime of a massive star. Considering the type of cloud core where a globular cluster might have formed, we follow the evolution and enrichment of the gas and the time dependence of stellar mass. A key assumption is that interactions between massive stars are important at high density, including interactions between massive stars and massive star binaries that can shred stellar envelopes. Massive-star interactions should also scatter low-mass stars out of the cluster. Reasonable agreement with the observations is obtained for a cloud core mass of $\sim4\times10^6\;M_\odot$ and a density of $\sim2\times10^6$ cm$^{-3}$. The results depend primarily on a few dimensionless parameters, including, most importantly, the ratio of the gas consumption time to the lifetime of a massive star, which has to be low, $\sim10$%, and the efficiency of scattering low-mass stars per unit dynamical time, which has to be relatively large, such as a few percent. Also for these conditions, the velocity dispersions of embedded globular clusters should be comparable to the high gas dispersions of galaxies at that time, so that stellar ejection by multi-star interactions could cause low-mass stars to leave a dwarf galaxy host altogether. This could solve the problem of missing first-generation stars in the halos of Fornax and WLM.'
author:
- 'Bruce G. Elmegreen'
title: 'Globular Cluster Formation at High Density: A model for Elemental Enrichment with Fast Recycling of Massive-Star Debris'
---
Introduction {#intro}
============
Most globular clusters (GCs) in the Milky Way have two populations of stars in approximately equal proportion with a first generation (G1) relatively abundant in Oxygen compared to Sodium and a second generation (G2) the reverse. This bimodality is presumably the result of contamination of the G2 stars by the products of p-process nucleosynthesis in the G1 stars, particularly the reactions $^{22}$Ne to $^{23}$Na along with the simultaneous destruction of $^{16}$O at a temperature of $2\times10^7$ K [@den90; @dec07a]. Other elemental anti-correlations include Mg with Al [e.g., @carretta14], explained by [@langer93] and others as a result of high temperature ($T>7\times10^7$ K) proton-capture plus beta-decay that transforms $^{24}$Mg into $^{25}$Mg, $^{26}$Mg and $^{27}$Al. Also in some GCs, stellar Nitrogen anti-correlates with Carbon and Oxygen [@dickens91], with a total C+N+O abundance that is about constant, suggesting a CNO cycle. Reviews of these abundance anomalies are in [@gratton04], [@char05], [@gratton12], [@renzini15], and [@bastian15].
Light element anomalies from the CNO cycle and the NeNa- and MgAl-chains are peculiar to the GCs and are not in field or halo stars [@gratton00; @char05; @prantzos06]. Because they involve high-temperature reactions and the stars in which they are observed today are too low in mass to have had such p-process reactions themselves [@gratton01], the Na-excess and Al-excess stars in current GCs had to form in gas that was pre-enriched with these elements [@cottrell81], most likely from the previous generation of stars as mentioned above. The debris from this previous generation may also have mixed with some left-over initial gas to explain an anticorrelation between Li, which is destroyed in stars, and Na, which is produced [@pasquini05; @bonifacio07; @dec07a], and to explain the difference between the high enrichment in the nuclear burning regions of massive stars and the observed enrichment in low mass stars [@dec07b table 4].
The first generation has been proposed to include normal massive stars [@cottrell81], such as rapidly spinning massive stars [@prantzos06], which, because of their rotation, bring p-processed material from the H-burning zone into the envelope and then shed it along the equator via centrifugal force and radiation pressure [@prantzos06; @dec07b]. Binary massive stars with Roche lobe overflow are also an option [@demink09]. AGB stars [@dercole08] shed processed gas at low speed too and may be a source of the anomalies, although the near-constant C+N+O does not look like a byproduct of AGB stars, which make carbon during Helium burning [@char05]; NGC 1851 may be an exception though [@yong14; @simpson16]. [@renzini15] discuss ways in which the AGB option might still be viable. Other models for the bimodality include proto-stellar disk accretion [@bastian13a], supermassive stars [@denis14; @denis15], GC-merging in dwarf galaxy hosts [@bekki16], and AGB wind re-collection in the pressurized cavity around the GC [@dercole16].
A problem with these models is that the stellar debris is only a small fraction of the total stellar mass in a normal IMF. This implies that because the current G1 and G2 masses are comparable to each other, the mass in the original G1 population had to be at least the inverse of this fraction times the current G1 mass. For the model with rapidly rotating massive stars and a normal IMF, the original cluster had to be $\sim20$ times the current G1 mass [@dec07b], and for the AGB model, the original cluster had to be $>10$ times the current mass [@dercole08]. The missing 90%-95% of the G1 stars had to escape, and even more had to escape if G2 stars escaped too, which is likely. These escaping stars presumably comprise the halo of the Milky Way [@prantzos06; @martell11], but such large halo amounts are not evident in the Fornax or WLM dwarf galaxies [@larsen12; @larsen14; @elmegreen12].
Another constraint on the models is that there is virtually no elemental contamination from G1 supernova inside the G2 stars [@char05; @renzini13]. In the model with rapidly rotating massive stars, this constraint implies that the G2 stars formed quickly, before the G1 supernovae exploded. A potential problem here is that without supernovae, the gas that formed G1 may not be cleared away from a massive cluster [@ginsburg16; @lin16], causing a high fraction of it to turn into stars. Such high efficiency prevents the cluster from shedding its excess G1 stars during rapid gas loss [@krause16], and then the final cluster is too massive and it has too many G1 stars compared to G2. Cluster clearing was tested by [@bastian14b], who found that clusters with masses up to $10^6\;M_\odot$ and ages between 20 and 30 Myr are often clear of gas. This age is old enough for some supernovae to have occurred. Without gas, star formation stops and a prolonged epoch of secondary star formation that feeds on G1 debris does not occur.
Secondary or delayed star formation as in the AGB model is also not observed in today’s massive clusters at the predicted age range of 10-1000 Myr [@bastian13b; @cabrera14], although it was reported by [@li16] whose result was then questioned by [@cabrera16a]. Neither do massive 100 Myr old clusters have obvious gas from accumulated winds or secondary accretion [@bastian14a].
Here we investigate further the model by [@prantzos06] and others, where p-process contamination from massive stars is quickly injected into a star-forming cloud core and incorporated into other forming stars. We consider conditions appropriate for the early Universe where the interstellar pressure was much higher than it is today. This pressure follows from the observation that star formation was $\sim10$ times more active per unit area than it is now [@genzel10] as a result of a factor of $\sim10$ higher gas column density [@tacconi10; @daddi10]. Because interstellar pressure scales with the square of the column density ($P\sim
G\Sigma^2$), the pressure was $\sim100\times$ larger when GCs formed than it is in typical star-forming regions today. Consequently, the core of the GC-forming cloud was likely to have a higher density and therefore a larger number of free-fall times, such as 10 or more, before the first supernovae occurred. Star formation could have occurred so quickly and the energy dissipation rate at high density could have been so high that feedback from massive stars had little effect on cloud dispersal before the supernova era [e.g., @wunsch15]. Also at high density, massive stars experience a significant drag force from gas and low-mass stars [@ostriker99], causing them to spiral into an even more compact configuration. High stellar densities lead to the dispersal of extrusion disks filled with stellar envelope material [@prantzos06], and also to close encounters between massive stars or between tight massive binaries and massive stars, which can shred the stellar envelopes [@gaburov10]. High stellar densities may also make supermassive stars by coalescence [@ebisuzaki01; @bally05], and these stars can produce the highest-temperature p-process elements [@denis15].
A related point is that at high density, a significant fraction of early-forming low-mass stars should have been ejected from the cloud core by interactions with massive binaries and massive multi-body collisional systems. This could occur long before “infant mortality,” when the final clearing of gas signals the end of the star formation process. Rapid and continuous stellar ejections by massive star interactions are well observed in numerical simulations of cluster formation [@reipurth01; @bate05; @fujii13]. Because massive three-body collisions can also dump nearly a supernova’s worth of kinetic energy into a region [@gaburov10; @umbreit08], the gas in the core should be continuously agitated. The corresponding changes in the gravitational potential energy of the gas should eject even more low-mass stars, in analogy to the proposed ejection of stars and dark matter from the cores of young star-bursting dwarf galaxies [@governato12; @elbadry16].
Dense cores are likely to continue accreting from the cloud envelope for many core dynamical times. If the core plus envelope gas mixes with stellar debris and forms new stars, then a succession of stellar populations will occur with ever-increasing levels of p-process elements. By the time the first massive stars begin to supernova and clear away the gas, some 3-7 Myr after star formation begins [@heger03], there should be a wide range of p-process elements in the stars that have formed. The model presented below shows that this range can reproduce the observations for certain values of dimensionless parameters.
A key observation is that elemental enrichment in GCs seems to be discrete [@marino11; @carretta12; @carretta14; @renzini15]. Discreteness requires burst-like contamination, and in the stellar interaction model here, that means intermittent events of catastrophic massive-star interactions. The interactions occur where the density is highest, so either the stellar density in the cloud core varies episodically with a burst of stellar collisions following each high-density phase, or there are several mass-segregated cores in a proto-GC [@mcmillan07; @fujii13] and each has its own burst of intense interactions. Both situations are likely and could contribute to discrete populations of stars. The temporal density variations would presumably follow from the time-changing gravitational potential in the cluster core, as stellar envelope mass is disbursed along with cloud mass through collisions, and then recollected in the core after a dynamical time because of self-gravity and background pressure.
The following sections examine this model in more detail. Section \[basicmodel\] outlines the basic model of GC formation at high density, section \[modeleq\] presents the equations that govern this model, giving some analytical solutions in section \[analyticalsol\], and section \[results\] shows the results. A conclusion that highlights the main assumptions of the model and their implications for high-density cluster formation is in Section \[conclu\].
Star Formation in Stellar Debris
================================
Basic Model {#basicmodel}
-----------
The basic scale of GC formation considered here involves a molecular cloud core of mass $\sim4\times10^6\;M_\odot$ in a spherical region of radius $\sim3$ pc. A larger region of lower-density gas should surround this core, possibly in the form of filaments or spokes which continuously deliver new gas [@klessen98; @myers09]. Perhaps the total mass involved with the proto-GC and its neighborhood is $\sim10^7\;M_\odot$ or more, as observed for a massive dense region in the Antenna galaxy [@herrera12; @johnson15]. The core molecular density for the above numbers is $2.2\times10^6$ cm$^{-3}$ and the free fall time ($=(3\pi/32 G \rho)^{0.5}$) is 0.03 Myr. The ratio of the core mass to the free fall time, $133\;M_\odot$ yr$^{-1}$, is a measure of the core accretion rate during core formation. If this core collapsed from the interstellar gas at the typical rate of $\sigma_{\rm ISM}^3/G$ for interstellar velocity dispersion $\sigma_{\rm
ISM}$, then $\sigma_{\rm ISM}=82$ km s$^{-1}$, which is not unusual for high-redshift disk galaxies [e.g., @forster09]. The accretion rate of low-density peripheral gas on to the core should be much less than the initial core formation rate.
The fiducial cloud core mass of $\sim4\times10^6\;M_\odot$ was chosen to produce a final GC mass of around $2\times10^5\;M_\odot$, which is at the peak of the GC mass distribution function [@harris79]. The first factor of $\sim10$ in mass reduction follows from our model including multiple generations of star formation in the core and mass loss from stellar ejection, as required by the observed spread in p-processed elemental abundance. Another factor of $\sim2$ reduction in the GC mass is likely from evaporation over a Hubble time [@mclaughlin08].
The average core surface density in this basic model, $1.4\times10^5\;M_\odot$ pc$^{-2}$, is comparable to the maximum for stellar systems [@hopkins10; @walker16]. The core velocity dispersion is $\sim76$ km s$^{-1}$, which is a factor of 2 higher than for massive clusters today, but not unreasonable for a young galaxy where the gas turbulence speed is this high. Over time, the cluster should expand and the dispersion decrease [e.g. @gieles16]. The original cluster dispersion is high enough to make feedback-driven gas loss difficult before the supernova era [@matzner15; @krause16]. We consider that because of this difficulty, the efficiency of star formation per unit free-fall time might be relatively high, $\sim10$%, instead of the usual 1% [@krumholz07]. Then the gas consumption time, which is the free-fall time divided by the efficiency, is 0.3 Myr. The significant point here is that this consumption time is much less than the evolution time of a high-mass star.
One potential implication of the high velocity dispersion in GC-forming cloud cores is that stars ejected by time-changing gravitational potentials (e.g., binary or multi-star interactions and the induced rapid gas motions) should also have fairly high velocities. Dynamical ejection processes can be much more energetic than thermal evaporation from a random walk. This offers an intriguing solution to the problem stated in the introduction that Fornax and WLM do not have stellar halos massive enough to include all of the required mass of G1 stars that should have been ejected from their GCs. In fact these are dwarf galaxies with very slow internal motions: the Fornax dwarf spheroidal galaxy has a $\sim12$ km s$^{-1}$ internal velocity dispersion and a much lower rotation speed [@walker06], while WLM has a 36 km s$^{-1}$ rotation speed [@leaman12]. Stars that are ejected at a factor of 1.5 to 2 times the escape speed of the GC could leave the galaxy. This possibility leads to the prediction that [*galaxies with slow internal motions should have a systematic depletion in halo stars from the G1 population that escaped their GCs*]{}. A similar conclusion was reached by [@khalaj16] based on stellar loss from GCs during gas expulsion.
The IMF for all star formation is assumed to be fully populated and given by the log-normal distribution in [@paresce00] for stellar mass $0.1\;M_\odot<M<0.8\;M_\odot$, with mass at the peak $M_C= 0.33\;M_\odot$ and dispersion $\sigma= 0.34\;M_\odot$, and by a power law with the Salpeter slope $-2.35$ above $0.8\;M_\odot$ [see also @prantzos06]. The upper limit to the stellar mass will be varied from $M_{\rm upper}=100\;M_\odot$ to $300\;M_\odot$, with the high value considered because of stellar coalescence. Note that a $320\;M_\odot$ star has been suggested for the dense massive cluster R136 in the LMC [@crowther10; @crowther16], and stars more massive than $100\;M_\odot$ were found in the dense cluster in NGC 5253 [@smith16]. Stars with masses larger than $20\;M_\odot$ are assumed to make p-process elements [@dec07a] and mix them into their stellar envelopes, which are defined to be all of the stellar mass outside of the He core, as given by [@prantzos06]. For $M_{\rm
upper}=100\;M_\odot$, the fraction of the total stellar mass in the form of these envelopes is $f_{\rm env}= 7.9$%, which is the product of the fraction of the IMF in stars with $M>20\;M_\odot$ (12.1%) and the average fraction of this stellar mass in the form of envelopes (65.1%). For $M_{\rm upper}=300\;M_\odot$, $f_{\rm
env}= 9.3$% (16.4% of the IMF is in $M>20\;M_\odot$ stars and 56.8% of that mass is in envelopes). Also for these two upper masses, the fraction of the total stellar mass in long-lived, low-mass stars ($M<0.8\;M_\odot$), is $f_{\rm LM}=
31.2$% and 29.7%, respectively. The formation of p-process elements and the delivery of these elements into the stellar envelopes and equatorial disks is assumed to proceed at a steady rate with an average timescale $t_{\rm evol}=3$ Myr, the lifetime of a high-mass star.
Regarding the contamination by p-processed elements, we note that the \[O/Na\] ratio in GCs varies from the large value of $\sim 0.4$ in G1 to the small value of $\sim
-1.4$ in G2, depending on the GC [$-1.4$ is for NGC 2802; @carretta06; @prantzos06; @gratton12]. The maximum value is about the same as in field stars and the minimum value is close what is expected in the envelope of a massive star near the end of its pre-main sequence phase [@prantzos06]. The range of observed values, 1.8 dex, is less than the change expected inside the nuclear burning regions of massive stars, which is 2.8 to 3.4 dex (depending on reaction rates) in [@dec07a]. This difference allows for some dilution of the core region with the envelope of the star before dispersal. The full range in today’s stars therefore extends from the presumed initial condition when the first generation formed, to a value that represents the near-complete conversion of a massive stellar envelope into one or more G2 stars. Values of \[O/Na\] between these extremes correspond to some combination of incomplete mixing of the processed debris with first generation gas, and partially processed gas from stars that have not yet finished their main sequence evolution [Scenarios II and I in @prantzos06].
The dilution proportions of processed stellar envelope gas and original cloud gas has also been estimated from the relative Li abundance in the G2 stars, considering that Li will be destroyed in the massive G1 stars. The observations suggest that up to 70% of the mass of a G2 star could come from the debris of G1 stars [@dec07b]. This high fraction limits the amount of accretion from the cloud envelope during the star formation process.
As mentioned above, the enriched gas is assumed to come from stellar equatorial disks [@prantzos06] and stellar debris generated by close interactions. For example, [@gaburov10] show that massive binaries that collide with another star can merge and puff up [see also @fregeau04], and that continued collisions with the third star before merging can shred the common envelope. Mass loss fractions of $\sim10$% are feasible in this situation. [@umbreit08] also consider the implications of collisions and suggest that the kinetic energy from collision-induced mass loss can clear gas away every few million years from aging globular clusters. This kinetic energy is comparable to that of a supernova [@gaburov10]. For the dense cloud cores considered here, that is not enough energy to clear away the gas, which is highly dissipative, but it could be enough for later stages of cluster formation after a significant amount of gas has turned into stars, and that is also when the supernovae themselves begin to clear the clusters of residual gas. We assume here that all of the stellar debris, i.e., from massive-star equatorial disks [@prantzos06], massive star collisions, and Roche lobe overflow around massive binaries [@demink09], brings p-processed material from the envelopes of massive stars into the dense cloud core where it mixes with existing and newly accreted core gas on the turbulent crossing time ($\sim0.1$ Myr), and is incorporated into new stars. This process continues for $\sim3$ Myr, forming the whole final cluster, at which point supernovae begin to remove the remaining gas.
Model Equations {#modeleq}
---------------
Consistent with these assumptions, we consider an initial cloud core of mass $M_{\rm gas}(t=0)$ in which stars begin to form, and a continuous accretion of new cloud gas onto this core at a rate $R_{\rm acc}$. Stars are assumed to form in the core with a constant consumption time, $t_{\rm consume}$ (equal to $t_{\rm ff}$ divided by the efficiency per free fall time). The effect of varying the consumption time will be discussed below. The formation rates of low ($<0.8\;M_\odot$), intermediate ($0.8\;M_\odot-20\;M_\odot$) and high ($>20\;M_\odot$) mass stars are given by the star formation rate in the core multiplied by the fractions of the IMF in these three mass intervals: $$\begin{aligned}
dM_{\rm star,LM}/dt= f_{\rm LM}M_{\rm gas}(t)/t_{\rm consume}\label{e1}\\
dM_{\rm star,IM}/dt= f_{\rm IM}M_{\rm gas}(t)/t_{\rm consume}\label{e2}\\
dM_{\rm star,HM}/dt= f_{\rm HM}M_{\rm gas}(t)/t_{\rm consume},
\label{e3}\end{aligned}$$ where $f_{\rm LM}=0.312$, $f_{\rm IM}=0.567$, and $f_{\rm HM}=0.121$ for an IMF with a most massive star of $100\;M_\odot$, and where $f_{\rm LM}=0.297$, $f_{\rm
IM}=0.540$, and $f_{\rm HM}=0.164$ for an IMF with a most massive star of $300\;M_\odot$ (Sect. \[basicmodel\]). The total mass formed in these stars is $$M_{\rm star,LM}(t)=\int_0^t {\dot M}_{\rm star,LM} dt, ...$$ and so on for the other mass ranges. Here we use the notation ${\dot M}=dM/dt$. The total mass for all stars is $M_{\rm star}=M_{\rm star,LM}+M_{\rm star,IM}+M_{\rm
star,HM}$.
For the gas, we track the primordial and enriched gas masses separately. The primordial gas is a combination of what was originally in the cloud core plus what gets accreted after star formation begins, and it also includes the part of the stellar envelope debris that was not converted into p-processed elements. We assume that massive stars make p-process elements continuously and that mixing from rapid rotation puts these elements into the stellar envelops continuously. Thus we conceptually divide the envelope mass into an unprocessed fraction at the original primordial abundance, and a completely processed, or enriched, fraction at the abundance of fully processed material. The total envelope is a combination of these, giving a partially-processed elemental abundance that comes from the dilution of fully processed material by the mass that is still in an unprocessed form.
With these assumptions, the rate of change of the primordial (1st generation) gas mass in the cloud core, $M_{\rm gas,1}$, is the increase from stellar debris and envelope accretion minus what goes into stars, $$\begin{aligned}
{\dot M}_{\rm gas,1}(t)=\int_0^t \left({{f_{\rm debr}{\dot M}_{\rm star,HM}(t^\prime)}\over{t_{\rm evol}}}
\right)\left(1-f_{\rm p}(t^\prime)\right)\left[1-{{t-t^\prime}\over{t_{\rm evol}}}\right]
dt^\prime\label{md1}\\
+R_{\rm acc}(t)-(1-f_{\rm p}(t)){\dot M}_{\rm star}(t).
\nonumber\end{aligned}$$ The first term in the integral is from ejection of stellar debris (equatorial disks, collisional debris, Roche-lobe overflow). The term $f_{\rm debr}$ is the average fraction of the mass of a high-mass star that is in the envelope and can be ejected. It equals $0.651$ for $M_{\rm upper}=100\;M_\odot$ and $0.568$ for $M_{\rm
upper}=300\;M_\odot$ [@prantzos06]. Division by $t_{\rm evol}$ indicates that we assume this debris is ejected steadily over the evolution time of the massive star, nominally assumed to be $t_{\rm evol}=3$ Myr. The quantity $f_{\rm
p}(t^\prime)$ is the processed fraction in the gas at time $t^\prime$, and therefore also the processed fraction in stars that form at time $t^\prime$, assuming rapid mixing. Thus, $1-f_{\rm p}(t^\prime)$ is the unprocessed fraction of the mass of the star that previously formed at $t^\prime$. The last term, $1-([t-t^\prime]/t_{\rm evol})$, tracks the remaining unprocessed fraction in the stellar envelope as the concentration of processed material increases linearly with time. This linear increase assumes the p-process elements from the stellar core mix into the stellar envelope at a steady rate. In addition to the integral that represents debris output, the unprocessed gas mass also increases by accretion at the rate $R_{\rm acc}$. From these additions we subtract the unprocessed gas mass in the cloud core that goes into stars.
The rate of change of processed gas mass in the cloud core, $M_{\rm gas,2}$ (2nd generation), is from the addition of stellar debris minus what goes into stars: $$\begin{aligned}
\label{md2}
{\dot M}_{\rm gas,2}(t)=\int_0^t \left({{f_{\rm debr}{\dot M}_{\rm star,HM}(t^\prime)}\over{t_{\rm evol}}}
\right)\left(1-f_{\rm p}(t^\prime)\right)\left({{t-t^\prime}\over{t_{\rm evol}}}\right)dt^\prime\\
\nonumber
+\int_0^t \left({{f_{\rm debr}{\dot M}_{\rm star,HM}(t^\prime)}\over{t_{\rm evol}}}
\right)f_{\rm p}(t^\prime) dt^\prime -f_{\rm p}(t){\dot M}_{\rm star}(t).\end{aligned}$$ The first integral represents the originally unprocessed mass fraction in the star when it formed, $(1-f_{\rm p}(t^\prime))$, that came out as debris and became more and more contaminated with time (as measured by $[t-t^\prime]/t_{\rm evol}$), and the second integral represents the return of originally processed mass (at fraction $f_{\rm p}$) into the cloud core. Note that the sum of the processed and unprocessed gas mass rates from equations (\[md1\]) and (\[md2\]) equals $\int_0^t \left[f_{\rm debr}M_{\rm HM}(t^\prime)/t_{\rm
evol}\right]dt^\prime+R_{\rm acc}-dM_{\rm star}/dt$, which is the total debris rate plus the accretion rate minus the star formation rate.
Now we determine the masses of primordial and enriched gas in the star-forming cloud core by integration, $$M_{\rm gas,1}(t)=\int_0^t {\dot M}_{\rm gas,1}(t)dt$$ $$M_{\rm gas,2}(t)=\int_0^t {\dot M}_{\rm gas,2}(t)dt,$$ we combine these to get the total gas mass, $$M_{\rm gas}=M_{\rm gas,1}+M_{\rm gas,2},$$ and we determine the mass fractions of enriched gas used above. $$f_{\rm p}(t)={{M_{\rm gas,2}}\over{M_{\rm gas}}}.
\label{fp}$$
The low mass stars do not contribute to the above equations except as a long-term sink for stellar mass. However, these stars are important for the observation of GCs today, and this is where stellar ejection and evaporation come in. We assume here that only low and intermediate mass stars leave the cluster by these processes, and we trace only the low mass stellar loss because the intermediate mass stars will have disappeared by now anyway, except as residual collapsed remnants. The high mass stars are assumed to segregate to the center of the GC where essentially all of their p-processes elements are available for gas contamination, as written in the above equations. Thus we need to model the escape of low mass stars.
The discussion in Section \[intro\] suggests that multi-star interactions and gas motions in the GC core occasionally accelerate low mass stars up to escape speed or beyond. Thus the rate of stellar ejection depends on the dynamical rate in the core, and this is directly proportional to both the free-fall rate and the consumption rate in the basic model. This implies that there is an additional rate of change of the mass of low-mass stars, so equation (\[e1\]) should be revised to contain an additional term, $$dM_{\rm star,LM}/dt= f_{\rm LM} M_{\rm gas}(t)/t_{\rm consume}-
f_{\rm eject} M_{\rm star,LM}/t_{\rm consume}
\label{eject}$$ for $f_{\rm eject}$ a number less than unity. Recall that $t_{\rm
consume}\sim10t_{\rm ff}$, so the ejection time for $f_{\rm eject}=0.4$ is 25 free fall times.
Analytical solution to a dimensionless model {#analyticalsol}
--------------------------------------------
The time dependence of the star formation rate and cloud mass have analytical solutions that are conveniently written using the above equations in dimensionless form. We normalize the gas and stellar masses to the initial cloud core mass, $M_{\rm gas}(t=0)$, and the time to the consumption time, $t_{\rm consume}$, assumed to be constant. Normalized quantities are denoted with a tilde. Then the star formation rate is $${{d{\tilde M}_{\rm star}}\over{d{\tilde t}}}={\tilde M}_{\rm gas}.
\label{dless1}$$ The gas mass changes from the addition of stellar debris from high mass stars and cloud envelope accretion and the subtraction of new stars: $${{d{\tilde M}_{\rm gas}}\over{d{\tilde t}}}={{f_{\rm debr}f_{\rm HM}}\over{{\tilde t}_{\rm evol}}}
\int_0^{\tilde t} {{{\tilde M}_{\rm star}(t^\prime)}\over{d{\tilde t}}}dt^\prime + {\tilde R}_{\rm acc}
-{{d{\tilde M}_{\rm star}}\over{d{\tilde t}}}.
\label{dless2}$$ Here, ${\tilde R}_{\rm acc}=R_{\rm acc}t_{\rm consume}/M_{gas}(t=0)$.
Equation (\[dless2\]) can be differentiated with respect to time and equation (\[dless1\]) can be substituted to give a second order linear differential equation, $${{d^2{\tilde M}_{\rm gas}}\over{d{\tilde t}^2}}+
{{d{\tilde M}_{\rm gas}}\over{d{\tilde t}}}-\gamma{\tilde M}_{\rm gas}=0$$ where $$\gamma={{f_{\rm debr}f_{\rm HM}t_{\rm consume}}\over{t_{\rm evol}}}
\label{secondorder}$$ is a dimensionless measure of the relative rate of return of processed gas. For the numbers in the basic model, $\gamma\sim10^{-2}$. The general solution of equation (\[secondorder\]) is the sum of ${\tilde M}_{\rm gas}\propto\exp\left({-\alpha
{\tilde t}}\right)$ with two values of $\alpha$, $$\alpha=0.5\left(1\pm\left[1+4\gamma\right]^{0.5}\right).$$ Using equation (\[dless2\]) again to fit $d{\tilde M}_{\rm gas}/d{\tilde t}$ at ${\tilde t}=0$, and defining $\Gamma=(1+4\gamma)^{0.5}$, we obtain the solution for normalized gas mass, $${\tilde M}_{\rm gas}({\tilde t})=\left({{\Gamma+0.5-{\tilde R}_{\rm acc}}\over{\Gamma}}\right)
e^{-0.5(1+\Gamma){\tilde t}}
+\left({{\Gamma-0.5+{\tilde R}_{\rm acc}}\over{\Gamma}}\right)e^{0.5(\Gamma-1){\tilde t}}.
\label{total}$$ The normalized stellar mass is the time integral over the normalized gas mass, or, ignoring stellar ejection, $${\tilde M}_{\rm stars}({\tilde t})=
\left({{\Gamma+0.5-{\tilde R}_{\rm acc}}\over{0.5\Gamma(1+\Gamma)}}\right)
\left(1-e^{-0.5(1+\Gamma){\tilde t}}\right)
+\left({{\Gamma-0.5+{\tilde R}_{\rm acc}}\over{0.5\Gamma(\Gamma-1)}}\right)
\left(e^{0.5(\Gamma-1){\tilde t}}-1\right).$$ This solution consists of an exponentially decaying initial burst of star formation and gas depletion with a timescale approximately equal to the consumption time ($0.5(1+\Gamma)\sim1$ in normalized units), followed by a slowly growing gas mass and star formation rate as accretion and stellar debris add to the cloud core (with growth rate $0.5(\Gamma-1)\sim\gamma<<1)$.
Results
=======
Figure \[walch16sfr2gasmass\] shows numerical solutions to equations (\[e1\]) to (\[fp\]) with mass and time normalized as above. The three curves are for different normalized accretion rates, as indicated by their colors and the labels for ${\tilde R}_{\rm acc}$. The solutions have the parameters discussed above for an IMF extending out to $300\;M_\odot$, i.e., $f_{\rm HM}=0.164$ and $f_{\rm
debr}=0.568$, and they assume $t_{\rm consume}/t_{\rm evol}=0.1$, which together make $\gamma=0.0093$. The top left panel shows the cloud core gas mass as a function of time and the top right panel shows the mass of low mass stars without (dashed curves) and with (solid curves) ejection assuming $f_{\rm eject}=0.4$. The lower left panel shows the processed fraction in the cloud core, which is also the processed fraction in the stars that form at that time. The analytical solutions to the gas and stellar masses fit exactly on top of the numerical solutions and are not shown in the figure. However, the top left panel has each exponential function from the analytical solution plotted separately as a dotted line; the total solution is the sum of these, as in equation (\[total\]).
The lower right panel shows the distribution of the processed fraction, $f_{\rm
p}$, among low-mass stars after star formation is assumed to stop, which is at the time $t_{\rm evol}$ in these models to avoid supernova contamination. All three cases assume stellar ejection with $f_{\rm eject}=0.4$. Recall that in observed GCs, $f_{\rm p}$ can range up to 0.7 from the Li abundance [@dec07b], and that is how far the blue histogram goes (which is for ${\tilde R}_{\rm acc}=0$). Also from observations, the ratio of stellar mass in the second generation to total stellar mass ranges from $\sim0.5$ to $\sim0.8$. For the 3 histograms in Figure \[walch16sfr2gasmass\], the ratios of the masses in all but the lowest-$f_{\rm
p}$ interval (i.e., the G2 stars) to the sum of all the masses (the G1+G2 stars) is about the same as the observations: 0.63 for the red curve (${\tilde R}_{\rm
acc}=0.03$), 0.50 for the green curve (${\tilde R}_{\rm acc}=0.003$) and 0.43 for the blue curve (${\tilde R}_{\rm acc}=0$). The processed fraction $f_{\rm p}$ increases with time as more and more massive stellar envelopes with their ever-increasing p-process contaminations disperse inside the cloud core (lower left panel). The maximum value of $f_{\rm p}$ at the end of the star formation time (right-hand limit to the curves in the lower-left panel) increases with decreasing ${\tilde R}_{\rm acc}$ because there is less dilution of the cloud core gas with pristine infall from the cloud envelope.
In equation (\[eject\]), ejection operates on all low-mass stars regardless of when they form, but it tends to remove a higher proportion of the first stars that form, i.e., weighted toward the G1 stars, than the later stars that form because the first stars have been exposed for the longest time to the multi-star interactions and gas motions that cause ejection. The top right panel of Figure \[walch16sfr2gasmass\] shows that ejection with the assumed $f_{\rm eject}=0.4$ reduces the mass in low mass stars by about a factor of 10. Without such stellar loss, the relative proportion of G2 stars shown in the lower right would be much less; i.e., the lowest $f_{\rm p}$ bin would be much higher compared to the others, although the $f_{\rm p}$ range in the histogram would be the same.
Related to this age-dependence of stellar ejection is a prediction from this model that the G1 stars should on average be less concentrated in the GC than the G2 stars. This is because the G1 stars have had more opportunities to absorb kinetic energy from multi-stellar interactions in the core than the G2 stars, regardless of whether the stars escape the cluster. Such a central concentration of G2 stars is observed [@gratton12].
The left-hand panel of Figure \[walch16g2g1\] shows the effect of stellar ejection through the parameter $f_{\rm eject}$ on the fraction of the mass in the G2 population. The blue curves assume the usual $t_{\rm consume}/t_{\rm evol}=0.1$ with a range of $f_{\rm eject}$ in equal steps between 0.1 and 0.7. The dashed blue curve with $f_{\rm eject}=0.3$ is one of this sequence but it is highlighted to contrast it with the red and green curves, which assume $f_{\rm eject}=0.3$ but also $t_{\rm consume}/t_{\rm evol}=0.05$ and 0.2, respectively, for the same value of $\gamma=f_{\rm debr}f_{\rm HM}t_{\rm consume}/t_{\rm evol}$ (i.e., $f_{\rm
debr}$ is $2\times$ larger and smaller when $t_{\rm evol}$ is $2\times$ larger and smaller, respectively, to keep $\gamma$ the same). There are three important dependencies shown in this figure: (1) the G2 mass fraction is low for both low accretion and high accretion rates, with a peak at ${\tilde R}_{\rm acc}\sim0.03$; (2) the G2 mass fraction increases with higher ejection parameter, and (3) the G2 mass fraction depends strongly on the ratio of the gas consumption time to the stellar evolution time.
The first result arises because high ${\tilde R}_{\rm acc}$ causes severe dilution of the p-processed material for later stars that form, and because low ${\tilde
R}_{\rm acc}$ causes most of the stars to form quickly and deplete the gas before p-processed elements have time to get into the cloud core. The second result is a consequence of ejecting more early-forming stars than late-forming stars because of the greater exposure of early-forming stars to time-changing gravitational forces.
The third result indicates the importance of the dimensionless parameter $t_{\rm
consume}/t_{\rm evol}$. A low value means there are more dynamical times available for stellar ejection before the supernova era begins (at $t=t_{\rm evol}$), so the first-forming stars become much more depleted compared to the last-forming stars. This increases $M_{\rm G2}/M_{\rm stars}$, as shown on the left of Figure \[walch16g2g1\]. It also means there are more massive stars at early times compared to late times, because of the more rapid drop in the star formation rate from faster gas consumption at small $t_{\rm consume}$. These greater numbers of early massive stars send relatively more p-processed matter into the smaller amount of remaining gas mass in the cloud core, increasing the maximum $f_{\rm p}$. These trends with lower $t_{\rm consume}/t_{\rm evol}$ occur when the density of the cloud core increases, because that lowers the free fall time and, correspondingly, $t_{\rm consume}$, at a fixed stellar evolution time. Thus we have the important result that [*p-process contamination is larger and involves a higher fraction of remaining long-lived stars in a globular cluster when the initial cluster gas density is higher.*]{} This may be the critical clue to the distinct origin of old globular clusters.
Figure \[walch16g2g1\] suggests that realistic results require a fairly high ejection parameter $f_{\rm eject}$, on the order of several tenths, which corresponds to a mean time before ejection of low mass stars equal to several tens of dynamical times ($=t_{\rm consume}/(t_{\rm ff}f_{\rm eject})$). [@fujii13] follow the ejection of stars of various masses from the collapsed core of a cluster owing to binary star interactions. The number of ejected high mass stars per cluster is fairly low, although it can account well enough for runaway OB stars in the field. The number of ejected low mass stars, which are of greatest interest in the present context, is much higher, following the stellar IMF (see their figure 1). They did not do a simulation that is exactly what we need, which would involve a lowest stellar mass less than $0.1\;M_\odot$ and a cluster more massive than $10^6\;M_\odot$. In their most relevant case, which had the lowest mass stars, the fraction of stars ejected was 0.4%. That is only for the binary star mechanism, however. Here we envision a much more dynamic environment in which gaseous mass motions from multi-star interactions and winds stochastically change the whole cluster core potential, as in simulations of dwarf galaxy nuclei [@elbadry16]. That second process, in addition to the possible presence of several dense cores in a proto-GC, each of which has one or more important massive binaries for stellar ejection, could possibly bring the ejection fraction per dynamical time up to the required value.
The dependence of relative G2 mass and processed fraction, $f_{\rm p}$, on $t_{\rm
consume}$ is shown in Figure \[walch16consume\], where we assume fixed $f_{\rm
debr}$, $f_{\rm HM}$, $f_{\rm eject}$, and $t_{\rm evol}$ and vary only $t_{\rm
consume}$. In this case the input parameters are dimensional, with $t_{\rm
consume}$ up to $\sim1.4$ Myr and $R_{\rm acc}=0$ (blue curves) and $4\times10^5\;M_\odot$ Myr$^{-1}$ (red curves). There are sharp decreases in the relative mass of the second generation (left panel) and the maximum processed fraction (right panel) as the consumption time increases, which corresponds to a decrease in cloud core density. Because the bulk peculiarity of old GCs is in these two quantities, it seems evident that [*the primary distinction between clusters that formed in the early universe and most of the main-disk clusters forming today is cloud-core density*]{}.
Conclusions {#conclu}
===========
We present a model of star formation in massive dense clusters that may be relevant to the formation of old globular clusters. This model has several key features that should cause a cluster to form stars with a wide range of p-processed elements, in agreement with observations. These features are:
- A $10^6-10^7\;M_\odot$ cloud core with a density of $\sim 10^6$ cm$^{-3}$, giving a free fall time of $\sim0.03$ Myr. The corresponding virial velocity is $\sim 80$ km s$^{-1}$, which is comparable to the turbulent speed in an L\* galaxy at the same redshift where the GC forms. Such a core is tightly bound and should form stars with an elevated star formation efficiency per unit free fall time, assumed here to be $\sim10$%. The free fall time and efficiency combine to give a gas consumption time that is much shorter than the evolution time of a high mass star. Then many generations of stars form in the gaseous debris of earlier generations.
- Interactions between massive stars at close range, including massive binary star interactions with single stars and massive stellar mergers, which lead to the swelling and dispersal of massive stellar envelopes and the dispersal of extruded equatorial disks around rapidly rotating massive stars. The dispersed envelope gas carries p-processed elements into the cloud core where it gets incorporated into new stars.
- Rapid stellar ejection driven by the time-changing gravitational potential of multi-star interactions and pressurized gas motions. The ejection rate is assumed to scale with the dynamical rate in the cluster. We point out that stars ejected from a cluster with a terminal velocity equal to only several tenths of the cloud core virial speed can move through the host galaxy at greater than the galaxy’s escape speed and thereby leave the host altogether. Such stellar loss may solve the problem of missing halo stars in Fornax and WLM, which have very low escape speeds.
- Star formation in the cloud core using gas that is a combination of original core gas, stellar debris that becomes more and more contaminated by p-processed elements over time, and newly accreted gas with the original abundances. Because the consumption time at high density is much less than the lifetime of a massive star, many generations of stars form before supernovae finally clear the gas away. We assume that each generation has a complete and normal IMF. We track the high mass stars for the production of p-processed elements, and the low mass stars for comparison to GCs today.
As a result of these assumptions, we produce GC stellar populations with approximately the observed range, mass, and radial distributions of p-process contamination. We find that the mass fraction of the final GC in the form of contaminated stars (i.e., the “2nd generation” fraction), and the maximum amount of p-process contamination in the late-forming stars, both increase strongly with cloud core density. This result suggests that the primary difference between old GCs with their significant amounts of p-process contamination and young super-massive star clusters that show little evidence for self-enrichment is that the cluster-forming clouds at high redshift have higher densities. This is to be expected because high redshift galaxies have larger gas fractions, higher gas surface densities and faster turbulent speeds than all but the most active galaxies today. As a result, the pressures in high-redshift galaxies are high, and the densities in the star-forming cloud cores should be high too.
This paper benefited from interesting discussions with Dr. Stefanie Walch at an early stage of this investigation, and from comments by the referee.
Bally, J., & Zinnecker, H. 2005, AJ, 129 2281
Bastian, N., Lamers, H. J. G. L. M., de Mink, S.E., Longmore, S.N., Goodwin, S. P., & Gieles, M. 2013, MNRAS, 436, 2398 Bastian, N. Cabrera-Ziri, I., Davies, B., & Larsen, S.S. 2013, MNRAS, 436, 2852 Bastian N. & Strader J., 2014, MNRAS, 443, 3594 Bastian, N., Hollyhead, K., & Cabrera-Ziri, I. 2014, MNRAS, 445, 378 Bastian, N. 2015, in IAUS 316, arXiv.1510.01330
Bate, M.R., & Bonnell, I.A. 2005, MNRAS, 356, 1201
Bekki, K., & Tsujimoto, T. 2016, ApJ, 831, 70
Bonifacio, P., Pasquini, L., Molaro, P., et al. 2007, A&A, 470, 153
Cabrera-Ziri I., Bastian N., Davies B., Magris G., Bruzual G., Schweizer F., 2014, MNRAS, 441, 2754 Cabrera-Ziri, I., Niederhofer, F., Bastian, N., et al. 2016, MNRAS, 459, 4218
Carretta E. 2006, AJ, 131, 1766
Carretta, E., Bragaglia, A., Gratton, R. G., Lucatello, S., & D’Orazi, V. 2012, ApJL, 750, L14
Carretta E., 2014, ApJ, 795, L28
Carretta, E., Bragaglia, A., Gratton, R. G., D’Orazi, V., Lucatello, S., Momany, Y., Sollima, A., Bellazzini, M., Catanzaro, G., & Leone, F. 2014, A&A, 564A, 60
Charbonnel, C. 2005, in IAU Symp., 228, ed. Hill, V. Francois, P., & Primas, F., pp.347
Cottrell, P. L., & Da Costa, G. S. 1981, ApJ, 245, L79
Crowther, P. A., Schnurr, O., Hirschi, R., et al. 2010, MNRAS, 408, 731
Crowther, P. A., Caballero-Nieves, S. M., Bostroem, K. A., et al. 2016, MNRAS, 458, 624
Daddi, E., Bournaud, F., Walter, F., et al. 2010, ApJ, 713, 686
Decressin, T., Meynet, G., Charbonnel, C., Prantzos, N., & Ekström, S. 2007, A&A, 464, 1029
Decressin T., Charbonnel C., & Meynet G. 2007b, A&A, 475, 859
de Mink S. E., Pols O. R., Langer N., & Izzard R.G., 2009, A&A, 507, L1
Denissenkov, P. A., & Denissenkova, S. N. 1990, Sov. A Lett., 16, 275
Denissenkov, P. A., & Hartwick, F. D. A. 2014, MNRAS, 437, 21
Denissenkov, P. A., VandenBerg, D. A., Hartwick, F. D. A., et al. 2015, MNRAS, 448, 3314
D’Ercole A., Vesperini E., D’Antona F., McMillan S.L.W., & Recchi S., 2008, MNRAS, 391, 825
D’Ercole, A., D’Antona, F., & Vesperini, E.2016, MNRAS, 461, 4088
Dickens, R. J., Croke, B. F. W., Cannon, R.D., & Bell, R. A. 1991, Nature, 351, 212
Ebisuzaki, T., Makino, J., Tsuru, T.Go., Funato, Y., Portegies Zwart, S., Hut, P., McMillan, S., Matsushita, S., Matsumoto, H., & Kawabe, R, 2001, ApJ, 562, L19
El-Badry, K., Wetzel, A., Geha, M., Hopkins, P., Kereš, D., Chan, T., & Faucher-Giguère, C.-A. 2016, ApJ, 820, 131
Elmegreen, B.G., Malhotra, S., Rhoads, J.ApJ, 757, 9
Förster Schreiber, N. M., Genzel, R., Bouché, N., et al. 2009, ApJ, 706, 1364
Fregeau, J. M., Cheung, P., Portegies Zwart, S. F., Rasio, F. A. 2004, MNRAS, 352, 1
Fujii, M. S., & Portegies Zwart, S. 2013, MNRAS, 430, 1018
Gaburov, E., Lombardi, J.C., Jr., & Portegies Zwart, S. 2010, MNRAS, 402, 105
Genzel, R., Tacconi, L. J., Gracia-Carpio, J., et al. 2010, MNRAS, 407, 2091
Gieles, M. & Renaud, F. 2016, MNRAS, 463, L103
Ginsburg, A., Goss, W. M., Goddi, C., et al. 2016, A&A, 595A, 27
Governato, F., Zolotov, A., Pontzen, A., Christensen, C., Oh, S. H., Brooks, A. M., Quinn, T., Shen, S., & Wadsley, J. 2012, MNRAS, 422, 1231
Gratton R. G., Sneden C., Carretta E., & Bragaglia A. 2000, A&A, 354, 169
Gratton, R., Bonifacio, P., Bragaglia, A., et al. 2001, A&A, 369, 87
Gratton, R., Sneden, C., & Carretta, E. 2004, ARA&A, 42, 385
Gratton, R.G., Carretta, E., Bragaglia, A. 2012, A&A Rev., 20, 50
Harris,W. E. & Racine, R. 1979, ARA&A, 17, 241
Heger A., Fryer C. L., Woosley S. E., Langer N., Hartmann D. H., 2003, ApJ, 591, 288
Herrera, C. N., Boulanger, F., Nesvadba, N.P.H., & Falgarone, E. 2012, A&A, 538, L9
Hopkins, P.F., Murray, N., Quataert, E., Thompson, T.A. 2010, MNRAS, 401, L19
Johnson, K. E., Leroy, A. K., Indebetouw, R., Brogan, C.L., Whitmore, B.C., Hibbard, J., Sheth, K., & Evans, A. S. 2015, ApJ, 806, 35
Khalaj, P., & Baumgardt, H. 2016, MNRAS, 457, 479
Klessen, R.S., Burkert, A., Bate, M.R. 1998, ApJ, 501, L205
Krause, M.G.H., Charbonnel, C., Bastian, N., & Diehl, R. 2016, A&A, 587A, 53
Krumholz, M.R. & Tan, J.C. 2007, ApJ, 654, 304
Langer, G. E., Hoffman, R., & Sneden, C. 1993, PASP, 105, 301
Larsen S.S., Brodie J. P., Strader J., 2012, A&A, 546, A53
Larsen S. S., Brodie J. P., Forbes D. A., Strader J., 2014, A&A, 565, A98
Leaman, R., Venn, K.A., Brooks, A.M., Battaglia, G., Cole, A.A., Ibata, R.A., Irwin, M.J., McConnachie, A.W., Mendel, J.T., & Tolstoy, E. 2012, ApJ, 750, 33
Li, C., de Grijs, R., Deng, L., Geller, A.M., Xin, Y., Hu, Y., & Faucher-Giguère, C.-A. 2016, Nature, 529, 502
Lin, Y., Liu, H.B., Li, D. et al. 2016, ApJ, 828, 32
Marino, A.F., Villanova, S., Milone, A. P., Piotto, G., Lind, K., Geisler, D., & Stetson, P. B. 2011, ApJ, 730, L16
Martell, S. L., Smolinski, J. P., Beers, T. C., Grebel, E. K. 2011, A&A, 534A, 136
Matzner, C.D., Jumper, P.H. 2015, ApJ, 815, 68
McLaughlin, D.E. & Fall, S. M. 2008, ApJ, 679, 1272
McMillan, S. L. W., Vesperini, E., & Portegies Zwart, S. F. 2007, ApJL, 655, L45
Myers P. C. 2009, ApJ, 700, 1609
Ostriker, Eve C. 1999, ApJ, 513, 252
Paresce, F. & De Marchi, G. 2000, ApJ, 534, 870
Pasquini, L., Bonifacio, P., Molaro, P., et al. 2005, A&A, 441, 549
Prantzos, N., & Charbonnel, C. 2006, A&A, 458, 135
Reipurth B.. & Clarke C. 2001, AJ, 122, 432.
Renzini A., 2013, Mem. Soc. Astron. Ital., 84, 162
Renzini, A., D’Antona, F., Cassisi, S. et al. 2015, MNRAS, 454, 4197
Simpson, J.D., Martell, S.L., Navin, C.A. 2017, MNRAS, 465, 1123
Smith, L. J., Crowther, P. A., Calzetti, D., Sidoli, F. 2016, ApJ, 823, 38
Tacconi, L. J., Genzel, R., Neri, R. et al. 2010, Nature, 463, 781
Umbreit, S., Chatterjee, S., Rasio, F.A. 2008, ApJ, 680, L113
Walker, M.G., Mateo, M., Olszewski, E.W., Bernstein, R., Wang, X., & Woodroofe, M. 2006, AJ, 131, 2114
Walker, D. L., Longmore, S. N., Bastian, N., Kruijssen, J. M. D., Rathborne, J. M., Galván-Madrid, R., & Liu, H. B. 2016, MNRAS, 457, 4536
Wunsch, R., Palous, J., Tenorio-Tagle, G., Munoz-Tunon, C., & Ehlerova, S. 2015, IAU General Assembly, Meeting 29, id.2254808
Yong D., Grundahl F., Norris J. E., 2014, MNRAS, 13, 3319
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We exhibit the first examples of contact structures on $S^{2n-1}$ with $n\geq 4$ and on $S^3\times S^2$, all equipped with their standard smooth structures, for which every Reeb flow has positive topological entropy. As a new technical tool for the study of the volume growth of Reeb flows we introduce the notion of algebraic growth of wrapped Floer homology. Its power stems from its stability under several geometric operations on Liouville domains.'
address:
- |
Institut de mathématiques\
Université de Neuchâtel
- |
Department of Mathematics\
Universität Heidelberg
author:
- 'Marcelo R.R. Alves'
- Matthias Meiwes
bibliography:
- 'References.bib'
date:
-
-
title: 'Dynamically exotic contact spheres in dimensions $\geq 7$'
---
Introduction
============
On a contact manifold there exists a natural class of flows, the so-called Reeb flows. Although the dynamics of distinct Reeb flows on the same contact manifold can be very different, there are dynamical properties which are common to all Reeb flows on a given contact manifold. For instance, the combined works of Hofer [@Hofer] and Taubes [@Taubes] imply that on a closed contact 3-manifold all Reeb flows have at least one periodic orbit. In this paper we construct a large class of contact manifolds on which all Reeb flows have chaotic dynamics. Surprisingly, some of the contact manifolds we construct have a very simple topology, which contrasts with the complicated dynamics of their Reeb flows.
A contact structure is said to have positive entropy if all Reeb flows associated to this contact structure have positive topological entropy. We show that there exist contact structures with positive entropy on spheres of dimension $\geq7$ and on $S^3\times S^2$. As a consequence we prove that every manifold of dimension $\geq 7$ that admits an exactly fillable contact structure also admits a (possibly different) contact structure with positive entropy. Our approach to prove these results is based on wrapped Floer homology and uses in an essential way its product structure. This product structure enables us to define the notion of algebraic growth of wrapped Floer homology, and we relate this growth to the volume growth of Reeb flows. Even though the richer algebraic structures in Floer homology were studied extensively, so far they lead to only very few applications in dynamics: the ones we are aware of are Viterbo’s result [@Viterbo1999] on the existence of one closed Reeb orbit on hypersurfaces of restricted contact type in Liouville domains with vanishing symplectic homology, and Ritter’s result [@Ritter2013] on the existence of Reeb chords for exactly fillable Legendrian submanifolds on Liouville domains with vanishing symplectic homology.
Basic notions
-------------
An important measure of the complexity of a dynamical system on a manifold $M$ is the topological entropy $h_{top}$ which quantifies in a single number the exponential complexity of the system. We refer the reader to [@Hasselblat-Katok] for the definition and basic properties of $h_{{\mathrm{top}}}$. By deep results of Yomdin and Newhouse, $h_{{\mathrm{top}}}(\phi)$ for a $C^{\infty}$-flow $\phi = (\phi^t)_{t\in{\mathbbm{R}}}$ equals the exponential growth rate of volume $$v(\phi) = \sup_{N\subset M} v(\phi,N), \text{ where }$$ $$\label{Yomdin}
v(\phi, N) = \limsup_{t\to \infty} \frac{\log {\operatorname{Vol}}_g^n(\phi^t(N))}{t}.$$ Here, $n = \dim N$, the supremum is taken over all submanifolds $N \subset M$, and ${\operatorname{Vol}}^n_g$ is the $n$-dimensional volume with respect to some Riemannian metric $g$ on $M$.
In this paper we study the topological entropy for Reeb flows of contact manifolds. Recall that a *(co-oriented) contact manifold* $(\Sigma, \xi)$ is a compact odd-dimensional manifold $\Sigma^{2n-1}$ equipped with a contact structure $\xi$, that is, a hyperplane distribution on $\Sigma$ which is given by $\xi = \ker \alpha$ for a $1$-form $\alpha$ with $\alpha \wedge (d\alpha)^{n-1} \neq 0$. Such an $\alpha$ is called a *contact form* on $(\Sigma,\xi)$, and we can associate to it the *Reeb vector field* $X_{\alpha}$ defined by $\iota_{X_{\alpha}}d\alpha = 0$, $\alpha(X_{\alpha})= 1$. Denote the flow of $X_\alpha$, the *Reeb flow* of $\alpha$, by $\phi_{\alpha}= (\phi_{\alpha}^t)_{t\in {\mathbbm{R}}}$. An *isotropic submanifold* of $\Sigma^{2n-1}$ is one whose tangent space is contained in $\xi$; isotropic submanifolds of dimension $n-1$ are called *Legendrian submanifolds*.
Main results
------------
The main result of this paper is the existence of contact structures with positive entropy on high dimensional manifolds.
\[spheres\]
- \[itm:spheres\] Let $S^{2n-1}$ be the $(2n-1)$ - dimensional sphere with its standard smooth structure. For $n\geq 4$ there exists a contact structure on $S^{2n-1}$ with positive entropy.
- \[itm:S3S2\] There exists a contact structure on $S^3\times S^2$ with positive entropy.
Recall that a contact manifold is said to be exactly fillable if it is the boundary of a Liouville domain. From Theorem \[spheres\] and the methods developed in this paper we obtain the following more general result.
\[maincorollary\]
- If $V$ is a manifold of dimension $2n-1\geq 7$ that admits an exactly fillable contact structure, then $V$ admits a contact structure with positive entropy.
- If $V$ is a $5$-manifold that admits an exactly fillable contact structure, then the connected sum $V \# (S^3\times S^2)$ admits a contact structure with positive entropy.
Note that the standard contact structure on spheres as well as the canonical contact structure on $S^{*}S^3 \cong S^3 \times S^2$ have a contact form with periodic Reeb flow. In particular these are not diffeomorphic to the contact structures in Theorem \[spheres\]. Other exotic contact spheres have been constructed by several authors, see [@Eliashberg1991; @Ustilovsky1999; @GeigesDing2004; @McLean2011]. The contact spheres constructed in this paper are, from our perspective, the “most exotic” ones. From the dynamical point of view they are the most remote from the standard contact spheres since they admit Legendrian submanifolds that have exponential volume growth under every Reeb flow. It would be interesting to relate our examples of exotic contact spheres to others that were constructed so far.
In order to explain further the relevance of these results we recall what is known about the topological entropy of Reeb flows. Motivated by results on topological entropy for geodesic flows (see [@Paternain]), combined with the geometric ideas of [@FrauenfelderSchlenk2006], Macarini and Schlenk proved in [@MacariniSchlenk2011] that for various manifolds $Q$ the unit cotangent bundle $(S^{*}Q, \xi)$ equipped with the canonical contact structure $\xi$ has positive entropy[^1].
In previous works of the first author, different examples of contact 3-manifolds with positive entropy were discovered. In [@Alves-Cylindrical; @Alves-Anosov; @Alves-Legendrian] it was shown that contact 3-manifolds with positive entropy exist in abundance: there exist hyperbolic contact 3-manifolds with positive entropy (see also [@ACH]), non-fillable contact 3-manifolds with positive entropy, and even 3-manifolds which admit infinitely many non-diffeomorphic contact structures with positive entropy. This shows that the class of contact manifolds with positive entropy is much larger than the class of unit cotangent bundles over surfaces with positive entropy, which were studied in [@MacariniSchlenk2011]. One common feature of all known examples of contact 3-manifolds with positive entropy is that the fundamental group of the underlying smooth 3-manifold has exponential growth. We expect this to be always the case:
\[conjecture3dim\] If a contact 3-manifold $(\Sigma,\xi)$ has positive entropy, then $\pi_1(\Sigma)$ grows exponentially.
Already from the unit cotangent bundles of simply connected rationally hyperbolic manifolds, which were considered in [@MacariniSchlenk2011], we know that Conjecture \[conjecture3dim\] is false in higher dimensions. However it is natural to ask if there are restrictions on the smooth topology of contact manifolds with positive entropy in higher dimensions. Theorem \[spheres\] shows that in contrast to what happens in dimension three, the phenomenon in higher dimensions is quite flexible from the topological point of view.
[Examples of contact manifolds of dimension $\geq 9$ which have positive entropy and are not unit cotangent bundles are also constructed using connected sums in an ongoing work of the first author and Macarini [@connected], following an idea of Schlenk. However, these contact manifolds have very complicated smooth topology, in the sense that the underlying smooth manifolds are rationally hyperbolic. For this reason they are much less surprising than the ones obtained in the present paper.]{}
Let us now explain our approach to establishing these results.
Symplectic and algebraic growth
-------------------------------
To establish our results we introduce the notion of algebraic growth of wrapped Floer homology. This notion is useful because, on one hand, it gives a lower bound for the growth rate of wrapped Floer homology defined using its action filtration and, on the other hand, it is stable under several geometric modifications of Liouville domains.
The contact manifolds we consider in this paper arise as boundaries of Liouville domains. Recall that a Liouville domain $M=(Y,\omega, \lambda)$ is an exact symplectic manifold $(Y,\omega)$ with boundary $\Sigma = \partial Y$ and a primitive $\lambda$ of $\omega$ such that $\alpha_M := \lambda|_{\Sigma}$ is a contact form on $\Sigma$: we let $ \xi_{M}= \ker \alpha_M$ be the contact structure induced by $M$ on $\Sigma$. For two exact Lagrangians $L_0$ and $L_1$ in $M$ that are asymptotically conical, i.e. conical near $\partial Y$ with Legendrian boundaries $\Lambda_0$ and $\Lambda_1$ in $(\Sigma,\xi_M)$, we consider the wrapped Floer homology of $(M,L_0,L_1)$ with ${\mathbbm{Z}}_2$-coefficients denoted by ${\mathrm{HW}}(M,L_0 \to L_1)$, whose underlying chain complex is, informally speaking, generated by Reeb chords from $\Lambda_0$ to $\Lambda_1$ and intersections of $L_0$ and $L_1$. We write ${\mathrm{HW}}(M,L)$ for ${\mathrm{HW}}(M,L \to L)$, see Section \[subsec:def\].
Results on positive entropy can be obtained from the exponential **symplectic growth** of wrapped Floer homology, which is defined as follows. By considering only critical points below an action value $a$, one obtains the filtered Floer homology ${\mathrm{HW}}^a(M,L_0 \to L_1)$. The homologies ${\mathrm{HW}}^a(M,L_0 \to L_1)$ form a natural filtration of ${\mathrm{HW}}(M,L_0 \to L_1)$, and they come with natural maps $\iota_a: {\mathrm{HW}}^a(M,L_0 \to L_1) \rightarrow {\mathrm{HW}}(M,L_0 \to L_1)$ into the (unfiltered) Floer homology. The *exponential symplectic growth rate* ${\Gamma}^{{\mathrm{symp}}}(M,L_0 \to L_1)$ of ${\mathrm{HW}}(M,L_0 \to L_1)$ is given by $$\begin{aligned}
\label{Gamma_symp}
\Gamma^{{\mathrm{symp}}}(M,L_0 \to L_1) = \limsup_{a \to \infty} \frac{\log (\dim {\mathrm{Im}}\ \iota_a)}{a};\end{aligned}$$ see Section \[subsec:def\] and Definition \[defi:Gamma\_symp\]. Since the generators of ${\mathrm{HW}}(M,L_0 \to L_1)$ correspond essentially to Reeb chords from $\Lambda_0$ to $\Lambda_1$, the symplectic growth gives a lower bound on the growth of Reeb chords with respect to their action. Assuming that $\Lambda_1$ is a sphere, we adapt the ideas of the first author in [@Alves-Legendrian] to get lower bounds for the volume growth $v(\phi_{\alpha},\Lambda_0)$ in terms of the exponential symplectic growth rate of ${\mathrm{HW}}(M,L_0,L_1)$ for every contact form $\alpha$ on $\xi_{M}$.
A *topological operation* on a Liouville domain $M$ is a recipe for producing a new Liouville domain $N$ from $M$. To obtain examples of contact manifolds with positive entropy we perform certain topological operations on Liouville domains. The operations we consider are: attaching symplectic handles on $M$ and, in the case $M$ is the unit disk bundle of a manifold, plumbing $M$ with the unit disk bundle of another manifold. Although one can understand the change or invariance of the (unfiltered) wrapped Floer homology under these operations, it is often much harder or not even possible to understand the effect of these operations on the symplectic growth. For instance, by an adaptation of a theorem of Cieliebak [@Cieliebak2001] we show that ${\mathrm{HW}}({M}',L)$ is isomorphic to ${\mathrm{HW}}(M,L\cap M)$ if ${M}'$ is obtained by subcritical handle attachment on $M$ (Theorem \[Viterbo\_iso\]). By contrast it is much harder to control the filtered Floer homology under this operation, see [@McLean2011] for an approach in the case of symplectic homology. In the case of plumbings of two cotangent bundles the computational results of a relevant part of the unfiltered wrapped Floer homology obtained by [@AbouzaidSmith2012] do not carry over to the symplectic growth rates of the plumbing. To overcome this difficulty we look at a notion of growth that is defined purely in terms the algebraic structure on wrapped Floer homology, the **algebraic growth**. Let us explain this briefly. Let $A$ be a (not necessarily unital) $K$-algebra with multiplication $\star$ and $S \subset A$ a finite set of elements of $A$. Given $j\geq 0$, let $N_S(j) = \{ a \in A \, \mid \, a = s_1\star s_2 \star \cdots \star s_j; \, s_1, \dots, s_j \in S \}$; i.e. $N_S(j)$ is the set of elements of $A$ that can be written as a product of $j$ not necessarily distinct elements of $S$. We define $W_S(n) \subset A$ to be the smallest $K$-vector space that contains the union $\bigcup_{j=1}^{n} N_S(j)$. The *exponential algebraic growth rate* of the pair $(A,S)$ is defined as $$\Gamma^{{\mathrm{alg}}}_S(A) = \limsup_{n\to \infty}\frac{1}{n}\log\dim_K W_S(n) \in [0,\infty).$$ In case $A = K \langle G \rangle$ is the group algebra over a finitely generated group $(G = \langle S \rangle ,\star) $, it is elementary to see that $\Gamma^{{\mathrm{alg}}}_S(A)$ coincides with the exponential algebraic growth of $G$ in the usual geometric group theoretical sense. Now, induced by the triangle product in Floer homology, ${\mathrm{HW}}(M,L)$ is equipped with a ring structure $\star$ turning it into a ${\mathbbm{Z}}_2$-algebra. Given a finite set $S$ of ${\mathrm{HW}}(M,L)$ we define (cf. Definition \[defi:growthHW\]) $\Gamma_{S}^{{\mathrm{alg}}}(M,L):= \Gamma_{S}^{{\mathrm{alg}}}({\mathrm{HW}}(M,L)) $. We say that ${\mathrm{HW}}(M,L)$ has *exponential algebraic growth* if there exists a finite subset $S$ of ${\mathrm{HW}}(M,L)$ such that $\Gamma_{S}^{{\mathrm{alg}}}(M,L) > 0$. Our main motivation for studying the exponential algebraic growth of ${\mathrm{HW}}$ is the following
\[prop:operations\] Let $M$ be a Liouville domain and $L$ be an asymptotically conical exact Lagrangian in it, and assume that ${\mathrm{HW}}(M, L)$ has exponential algebraic growth. Then we have:
- The Liouville domain ${M}'$ obtained by attaching subcritical handles to $M$ has exponential algebraic growth of ${\mathrm{HW}}$. More precisely, if the attachments are made away from $L$ (so that $L$ survives as an asymptotically conical exact Lagrangian submanifold of ${M}'$) then ${\mathrm{HW}}({M}',L)$ has exponential algebraic growth.
- If $M$ is the unit disk bundle of a closed orientable manifold $Q^n$ whose fundamental group grows exponentially, and ${M}'$ is obtained by a plumbing whose graph is a tree and one of the vertices is $M$, then ${M}'$ has exponential algebraic growth of ${\mathrm{HW}}$. More precisely, if $L_q$ is a unit disk fibre in $M$ and the plumbing is done away from $L_q$ then ${\mathrm{HW}}({M}',L_q)$ has exponential algebraic growth.
This result essentially says that plumbing and subcritical surgeries are topological operations that preserve exponential algebraic growth of ${\mathrm{HW}}$, and will allow us to construct many examples of Liouville domains which admit asymptotically conical exact Lagrangian disks with exponential algebraic growth of ${\mathrm{HW}}$.
The exponential algebraic growth of our examples stems from the algebraic growth of the homology of the based loop space $H_{*}(\Omega Q)$ equipped with the Pontrjagin product, where $Q$ is a compact manifold. In fact, we will only use the degree $0$ part whose algebraic growth is that of $\pi_1(Q)$.
The exponential algebraic growth of symplectic homology always vanishes since its product is commutative. Thus our approach is specifically designed for the open string case.
In order to obtain our main results we will bound the topological entropy of Reeb flows from below in terms of the algebraic growth of ${\mathrm{HW}}(M,L)$. For that we will use the crucial fact that the spectral number $c: {\mathrm{HW}}(M,L) \rightarrow {\mathbbm{R}}_+$ defined by $c(x) = \inf\{a \in {\mathbbm{R}}\, |\, x \in {\mathrm{Im}}\, i_a \}$ is subadditive, i.e. $c(x \star y) \leq c(x) + c(y)$ for all $x,y \in {\mathrm{HW}}(M,L)$. It follows (see Proposition \[prop:alg\_symp\]) that for any finite $ S \subset {\mathrm{HW}}(M,L)$ we have $$\label{algebraic_symplectic}
\Gamma^{{\mathrm{symp}}}(M,L) \geq \frac{1}{\rho(S)}\Gamma^{{\mathrm{alg}}}_S(M,L),$$ where $\rho(S) = \max_{s\in S} c(s)$. By using that ${\mathrm{HW}}(M,L \to L_1)$ is a module over $\left({\mathrm{HW}}(M,L), \star \right)$, this lower bound can be extended to $\Gamma^{{\mathrm{symp}}}(M,L \to L_1)$ for all $L_1$ that are exact Lagrangian isotopic to $L$, see Lemma \[lemmaprelim\]. In other words, exponential algebraic growth of ${\mathrm{HW}}(M,L)$ implies positive symplectic growth of ${\mathrm{HW}}(M,L \to L_1)$. This, combined with ideas from [@Alves-Legendrian], leads to
\[theorementropy\] Let $L$ be an asymptotically conical exact Lagrangian on a Liouville domain $M=(Y,\omega,\lambda)$, $\Sigma:= \partial Y$ and $ \alpha_{M}:= \lambda|_\Sigma$. We denote by $\xi_{M}:=\ker \alpha_{M})$ the contact structure induced by $M$ on $\Sigma$. Assume that there is a finite set $ S \subset {\mathrm{HW}}(M, L)$ such that $\Gamma^{{\mathrm{alg}}}_S(M,L) >0$ and that $\Lambda = \partial L$ is a sphere. Then, for every contact form $\alpha$ on $(\Sigma ,\xi_{M})$ the topological entropy of the Reeb flow $\phi_{\alpha}$ is positive. Moreover, if $\mathsf{f}_{\alpha}$ is the function such that $\mathsf{f}_{\alpha}\alpha_{M} = \alpha$ then $$h_{{\mathrm{top}}}(\phi_{\alpha}) \geq \frac{\Gamma^{{\mathrm{alg}}}_S(M,L)}{\rho(S) \max(\mathsf{f}_{\alpha})}.$$
Our paper is organised as follows. In section \[sec:Floer\] we consider the algebraic growth and the growth of filtered directed systems in general, and then we recall the definition of wrapped Floer homology together with its product structure. In section \[sec:Viterbo\] we present the construction of the Viterbo map and derive some of its properties. Section \[sec:entropy\] establishes implications of the growth properties of ${\mathrm{HW}}$ to topological entropy. In section \[sec:ring\_module\] we recall the computation of the algebra structure of the Floer homology of unit disk bundles and in section \[sec:top\_operations\] we give a proof of the invariance of ${\mathrm{HW}}$ under subcritical handle attachment, recollect a result on ${\mathrm{HW}}$ of plumbings and prove Proposition \[prop:operations\]. Finally, in section \[sec:constructions\], we construct our examples and prove the main theorems. The Appendix contains a construction of exact Lagrangian cobordisms used in the paper.
[Acknowledgement]{}: Most of this work was done when the second author visited the Université of Neuchâtel supported by the Erasmus mobility program, and the first author visited the Universität Münster supported by the SFB/TR 191. This work greatly benefited from discussions with Felix Schlenk and Peter Albers: we thank them for their interest in this work and their many suggestions. We also thank Lucas Dahinden for carefully reading the manuscript.
Wrapped Floer homology and its growth {#sec:Floer}
=====================================
As explained in the introduction, two features of wrapped Floer homology are crucial in this paper.
First, its natural filtration by action gives the wrapped Floer homology ${\mathrm{HW}}$ the structure of a filtered directed system and allows one to define the spectral value of elements of ${\mathrm{HW}}$. These give rise to the notion of symplectic growth[^2] of ${\mathrm{HW}}$; this is explained in Section \[sec:wrapped\].
Second, the product structure of ${\mathrm{HW}}$ gives it the structure of an algebra and gives rise to the notion of algebraic growth of ${\mathrm{HW}}$. This is explained in Section \[sec:productonHW\]. The link between these notions is given by the crucial fact that the spectral number is subadditive with respect to the product structure on ${\mathrm{HW}}$, see also Section \[sec:productonHW\].
We first recall the relevant algebraic notions and deduce some direct consequences.
Algebraic growth and growth of filtered directed systems
--------------------------------------------------------
Fix a field $K$. We use the convention that $\log(0) := 0$.
### Filtered directed systems and growth {#subsubsec:fds}
A *filtered directed system* over ${\mathbbm{R}}_{+} = [0,\infty)$ or for short *f.d.s.* is a pair $(V,\pi)$ where
- $V_t$, $t \in [0,\infty)$, are finite dimensional $K$-vector spaces.
- $\pi_{s \rightarrow t}: V_s \rightarrow V_t$, for $s\leq t$ are homomorphisms (*persistence homomorphisms*), such that $\pi_{s \rightarrow t}\circ\pi_{r \rightarrow s}=\pi_{r \rightarrow t}$ for $r\leq s\leq t$, and $\pi_{t\rightarrow t} = {id}_{V_t}$ for all $t \in {\mathbbm{R}}_{+}$.
Let $\mathfrak{J}$ be the smallest vector space of $\bigoplus_{t\in {\mathbbm{R}}_+} V_t$ containing $\bigcup_{s \leq t} \{\pi_{s\rightarrow t}(x_s) - x_s\}$. The *direct limit* ${\underrightarrow{\lim}\,}V$ of $V$ is defined by ${\underrightarrow{\lim}\,}V := \bigoplus_{t\in {\mathbbm{R}}_+} V_t/\mathfrak{J}$. The inclusions $V_t \hookrightarrow \bigoplus_{t\in {\mathbbm{R}}_+} V_t$ induce maps to ${\underrightarrow{\lim}\,}V$ which we denote by $i_t$. The *spectral number* $c_V$, or just $c$ if the context is clear, of an element $x \in {\underrightarrow{\lim}\,}V$ is $$c_V(x) := \inf\{t \in [0,\infty) \, | \, \exists x_t \in V_t \text{ such that } i_t(x_t) = x \}.$$ It is clear from the definition of $c_V$ that if $x_1,..,x_n \in V$ and $k_1,...,k_n\in K$ we have $$\label{spectralsuminequality}
c_V\bigg( \sum_{i=1}^n k_i x_i \bigg) \leq \max_{1 \leq i \leq n}{c_V(x_i)}.$$
\[fds\_growth\] Let $d_t^V := \dim \{x \, | \, c_V(x) \leq t\}$. The *exponential growth rate* of the f.d.s. $V$ is $$\widetilde{\Gamma}(V) := \limsup_{t \rightarrow \infty}\frac{1}{t} \log d^V_t.$$ We say that $V$ has exponential growth if $0 < \widetilde{\Gamma}(V) < \infty$.
A *morphism* between f.d.s. $(V,\pi)$ and $(V^{'},\pi^{'})$ is a collection of homomorphisms $f = (f_t)_{t\in[0, \infty)}$, $f_t : V_t \rightarrow V^{'}_t$, that are compatible with respect to the persistence homomorphisms: $$\label{morph}
f_t \circ \pi_{s \rightarrow t} = \pi^{'}_{s\rightarrow t}\circ f_s.$$ An *asymptotic morphism* is a collection of homomorphisms $f_t : V_t \rightarrow V^{'}_t$, $t\in (K, \infty)$, for some $K > 0$ such that holds for $K<s<t$.
Let $(V,\pi)$ be a f.d.s. and $\eta \geq 1$. We can dilate $V$ by $\eta$ to a filtered directed system $(V(\eta), \pi(\eta))$ given by $V({\eta})_t = V_{\eta t}$, $\pi(\eta)_{s\rightarrow t}= \pi_{\eta s \rightarrow \eta t}$. It follows that $\pi$ gives rise to a canonical morphism $\pi[\eta]:V \rightarrow V({\eta})$ by $\pi[\eta]_t = \pi_{t \to \eta t}$. For a morphism $f: V\rightarrow W$ we get a dilated morphism $f(\eta):V(\eta) \rightarrow W(\eta)$ by setting $f(\eta)_t = f_{\eta t}$.
\[interl\] Let $(V,\pi_V)$ and $(W,\pi_W)$ be f.d.s. We call them *$(\eta_1,\eta_2)$-interleaved*, or *interleaved*, if there are asymptotic morphisms $f:V\rightarrow W(\eta_1)$ and $g:W \rightarrow V(\eta_2)$ for two real numbers $\eta_1, \eta_2 \geq 1$ such that $$f(\eta_2)\circ g = \pi_{W}[\eta_1 \eta_2] \ \text{ and } \ g(\eta_1)\circ f = \pi_{V}[\eta_1 \eta_2].$$
The direct limits of interleaved f.d.s. are isomorphic. It is also easy to see the following
\[interl\_growth\] Let $V$ and $W$ be $(\eta_1, \eta_2)$-interleaved for some $\eta_1,\eta_2 \geq 1$. Then $$\widetilde{\Gamma}(V) \leq \eta_1 \widetilde{\Gamma}(W) \ \mbox{ and } \ \widetilde{\Gamma}(W) \leq \eta_2 \widetilde{\Gamma}(V).$$
The notion of interleaving comes from the theory of persistence modules (see [@PolterovichShelukhin2016] for applications of persistence modules and interleaving distance in symplectic geometry).
### Algebras and their algebraic growth {#subsec:algebra}
We recall from the introduction the definition of the algebraic growth of a $K$-algebra $A$ and a finite subset $S \subset A$. Given $j\geq 0$ let $N_S(j) = \{ a \in A \, \mid \, a = s_1\star s_2 \star \cdots \star s_j; \, s_1, \dots, s_j \in S \}$; i.e. $N_S(j)$ is the set of elements of $A$ that can be written as a product of $j$, not necessarily distinct, elements of $S$. We define $W_S(n) \subset A$ to be the smallest $K$-vector space that contains the union $\bigcup_{j=1}^{n} N_S(j)$. The exponential algebraic growth rate of the pair $(A,S)$ is defined as $$\Gamma^{{\mathrm{alg}}}_S(A) = \limsup_{n\to \infty}\frac{1}{n}\log\dim_K W(n) \in [0,\infty).$$ We will need the following definition.
\[streched\] Let $M$ be a module over an algebra $A$ with scalar multiplication denoted by $\ast$. The module $M$ is called *stretched* if there exists an element $m_0 \in M$ such that for all elements $a \neq 0 \in A$ we have $a\ast m_0 \neq 0$. An element $m_0 \in M$ satisfying this condition is called a *stretching* element.
In the following let $V$ be a filtered directed system and assume that the vector space $A= {\underrightarrow{\lim}\,}V$ has a $K$-algebra structure with multiplication $\star$. We do not assume that $A$ is finitely generated. Furthermore, let $W$ be a filtered directed system, such that $M= {\underrightarrow{\lim}\,}W$ is a module over $A$ with multiplication $\ast$, i.e. a module over $(A,\star)$ with scalar multiplication $\ast$ which is compatible with the $K$-vector space structure of $A$ and $M$.
Furthermore assume that the spectral numbers $c_V$ and $c_W$ are *subadditive* with respect to $\star$ and $\ast$, i.e. $$\begin{aligned}
\label{subadd_ring}
c_V(a \star b) \leq c_V(a) + c_V(b), \text{ for all } a,b \in A,\end{aligned}$$ and $$\begin{aligned}
\label{subadd_module}
c_W(a \ast m) \leq c_V(a) + c_W(m), \text{ for all }a\in A \text{ and } m\in M.\end{aligned}$$
\[finite\_growth\] Let $V$ be a f.d.s. such that $A= {\underrightarrow{\lim}\,}V$ has a $K$-algebra structure with multiplication $\star$, and assume that $c_V$ is subadditive with respect to $\star$. Then for every finite subset $S \subset A$ we have $$\widetilde{\Gamma}(V) \geq \frac{1}{\rho(S)}\Gamma^{{\mathrm{alg}}}_S(A),$$ where $\rho(S) = \max_{x\in S}c_V(x)$.
*Proof:* From the subadditivity of $c_V$ with respect to $\star$ it follows that if $a = s_1 \star s_2 \star \cdots \star s_n$, $s_i \in S$, we have $$c_V(a) = c_V(s_1 \star \cdots \star s_n) \leq c_V(s_1) + \cdots + c_V(s_n) \leq \rho(S) n.$$ It then follows from that $W(n) \subset \{x \in A \, \mid c(x) \leq \rho(S)n \}$. We thus conclude that $$\begin{split}
\Gamma^{{\mathrm{alg}}}_S(A) &= \limsup_{n\to \infty}\frac{1}{n}\log \dim W(n) \\
&\leq \limsup_{n \to \infty}\frac{1}{n}\log \, \dim \{x \mid c(x) \leq \rho(S) n\} \leq
\rho(S) \widetilde{\Gamma}(V).
\end{split}$$
\[mod\_growth\] Let $V$ and $W$ be f.d.s. and assume that the vector space $A= {\underrightarrow{\lim}\,}V$ has an $K$-algebra structure with multiplication $\star$, and that $M := {\underrightarrow{\lim}\,}W$ has the structure of a module over $A$ with multiplication $\ast$. Assume that $c_V$ and $c_W$ are subadditive with respect to $\star$ and $\ast$, respectively, and that $M \neq 0 $ is a stretched module over the algebra $A$. Then $$\label{firsteq}
\widetilde{\Gamma}(W) \geq \widetilde{\Gamma}(V).$$ Moreover, for every finite set $S \subset A$ we have $$\label{seceq}
\widetilde{\Gamma}(W) \geq \frac{1}{\rho(S)}\Gamma^{{\mathrm{alg}}}_S(A).$$
*Proof:* Take a stretching element $m_0 \neq 0$ in $M$. We have $a \ast m_0 \neq b \ast m_0$ for $a \neq b$, $a,b \in A$. In particular $a \mapsto a \ast m_0$ is an injective homomorphism from $A$ to $M$. Therefore, by , $d_t^V = \dim \{a \in A \, |\, c_V(a) \leq t \} \leq \dim \{m \in M \, | c_W(m) \leq t + c_W(m_0) \} = d_{t+c_W(m_0)}^W$, for all $t >0$. We then get $$\begin{split}
\widetilde{\Gamma}(V) &= \limsup_{t \rightarrow \infty} \frac{\log \, d_t^V}{t} \leq \limsup_{t \rightarrow \infty}
\frac{\log \, d^W_{t+c_W(m_0)}}{t} \\ &=
\limsup_{t \rightarrow \infty} \frac{\log \, d^W_{t+c_W(m_0)}}{t +c_W(m_0)} \frac{t+c_W(m_0)}{t} =
\widetilde{\Gamma}(W).
\end{split}$$ This proves . Inequality is obtained by combining with Lemma \[finite\_growth\].
In order to get results on entropy, we will need the following notions.
Let $\mathcal{W} = W(i)_{i \in I}$ be a family of f.d.s. with direct limits $M(i)$ that are modules over $A:= {\underrightarrow{\lim}\,}V$. We say that the family $M(i)_{i \in I}$ is *uniformly stretched* if there exists a constant $B\geq 0$ such that for every $i \in I$ there exists a stretching element $m_i \in M(i)$ with $c_{M(i)}(m_i) \leq B$.
\[defi:fam\_growth\] Let $\mathcal{W} = W(i)_{i \in I}$ be a family of filtered directed systems. The *uniform exponential growth rate* of $\mathcal{W}$ is $$\widetilde{\Gamma}_{i\in I}(\mathcal{W}) := \limsup_{t \rightarrow \infty} \frac{1}{t} \log \left(\inf_{I} d_t^{W(i)}\right).$$
\[mod\_fam\_growth\] Let $V$ be a f.d.s. such that $A= {\underrightarrow{\lim}\,}V$ has a $K$-algebra structure with multiplication $\star$. Let $\mathcal{W} = W(i)_{i \in I}$ be a family of f.d.s. such that for every $i\in I$ the direct limit $M(i)={\underrightarrow{\lim}\,}W(i)$ is a module over $A$ with multiplication $\ast(i)$. Assume that $c_V$ is subadditive with respect to $\star$, that $c_{W(i)}$ is subadditive with respect to $\ast(i)$ for every $i\in I$, and that the family $M(i)_{i \in I}$ is uniformly stretched over the algebra $A$. Then $$\widetilde{\Gamma}_{i\in I}(\mathcal{W}) \geq \widetilde{\Gamma} (V).$$
*Proof:* Since $M(i)_{i \in I}$ is uniformly stretched there exists $B>0$ such that for every $i \in I$, we can find a stretching element $m_i \in M(i)$ with $c_{M(i)}(m_i) \leq B$. Hence we have by that $d_t^V \leq \inf_{I}d^{W(i)}_{t+B}$ and the result is obtained as in the proof of Lemma \[mod\_growth\].
Wrapped Floer homology {#subsec:def}
----------------------
In the following we give the definition and conventions for wrapped Floer homology used in this paper. This Floer type homology theory appeared in [@AS-iso] for contangent bundles, and the case of general Liouville domains can be found in [@AbouzaidSeidel2010]. We refer to these papers and [@Ritter2013 Section 4] for more details.
### Liouville domains and Lagrangians
A *Liouville domain* $M=(Y,\omega, \lambda)$ is an exact symplectic manifold $(Y, \omega)$ with boundary $\Sigma = \partial Y$ and a primitive $\lambda$ of $\omega$ such that $\alpha_M = \lambda|_{\Sigma}$ is a contact form on $\Sigma$. The *Liouville vector field* X, is given by $i_X\omega = \lambda$ and points outwards along $\Sigma$. Using the flow of X, one can attach an infinite cone to $M$ along $\Sigma$ that gives the *completion* $\widehat{M}:=\left(\widehat{Y}, \widehat{\omega}, \widehat{\lambda}\right)$ of $M$ with $\widehat{Y} = Y \cup_{\Sigma} \left([1, \infty) \times \Sigma \right)$, $\widehat{\lambda}|_{Y} = \lambda$, $\widehat{\lambda}|_{[1, \infty) \times \Sigma} = r\alpha_M$, and $\widehat{\omega}=d \widehat{\lambda}$.
In order to simplify notation, we will usually write $M$ and $\widehat{M}$ instead of $Y$ and $\widehat{Y}$, respectively, as the domain of Hamiltonian functions or the target space of Floer trajectories. This does not cause any confusion since the smooth manifolds $Y$ and $\widehat{Y}$ are part of the data defining $M$ and $\widehat{M}$, respectively. Similarly, when we write $\Sigma=\partial M$ it should be understood as $\Sigma= \partial Y$.
Let $\mathsf{f}: \partial Y \rightarrow (0, \infty)$ be a smooth function. Let $Y_{\mathsf{f}} = \widehat{Y} \setminus \{(r,x) \, | \, r > \mathsf{f}(x), x \in \partial Y\}$. It is easy to see that $M_{\mathsf{f}} = (Y_{\mathsf{f}},\widehat{\omega}|_{Y_{\mathsf{f}}},\widehat{\lambda}|_{Y_{\mathsf{f}}})$ is a Liouville domain. For example, given $\delta >0$ we denote by $M_{1+\delta}$ the Liouville domain $(Y_{1+\delta},\omega_{1+\delta},\lambda_{1+\delta})$ embedded in $\widehat{M}$ defined by ${Y}_{1+\delta} = Y \cup_{\Sigma} \left([1, 1+\delta] \times \Sigma\right)$, $\omega_{1+\delta} = \widehat{\omega}|_{Y_{1+\delta}}$, $\lambda_{1+\delta} = \widehat{\lambda}|_{Y_{1+\delta}}$.
In our paper we only consider Liouville domains that have vanishing first chern class $c_1(M) \subset {\mathrm{H}}^2(M;{\mathbbm{Z}})$.
We consider Lagrangians $(L, \partial L)$ in $(M, \Sigma)$ that are exact, i.e. $\lambda|{L} = df$, and that satisfy $$\label{asympt}
\begin{split}
&\Lambda = \partial L \text{ is a Legendrian submanifold in } (\Sigma,\xi_M), \\
& L \cap [1-\epsilon, 1] \times \Sigma = [1-\epsilon, 1] \times \Lambda \text{ for a sufficiently small } \epsilon >0 .
\end{split}$$ We will call a Lagrangian that satisfies *asymptotically conical*. We can extend it naturally to an exact Lagrangian $\widehat{L} = L \cup_{\Lambda} ([1,\infty) \times \Lambda)$ in $\widehat{M}$. We will refer to a Lagrangian in $\widehat{M}$ of this form also as *asymptotically conical (with respect to $M$)*. More generally, given a subset $U \subset \widehat{M}$ we say that $L$ is *conical in $U$* if the Liouville vector field is tangent to $L\cap \mathrm{int}(U)$ in the interior $ \mathrm{int}(U)$ of $U$.
### Wrapped Floer homology {#sec:wrapped}
For two asymptotically conical exact Lagrangians $L_0$ and $L_1$ in ${M}$ denote by $\mathcal{P}_{L_0 \to L_1} = \{\gamma : [0,1] \rightarrow \widehat{M} \, |\, \gamma(0) \in \widehat{L}_0, \, \gamma(1) \in \widehat{L}_1\}$ the space of (smooth) paths from $\widehat{L}_0$ to $\widehat{L}_1$.
Denote by $X_{\alpha_M}$ the Reeb vector field on the boundary $(\Sigma, \xi_{M} = \ker{\alpha_M})$. A *Reeb chord of length $T$* of $\alpha_M$ from $\Lambda_0 = \partial L_0$ to $\Lambda_1 = \partial L_1$ is a map $\gamma: [0,T] \to \Sigma$ with $\dot{\gamma}(t) = X_{\alpha_M}(\gamma(t))$ with $\gamma(0) \in \Lambda_0$ and $\gamma(T) \in \Lambda_1$. Denote the set of Reeb chords of length $<T$ by $\mathcal{T}^{T}_{\Lambda_0 \to \Lambda_1}(\alpha_M)$, and the set of all Reeb chords by $\mathcal{T}_{\Lambda_0 \to \Lambda_1}(\alpha_M)$. The Reeb chord $\gamma$ of length $T$ of $\alpha_M$ from $\Lambda_0$ to $\Lambda_1$ is said to be *transverse* if the subspaces $T_{\gamma(1)}(\phi_{X_{\alpha_M}}^{T}(\Lambda_0))$ and $T_{\gamma(1)} \Lambda_1 $ of $T_{\gamma(1)}\Sigma$ intersect at only one point. The *spectrum* of the triple $(M,L_0 \to L_1)$, denoted by $\mathcal{S}(M,L_0 \to L_1)$, is the set of lengths of Reeb chords from $\Lambda_{0}$ to $\Lambda_{1}$ in $\Sigma$. It is a nowhere dense set in $[0,\infty)$.
Given a contact form $\alpha$ on $(\Sigma, \xi_{M})$ and a pair of Legendrian submanifolds $(\Lambda_0, \Lambda_1)$ on $(\Sigma, \xi_{M})$, we say that the triple $(\alpha,\Lambda_0 \to \Lambda_1)$ is *regular* if all Reeb chords of $\alpha$ from $\Lambda_0$ to $\Lambda_1$ are transverse. We say that $(M,L_0 \to L_1)$ is *regular* if $(\lambda_{\Sigma}, \Lambda_0 \to \Lambda_1 )$ is regular and $L_0$ and $L_1$ intersect transversely. From now on, we assume that for the contact form $\alpha_M$ induced by $M$ on $(\Sigma,\xi_M)$ the triple $(\alpha_M,\Lambda_0 \to \Lambda_1)$ is regular.
An autonomous Hamiltonians $H: \widehat{M} \to {\mathbbm{R}}$ is called *admissible* if
- $H < 0$ on $M$,
- and there exist constants $\mu>0$ and $b\leq - \mu$ such that $H(x,r) = h(r) = \mu r + b$ on $[1, \infty) \times \partial M$.
If $H: \widehat{M} \to {\mathbbm{R}}$ is admissible and satisfies $H(x,r) = \mu r + b$ on $[1, \infty) \times \partial M$ we say that $H$ is admissible with *slope* $\mu$ *(at infinity)*.
Define the action functional $\mathcal{A}^{L_0 \to L_1}_H = \mathcal{A}_H : \mathcal{P}_{L_0 \to L_1} \rightarrow {\mathbbm{R}}$ by $$\mathcal{A}_H(\gamma) = f_0(x(0)) - f_1(x(1)) + \int_{0}^{1} \gamma^{*}\lambda - \int_{0}^{1} H(\gamma(t))dt,$$ where $f_0$ and $f_1$ are functions on $L_0$ and $L_1$ respectively with $df_i = \lambda|_{\widehat{L}_i}, i= 0,1$. The critical points of $\mathcal{A}_H$ are Hamiltonian chords from $\widehat{L}_0$ to $\widehat{L}_1$ that reach $\widehat{L}_1$ at time $1$. We define $$\mathcal{T}_{L_0\to L_1}(H):={\mathrm{Crit\,}}{\mathcal{A}_H} = \{\gamma \in \mathcal{P}_{L_0 \to L_1} \mid \dot{\gamma}(t) = X_H(\gamma(t))\},$$ and write $\mathcal{T}_{L}(H)$ instead of $\mathcal{T}_{L \to L}(H)$. Here $X_H$ is the Hamiltonian vector field defined by $\omega(X_H, \cdot ) = -dH$. We call an admissible Hamiltonian *non-degenerate* for $L_0 \to L_1$ if all elements in $\mathcal{T}_{L_0\to L_1}(H)$ are non-degenerate, i.e. $\phi^1_{X_H}(\widehat{L_0})$ is transverse to $\widehat{L_1}$. Such a Hamiltonian must have slope $\mu \notin \mathcal{S}(M,L_0 \to L_1)$. Note that every admissible Hamiltonian can be made non-degenerate for $L_0 \to L_1$ after a generic perturbation ([@AbouzaidSeidel2010 Lemma 8.1]). We denote by $$\mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1)$$ the set of admissible Hamiltonians which are non-degenerate for $L_0 \to L_1$. For a Hamiltonian $H\in \mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1) $ all elements in $\mathcal{T}_{L_0 \to L_1}(H)$ have their image contained in $M$.
For admissible Hamiltonians $H$ with slope $\mu \notin \mathcal{S}$ that are constant in $M$ away from the boundary, depend on $r$ and increase sharply near $\partial M$, $\mathcal{T}_{L_0\to L_1}(H)$ corresponds to $\mathcal{T}^{\mu}_{\Lambda_0 \to \Lambda_1}(\alpha_M)$ and intersection points of $L_0$ and $L_1$ in $M$. If $(M,L_0 \to L_1)$ is regular, such Hamiltonians belong to the set $ \mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1)$.
If $(\alpha_M, \Lambda_0 \to \Lambda_1)$ is regular but $(M,L_0 \to L_1)$ is not, we can take $H \in \mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1)$ to be a $C^2$-small negative function away from the boundary of $M$, and to depend only on $r$ and increase sharply near $\partial M$. Then, $\mathcal{T}_{L_0\to L_1}(H)$ will correspond to $\mathcal{T}^{\mu}_{\Lambda_0 \to \Lambda_1}(\alpha_M)$ and intersection points of $L_0$ and $L_1$ in $M$ that are not destroyed by the Hamiltonian flow of $H$.
An almost complex structure $J$ on $\left((0,\infty) \times \partial M, \lambda = r\alpha_M \right)$ is called *cylindrical* if it preserves $\xi_{M}= \ker \alpha_M$, if $J|_{\xi_{M}}$ is independent of $r$ and compatible with $d(r\alpha_M)|_{\xi_{M}}$, and if $J X_{\alpha_M}=r\partial_{r} $. In the following we take almost complex structures $J$ on $\widehat{M}$ that are *asymptotically cylindrical*, i.e. cylindrical on $[r, \infty) \times \partial M$ for some $r>1$. The $L^2$-gradient of the action functional with respect to the Riemannian metric given by $d\lambda(J \cdot,\cdot) = g(\cdot, \cdot)$ is given by $$\nabla \mathcal{A}_H(\gamma) = -J(\gamma)\left(\partial_t \gamma - X_H(\gamma) \right),$$ and we interpret the negative gradient flow lines as Floer strips $$\begin{split}\label{Floer}
&u:{\mathbbm{R}}\times [0,1] \rightarrow \widehat{M}, \\
&\overline{\partial}_{J,H}(u)=\partial_s u + J(u)(\partial_t u - X_H(u) ) = 0, \\
&u(\cdot,0) \in \widehat{L}_0, \mbox{ and } u( \cdot ,1) \in \widehat{L}_1.
\end{split}$$ We define the moduli space of parametrized Floer strips connecting two critical points $x$ and $y$ of $\mathcal{A}_H$ $$\begin{split}
\widetilde{\mathcal{M}}(x,y, H, J) = \{ u: {\mathbbm{R}}\times [0,1] \rightarrow \widehat{M} \, | \, u \text{ satisfies \eqref{Floer} }, \lim_{s\rightarrow -\infty} = x \mbox{ and } \lim_{s\rightarrow +\infty} = y \}.
\end{split}$$ There is a natural ${\mathbbm{R}}$-action on $\mathcal{M}(x,y, H, J)$ coming from the translations in the domain. Letting $\widetilde{\mathcal{M}}^1(x,y, H, J) $ be the set of elements of $\widetilde{\mathcal{M}}(x,y, H, J) $ that have Fredholm index 1 we write $$\begin{split}
& \mathcal{M}^0(x,y, H, J) := \widetilde{\mathcal{M}}^1(x,y, H, J) / {\mathbbm{R}}.
\end{split}$$ where the quotient is taken with respect to the ${\mathbbm{R}}$-action mentioned above.The *energy* of an element $u$ is $$E(u) := \int_{-\infty}^{\infty}|\nabla \mathcal{A}_H|_{L^2}^2 \, ds
= \mathcal{A}_H(x)-\mathcal{A}_H(y).$$
For a generic $J$ and non-degenerate admissible $H$ define the wrapped Floer chain complex $${\mathrm{CW}}(H,L_0 \to L_1) = \bigoplus_{x \in {\mathrm{Crit\,}}(\mathcal{A}_H)} {\mathbbm{Z}}_2 \cdot x,$$ with differential $\partial : {\mathrm{CW}}(H,L_0 \to L_1) \rightarrow {\mathrm{CW}}(H,L_0 \to L_1)$ given by $$\partial (x) = \sum_{y \in {\mathrm{Crit\,}}(\mathcal{A}_H)} \#_{{\mathbbm{Z}}_2} \mathcal{M}^0(x,y,H,J) \cdot y.$$ For generic $J$ the differential is well-defined and moreover $\partial^2= 0$. For simplicity we will write ${\mathrm{CW}}(H)$ instead of ${\mathrm{CW}}(H,L_0 \to L_1)$ when there is no possibility of confusion. In this paper we are not concerned with gradings in ${\mathrm{CW}}$. The homology of $({\mathrm{CW}}(H,L_0 \to L_1),\partial)$ is called the wrapped Floer homology of $(H,L_0 \to L_1)$ and is denoted by ${\mathrm{HW}}(H;L_0 \to L_1)$, or in short ${\mathrm{HW}}(H)$.
Next we consider continuation maps. Let $H_-$ and $H_+$ be non-degenerate admissible Hamiltonians with $H_+(x) \geq H_-(x)$ for all $x \in \widehat{M}$, in short $H_+ \succ H_-$. Take an increasing homotopy through admissible Hamiltonians $(H_s)_{s\in {\mathbbm{R}}}$, $\partial_s H_s \geq 0$, with $H_s = H_{\pm}$ near $\pm \infty$. For elements in $\mathcal{M}^0(x_-,x_+,H_s,J)$, i.e. Floer strips $$\begin{split}\label{contin}
&u:{\mathbbm{R}}\times [0,1] \rightarrow \widehat{M}, \\
&\overline{\partial}_{J,H_s}(u) := \partial_s u + J(\partial_t u - X_{H_{s}}(u)) = 0, \\
&\lim_{s\rightarrow \pm \infty} u(s,t) = x_{\pm}, \\
&u(\cdot, 0) \in \widehat{L}_0, \mbox{ and } u(\cdot, 1) \in \widehat{L}_1,
\end{split}$$ with Fredholm index 0 connecting $x_- \in {\mathrm{Crit\,}}(\mathcal{A}_{H_-})$ and $x_+ \in {\mathrm{Crit\,}}(\mathcal{A}_{H_+})$, the action difference is $$\mathcal{A}_{H_-}(x_-) -\mathcal{A}_{H_+}(x_+) = E(u) + \int_{{\mathbbm{R}}\times [0,1]}\partial_s H_s(u).$$ Hence the action decreases under the continuation maps $$\iota^{H_-,H_+} : {\mathrm{CW}}(H_-) \rightarrow {\mathrm{CW}}(H_+),$$ given by $$\iota^{H_-,H_+} (x_-) = \sum_{x^+ {\mathrm{Crit\,}}(\mathcal{A}_{H_+})} \#_{{\mathbbm{Z}}_2} \mathcal{M}^0(x_-,x_+,H_s,J) \cdot x_+.$$ Define the wrapped Floer homology ${\mathrm{HW}}(M,L_0 \to L_1) := {\underrightarrow{\lim}\,}_H{\mathrm{HW}}(H;L_0 \to L_1)$, where the direct limit is taken over all $H\in \mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1) $.
The homology ${\mathrm{HW}}(M,L_0 \to L_1)$ is the direct limit of the filtered directed system $\widetilde{{\mathrm{HW}}}(M,L_0 \to L_1) = \left({\mathrm{HW}}^a(M,L_0 \to L_1)\right)_{a \in (0,\infty)}$. Here $${\mathrm{HW}}^a(M,L_0 \to L_1) := {\underrightarrow{\lim}\,}_{H} {{\mathrm{HW}}^{a}(H;L_0 \to L_1)},$$ where ${\mathrm{HW}}^{a}(H;L_0 \to L_1) $ is the homology of the Floer chain complex restricted to critical points of action less than $a$. The persistence maps $\iota_{a \rightarrow b}: {{\mathrm{HW}}^a(M,L_0 \to L_1)} \to {{\mathrm{HW}}^b(M,L_0 \to L_1)}$ are induced by the natural maps ${\mathrm{HW}}^a(H,L_0 \to L_1) \rightarrow {\mathrm{HW}}^b(H,L_0 \to L_1)$ that come from inclusions. We write $\iota_{a}: {{\mathrm{HW}}^a(M,L_0 \to L_1)} \to {{\mathrm{HW}}(M,L_0 \to L_1)}$ for the induced map from ${\mathrm{HW}}^a(M,L_0 \to L_1)$ to the direct limit ${{\mathrm{HW}}(M,L_0 \to L_1)}$.
Let $H'\geq H$ be Hamiltonians in $ \mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1) $ and $b\geq a$. Let $ \iota^{H,H'}_{a \to b}: {\mathrm{HW}}^a(H, L_0 \to L_1) \to {\mathrm{HW}}^b(H', L_0 \to L_1)$ by the continuation map induced by any non-decreasing homotopy from $H$ to $H'$. In case $b=+\infty$ we write $ \iota^{H,H'}_{a}: {\mathrm{HW}}^a(H, L_0 \to L_1) \to {\mathrm{HW}}(H', L_0 \to L_1)$.
\[rem:stationary\] Notice that $ \iota^{H,H}_{a \to b}$ is the map induced by the chain level inclusion ${\mathrm{CW}}^a(H,L_0 \to L_1) \hookrightarrow {\mathrm{CW}}^b(H,L_0 \to L_1)$. For this reason we will also denote this inclusion also by $\iota^{H,H}_{a \to b}: {\mathrm{CW}}^a(H,L_0 \to L_1) \hookrightarrow {\mathrm{CW}}^b(H,L_0 \to L_1)$.
By the construction of $\widetilde{{\mathrm{HW}}}(M,L_0 \to L_1)$ presented above we have for every number $a\geq 0$ and Hamiltonian $H\in \mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1) $ a map $$\chi^H_{a \to a}:{\mathrm{HW}}^a(H, L_0 \to L_1) \to {\mathrm{HW}}^a(M, L_0 \to L_1).$$ This allows us to define for every Hamiltonian $H\in \mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1) $ and numbers $b\geq a$ the map $$\chi^H_{a \to b}:=\iota_{a\to b} \circ \chi^{H}_{a \to a} = {\mathrm{HW}}^a(H, L_0 \to L_1) \to {\mathrm{HW}}^b(M, L_0 \to L_1).$$ Using functoriality properties of continuation maps it is straightforward to check that $$\chi^H_{a \to b} =\chi^H_{b \to b} \circ \iota^{H,H}_{a \to b }.$$ For simplicity, in the case $b=+\infty$ we write $$\chi^H_{a } = \iota_{a} \circ \chi^{H}_{a \to a} : {\mathrm{HW}}^a(H, L_0 \to L_1) \to {\mathrm{HW}}(M, L_0 \to L_1).$$
For each $H\in \mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1) $ we also have a map $$\chi^H: {\mathrm{HW}}(H, L_0 \to L_1) \to {\mathrm{HW}}(M, L_0 \to L_1).$$ To define it, we first notice that since $ \mathcal{T}_{L_0 \to L_1}(H)$ is a finite set we can choose a number $ a_H>\max_{x \in \mathcal{T}_{L_0 \to L_1}(H)}\{\mathcal{A}(x)\}$. For this choice of $a_H$ the chain complexes $ ({\mathrm{CW}}(H, L_0 \to L_1),\partial) $ and $ ( {\mathrm{CW}}^{a_H}(H, L_0 \to L_1),\partial)$ are identical, and we get $ {\mathrm{HW}}(H, L_0 \to L_1) = {\mathrm{HW}}^{a_H}(H, L_0 \to L_1)$. We then define $\chi^H := \chi^H_{a_H}$. It is an elementary exercise to check that the definition of $\chi^H $ does not depend on the choice of $a_H>\max_{x \in \mathcal{T}_{L_0 \to L_1}(H)}\{\mathcal{A}(x)\}$. In the same way we can construct for each $b >\max_{x \in \mathcal{T}_{L_0 \to L_1}(H)}\{\mathcal{A}(x)\}$ a map $$\chi^H_{\to b}: {\mathrm{HW}}(H, L_0 \to L_1) \to {\mathrm{HW}}^b(M, L_0 \to L_1).$$
These maps are useful for the study of spectral numbers done in the next section. We will need the identity $$\label{eq:naturality'}
\chi^H_a = \chi^{H'} \circ \iota^{H,H'}_a,$$ which is established in an elementary way from the functoriality properties of continuation maps. In particular, we have $$\label{eq:lepolepo}
\chi^H = \chi^{H'} \circ \iota^{H,H'},$$ and $$\label{eq:naturality}
\chi^H_a = \chi^{H} \circ \iota^{H,H}_a.$$
We will now define the symplectic growth rate of ${\mathrm{HW}}$.
\[defi:Gamma\_symp\] The exponential symplectic growth rate $\Gamma^{{\mathrm{symp}}}(M,L_0 \to L_1)$ is defined by $$\Gamma^{{\mathrm{symp}}}(M,L_0 \to L_1) := \limsup_{a \to \infty} \frac{\log (\dim {\mathrm{Im}}\ \iota_a)}{a} = \widetilde{\Gamma}(\widetilde{{\mathrm{HW}}}(M,L_0 \to L_1)).$$ Analogously, given a family $(L_i)_{i\in I}$ of asymptotically conical exact Lagrangians in $M$ we define $\Gamma_{i\in I}^{\mathrm{symp}}(M,L_0 \to L_{i}) := \widetilde{\Gamma}_{i\in I}(\widetilde{{\mathrm{HW}}}(M,L_0 \to L_i)_{i\in I})$, where $\widetilde{\Gamma}_{i\in I}(\widetilde{{\mathrm{HW}}}(M,L_0 \to L_i)_{i\in I})$ is defined as in Definition \[defi:fam\_growth\].
### Spectral numbers in ${\mathrm{HW}}$
As ${{\mathrm{HW}}(M,L_0 \to L_1)}$ is the direct limit of the the f.d.s. $\widetilde{{\mathrm{HW}}}(M,L_0 \to L_1) $, we define the spectral number $c$ of elements of ${{\mathrm{HW}}(M,L_0 \to L_1)}$ via the recipe given in Section \[subsec:algebra\].
We now present an equivalent definition of $c$ which is more geometrical. Given $H\in \mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1) $, and a cycle $w \in {\mathrm{CW}}(H, L_0 \to L_1)$ we denote by $[w] \in {\mathrm{HW}}(H, L_0 \to L_1)$ the homology class of $w$ in ${\mathrm{HW}}(H, L_0 \to L_1)$. The cycle $w$ can be expressed in a unique way as a sum of elements of $\mathcal{T}_{L_0 \to L_1}(H)$ and we denote by $\mathcal{A}(w)$ the maximum of the actions of these elements.
If $w' \in {\mathrm{CW}}^a(H, L_0 \to L_1)$, then it can be expressed in a unique way as a sum of elements in $\mathcal{T}^a_{L_0 \to L_1}(H)$. This expression is identical to the one of $\iota_a^{H,H}(w')$, from what we conclude $$\mathcal{A}(\iota_a^{H,H}(w')) < a \mbox{ for all } w' \in {\mathrm{CW}}^a(H, L_0 \to L_1).$$ The right hand side in the following identity is often taken as the definition of the spectral number $c(\mathfrak{h})$.
\[lem:minimax\] For a homology class $\mathfrak{h}\in{\mathrm{HW}}(M, L_0 \to L_1)$ we have $$\label{eq:minmax}
c(\mathfrak{h}) = \inf_{H\in \mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1)} \{\mathcal{A}(w) \mid w\in {\mathrm{CW}}(H, L_0 \to L_1) \mbox{ is a cycle with } \chi^H ([w]) = \mathfrak{h} \}.$$
*Proof:* Let $H\in \mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1) $ and $w\in {\mathrm{CW}}(H, L_0 \to L_1)$ be a cycle with $\chi^H ([w]) = \mathfrak{h}$. For each $a > \mathcal{A}(w) $ we know that there exists a cycle $w' \in {\mathrm{CW}}^a(H, L_0 \to L_1)$ such that $\iota_a^{H,H}(w')=w$. Using we obtain $$\chi^{H}_a([w'])= \chi^{H}\circ\iota^{H,H}_{a}( [w']) = \chi^{H} ( [w]) = \mathfrak{h}.$$ This implies that $\mathfrak{h}$ is in the image of $\chi^{H}_a$ and thus in the image of $\iota_a$, from what we get $c(\mathfrak{h})\leq a$. Since this is valid for each $a > \mathcal{A}(w) $ we obtain that $c(\mathfrak{h}) \leq \mathcal{A}(w)$, and it follows that $$\label{eq:minmaxll}
c(\mathfrak{h}) \leq \inf_{H\in \mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1)} \{\mathcal{A}(w) \mid w\in {\mathrm{CW}}(H, L_0 \to L_1) \mbox{ is a cycle with } \chi^H ([w]) = \mathfrak{h} \}.$$
To obtain the reverse inequality let $a > c(\mathfrak{h})$. Then there exists $\beta \in {\mathrm{HW}}^a(M, L_0 \to L_1)$ such that $\iota_a(\beta) = \mathfrak{h}$. By the construction of ${\mathrm{HW}}^a(M, L_0 \to L_1)$ we know that there exists $H \in \mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1)$ and a cycle $w' \in {\mathrm{CW}}^a(H, L_0 \to L_1)$ such that $\chi^H_{a \to a} ([w'])=\beta$. It follows that $$\chi^H_a ([w'])=\iota_a \circ \chi^H_{a \to a} ([w']) = \iota_a(\beta)=\mathfrak{h}.$$ Let $w:= \iota^{H,H}_{a}( w')$. By the observation we made before the lemma we have $\mathcal{A}(w) < a$. Using we obtain $$\chi^H([w])= \chi^H(\iota^{H,H}_a([w'])= \chi^H_a([w'])= \mathfrak{h}.$$ We have shown that for each $a > c(\mathfrak{h})$ there exists $H \in \mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1)$ and a cycle $w \in {\mathrm{CW}}(H, L_0 \to L_1)$ such that $\mathcal{A}(w) < a$ and $\chi^H([w]) = \mathfrak{h}$. It follows that $$c(\mathfrak{h}) \geq\inf_{H\in \mathcal{H}_{{\mathrm{reg}}}(M,L_0 \to L_1)} \{\mathcal{A}(w) \mid w\in {\mathrm{CW}}(H, L_0 \to L_1) \mbox{ is a cycle with } \chi^H ([w]) = \mathfrak{h} \}.$$
### A special type of Hamiltonians {#rem:rem1}
Given an admissible Hamiltonian $H$ in $M$ and a number $a> 0$ we write $H\prec a$ if the slope of $H$ is $<a$. We first define $$\label{K(M,L)}
\mathsf{K}(M,L_0 \to L_1) := \max \{ \max\{f_0(x) - f_1(x) \, | \, x \in L_0 \cap L_1\},\, 0\}.$$
For $a > \mathsf{K}(M,L_0 \to L_1)$ a careful choice of a cofinal family of Hamiltonians shows that ${\mathrm{HW}}^a(M,L_0 \to L_1)$ is isomorphic to ${\underrightarrow{\lim}\,}_{H\prec a}{\mathrm{HW}}(H; L_0 \to L_1)$, where the direct limit is taken only over all non-degenerate admissible $H$ with slope less than $a$.
To explain this we first take a collar neighbourhood $\mathfrak{V}=([1-\delta,1] \times \Sigma) \subset M$ of $\partial M$ on which $L_0$ and $L_1$ are conical, and $\lambda$ is given by $r\alpha_M$. Since $a> \mathsf{K}(M,L_0 \to L_1)$ we can choose $ \mathsf{K}(M,L_0 \to L_1)<\mu<a$, such that there is no element in $\mathcal{T}_{\Lambda_0 \to \Lambda_1}(\alpha_M)$ with length in the interval $[\mu,a)$. We now choose an admissible Hamiltonian $H^\mu$ in $\widehat{M}$ with slope $\mu$ such that
- $H^\mu$ is a negative constant $-\mathrm{k}$ in $M\setminus \mathfrak{V}$, with $\mathrm{k}$ small,
- $H^\mu$ depends only on $r$ in $\mathfrak{V}$, and is a convex function of $r$ that increases sharply close to $\partial M$.
If $\mathrm{k}$ is small enough, and $H^\mu$ increases sharply enough close to $\partial M$ then we have
- the action of all elements of $\mathcal{T}_{L_0 \to L_1}(H^\mu)$ have action $<a$;
see for example [@Ritter2013 Lemma 9.8].
If $(M,L_0 \to L_1)$ is regular then $H^\mu \in \mathcal{H}_{{\mathrm{reg}}}(M, L_0 \to L_1)$. In this case we have $$\mbox{the set } \mathcal{T}_{L_0 \to L_1}(H^\mu) \mbox{ is in bijective correspondence with }\mathcal{T}^a_{\Lambda_0 \to \Lambda_1}(\alpha_M) \cup (L_0\cap L_1).$$ In case $(\alpha_M,\Lambda_0 \to L_1)$ is regular but $(M,L_0 \to L_1)$ is not, we can make a $C^\infty$-small perturbation of $H^\mu$ inside $M$ that still satisfies $\blacksquare$ and is in $\mathcal{H}_{{\mathrm{reg}}}(M, L_0 \to L_1)$; for simplicity we still denote this perturbation by $H^\mu$. In both cases, the reasoning used to prove [@Viterbo1999 Lemma 1.5] gives $$\label{muitoimportante}
\chi^{H^\mu}_a: {\mathrm{HW}}(H^\mu,L_0 \to L_1) \to {\mathrm{HW}}^a(M,L_0 \to L_1) \mbox{ is an isomorphism.}$$ It follows that for $a > \mathsf{K}(M,L_0 \to L_1)$ we have $$\label{eq:coisabonita}
{\mathrm{HW}}^a(M,L_0 \to L_1)\cong {\underrightarrow{\lim}\,}_{H\prec a}{\mathrm{HW}}(H; L_0 \to L_1)$$
Algebra and module structures on wrapped Floer homology {#sec:productonHW}
-------------------------------------------------------
### Algebra structure in ${\mathrm{HW}}$
Let $L$ be an exact asymptotically conical Lagrangian on a Liouville domain $M$. We endow $M$ with an asymptotically cylindrical almost complex structure as in Section \[subsec:def\]. We recall the definition of the triangle product in the wrapped Floer homology ${\mathrm{HW}}(M,L)$, and follow the conventions of [@AS-product].
We first define the triangle $\Delta$. One first takes the disjoint union $\mathbb{R} \times [-1,0] \cup \mathbb{R} \times [0,1]$. We identify the points $(s,0^-) \in \mathbb{R} \times [-1,0] $ and $(s,0^+) \in \mathbb{R} \times [0,1] $ for all $s \geq 0$, and denote the resulting space by $\Delta$. Let $p_{\mathrm{sing}}$ be the point in $\Delta$ which comes from the points $(0,0^-) \in \mathbb{R} \times [-1,0]$ and $(0,0^+) \in \mathbb{R} \times [0,1] $.
The interior of $\Delta$ coincides with $(\mathbb{R} \times (-1,1)) \setminus ((-\infty,0] \times \{0\})$. As $(\mathbb{R} \times (-1,1)) \setminus ((-\infty,0] \times \{0\})$ is a subset of $\mathbb{C}$ we can restrict the complex structure of $\mathbb{C}$ to $(\mathbb{R} \times (-1,1)) \setminus ((-\infty,0] \times \{0\})$. We then obtain a complex structure $j$ in the interior of $\Delta$. This extends to a complex structure on $\Delta\setminus p_{\mathrm{sing}}$. Using again that the interior of $\Delta$ coincides with $(\mathbb{R} \times (-1,1)) \setminus ((-\infty,0] \times \{0\})$, we can define global coordinates $(s,t)$ on $\Delta \setminus p_{\mathrm{sing}}$.
For an admissible Hamiltonian $H$ on $\widehat{M}$, the solutions of the Floer equation on $\Delta$ are maps $u: \Delta \to \widehat{M}$ that satisfy $$\overline{\partial}_{J,H}(u):= \partial_s u + J(u)(\partial_t u - X_H(t,u))=0.$$ We write $\widehat{H} =2H \in C^{\infty}(M)$.
Given $x_1, x_2\in \mathcal{T}_{L}(H)$ and $y \in \mathcal{T}_{L}(\widehat{H})$ we let $\mathcal{M}(x_1,x_2;y,L,J)$ be the space of maps $u: \Delta \to \widehat{M}$ that satisfy $\overline{\partial}_{J,H}(u)=0$ and such that $u(z) \in L$ for all $z\in \partial(\Delta)$, $\lim_{s\to -\infty} u(s,t-1)=x_1(t)$ for $t \in [0,1]$, $\lim_{s\to -\infty} u(s,t)=x_2(t)$ for $t \in [0,1]$, and $\lim_{s\to +\infty} u(s,2t-1)=y(t)$ for $t \in [0,1]$. Define $n(x_1,x_2;y)$ as the number of elements of $\mathcal{M}(x_1,x_2;y,L,J)$ which have Fredholm index $0$. If the moduli spaces $\mathcal{M}(x_1,x_2;y,L,J)$ are transversely cut out, something that can be achieved by perturbing $H$ and $J$, the numbers $n(x_1,x_2;y)$ are always finite.
Define $\Upsilon_L: {\mathrm{CW}}(H,L) \otimes {\mathrm{CW}}(H,L) \to {\mathrm{CW}}(\widehat{H},L) $ by $$\Upsilon_L(x_1,x_2)= \sum_{y \in \mathcal{T}_L(\widehat{H})}(n(x_1,x_2;y)\mod 2) y$$ for $x_1,x_2\in \mathcal{T}_{L}(H)$, and extending it linearly to $ {\mathrm{CW}}(H,L) \otimes {\mathrm{CW}}(H,L)$. It is proved in [@AS-product] that $\Upsilon_L$ descends to a map $H \Upsilon_L: {\mathrm{HW}}(H,L) \otimes {\mathrm{HW}}(H,L) \to {\mathrm{HW}}(\widehat{H},L)$, that endows $ {\mathrm{HW}}(H,L)$ with a product which we denote by $\star$. It is compatible with the continuation maps, as follows by the results in [@Schwarz-thesis Chapter 5], and passing to the direct limit $H\Upsilon_L$ endows $ {\mathrm{HW}}(M,L)$ with a product. For homology classes $\mathfrak{h},\mathfrak{h}' \in {\mathrm{HW}}(M,L) $ we will also denote their product by $\mathfrak{h} \star \mathfrak{h}'$. The product $\star$ is associative: the proof is identical to the proof in [@Schwarz-thesis] that the pair of pants product in Floer homology is associative. As $\star$ is distributive with respect to the vector space structure of $ {\mathrm{HW}}(M,L)$ it gives $ {\mathrm{HW}}(M,L)$ the structure of a ring. Since we defined ${\mathrm{HW}}(M,L)$ with coefficients in $\mathbb{Z}_2$ the product $\star$ actually endows ${\mathrm{HW}}(M,L)$ with the structure of an algebra.
It was proved in [@AS-product] that in the case where $M =T^*Q$ of a compact manifold $Q$ and $L=T_q Q$ for some point $q\in Q$, the triangle product coincides with the Pontrjagin product.
An important property of the triangle product is given by
\[lem:subad\] The spectral numbers $c$ of ${{\mathrm{HW}}}(M,L)$ are subadditive with respect to $\star$.
*Proof:* We will need the triangle inequality $$\label{triangle_inequ}
\mathcal{A}_{\widehat{H}}(y) \leq \mathcal{A}_{H}(x_1) + \mathcal{A}_{H}(x_2),$$ that must be satisfied by the actions of $x_1,x_2 \in \mathcal{T}_{L}(H)$ and $y \in \mathcal{T}_{L}(\widehat{H})$ if the moduli space $\mathcal{M}(x_1,x_2;y,L,J) \neq \emptyset$ (see [@AS-product Formula 3.18]).
Let $\mathfrak{h}_1,\mathfrak{h}_2 \in {{\mathrm{HW}}}(M,L)$. Given $\delta>0$, we know from Lemma \[lem:minimax\] that there exits Hamiltonians $H_1, H_2 \in \mathcal{H}_{{\mathrm{reg}}}(M,L)$ and cycles $w'_i \in {\mathrm{CW}}(H_i,L)$ such that $$\begin{aligned}
\chi^{H_i}([w'_i]) = \mathfrak{h}_i \ \mbox{ and } \ \mathcal{A}(w'_i)< c(\mathfrak{h}_i) +\frac{\delta}{2}\end{aligned}$$ for $i=1,2$. Let now $H\in \mathcal{H}_{{\mathrm{reg}}}(M,L)$ such that $H\geq H_1$ and $H\geq H_2$. We define $w_i:= \iota^{H_i,H}(w'_i)$ for $i=1,2$. Since the action decreases under the continuation maps $\iota^{H_i,H}$ we have $\mathcal{A}(w_i)< c(\mathfrak{h}_i) + \frac{\delta}{2}$, and using we obtain $$\chi^H([w_i])=\chi^{H} ( \iota^{H_i,H}([w'_i])) = \chi^{H_i}([w'_i])=\mathfrak{h}_i,$$ for $i=1,2$. By we have $\mathcal{A}(\Upsilon_L(w_1 \otimes w_2))\leq c(\mathfrak{h}_1) +c(\mathfrak{h}_2) +\delta$. By definition $[\Upsilon_L(w_1 \otimes w_2)]= [w_1]\star[w_2]$, and by our construction of $\star$ in ${\mathrm{HW}}(M,L)$ we have $$\chi^{\widehat{H}}([w_1]\star[w_2]) = \chi^H([w_1])\star \chi^H([w_2]) = \mathfrak{h}_1 \star \mathfrak{h}_2.$$ It then follows from Lemma \[lem:minimax\] that $c( \mathfrak{h}_1 \star \mathfrak{h}_2) \leq \mathcal{A}(\Upsilon_L(w_1 \otimes w_2)) \leq c(\mathfrak{h}_1) +c(\mathfrak{h}_2) +\delta$.
Summing up, we have shown that for any $\delta>0$ we have $c( \mathfrak{h}_1 \star \mathfrak{h}_2) <c(\mathfrak{h}_1) +c(\mathfrak{h}_2) +\delta$, which implies $$c( \mathfrak{h}_1 \star \mathfrak{h}_2) \leq c(\mathfrak{h}_1) +c(\mathfrak{h}_2).$$
We are ready to define the algebraic growth of ${\mathrm{HW}}$.
\[defi:growthHW\] Let $S$ be a finite set of elements of ${\mathrm{HW}}(M,L)$. We define $$\Gamma^{{\mathrm{alg}}}_S(M,L):=\Gamma^{{\mathrm{alg}}}_S({\mathrm{HW}}(M,L)).$$
Combining Lemma \[finite\_growth\] and Lemma \[lem:subad\] we obtain:
\[prop:alg\_symp\] For every finite set $S$ of ${\mathrm{HW}}(M,L)$ we have $$\Gamma^{{\mathrm{symp}}}(M,L) \geq \frac{\Gamma^{{\mathrm{alg}}}_S(M,L)}{\rho(S)}.$$
### ${\mathrm{HW}}(M,L \to L')$ as a module over ${\mathrm{HW}}(M,L)$
We start by picking two exact asymptotically conical Lagrangians $L$ and $L'$ on $(M,\omega,\lambda)$. The boundary $\partial(\Delta)$ contains three connected components: the component $\mathcal{D}_{{\mathrm{left}}}$ which is equal to $\mathbb{R}\times \{-1\}$, the component $\mathcal{D}_{{\mathrm{mid}}}$ which contains the singular point, and the component $ \mathcal{D}_{{\mathrm{right}}}$ which is equal to $\mathbb{R}\times \{1\}$.
Let $x \in \mathcal{T}_{L}(H)$, $z \in \mathcal{T}_{L \to L'}(H)$ and $\widetilde{z} \in \mathcal{T}_{L \to L'}(H)$. We let $\mathcal{M}(x;z,\widetilde{z},J,H)$ be the moduli space of maps $u : \Delta \to \widehat{M}$ which satisfy and such that $u(\mathcal{D}_{{\mathrm{left}}}) \subset L$, $u(\mathcal{D}_{{\mathrm{mid}}}) \subset L$, $u(\mathcal{D}_{{\mathrm{right}}}) \subset L'$, and $\lim_{s\to -\infty}u(s,t-1) = x(t)$ for $t\in [0,1]$, $\lim_{s\to -\infty}u(s,t) = z(t)$ for $t\in [0,1]$, and $\lim_{s\to +\infty}u(s,2t-1) = \widetilde{z}(t)$ for $t\in [0,1]$. Let $n(x;z,\widetilde{z})$ be the number of elements in $\mathcal{M}(x;z,\widetilde{z},J,H)$ that have Fredholm index $0$. For non-degenerate $H$ and a generic choice of $J$, all the spaces $\mathcal{M}(x;z,\widetilde{z},J,H)$ are transversely cut out, and therefore the numbers $n(x;z,\widetilde{z})$ are all finite.
We then define a map $\Theta_{L,L'}: {\mathrm{CW}}(H;L) \otimes {\mathrm{CW}}(H; L\to L') \to {\mathrm{CW}}(\widehat{H} ; L \to L') $ by letting $$\Theta_{L,L'}(x\otimes z) = \sum_{\widetilde{z} \in \mathcal{T}_{L \to L'}(H)} (n(x;z,\widetilde{z})\mod 2 )\widetilde{z},$$ for $x \in \mathcal{T}_{L}(H)$, $z \in \mathcal{T}_{L \to L'}(H)$ , and extending it linearly to ${\mathrm{CW}}(H;L) \otimes {\mathrm{CW}}(H; L\to L')$.
The map $\Theta_{L,L'}$ descends to a map $H\Theta_{L,L'}: {\mathrm{HW}}(H;L) \otimes {\mathrm{HW}}(H; L\to L') \to {\mathrm{HW}}(\widehat{H} ; L \to L') $. The proof is again identical to the one used in [@Schwarz-thesis] to show that the pair of pants product descends to the Floer homology. Taking direct limits we obtain a product $H\Theta_{L,L'}: {\mathrm{HW}}(M,L) \otimes {\mathrm{HW}}(M,L \to L') \to {\mathrm{HW}}(M,L \to L')$. We will use the notation $H\Theta_{L,L'}(\mathfrak{h},\mathfrak{m}) = \mathfrak{h} \ast \mathfrak{m}$.
In order to conclude that ${\mathrm{HW}}(M; L\to L')$ is a module over the algebra ${\mathrm{HW}}(M; L)$ we must prove that:
1. $ \mathfrak{h} \ast (\mathfrak{m}_1 + \mathfrak{m}_2) = \mathfrak{h} \ast \mathfrak{m}_1 + \mathfrak{h} \ast \mathfrak{m}_2$ for all $\mathfrak{h} \in {\mathrm{HW}}(H;L)$ and $ \mathfrak{m}_1, \mathfrak{m}_2 \in {\mathrm{HW}}(H ; L \to L') $,
2. $ (\mathfrak{h}_1 + \mathfrak{h}_2) \ast \mathfrak{m} = \mathfrak{h}_1 \ast \mathfrak{m} +\mathfrak{h}_2 \ast \mathfrak{m}$ for all $\mathfrak{h}_1, \mathfrak{h}_2 \in {\mathrm{HW}}(H;L)$ and $\mathfrak{m} \in {\mathrm{HW}}(H ; L \to L') $,
3. $( \mathfrak{h}_1 \star \mathfrak{h}_2) \ast \mathfrak{m} = \mathfrak{h}_1 \ast (\mathfrak{h}_1 \ast \mathfrak{m} )$ for all $\mathfrak{h}_1, \mathfrak{h}_2 \in {\mathrm{HW}}(H;L)$ and $\mathfrak{m} \in {\mathrm{HW}}(H ; L \to L') $.
The first two properties follow from the linearity of $H\Theta_{L,L'}$. The proof of the third one is a cobordism argument identical to the of [@Schwarz-thesis Chapter 5] that proves the associativity the associativity of the triangle product $\star$. An argument identical to one used to prove Lemma \[lem:subad\] gives
\[lem:subad2\] The spectral numbers $c$ are subadditive with respect to $\ast$.
Viterbo functoriality {#sec:Viterbo}
=====================
The Viterbo transfer map on ${\mathrm{HW}}$ will be described. As first applications we then deduce invariance properties under a graphical change of the boundary of the Liouville domain in the completion.
The Viterbo transfer map on ${\mathrm{HW}}$
-------------------------------------------
The Viterbo transfer map was first introduced as a map for symplectic homology in [@Viterbo1999], see also [@Cieliebak2001; @McLean2009]. The analogous map in wrapped Floer homology was studied by [@AbouzaidSeidel2010], see also [@Ritter2013]. Our focus lies on its compatibility with the action filtration.
Let $M:=(Y_M, \omega_M, \lambda_M)$ be a Liouville domain and let $j: W \rightarrow M$ be a codimension $0$ exact embedding of a Liouville domain $W:=(Y_W, \omega_W, \lambda_W)$ into $M$, i.e. $j^{*}\lambda_M = \lambda_W$. Let $L_0$ and $L_1$ be asymptotically conical exact Lagrangians in $M$, and assume $L^{'}_0 := L_0 \cap W$ and $L^{'}_1 := L_1 \cap W$ are asymptotically conical in $W$. Additionally assume that $L_0$ is also conical on $M \setminus W$ and $L_1$ satisfies the property $$\begin{aligned}
\begin{split}\label{transfer_admissible}
&\lambda|_{L \setminus L^{'}} \text{ vanishes on the boundary }\partial (L\setminus L^{'}) = \partial L \cup \partial L^{'}, \text{ and}\\
&\text{one can write }\lambda|_{L \setminus L^{'}} = df, \text{ where }f \text{ vanishes near } \partial L \cup \partial L^{'}.
\end{split}\end{aligned}$$ We will call a Lagrangian with this property *transfer admissible* for the pair $(M,W)$. See [@AbouzaidSeidel2010] for a discussion of that condition and why the transfer map can in general not be defined if one removes this condition. We give the construction of the *Viterbo transfer map* as an asymptotical morphism of filtered directed systems ${j}(L_0,L_1): \widetilde{{\mathrm{HW}}}(M,L_0 \to L_1) \to \widetilde{{\mathrm{HW}}}(W,L^{'}_0 \to L^{'}_1)$. More precisely we get for $a > K = \mathsf{K}(M,L_0 \to L_1)$, defined in , homomorphisms $$j_{!}(L_0,L_1)_a : {\mathrm{HW}}^a(M,L_0 \to L_1) \rightarrow {\mathrm{HW}}^a(W, L^{'}_{0} \to L^{'}_{1})$$ that are compatible with the persistence morphisms $\iota_{a\rightarrow b}$, for $K< a < b$. Moreover, the homomorphisms are functorial with respect to a composition of embeddings $W_1 \subset W_2 \subset M$ and the induced maps in the direct limit $$\bar{j}_{!}(L_0) = \bar{j}_{!}(L_0,L_0): {\mathrm{HW}}(M,L_0) \rightarrow {\mathrm{HW}}(W,L^{'}_0), \text{ and}$$ $$\bar{j}_{!}(L_0,L_1): {\mathrm{HW}}(M,L_0 \to L_1) \rightarrow {\mathrm{HW}}(W, L^{'}_0 \to L^{'}_1)$$ are compatible with the algebra and module structure, i.e. $$\label{vitproduct}
\bar{j}_{!}(L_0) (x \star y) = \bar{j}_{!}(L_0)(x) \star \bar{j}_{!}(L_0)(y)$$ and $$\label{vitmodule}
\bar{j}_{!}(L_0,L_1)(x \ast z) = \bar{j}_{!}(L_0)(x) \ast \bar{j}_{!}(L_0,L_1)(z)$$ for all $x, y \in {\mathrm{HW}}(M,L_0)$ and $z \in {\mathrm{HW}}(M,L_0, L_1)$.
We first give the definition of ${j}_{!}(L_0,L_1)$. We may assume that $(M,L_0 \to L_1)$ and $(W,L^{'}_0 \to L^{'}_1)$ are regular. Otherwise we can perform the construction considering suitable compactly supported Hamiltonian perturbations of $L_0$ and $L_1$. Let $\mathcal{S} := \mathcal{S}(M,L_0 \to L_1) \cup \mathcal{S}(W,L^{'}_0 \to L^{'}_1)$. We furthermore assume that actually $W \subset M_{\tau^2}$ for some $\tau < 1$, sufficiently close to $1$. One can get the maps for general $W \subset M$ by an inverse limit.
First of all, for every $R > 1$ one can construct a compactly supported Hamiltonian isotopy $(\psi^R_t)_{t\in [0,1]}$ on $\widehat{M}$, ($\psi^R_0 = {id}$, $\psi:= \psi^R_1$) that leaves $\widehat{L}_0$ invariant and maps $\widehat{L}_1$ to a Lagrangian $\widehat{L}^R_1$ that is conical on $(\widehat{M} \setminus M_R) \cup (W_R \setminus W)$ and that is transfer admissible for the pair $(M_R, W_R)$ as follows. Map $L_1 \setminus W$ by the Liouville flow $(\phi_{\log t})_{t\in [1,R]}$ into $A_R = M_R \setminus W_R$. Since $L_1$ is conical near $\partial W$, we can extend $\left(L^{'}_1 \cup \phi_{\log{t}}(L_1 \setminus W)\right)_{t\in [1,R]}$ to a 1-parameter family of exact Lagrangians interpolating between $\widehat{L}_1$ and a Lagrangian $\widehat{L}^R_1$. Therefore we can choose a Hamiltonian isotopy $(\psi^R_t)_{t \in [0,1]}$ in $\widehat{M}$ that realizes this Lagrangian isotopy and is supported in $M_{\frac{1}{\tau}R} \setminus W_{\tau}$. Since $\widehat{L}_0$ is conical outside $W$, we can choose the isotopy to leave $\widehat{L}_0$ invariant. We can choose the isotopy such that $(\psi\circ\zeta)^{*}\lambda = R\zeta^{*}\lambda$, where $\zeta: {L}_1 \setminus W \hookrightarrow \widehat{M}$ is the embedding of $L_1$ restricted to $L_1 \setminus W$. The function $$f_R: \widehat{L}^R_1 \to {\mathbbm{R}}, \text{ with } f_R(x) = \begin{cases} f_1(x), &\text {if } x\in L_1 = \widehat{L}^R_1\cap W, \\
Rf_1(\psi^{-1}x), &\text{ elsewhere}
\end{cases}$$ is a primitive of $\lambda|_{\widehat{L}^R_1}$. We now carefully choose for every $\mu \notin \mathcal{S}$ sufficiently large a step-shaped Hamiltonian $H^{step}_{\mu}$ on $\widehat{M}$. Let $k_W := \min\{f_0(x) - f_1(x) \, | \, x \in L_0 \cap L_1 \cap W \}$ where $f_i$ are the primitives of $\lambda|_{L_i}$, $i=0,1$. Let $\widetilde{k} = \max\{-k_W, 0\}$. Let $$\widetilde{K} = \mathsf{K}(M,W,L_0 \to L_1) = \max\{ \max \{f_0(x) - f_1(x) \, | \, x \in L_0 \cap L_1 \cap M\setminus W\}, 0\}.$$ Choose a small $\epsilon > 0$. Let $\mu > \widetilde{K}$, $\mu \notin \mathcal{S}$, and let $\delta_{\mu} = \min \{{\mathrm{dist\,}}(\mu,\mathcal{S}), \mu - \widetilde{K} \}$. Choose $R > \frac{\widetilde{k} + \mu + 4\epsilon}{\delta_{\mu}}$.
We choose a smooth function $H^{step}_{\mu}: \widehat{M} \rightarrow {\mathbbm{R}}$ that only depends on the radial coordinate $r= r_W$ in $(0,R) \times \partial W$ and only on the radial coordinate $r=r_M$ in $(\tau R,\infty) \times \partial M$, and such that $$\label{step}
H^{step}_{\mu}(x) = \begin{cases}
-\epsilon,&\text{ if }x \in W_{\tau} \\
\frac{\partial^2 H}{\partial r} \geq 0, &\text{ if } x = (r,y) \in W \setminus W_{\tau} \\
\mu r - \mu, &\text{ if } x=(r,y) \in W_{\tau R} \setminus W \\
\frac{\partial^2 H}{\partial r} \leq 0, &\text{ if } x = (r,y) \in W_R \setminus W_{\tau R} \\
(R-1)\mu - \epsilon, &\text{ if } x \in M_{\tau R} \setminus W_R \\
\frac{\partial^2 H}{\partial r} \geq 0, &\text{ if } x = (r,y) \in M_R \setminus M_{\tau R} \\
\mu r -\mu, &\text{ if }x = (r,y) \in \widehat{M} \setminus M_R.
\end{cases}$$
We divide the critical points of the action functional $\mathcal{A} := \mathcal{A}_{H_{\mu}^{step}}^{\widehat{L}_0 \to \widehat{L}^R_1}$ of $H_{\mu}^{step}$ with respect to $\widehat{L}_0$ and $\widehat{L}^R_0$ into four classes: Intersections of $L_0$ and $L_1$ in $W_{\tau}$ denoted by $\mathfrak{A}^{*}$, Hamiltonian chords close to $\partial W$ denoted by $\mathfrak{A}^{**}$, intersections of $\widehat{L}_0$ and $\widehat{L}^R_1$ in $M_R \setminus W_R$ denoted by $\mathfrak{B}^{*}$, and chords close to $\partial W_R$ and $\partial M_R$ denoted by $\mathfrak{B}^{**}$. We can estimate the action values as follows. $$\begin{aligned}
\mathcal{A}(x) &\geq k_w - \epsilon \geq -\widetilde{k} - \epsilon, \text{ if } x \in \mathfrak{A}^{*}, \\
\mathcal{A}(x) &> -\epsilon \geq -\widetilde{k} - \epsilon, \text{ if } x \in \mathfrak{A}^{**}, \\
\mathcal{A}(x) &\leq R\widetilde{K} - ((R-1)\mu -\epsilon) < - \widetilde{k} - 3\epsilon, \text{ if } x \in \mathfrak{B}^{*}, \text{ and} \label{B*}\\
\mathcal{A}(x) &< (\mu - {\mathrm{dist\,}}(\mu, \mathcal{S}))R - ((R-1)\mu -\epsilon) < -\widetilde{k} - 3\epsilon, \text{ if } x \in \mathfrak{B}^{**}.\end{aligned}$$
In we use that $f_0(x) - f_R(x) \leq \widetilde{K}R$ for every $x \in \mathfrak{B}^{*}$.
Altogether we get that $\mathcal{A}(x) \geq -\widetilde{k} - \epsilon $, if $x\in \mathfrak{A} = \mathfrak{A}^{*} \cup \mathfrak{A}^{**}$ and $\mathcal{A}(x) < -\widetilde{k} - 3\epsilon$, if $x\in \mathfrak{B} = \mathfrak{B}^{*} \cup \mathfrak{B}^{**}$. Hence there are no Floer trajectories from $\mathfrak{B}$ to $\mathfrak{A}$. So ${\mathrm{CW}}^{(-\widetilde{k} - 2\epsilon, +\infty)}_{*}(H^{step}_{\mu}; \widehat{L}_0 \to \widehat{L}^R_1) = {\mathrm{CW}}_*(H^{step}_{\mu})/{\mathrm{CW}}^{(-\infty, -\widetilde{k} - 2\epsilon)}_{*}(H^{step}_{\mu})$ generated by elements of action larger then $-\widetilde{k} - 2\epsilon$ is a chain complex, and the projection ${\mathrm{CW}}(H^{step}_{\mu}) \rightarrow {\mathrm{CW}}^{(-\widetilde{k} - 2\epsilon, +\infty)}(H^{step}_{\mu})$ induces a map $$\label{vit}
{\mathrm{HW}}(H^{step}_{\mu}; \widehat{L}_0 \to \widehat{L}^R_1) \rightarrow {\mathrm{HW}}^{(-\widetilde{k} - 2\epsilon,+\infty)}(H^{step}_{\mu}; \widehat{L}_0 \to \widehat{L}^R_1)$$ on homology.
Let now $H^M_{\mu}$ be a non-degenerate admissible Hamiltonian with respect to $M$ on $\widehat{M}$ with slope $\mu$, and $H^W_{\mu}$ a non-degenerate admissible Hamiltonian with respect to $W$ on $\widehat{W}$ with slope $\mu$. We have the isomorphisms $$\begin{aligned}
&{\mathrm{HW}}(H_{\mu}^M; L_0 \to L_1) \overset{\cong}\rightarrow {\mathrm{HW}}((\psi^{-1})^{*}H_{\mu}^M; \widehat{L}_0 \rightarrow \widehat{L}^R_1) \overset{\cong}\rightarrow {\mathrm{HW}}(H^{step}_{\mu}; \widehat{L}_0 \to \widehat{L}^R_1), \text{ and} \label{H^M}\\
&{\mathrm{HW}}^{(-\widetilde{k} - 2\epsilon,+\infty)}(H^{step}_{\mu}; \widehat{L}_0 \to \widehat{L}^R_1) \overset{\cong}\rightarrow {\mathrm{HW}}(H_{\mu}^W; L^{'}_0 \to L^{'}_1). \label{H^W} \end{aligned}$$
Here, the second isomorphism in holds, since $(\psi^{-1})^{*}H_{\mu}^M$ and $H^{step}_{\mu}$ can be connected by a compactly supported homotopy of Hamiltonians. To get the isomorphism in $\eqref{H^W}$ we choose a conical almost complex structure near $\partial W$. By [@AbouzaidSeidel2010 Lemma 7.2], see also [@Ritter2013 Appendix D] there are no Floer trajectories with asymptotics in $W$ that leave $W$ and hence the differential of ${\mathrm{CW}}^{(-\widetilde{k} - 2\epsilon,+\infty)}(H^{step}_{\mu}, \widehat{L}_0 \to \widehat{L}^R_1)$ only counts Floer trajectories that map into $W$.
Combining , , and gives maps $$\label{vit1}
j_{\mu}: {\mathrm{HW}}(H^M_{\mu}; L_0 \to L_1) \rightarrow {\mathrm{HW}}(H^W_{\mu}; L^{'}_0 \to L^{'}_1)$$ for any $\mu > \widetilde{K}$, $\mu \notin \mathcal{S}$. The isomorphisms , , and are all compatible with Floer continuation maps induced by monotone increasing homotopies of the corresponding Hamiltonians. We do not give the details here and refer the reader to [@Ritter2013 Theorem 9.8]. We thus get commutative diagrams $$\begin{CD}
{\mathrm{HW}}(H^M_{\mu}; L_0 \to L_1) @>{j_{\mu}}>> {\mathrm{HW}}(H^W_{\mu}; L^{'}_0 \to L^{'}_1) \\
@V{\iota^{H^M_{\mu}, H^M_{\eta}}}VV @V{\iota^{H^W_{\mu}, H^W_{\eta}}}VV \\
{\mathrm{HW}}(H^M_{\eta}; L_0 \to L_1) @>{j_{\eta}}>> {\mathrm{HW}}(H^W_{\eta}; L^{'}_0 \to L^{'}_1)
\end{CD}$$ for any $\eta > \mu > \widetilde{K}$, $\mu, \eta \notin \mathcal{S}$.
Hence, for any $a > K=\mathsf{K}(M,L_0 \to L_1) \geq \widetilde{K}$ one obtains, because of the construction in Section \[rem:rem1\], a map $$j_{!}(L_0,L_1)_a : {\mathrm{HW}}^a(M, L_0 \to L_1) \rightarrow {\mathrm{HW}}^a(W, L^{'}_0 \to L^{'}_1)$$ induced in the direct limit taken over all non-degenerate admissible Hamiltonians with slope $\mu$, $K <\mu < a$. By the construction these maps are compatible with the persistence morphisms $\iota_{a \rightarrow b}$, for $K < a < b$.
By a standard compactness-cobordism argument, and by using once again the non-escaping result [@AbouzaidSeidel2010 Lemma 7.2] one can show the compatibility of the algebra and module structure with the Viterbo transfer maps and ; for this see [@Ritter2013].
Change of the contact hypersurface $\partial M$
-----------------------------------------------
From the Viterbo transfer one can deduce invariance properties of ${\mathrm{HW}}$ under a graphical change of $\partial M$ in $\widehat{M}$. This will be used to bound the growth rate of Reeb chords for different choices of contact forms on $(\partial M, \xi_{M})$. Let $M$ be a Liouville domain with asymptotically conical exact Lagrangians $L_0$ and $L_1$ as above, let $0 < \epsilon < 1$.
\[M\_epsilon\] Assume that $L_i$, $i=0,1$, are conical on $M\setminus M_{\epsilon}$. Then, for $a > K = \mathsf{K}(M,L_0 \to L_1)$, we have ${\mathrm{HW}}^a(M_{\epsilon},L_0 \cap M_{\epsilon} \to L_1 \cap M_{\epsilon}) \overset{\varphi_a}{\cong} {\mathrm{HW}}^{\frac{1}{\epsilon}a}(M,L_0 \to L_1)$. Moreover, the Viterbo map ${\mathrm{HW}}^a(M, L_0 \to L_1) \rightarrow {\mathrm{HW}}^a(M_{\epsilon},L_0 \cap M_{\epsilon} \to L_1 \cap M_{\epsilon})$ composed with $\varphi_a$ is the persistence morphism ${\mathrm{HW}}^a(M, L_0 \to L_1) \rightarrow {\mathrm{HW}}^{\frac{1}{\epsilon}a}(M,L_0 \to L_1)$.
*Proof:* Note, that adding a constant to any Hamiltonian $H$ or applying a compactly supported deformation to $H$ does not change its Floer homology. Let $H$ be an admissible Hamiltonian with slope $\mu$ with respect to $M_{\epsilon}$. Then $H-\mu(\frac{1}{\epsilon}-1)$ is an admissible Hamiltonian with slope $\frac{1}{\epsilon} \mu$ with respect to $M$. Moreover, if one chooses a cofinal sequence of Hamiltonians of the first kind with slopes $K< \mu < a$, there are compactly supported homotopies of the shifted Hamiltonians to a cofinal sequence with respect to $M$ with slopes $\frac{1}{\epsilon}\mu$. This gives the first statement.
Observe, that both the Viterbo transfer map in the present situation and the persistence morphisms are given by a continuation map induced by a monotone homotopy. One can apply a usual chain homotopy argument in Floer homology to see the second statement.
Let $\mathsf{f}: \partial M \rightarrow [1, \infty)$ be a smooth function. Recall that $M_{\mathsf{f}} = \widehat{M} \setminus \{(r,x) \, | \, r > \mathsf{f}(x), x \in \partial M\}$. Let $\zeta = \max_{\partial M} \mathsf{f}$.
\[lem:changeofhypersurface\] The filtered directed systems $({\mathrm{HW}}^a(M,L_0 \to L_1))_{a\in (0,\infty)}$ and $({\mathrm{HW}}^a(M_{\mathsf{f}},\widehat{L}_0 \cap M_{\mathsf{f}} \to \widehat{L}_0 \cap M_{\mathsf{f}} ))_{a\in (0, \infty)}$ are $(\zeta, 1)$-interleaved.
*Proof:* The morphisms of filtered directed systems $f$ and $g$, with $f_a: {\mathrm{HW}}^a(M,L_0 \to L_1) \cong {\mathrm{HW}}^{\zeta a}(M_{\zeta}, L_0 \to L_1) \to {\mathrm{HW}}^ {\zeta a}(M_{\mathsf{f}},\widehat{L}_0 \cap M_{\mathsf{f}}, \widehat{L}_1 \cap M_{\mathsf{f}})$ and $g_a:{\mathrm{HW}}^a(M_{\mathsf{f}},\widehat{L}_0 \cap M_{\mathsf{f}}, \widehat{L}_1 \cap M_{\mathsf{f}}) \to {\mathrm{HW}}^a(M,L_0 \to L_1)$, given by Viterbo maps, yield by functoriality of Viterbo maps and Lemma \[M\_epsilon\] the $(\zeta,1)$- interleaving of $({\mathrm{HW}}^a(M,L_0 \to L_1))_{a\in (0,\infty)}$ and $({\mathrm{HW}}^a(M_{\mathsf{f}},\widehat{L}_0 \cap M_{\mathsf{f}} \to \widehat{L}_0 \cap M_{\mathsf{f}} ))_{a\in (0, \infty)}$.
From algebraic growth to positivity of topological entropy {#sec:entropy}
==========================================================
In this section we prove Theorem \[theorementropy\].
Legendrian isotopies, transfer admissible Lagrangians and growth {#subsec:technical}
----------------------------------------------------------------
We start by introducing some notation. Let $M=(Y,\omega,\lambda)$ be a Liouville domain and $L$ be an asymptotically conical exact Lagrangian disk in $M$. We denote by $\Lambda$ the Legendrian sphere $\partial L$. Letting $\Sigma := \partial M$ and $\alpha_{M} := \lambda |_{\Sigma}$ be the contact form induced by $M$ on $\Sigma$ we assume that $(\alpha_{M}, \Lambda \to \Lambda)$ is regular. As usually, we denote by $\xi_{M}$ the contact structure $\ker \alpha_{M}$.
Our approach to prove invariance of the exponential symplectic growth of ${\mathrm{HW}}$ differs from the ones developed by [@MacariniSchlenk2011; @McLean2012]. It makes extensive use of the module and algebra structures that exist on ${\mathrm{HW}}$. We will need the following
Let $\mu>0$ and $\Lambda_0$ be a Legendrian sphere in $(\Sigma, \xi_{M})$. Assume that $\Lambda_1$ is Legendrian isotopic to $\Lambda_0$. We say that $\Lambda_1$ is *$\mu$-close to $\Lambda_0$ in the $C^3$-sense* if there exists a Legendrian isotopy $\theta:[-1,1] \times S^{n-1} \to (\Sigma, \xi_{M})$ from $\Lambda_0$ to $\Lambda_1$ whose $C^3$-norm is $<\mu$, and which is stationary in the first coordinate outside a compact subset of $(-1,1)$.
Recall that the symplectisation of a contact form $\alpha$ on $(\Sigma, \xi_{M})$ is the exact symplectic manifold $((0,+\infty)\times \Sigma, dr\alpha, r\alpha)$ where $r$ denotes the first coordinate in $(0,+\infty)\times \Sigma$. The following lemma is essentially due to Chantraine [@Baptiste] and is proved in Appendix B.
\[lemmaBaptiste\] Fix a constant $\epsilon >0$, a contact form $\alpha$ on $(\Sigma,\xi)$, a Legendrian $\Lambda_0$ in $(\Sigma,\xi)$, and a tubular neighbourhood $U(\Lambda_0)$ of $\Lambda_0$ in $\Sigma$. Then there exists $\delta>0$ such that if $\Lambda_1$ is $\delta$-close to $\Lambda_0$ in the $C^3$-sense, then there exist exact Lagrangian cobordisms $\mathcal{L}^-$ from $\Lambda_1$ to $\Lambda_0$ and $\mathcal{L}^+$ from $\Lambda_0$ to $\Lambda_1$ in the symplectization of $\alpha$ satisfying:
- $\mathcal{L}^- $ is conical outside $[1-\frac{\epsilon}{2}, 1-\frac{\epsilon}{4}] \times \Sigma$,
- $\mathcal{L}^+$ is conical outside $[1+\frac{\epsilon}{4}, 1+ \frac{\epsilon}{2}] \times \Sigma$,
- the projections of $\mathcal{L}^+$ and $\mathcal{L}^-$ to $\Sigma$ are completely contained in $U(\Lambda_0)$,
- the primitives $f^\pm$ of $(r\alpha) |_{\mathcal{L^\pm}}$ have support in $[1-\frac{\epsilon}{2}, 1-\frac{\epsilon}{4}] \times \Sigma$ and $[1+\frac{\epsilon}{4}, 1+ \frac{\epsilon}{2}] \times \Sigma$, respectively, and $| f^\pm |_{C^0} < \epsilon$.
Moreover if $\mathcal{L}$ is the exact Lagrangian cylinder obtained by gluing $\mathcal{L}^+ \cap [1,+\infty) \times \Sigma)$ on top of $\mathcal{L}^- \cap ((0,1] \times \Sigma)$ we have that
- $\mathcal{L}$ is Hamiltonian isotopic to $\mathbb{R}\times \Lambda_0$ in the symplectization of $\alpha$, and the Hamiltonian producing the isotopy can be taken to have support in $[1- \frac{\epsilon}{2},1+ \frac{\epsilon}{2}] \times \Sigma$.
We now fix $\epsilon > 0$ such that $L$ is conical on $M \setminus M_{1 - 2\epsilon}$. We choose a Legendrian tubular neighbourhood $\mathcal{U}(\Lambda)$ of $\Lambda$ on $(\Sigma,\xi_M) $. For these choices of $\epsilon>0$ and $\mathcal{U}(\Lambda)$, we choose $\delta_1>0$ given by Lemma \[lemmaBaptiste\].
We then choose a Legendrian sphere $\Lambda_1$ which is $\delta_1$-close to $\Lambda$ in the $C^3$ sense, is disjoint from $\Lambda$, and satisfies that $(\alpha_{M}, \Lambda \to \Lambda_1)$ is regular.
It follows from Lemma \[lemmaBaptiste\] that there exists an exact Lagrangian cobordism $\mathcal{L}^-$ from $\Lambda_1$ to $\Lambda$ in the symplectization of $\alpha_{M}$ which is conical outside $[1-\frac{\epsilon}{2},1-\frac{\epsilon}{4}] \times \Sigma$. We can then glue $\mathcal{L}^- \cap[1-\frac{\epsilon}{2},1] \times \Sigma$ to $L \cap M_{1-\frac{\epsilon}{2}}$ to obtain an exact Lagrangian submanifold $L_1$ in $M$. The Lagrangian $L_1$ is an exact filling of $\Lambda_1$. Let $f_L$ be the primitive of $\lambda\mid_{L}$ which vanishes in $\Lambda$. Using Lemma \[lemmaBaptiste\] we can glue $f^-$ to the restriction of $f_L$ to $L \cap M_{1-\frac{\epsilon}{2}}$ to obtain primitive of $f_{L_1}$ of $\lambda\mid_{L_1}$ which vanishes in $\Lambda_1$.
Because of the control given by Lemma \[lemmaBaptiste\] on the function $| f^- |_{C^0}$ on $\mathcal{L}^-$, and the facts that $L$ and $L_1$ coincide on $M_{1-\frac{\epsilon}{2}}$ and $f_L$ vanishes on $L \cap(M\setminus M_{1-\frac{\epsilon}{2}})$ we have $$\label{eq:control:LtoL_1}
\mathsf{K}(M,L \to L_1) < \epsilon.$$
By Lemma \[lemmaBaptiste\] d) the Lagrangian $L_1$ is transfer admissible for the pair $(M,M_{1-\epsilon})$. Combining this with we obtain for each $a>\epsilon\geq \mathsf{K}(M,L \to L_1)$ a Viterbo map $\Psi^a_{\mathcal{L}^-} : {\mathrm{HW}}^a(M,L \to L_1) \to {\mathrm{HW}}^a(M_{1-\epsilon},L)$, where to simplify notation we keep denoting by $L$ and $L_1$ the restrictions of $L$ and $L_1$ to $M_{1-\epsilon}$. Passing to the direct limit we obtain a map $\Psi_{\mathcal{L}^-} : {\mathrm{HW}}(M,L \to L_1) \to {\mathrm{HW}}(M_{1-\epsilon},L)$.
By Lemma \[lemmaBaptiste\] we also have an exact Lagrangian cobordism $\mathcal{L}^+$ from $\Lambda$ to $\Lambda_1$, which is diffeomorphic to $\mathbb{R}\times S^{n-1}$, and is conical over $\Lambda$ for $r\geq 1 + \frac{\epsilon}{2}$ and conical over $\Lambda_1$ for $r \leq 1 + \frac{\epsilon}{4} $. By gluing $\mathcal{L}^+ \cap ([1, 1+ \epsilon] \times \Sigma)$ to $L_1$ we obtain an exact Lagrangian $\overline{L}$ in $M_{1+\epsilon}$. By Lemma \[lemmaBaptiste\] d) the Lagrangian $\overline{L}$ is transfer admissible for the pair $(M_{1+\epsilon},M)$. By gluing $f^+$ to $f_{L_1}$ we obtain a primitive $f_{\overline{L}}$ of $\lambda\mid_{\overline{L}}$. Reasoning as in the proof of one obtains $$\label{eq:control:LtoL_1above}
\mathsf{K}(M_{1+\epsilon},L \to \overline{L}) < \epsilon.$$ We thus obtain for each $a>\epsilon$ a Viterbo map $\Psi^a_{\mathcal{L}^+} : {\mathrm{HW}}^a(M_{1+\epsilon},L \to \overline{L}) \to {\mathrm{HW}}^a(M,L \to L_1 )$, where by abuse of notation we denote by $L$ the conical extension of $L$ to $M_{1+\epsilon}$. Passing to the direct limit we obtain a map $\Psi_{\mathcal{L}^+} : {\mathrm{HW}}(M_{1+\epsilon},L \to \overline{L}) \to {\mathrm{HW}}(M,L \to L_1 )$.
By Lemma \[lemmaBaptiste\], $\overline{L}$ is Hamiltonian isotopic to the conical extension of $L$ to $M_{1+\epsilon}$, which we will still denote by $L$, for a Hamiltonian function which vanishes outside $M_{1+\frac{\epsilon}{2}} \setminus M_{1- \frac{\epsilon}{2}}$. A continuation argument then implies that for each admissible Hamiltonian $H$ that is regular for both $(M_{1+\epsilon},L \to \overline{L})$ and $(M_{1+\epsilon}, L)$ and has slope $>\epsilon$ we have that ${\mathrm{HW}}(H,L \to \overline{L})$ and ${\mathrm{HW}}(H, L)$ are isomorphic. By Section \[rem:rem1\] we conclude that for each $a>\epsilon$ the wrapped Floer homologies $$\label{eq:isocont}
{\mathrm{HW}}^a(M_{1+\epsilon},L \to \overline{L}) \mbox{ and } {\mathrm{HW}}^a(M_{1+\epsilon}, L) \mbox{ are isomorphic}.$$ This induces an isomorphism $\Phi: {\mathrm{HW}}(M_{1+\epsilon}, L) \to {\mathrm{HW}}(M_{1+\epsilon},L \to \overline{L})$.
Since $L$ is conical on $M_{1+\epsilon} \setminus M_{1-\epsilon}$, $M \setminus M_{1-\epsilon}$ and $M_{1+\epsilon} \setminus M$, we have transfer maps
- $\Psi^{\pm}_{L} : {\mathrm{HW}}(M_{1+\epsilon}, L) \to {\mathrm{HW}}(M_{1-\epsilon}, L)$,
- $\Psi^{-}_{L} : {\mathrm{HW}}(M, L) \to {\mathrm{HW}}(M_{1-\epsilon}, L)$,
- $\Psi^{+}_{L} : {\mathrm{HW}}(M_{1+\epsilon}, L) \to {\mathrm{HW}}(M, L)$.
We notice that the contact forms induced by $\lambda$ on $\{1-\epsilon\} \times \Sigma$ and $\{1+\epsilon\} \times \Sigma$ are $\frac{\alpha_{M}}{1-\epsilon}$ and $\frac{\alpha_{M}}{1+\epsilon}$, respectively. Thus, as explained in Lemma \[M\_epsilon\], the maps $\Psi^\pm_L$, $\Psi^-_L$ and $\Psi^+_L$ are induced by asymptotic isomorphisms of f.d.s. For this reason we will denote by $A_L$ the algebras $ {\mathrm{HW}}(M_{1+\epsilon}, L)$, ${\mathrm{HW}}(M_{1-\epsilon}, L)$ and $ {\mathrm{HW}}(M, L)$. More generally, the same reasoning shows that for any $\zeta > -\epsilon$ the algebra ${\mathrm{HW}}(M_{1+\zeta}, L)$ is isomorphic to $ {\mathrm{HW}}(M_{1-\epsilon}, L)$ by an asymptotic isomorphism.
The homologies ${\mathrm{HW}}(M, L \to L_1)$, ${\mathrm{HW}}(M_{1+\epsilon}, L \to \overline{L})$ and ${\mathrm{HW}}(M_{1-\epsilon}, L )$ are modules over the algebras ${\mathrm{HW}}(M, L)$, ${\mathrm{HW}}(M_{1+\epsilon}, L)$ and ${\mathrm{HW}}(M_{1-\epsilon}, L)$, respectively: they are therefore $A_L$-modules. By this discussion and in section \[sec:Viterbo\] the maps $\Phi$, $\Psi_{\mathcal{L}^-}$ and $\Psi_{\mathcal{L}^+}$ are $A_L$-module homomorphisms.
By functoriality of continuation maps, the diagram\
$$\begin{CD}
{\mathrm{HW}}(M_{1+\epsilon}, L\to \overline{L}) @<\Phi<< {\mathrm{HW}}(M_{1+\epsilon},L ) \\
@V{\Psi_{\mathcal{L}^-}\circ\Psi_{\mathcal{L}^+}}VV @V\Psi^{\pm}_{L}VV \\
{\mathrm{HW}}(M_{1-\epsilon}, L ) @<\mathrm{id}<<{\mathrm{HW}}(M_{1-\epsilon}, L) \\
\end{CD}$$\
is commutative. It thus follows that the map ${\Psi_{\mathcal{L}^-}\circ\Psi_{\mathcal{L}^+}}$ is an $A_L$-module isomorphism. We thus conclude that $\Psi_{\mathcal{L}^+}$ is injective. Let $\mathbf{1}_L$ be the unit in ${\mathrm{HW}}(M_{1+\epsilon},L )$. As $\Phi$ is an $A_L$-module isomorphism and $\Psi_{\mathcal{L}^+}$ is an injective $A_L$-module homomorphism we know that the element $m_{L_1} :=\Psi_{\mathcal{L}^+} \circ \Phi( \mathbf{1}_L)$ in ${\mathrm{HW}}(M,L \to L_1 )$ is a stretching element. We have thus proved the following:
\[lemmaprelim\] The wrapped Floer homology ${\mathrm{HW}}(M,L \to L_1 )$ is a stretched module over ${\mathrm{HW}}(M,L)$. It follows from Lemma \[mod\_growth\], Lemma \[lem:subad\], and Lemma \[lem:subad2\] that $$\Gamma^{{\mathrm{symp}}}(M,L \to L_1 ) \geq {\Gamma^{{\mathrm{symp}}}(M,L )}.$$
Recall that our Legendrian sphere $\Lambda_1$ was chosen disjoint from $\Lambda$. It follows that intersections of the Lagrangian disk $L_1$ and $L$ are contained in $M_{1-\frac{\epsilon}{5}}$. By a small Hamiltonian isotopy supported inside $M_{1-\frac{\epsilon}{5}}$ we can perturb $L_1$ to an exact Lagrangian $L'_1$ that is transverse to $L$. We take the perturbation to be small enough so that there is a primitive $f_{L'_1}$ of $\lambda\mid_{L'_1}$ which vanishes in $\partial L_1'$ and satisfies $$\label{eq:control:LtoL'_1}
\mathsf{K}(M_{1},L \to L'_1) < \epsilon.$$ A continuation argument identical to the one used in the proof of implies that for $a>\epsilon$ the homologies ${\mathrm{HW}}^a(M, L \to L_1)$ and ${\mathrm{HW}}^a(M, L \to L'_1)$ are isomorphic. We let $$\mathsf{C}_{{\mathrm{regium}}}:= \#( L'_1 \cap L).$$ This number will be useful later for estimates of the growth of the number of Reeb chords.
We now consider a tubular neighbourhood $\widetilde{\mathcal{U}}(\Lambda_1) $ which does not intersect $\Lambda$. By Lemma \[lemmaBaptiste\] there exists $\delta_2>0$ such that if a Legendrian sphere $\Lambda_2$ is $\delta_2$-close to $\Lambda_1$ in the $C^3$-sense, then there exist exact Lagrangian cobordisms $\mathcal{L}_{2 \to 1}$ from $\Lambda_2$ to $\Lambda_1$, and $\mathcal{L}_{1 \to 2}$, from $\Lambda_1$ to $\Lambda_2$, both contained in the symplectization of $\alpha_{M}$. It follows from Lemma \[lemmaBaptiste\] that by taking $\delta_2>0$ smaller, if necessary, we can guarantee that
- $\mathcal{L}_{2 \to 1}$ is conical outside $[1-\frac{\epsilon}{5}, 1-\frac{\epsilon}{6}] \times \Sigma$,
- $\mathcal{L}_{1 \to 2}$ is conical outside $[1+\frac{\epsilon}{6}, 1+\frac{\epsilon}{5}] \times \Sigma$,
- the projections of $\mathcal{L}_{2 \to 1}$ and $\mathcal{L}_{1 \to 2}$ to $\Sigma$ are contained in $\widetilde{\mathcal{U}}(\Lambda_1) $,
- there exist primitives $f_{2 \to 1}$ and $f_{1 \to 2}$ of $r\alpha_{M} |_{\mathcal{L}_{2 \to 1}}$ and $r\alpha_{M} |_{\mathcal{L}_{1 \to 2}}$, respectively, with support in $[1-\frac{\epsilon}{5}, 1-\frac{\epsilon}{6}] \times \Sigma$ and $[1+\frac{\epsilon}{6}, 1+\frac{\epsilon}{5}] \times \Sigma$, respectively, such that $|f_{2 \to 1}|_{C^0}<\epsilon$ and $|f_{1 \to 2}|<\epsilon$,
- the exact Lagrangian $\mathcal{L}_{1 \to 1}$ in the symplectisation of $\alpha_{M}$ obtained by gluing $\mathcal{L}_{1 \to 2} \cap ({[1,+\infty)\times \Sigma})$ on top of $\mathcal{L}_{2 \to 1} \cap ((0,1]\times \Sigma)$ is Hamiltonian isotopic to $(0,+\infty)\times \Lambda_1$ for an isotopy which is stationary outside ${(1-\frac{\epsilon}{5},1+\frac{\epsilon}{5})\times \Sigma}$.
It is clear that one can glue $f_{1 \to 2}$ and $f_{2 \to 1}$ to obtain a primitive $f_{1 \to 1}$ of $r\alpha_{M} |_{\mathcal{L}_{1 \to 1}}$ which satisfies $|f_{1 \to 1}| < \epsilon$.
We then glue $\mathcal{L}_{2 \to 1} \cap ([1-\frac{\epsilon}{5}, 1] \times \Sigma)$ on top of $L_1 \subset M_{1-\frac{\epsilon}{5}}$ to obtain an asymptotically conical exact Lagrangian $L_2$ with $L_2\cap \partial M= \Lambda_2$. Let $L'_2$ be the exact Lagrangian submanifold obtained from gluing $\mathcal{L}_{2 \to 1} \cap ([1-\frac{\epsilon}{5}, 1] \times \Sigma)$ on top of $L'_1 \subset M_{1-\frac{\epsilon}{5}}$. It is clear that $L'_2$ and $L_2$ are Hamiltonian isotopic for a Hamiltonian which has support contained in $M_{1 - \frac{\epsilon}{6}}$.
Notice that the intersection points of $L'_2 \cap L$ are the same as the intersection points of $ L'_1 \cap L$. We thus conclude: $$\#(L'_2 \cap L) = \mathsf{C}_{{\mathrm{regium}}}.$$ We can glue $f_{2 \to 1}$ to the restriction of $f_{L_1}$ to $L_1 \cap M_{1-\frac{\epsilon}{5}}$ to obtain a primitive $f_{L_2}$ of $\lambda\mid_{L_2}$ such that $$\label{eq:controlLtoL_2}
\mathsf{K}(M,L \to L_2) < \epsilon.$$ Similarly, one obtains a primitive $f_{L'_2}$ of $\lambda\mid_{L'_2}$ such that $$\label{eq:controlLtoL'_2}
\mathsf{K}(M,L \to L'_2) < \epsilon.$$ Assuming that $(\alpha_{M}, \Lambda \to \Lambda_2)$ is regular the Lagrangian $L_2$ is admissible for the pair $(M,M_{1-\frac{\epsilon}{5}})$. We then obtain for each $a>\epsilon$ a transfer map $\Psi^a_{\mathcal{L}_{2\to 1}}: {\mathrm{HW}}^a(M, L \to L_2) \to {\mathrm{HW}}^a(M_{1-\frac{\epsilon}{5}}, L \to L_1)$. These induce a map $ \Psi_{\mathcal{L}_{2\to 1}}: {\mathrm{HW}}(M, L \to L_2) \to {\mathrm{HW}}(M_{1-\frac{\epsilon}{5}}, L \to L_1)$.
By and and the fact that $L_2$ and $L'_2$ are Hamiltonian isotopic for an isotopy supported inside $M_{1-\frac{\epsilon}{5}}$, we can apply the reasoning used to prove to show that for each $a>\epsilon$ $$\label{equsefulestimate}
{\mathrm{HW}}^a(M, L \to L_2) \mbox{ and } {\mathrm{HW}}^a(M, L \to L'_2) \mbox{ are isomorphic.}$$
Gluing $\mathcal{L}_{1 \to 2} \cap ([1,1+\frac{\epsilon}{5}] \times \Sigma)$ on top of $L_2 \subset M$ we obtain an asymptotically conical Lagrangian $\widetilde{L}_1$ in $M_{1+\frac{\epsilon}{5}}$ which is transfer admissible for the pair $(M_{1+\frac{\epsilon}{5}},M)$. Reasoning as in the proof of we obtain that $$\mathsf{K}(M_{1+\frac{\epsilon}{5}},L \to \widetilde{L}_1) < \epsilon.$$ We thus get for each $a>\epsilon$ a transfer map $\Psi^a_{\mathcal{L}_{1\to 2}}: {\mathrm{HW}}^a(M_{1+\frac{\epsilon}{5}}, L \to \widetilde{L}_1) \to {\mathrm{HW}}^a(M, L \to L_2)$, and in the direct limit a homomorphism $\Psi_{\mathcal{L}_{1\to 2}}: {\mathrm{HW}}(M_{1+\frac{\epsilon}{5}}, L \to \widetilde{L}_1) \to {\mathrm{HW}}(M, L \to L_2)$.
We finally glue $\mathcal{L}^+ \cap ([1 +\frac{\epsilon}{5} ,1 + {\epsilon}]\times \Sigma)$ on top of $\widetilde{L}_1$ to obtain an asymptotically conical exact Lagrangian $\widetilde{L}$ on $M_{1+\epsilon}$. The Lagrangian $\widetilde{L}$ is an exact filling of $\Lambda$. It is clear from Lemma \[lemmaBaptiste\] that $\widetilde{L}$ is Hamiltonian isotopic to $L$, for a Hamiltonian which has support contained in $M_{1+\frac{\epsilon}{2}} \setminus M_{1-\frac{\epsilon}{2}}$. Reasoning as in the proof of we obtain a primitive $f_{\widetilde{L}}$ of $\lambda \mid_{\widetilde{L}}$ such that $$\label{eq:controlLtoLhat}
\mathsf{K}(M_{1+\epsilon},L \to \widetilde{L}) < \epsilon.$$ We claim that for every $a>\epsilon$ there exists an isomorphism $$\label{eq:isocrucial}
\Psi^a_{L, \widetilde{L}}: {\mathrm{HW}}^a(M_{1+\epsilon}, L ) \to {\mathrm{HW}}^a(M_{1+\epsilon}, L \to \widetilde{L}).$$ To establish this claim we first notice that if $H$ is a Hamiltonian in[^3] $\mathcal{H}_{{\mathrm{reg}}}(M_{1+\epsilon},L\to \widetilde{L}) \cap \mathcal{H}_{{\mathrm{reg}}}(M_{1+\epsilon},L)$ it follows from the fact that $\widetilde{L}$ is Hamiltonian isotopic to $L$ for a Hamiltonian which has support contained in $M_{1+\frac{\epsilon}{2}} \setminus M_{1-\frac{\epsilon}{2}}$ that there exists a continuation isomorphism $\Psi_{H,L, \widetilde{L}}: {\mathrm{HW}}(H, L ) \to {\mathrm{HW}}(H, L \to \widetilde{L})$. Equation then follows from combining these isomorphisms and the identifications ${\mathrm{HW}}^a(M_{1+\epsilon}, L ) \cong {\underrightarrow{\lim}\,}_{H\prec a}{\mathrm{HW}}(H; L)$ and ${\mathrm{HW}}^a(M_{1+\epsilon}, L ) \cong {\underrightarrow{\lim}\,}_{H\prec a}{\mathrm{HW}}(H; L \to \widetilde{L})$ for $a>\epsilon\geq \max\{\mathsf{K}(M_{1+\epsilon},L \to \widetilde{L}); \mathsf{K}(M_{1+\epsilon},L)\}$ which were established in . The maps $\Psi^a_{L, \widetilde{L}}$ are compatible with the persistence morphisms of the f.d.s. $\widetilde{{\mathrm{HW}}}(M_{1+\epsilon}, L )$ and $\widetilde{{\mathrm{HW}}}(M_{1+\epsilon}, L \to \widetilde{L})$ and induce an asymptotic morphism between them. On the direct limit we get a map $$\label{eq:asymp}
\Psi_{L, \widetilde{L}}: {\mathrm{HW}}(M_{1+\epsilon}, L ) \to {\mathrm{HW}}(M_{1+\epsilon}, L \to \widetilde{L}).$$
The succession of exact Lagrangian submanifolds we constructed is schematically presented in Figure 2.
\[figure3\]
Since $\widetilde{L}$ is transfer admissible for $(M_{1+\epsilon}, M_{1+\frac{\epsilon}{5}})$ we also obtain for each $a>\epsilon\geq \mathsf{K}(M_{1+\epsilon},L \to \widetilde{L})$ a transfer map $\Phi^a_{\mathcal{L}^+} : {\mathrm{HW}}^a(M_{1+\epsilon}, L \to \widetilde{L}) \to {\mathrm{HW}}^a(M_{1+\frac{\epsilon}{5}}, L \to \widetilde{L}_1) $. This induces a homomorphism $\Phi_{\mathcal{L}^+} : {\mathrm{HW}}(M_{1+\epsilon}, L \to \widetilde{L}) \to {\mathrm{HW}}(M_{1+\frac{\epsilon}{5}}, L \to \widetilde{L}_1) $.
Analogously, it follows from Lemma \[lemmaBaptiste\] that $L_1$ is transfer admissible for the pair $(M_{1-\frac{\epsilon}{5}}, M_{1-\epsilon})$, which gives us for each $a>\epsilon\geq \mathsf{K}(M,L \to L_1) \geq \mathsf{K}(M_{1-\frac{\epsilon}{5}},L \to L_1)$ a map $\Phi^a_{\mathcal{L}^-} : {\mathrm{HW}}^a(M_{1-\frac{\epsilon}{5}}, L \to {L}_1) \to {\mathrm{HW}}^a(M_{1-\epsilon}, L)$. These homomorphisms induce a homomorphism $\Phi_{\mathcal{L}^-} : {\mathrm{HW}}(M_{1-\frac{\epsilon}{5}}, L \to {L}_1) \to {\mathrm{HW}}(M_{1-\epsilon}, L)$.
The following lemma will be important for the study of the growth rate of ${\mathrm{HW}}(M, L \to L_2)$.
\[lem:cacareco\] For $0<\delta_1$ and $0<\delta_2$ chosen as above we have that the spectral number of $\Psi_{L, \widetilde{L}}( \mathbf{1}_L)$ is $\leq \epsilon$.
*Proof:* We know from [@Ritter2013] that $ c(\mathbf{1}_L) = 0$. This implies that for every $a\geq 0$ the element $\mathbf{1}_L $ is in the image of $\iota_a: {\mathrm{HW}}^a(M_{1+\epsilon},L) \to {\mathrm{HW}}(M_{1+\epsilon},L)$.
Let $a>\epsilon$. As remarked above, the maps $\Psi^a_{L, \widetilde{L}}$ are compatible with the persistence morphisms of $\widetilde{{\mathrm{HW}}}(M_{1+\epsilon}, L )$ and $\widetilde{{\mathrm{HW}}}(M_{1+\epsilon}, L \to \widetilde{L})$, which implies that the diagram\
$$\begin{CD}
{\mathrm{HW}}^a(M_{1+\epsilon}, L ) @>\Psi^a_{L, \widetilde{L}}>> {\mathrm{HW}}^a(M_{1+\epsilon}, L \to \widetilde{L}) \\
@V\iota_aVV @V\iota_aVV \\
{\mathrm{HW}}(M_{1+\epsilon}, L) @>\Psi_{L, \widetilde{L}}>>{\mathrm{HW}}(M_{1+\epsilon}, L \to \widetilde{L})
\end{CD}$$\
is commutative. It follows that $\Psi_{L, \widetilde{L}}(\mathbf{1}_L)$ is in the image of $\iota_a: {\mathrm{HW}}^a(M_{1+\epsilon}, L \to \widetilde{L}) \to {\mathrm{HW}}(M_{1+\epsilon}, L \to \widetilde{L})$, from what we obtain that $c(\Psi_{L, \widetilde{L}}(\mathbf{1}_L)) \leq a$. Since this is true for every $a>\epsilon$ we conclude that $c(\Psi_{L, \widetilde{L}}(\mathbf{1}_L)) \leq \epsilon$.
By our discussion so far we have transfer maps\
$$\begin{CD}
{\mathrm{HW}}(M_{1+\epsilon}, L \to \widetilde{L}) @>\Phi_{\mathcal{L}^+}>> {\mathrm{HW}}(M_{1+\frac{\epsilon}{5}}, L \to \widetilde{L}_1)
@>\Psi_{\mathcal{L}_{1 \to 2}}>> {\mathrm{HW}}(M, L \to {L}_2) \\ @. @. @V\Psi_{\mathcal{L}_{2\to 1}}VV \\ @. {\mathrm{HW}}(M_{1-\epsilon}, L) @<\Phi_{\mathcal{L}^-}<< {\mathrm{HW}}(M_{1-\frac{\epsilon}{5}}, L \to L_1)
\end{CD}$$\
Using the fact that $\widetilde{L}$ is Hamiltonian isotopic to $L$ by a Hamiltonian with support contained in $M_{1+\frac{\epsilon}{2}} \setminus M_{1-\frac{\epsilon}{2}}$ and reasoning identically as in the proof of Lemma \[lemmaprelim\] we conclude that the composition $\Phi_{\mathcal{L}^-} \circ \Psi_{\mathcal{L}_{2\to 1}} \circ \Psi_{\mathcal{L}_{1 \to 2}} \circ \Phi_{\mathcal{L}^+} \circ \Psi_{L, \widetilde{L}}$ is induced by an asymptotic isomorphism from ${\mathrm{HW}}(M_{1+\epsilon}, L )$ to ${\mathrm{HW}}(M_{1-\epsilon}, L )$. It follows that $\Psi_{\mathcal{L}_{1 \to 2}} \circ \Phi_{\mathcal{L}^+} : {\mathrm{HW}}(M_{1+\epsilon}, L \to \widetilde{L}) \to {\mathrm{HW}}(M, L \to {L}_2)$ is an injective $A_L$-module homomorphism. We define $m_{L_2}:= \Psi_{\mathcal{L}_{1 \to 2}} \circ \Phi_{\mathcal{L}^+}(\Psi_{L, \widetilde{L}}( \mathbf{1}_L))$. The element $m_{L_2} \in {\mathrm{HW}}(M, L \to {L}_2)$ is stretching since it is the image of a stretching element by an injective $A_L$-module homomorphism.
By the behaviour of spectral numbers under transfer maps, combined with and Lemma \[lem:cacareco\] we conclude that $$\label{uniformitymodules}
c(m_{L_2}) \leq \max\{c(\Psi_{L, \widetilde{L}}( \mathbf{1}_L)),\mathsf{K}(M_{1+\epsilon}, L \to \widetilde{L}) \} \leq \epsilon.$$
We denote by $\mathcal{V}_{\alpha_{M}}(\Lambda_1)$ the set of Legendrian spheres $\Lambda_2$ in the same Legendrian isotopy class of $\Lambda_1$ that are $\delta_2$-close to $\Lambda_1$ is the $C^3$-sense. Let $\mathcal{V}_{\alpha_{M}}^{ {\mathrm{reg}}}(\Lambda_1) \subset \mathcal{V}_{\alpha_{M}}(\Lambda_1)$ be the subset of these $\Lambda_2$ for which, in addition, $(\alpha_{M}, \Lambda \to \Lambda_2)$ is regular. We denote by $L_2$ the filling of $\Lambda_2$ constructed above. Our discussion so far implies the following
\[uniformmodules2\] The family $({\mathrm{HW}}(M,L \to L_2))_{\Lambda_2 \in \mathcal{V}_{\alpha_{M}}^{{\mathrm{reg}}}(\Lambda_1)}$ of $A_L$-modules is uniformly stretched. It follows from Lemma \[mod\_fam\_growth\], Lemma \[lem:subad\], and Lemma \[lem:subad2\] that $$\Gamma_{\Lambda_2 \in \mathcal{V}_{\alpha_{M}}^{{\mathrm{reg}}}(\Lambda_1)}^{{\mathrm{symp}}}(M,L \to L_2) \geq \Gamma^{{\mathrm{symp}}}(M, L).$$
*Proof:* The proposition follows directly from the fact that the element $m_{L_2} \in {\mathrm{HW}}(M,L \to L_2)$ is stretching and from .
Let $\alpha$ be a contact form on $(\Sigma,\xi_{M})$. We assume that the function $\mathsf{f}_\alpha$ defined by $\alpha= \mathsf{f}_\alpha \alpha_{M}$ satisfies $\mathsf{f}_\alpha \geq 1$. We thus have the inclusions $ M_{\mathsf{f}_\alpha} \subset M_{\max \mathsf{f}_\alpha}$ and $M \subset M_{\mathsf{f}_\alpha}$.
We denote by $\mathcal{V}_{\alpha_{M}}^{ \alpha-{\mathrm{reg}}}(\Lambda_1) \subset \mathcal{V}_{\alpha_{M}}^{ {\mathrm{reg}}}(\Lambda_1)$ the set of $\Lambda_2 \in \mathcal{V}_{\alpha_{M}}^{ {\mathrm{reg}}}(\Lambda_1)$ such that $(\alpha, \Lambda \to \Lambda_2)$ is regular.
Let $W_\alpha^+:= M_{\max \mathsf{f}_\alpha} \setminus M_{\mathsf{f}_\alpha}$ and $W_\alpha^-:= M_{ \mathsf{f}_\alpha} \setminus M$. Since the Lagrangians $L_1$ and $L_2$ are conical in $M_{\max \mathsf{f}_\alpha} \setminus M$ we obtain for elements $\Lambda_2 \in \mathcal{V}_{\alpha_{(\Sigma,\lambda)}}^{ \alpha-{\mathrm{reg}}}(\Lambda_1)$ transfer maps\
$$\begin{CD}
{\mathrm{HW}}(M_{\max \mathsf{f}_\alpha}, L\to L_2) @>\Phi_{W^+,L \to L_2}>> {\mathrm{HW}}(M_{\mathsf{f}_\alpha},L \to L_2) @>\Phi_{W^-,L \to L_2}>> {\mathrm{HW}}(M,L \to L_2).
\end{CD}$$\
By Lemma \[lem:changeofhypersurface\], the composition $\Phi_{W^-,L \to L_2}\circ \Phi_{W^+,L \to L_2}$ is induced by asymptotic morphisms, and the f.d.s. $ \widetilde{{\mathrm{HW}}}(M_{\mathsf{f}_\alpha},L \to L_2)$ and $\widetilde{{\mathrm{HW}}}(M,L \to L_2)$ are $(\max \mathsf{f}_{\alpha}, 1)$- interleaved.
The following proposition then follows from combining this observation and Proposition \[uniformmodules2\].
\[uniformgrowth\] Let $\alpha$ be a contact form on $(\Sigma,\xi_{M})$ and assume that the function $\mathsf{f}_\alpha$ defined by $\alpha=\mathsf{f}_\alpha \alpha_{M}$ is $\geq 1$. Then, the family of f.d.s. $(\widetilde{{\mathrm{HW}}}(M_{\mathsf{f}_\alpha},L \to L_2))_{\Lambda_2 \in \mathcal{V}_{\alpha_{M}}^{\alpha-{\mathrm{reg}}}(\Lambda_1)}$ satisfies $$\label{equniformgrowth1}
\Gamma^{{\mathrm{symp}}}_{\Lambda_2 \in \mathcal{V}_{\alpha_{M}}^{\alpha-{\mathrm{reg}}}(\Lambda_1)}(M_{\mathsf{f}_\alpha},L \to L_2) \geq \frac{\Gamma^{{\mathrm{symp}}}(M, L)}{\max \mathsf{f}_\alpha}.$$
A reasoning identical to the one used to establish shows that for every $\Lambda_2 \in \mathcal{V}_{\alpha_{M}}^{\alpha-{\mathrm{reg}}}(\Lambda_1)$ and for the exact filling $L'_2$ of $\Lambda_2$ constructed above we have $${\mathrm{HW}}^a(M_{\mathsf{f}_\alpha}, L \to L_2) \mbox{ and } {\mathrm{HW}}^a(M_{\mathsf{f}_\alpha}, L \to L'_2) \mbox{ are isomorphic.}$$ Combining this with Proposition \[uniformgrowth\] we have
\[coro:uniformgrowth\] Let $\alpha$ be a contact form on $(\Sigma,\xi_{M})$ and assume that the function $\mathsf{f}_\alpha$ defined by $\alpha=\mathsf{f}_\alpha \alpha_{M}$ is $\geq 1$. Then, the family of f.d.s. $(\widetilde{{\mathrm{HW}}}(M_{\mathsf{f}_\alpha},L \to L'_2))_{\Lambda_2 \in \mathcal{V}_{\alpha_{M}}^{\alpha-{\mathrm{reg}}}(\Lambda_1)}$ satisfies $$\label{equniformgrowth1}
\Gamma^{{\mathrm{symp}}}_{\Lambda_2 \in \mathcal{V}_{\alpha_{M}}^{\alpha-{\mathrm{reg}}}(\Lambda_1)}(M_{\mathsf{f}_\alpha},L \to L'_2) \geq \frac{\Gamma^{{\mathrm{symp}}}(M, L)}{\max \mathsf{f}_\alpha}.$$
Recall that for every $\Lambda_2 \in \mathcal{V}_{\alpha_{M}}^{\alpha-{\mathrm{reg}}}(\Lambda_1)$ the exact filling $L'_2$ of $\Lambda_2$ satisfies $$\label{bound}
\#(L'_2 \cap L) = \mathsf{C}_{{\mathrm{regium}}}.$$
Now, given a Legendrian $\Lambda_2 \in \mathcal{V}_{\alpha_{M}}^{\alpha-{\mathrm{reg}}}(\Lambda_1)$ let $\mathsf{N}_\alpha^a(\Lambda \to \Lambda_2)= \# \mathcal{T}^a_{\Lambda \to \Lambda_2}(\alpha)$. We define $$\mathsf{N}^a_\alpha( \Lambda \to \mathcal{V}_{\alpha_{M}}^{\alpha-{\mathrm{reg}}}(\Lambda_1)):=
\inf_{\Lambda_2 \in \mathcal{V}_{\alpha_{M}}^{\alpha-{\mathrm{reg}}}(\Lambda_1)} \{\mathsf{N}^a_\alpha(\Lambda \to \Lambda_2) \}.$$
Let $a > \epsilon$. By the results of Section \[rem:rem1\] there exists a Hamiltonian $H^a\in \mathcal{H}_{{\mathrm{reg}}}(M_{\mathsf{f}_\alpha},L \to L'_2)$ with slope $<a$ such that
- all elements in $\mathcal{T}_{L \to L'_2}(H^a)$ have action $<a$,
- there is a bijection between $\mathcal{T}_{L \to L'_2}(H^a)$ and $\mathcal{T}_{\Lambda \to \Lambda_2}(\alpha)\cup (L\cap L'_2)$,
- the map $\chi^{H^a}_{ \to a}:{\mathrm{HW}}(H^a,L\to L'_2) \to {\mathrm{HW}}^{a}(M_{f_\alpha},L\to L'_2)$ is isomorphism.
Combining (**p.3**) and Corollary \[coro:uniformgrowth\] we obtain that $$\label{eq:tatatadada}
\limsup_{a \to +\infty} \frac{\log (\inf_{\Lambda_2 \in \mathcal{V}_{\alpha_{M}}^{\alpha-{\mathrm{reg}}}(\Lambda_1)}\{\dim{\mathrm{CW}}(H^{a} ,L\to L'_2)\})}{a} \geq \frac{\Gamma^{{\mathrm{symp}}}(M, L)}{\max \mathsf{f}_\alpha}.$$ Since $\#(L\cap L'_2)=C_{{\mathrm{regium}}}$ it follows from (**p.2**) that $\dim({\mathrm{CW}}^{a}(H^{{a}},L\to L'_2)) - \mathsf{C}_{{\mathrm{regium}}} = \mathsf{N}_\alpha^a(\Lambda \to \Lambda_2)$. This together with gives
\[corocrucial\] The sequence of numbers $\mathsf{N}^a_\alpha( \Lambda \to \mathcal{V}_{\alpha_{M}}^{\alpha-{\mathrm{reg}}}(\Lambda_1))$ satisfies $$\limsup_{a \to +\infty} \frac{\log \mathsf{N}^a_\alpha( \Lambda \to \mathcal{V}_{\alpha_{M}}^{\alpha-{\mathrm{reg}}}(\Lambda_1))}{a} \geq \frac{\Gamma^{{\mathrm{symp}}}(M, L)}{\max \mathsf{f}_\alpha}.$$
This corollary will be crucial for the estimate of the topological entropy of $\phi_\alpha$ proved next.
From the growth of Reeb chords to topological entropy
-----------------------------------------------------
Let $\alpha$ be a contact form on a contact manifold $(\Sigma,\xi)$, and $X_\alpha$ be its Reeb vector field. Recall that a Riemannian metric $g$ on $X$ is said to be compatible with $\alpha$ if $g(X_\alpha,X_\alpha)=1$ and $X_\alpha$ is orthogonal to $\xi$ with respect to $g$.
We proceed by fixing some more notation. We denote by $\mathbb{D}^n(\rho)$ the $n$-dimensional disk of radius $\rho>0$ around the origin. We endow $\mathbb{D}^n(\rho)$ with the Euclidean metric, and consider on $T^*_1\mathbb{D}^n(\rho) = \mathbb{D}^n(\rho) \times S^{n-1}$ the contact form $\alpha_{{\mathrm{euc}}}$ associated to the Euclidean metric. For each $z\in \mathbb{D}^n(\rho)$ the sphere $S^{n-1}_z:= \{z\} \times S^{n-1}$ is Legendrian in $(\mathbb{D}^n(\rho) \times S^{n-1},\ker \alpha_{{\mathrm{euc}}}) $. Let $g_{{\mathrm{round}}}$ be the metric with constant curvature $1$ on $S^{n-1}$ and $g_{{\mathrm{euc}}}$ be the Euclidean metric on $\mathbb{D}^{n}(\rho)$. The metric $\widetilde{g}= g_{{\mathrm{euc}}} \oplus g_{{\mathrm{round}}}$ on $\mathbb{D}^n(\rho) \times S^{n-1}$ is compatible with the contact form $\alpha_{{\mathrm{euc}}}$; see [@calvaruso].
\[propgrowth\] Let $\alpha$ be a contact form on $(\Sigma,\xi_{M})$ and assume that we have $\Gamma^{{\mathrm{alg}}}_S(M_{f_\alpha}, L) >0$. Then there exists a Riemannian metric $g$ on $(\Sigma,\xi_{M})$ adapted to the $\alpha$, such that $$\limsup_{t \to +\infty} \frac{ \log {\operatorname{Vol}}_g^{n-1}(\phi_{\alpha}^t(\Lambda))}{t} \geq \frac{\Gamma^{{\mathrm{symp}}}(M, L)}{\max f_\alpha}>0,$$ where ${\operatorname{Vol}}_g^{n-1}$ is the $(n-1)$-dimensional volume with respect to $g$, and $f_\alpha$ is the function such that $\alpha=f_\alpha \alpha{(\Sigma,\lambda)}$.
*Proof:* The proof will consist of several steps.
**Step 1.** It suffices to prove the proposition for all contact forms $\alpha$ for which $\mathsf{f}_\alpha\geq1$. Indeed assume that the proposition holds for all such contact forms.
Take a contact form ${\alpha'}$ on $(\Sigma,\xi_{M})$. For the contact form $\widehat{\alpha}:= \frac{\alpha'}{\min \mathsf{f}_{\alpha'}}$ we have $\mathsf{f}_{\widehat{\alpha}}= \frac{\mathsf{f}_{\alpha'}}{\min \mathsf{f}_{\alpha'} }\geq 1$. By assumption there is a Riemannian metric $g$ on $\Sigma$ compatible with $\widehat{\alpha}$ and such that $$\limsup_{t \to +\infty} \frac{ \log {\operatorname{Vol}}_g^{n-1}(\phi_{\widehat{\alpha}}^t(\Lambda))}{t} \geq \frac{\Gamma^{{\mathrm{symp}}}(M, L)}{\max \mathsf{f}_{\widehat{\alpha}}}.$$
The Riemannian metric $g':=(\min \mathsf{f}_{{\alpha'}})^2g$ is compatible with $\alpha'$. A simple computation shows that $\limsup_{t \to +\infty} \frac{ \log {\operatorname{Vol}}_{g'}^{n-1}(\phi_{{\alpha'}}^t(\Lambda))}{t} \geq \frac{\Gamma^{{\mathrm{symp}}}(M, L)}{\max \mathsf{f}_{{\alpha'}}}$, as claimed. We thus fix from now on a contact form $\alpha$ on $(\Sigma,\xi_{M})$ with $\mathsf{f}_\alpha\geq1$.
**Step 2. A tubular neighbourhood of $\Lambda_1$ and construction of the metric $g$**\
It follows from the Legendrian neighbourhood theorem (see [@MichorKriegl Proposition 43.18]) that there exists a tubular neighbourhood $\mathcal{V}(\Lambda_1)$ of $(\Lambda_1)$ and a contactomorphism $\Upsilon : (\mathcal{V}(\Lambda_1), \xi_{M}) \to (\mathbb{D}^n(\rho) \times S^{n-1},\ker \alpha_{{\mathrm{euc}}}) $ that satisfies $$\begin{aligned}
\label{neighbourhood}
\Upsilon^* \alpha_{{\mathrm{euc}}} = \alpha, \\
\Upsilon(\Lambda_1) = \{0\} \times S^{n-1}.\end{aligned}$$
We extend the Riemannian metric $\Upsilon^*\widetilde{g}$, which is compatible with $\alpha$ on $\mathcal{V}(\Lambda_1)$, to a metric $g$ on $\Sigma$ which is compatible with the contact form $\alpha$.
After shrinking the neighbourhood $\mathcal{V}(\Lambda_1)$ and $\rho>0$, we can assume that for every $z \in \mathbb{D}^{n}(\rho)$ the Legendrian $\Lambda^z := \Upsilon^{-1}(\{z\} \times S^{n-1})$ is in the neighbourhood $\mathcal{V}_{\alpha_{M}}(\Lambda_1)$ constructed in Section \[subsec:technical\].
**Step 3.** For each $a>0$ we define the map $F^a_{\Lambda}: \Lambda \times [0,a] \to \Sigma$ by $$F^a_{\Lambda}(q,t) = \phi^t_\alpha(q).$$ Let ${\operatorname{Cyl}}^a_\alpha(\Lambda)$ be the image $F^a_{\Lambda}(\Lambda \times [0,a] )$. We want to estimate from below the $n$-dimensional volume ${\operatorname{Vol}}^{n}_g({\operatorname{Cyl}}^a_\alpha(\Lambda))$ of ${\operatorname{Cyl}}^a_\alpha(\Lambda)$ with respect to the Riemannian metric $g$. For this we define $\mathfrak{B}^a_\alpha(\Lambda):= \Upsilon({\operatorname{Cyl}}^a_\alpha(\Lambda) \cap \mathcal{V}(\Lambda_1)) $. We have $$\label{tititi}
{\operatorname{Vol}}^{n}_g({\operatorname{Cyl}}^a_\alpha(\Lambda)) \geq {\operatorname{Vol}}^{n}_{{g}}({\operatorname{Cyl}}^a_\alpha(\Lambda)\cap \mathcal{V}(\Lambda_1)) = {\operatorname{Vol}}^{n}_{\widetilde{g}}(\mathfrak{B}^a_\alpha(\Lambda)).$$
Let $\Pi: \mathbb{D}^{n}(\rho) \times S^{n-1} \to \mathbb{D}^{n}(\rho)$ be the projection to the first coordinate. Applying Sard’s Theorem to the map $\Pi \circ \Upsilon \circ F^a_\Lambda: (\{a\} \times \Lambda) \cap (F^a_\Lambda)^{-1}( \mathcal{V}(\Lambda_1)) \to \mathbb{D}^n(\rho)$ we conclude that the set $ \mathbb{D}^{n}(\rho) \setminus \Pi\circ \Upsilon(\phi^a_\alpha(\Lambda))$ is an open set of full Lebesgue measure in $ \mathbb{D}^{n}(\rho)$. We define the set $\mathfrak{U}^a_\alpha(\Lambda) \subset \mathbb{D}^{n}(\rho) \setminus \Pi\circ \Upsilon(\phi^a_\alpha(\Lambda) )$ by the property
- $z \in \mathfrak{U}^a_\alpha(\Lambda)$ if all $\alpha$-Reeb chords from $\Lambda$ to $\Lambda^z$ with length $<a$ are transverse.
The proof of the next lemma is identical to the one of [@Alves-Legendrian Lemma 4].
The set $\mathfrak{U}^a_\alpha(\Lambda)$ is an open subset of $\mathbb{D}^{n}(\rho)$ of full Lebesgue measure. The set $\widetilde{\mathfrak{U}}^a_\alpha(\Lambda) \subset \mathfrak{U}^a_\alpha(\Lambda)$ of elements $z \in \mathfrak{U}^a_\alpha(\Lambda)$ such that $\Lambda^z \in \mathcal{V}_{\alpha_{M}}^{\alpha-{\mathrm{reg}}}(\Lambda_1)$ is a dense subset of full Lebesgue measure in ${\mathfrak{U}}^a_\alpha(\Lambda)$.
**Step 4. A volume estimate**\
The function $h^a: \mathfrak{U}^a_\alpha(\Lambda) \to [0, +\infty)$ defined by $h^a(z) := \#(\mathcal{T}^a_{\Lambda \to \Lambda^z}(\alpha)) $ is locally constant on $\mathfrak{U}^a_\alpha(\Lambda)$ since it is continuous and takes only integer values.
We define $\mathfrak{R}^a_\alpha(\Lambda) := \Pi^{-1}(\mathfrak{U}^a_\alpha(\Lambda))\cap \mathfrak{B}^a_\alpha(\Lambda)$. Since $\mathfrak{R}^a_\alpha(\Lambda) \subset \mathfrak{B}^a_\alpha(\Lambda)$ we have ${\operatorname{Vol}}^{n}_{\widetilde{g}}(\mathfrak{B}^a_\alpha(\Lambda)) \geq {\operatorname{Vol}}^{n}_{\widetilde{g}}(\mathfrak{R}^a_\alpha(\Lambda))$. As the map $\Pi: \mathbb{D}^{n}(\rho) \times S^{n-1} \to \mathbb{D}^{n}(\rho)$ is a Riemannian submersion we have that ${\operatorname{Vol}}^{n}_{\widetilde{g}}(\mathfrak{R}^a_\alpha(\Lambda)) \geq {\operatorname{Vol}}^{n}_{g_{{\mathrm{euc}}}}(\Pi(\mathfrak{R}^a_\alpha(\Lambda)))$, where ${\operatorname{Vol}}^{n}_{g_{{\mathrm{euc}}}}(\Pi (\mathfrak{B}^a_\alpha(\Lambda)))$ is computed with multiplicities. If an open set is covered $k$-times by $\Pi : \mathfrak{R}^a_\alpha(\Lambda) \to \mathfrak{U}^a_\alpha(\Lambda)$, then its volume contributes $k$-times to ${\operatorname{Vol}}^{n}_{g_{{\mathrm{euc}}}}(\Pi (\mathfrak{R}^a_\alpha(\Lambda)))$.
For each $z \in \mathfrak{U}^a_\alpha(\Lambda)$ the number of times $\Pi: \mathfrak{R}^a_\alpha(\Lambda) \to \mathfrak{U}^a_\alpha(\Lambda)$ covers $z$ is $h^a(z) =\#(\mathcal{T}^a_{\Lambda \to \Lambda^z}(\alpha)) $. We thus obtain $${\operatorname{Vol}}^{n}_{g_{{\mathrm{euc}}}}(\Pi(\mathfrak{R}^a_\alpha(\Lambda))) = \int_{ \mathfrak{U}^a_\alpha(\Lambda)} h^a(z) \ d \mathsf{vol}_{g_{{\mathrm{euc}}}},$$ where $d \mathsf{vol}_{g_{{\mathrm{euc}}}}$ is the volume form generated by $g_{{\mathrm{euc}}}$ on $\mathbb{D}^n(\rho)$.
Since $\frac{\Gamma^{{\mathrm{symp}}}(M, L)}{\max \mathsf{f}_\alpha}>0$, we can fix $0<\eta <\frac{\Gamma^{{\mathrm{symp}}}(M, L)}{\max \mathsf{f}_\alpha}$. It follows from Corollary \[corocrucial\] that there exists a sequence $a_j \to +\infty$ such that $h^{a_j}(z) \geq e^{\eta a_j}$ for all $z \in \widetilde{\mathfrak{U}}^{a_j}_\alpha(\Lambda)$. Since $\widetilde{\mathfrak{U}}^{a_j}_\alpha(\Lambda)$ is dense in ${\mathfrak{U}}^{a_j}_\alpha(\Lambda)$ and $h^{a_j}$ is locally constant on ${\mathfrak{U}}^{a_j}_\alpha(\Lambda)$ we obtain $h^{a_j}(z) \geq e^{\eta a_j} \mbox{ for all } z \in {\mathfrak{U}}^{a_j}_\alpha(\Lambda)$ and all $a_j$. With it follows that $$\label{volgrowth}
{\operatorname{Vol}}^{n}_g({\operatorname{Cyl}}^{a_j}_\alpha(\Lambda)) \geq \int_{ \mathfrak{U}^{a_j}_\alpha(\Lambda)} h^{a_j}(z) \ d \mathsf{vol}_{g_{{\mathrm{euc}}}} \geq e^{\eta a_j} \uppi \rho^2$$ for every $a_j$.
**Step 5. A Fubini type equality** We define $\widehat{g}:= (F^a_\Lambda)^*g$. Then $$\label{changecoord}
{\operatorname{Vol}}^{n}_g({\operatorname{Cyl}}^{a}_\alpha(\Lambda))= \int_{ \Lambda \times [0,a]} d \mathsf{vol}_{\widehat{g}},$$ where $d \mathsf{vol}_{\widehat{g}}$ is the volume form associated to $\widehat{g}$. Since the metric $g$ is adapted to the contact form $\alpha$ the Reeb vector field has length $1$ and is orthogonal to the Legendrian spheres $F^a_\Lambda(t,\Lambda) = \phi^t_\alpha(\Lambda)$ for every $t\in [0,a]$. Letting $\partial_t$ be the tangent vector field on $[0,a] \times \Lambda$ associated to the first coordinate $t\in [0,a]$, and using the definition of $F^a_\Lambda$, it follows that $D(F^a_\Lambda) \partial_t = X_\alpha$. Therefore $\partial_t$ has $\widehat{g}$-norm equal to $1$ at every point in $[0,a] \times \Lambda$, and is orthogonal to the spheres $\{t\}\times \Lambda$. We thus conclude that $$\label{fubini}
{\operatorname{Vol}}^{n}_g({\operatorname{Cyl}}^{a}_\alpha(\Lambda))=\int_{ \Lambda \times [0,a]} d \mathsf{vol}_{\widehat{g}} = \int_0^{a} {\operatorname{Vol}}_{\widehat{g}}^{n-1}(\{t\} \times \Lambda) dt = \int_0^a {\operatorname{Vol}}_{g}^{n-1}( \phi^t_\alpha(\Lambda)) dt,$$ where ${\operatorname{Vol}}_{\widehat{g}}^{n-1}$ is the $(n-1)$-dimensional volume associated to $\widehat{g}$.
**Step 6. End of the proof.** To finish the proof we argue by contradiction and assume that $\limsup_{t \to +\infty } \frac{\log {\operatorname{Vol}}_{g}^{n-1}( \phi^t_\alpha(\Lambda))}{t} < \eta$. In this case, there exist $ a_0 >0$ and $\varepsilon>0$ such that for all $t\geq a_0$ we have ${\operatorname{Vol}}_{g}^{n-1}( \phi^t_\alpha(\Lambda)) \leq e^{t(\eta-\varepsilon)}$. Integrating both sides of this inequality from $0$ to $a\geq a_0$ and invoking we obtain $$\label{cacaca}
{\operatorname{Vol}}^{n}_g({\operatorname{Cyl}}^{a}_\alpha(\Lambda))\leq \frac{e^{a (\eta-\varepsilon)} - e^{a_0 (\eta-\varepsilon)}}{\eta-\varepsilon} + \int_0^{a_0} {\operatorname{Vol}}_{g}^{n-1}( \phi^t_\alpha(\Lambda)) dt.$$ For $a$ large enough the right hand side of is smaller than $e^{\eta a}\uppi \rho^2$, contradicting . We thus conclude that $$\limsup_{t \to +\infty } \frac{\log {\operatorname{Vol}}_{g}^{n-1}( \phi^t_\alpha(\Lambda))}{t} \geq \eta.$$
Since this is valid for any $\eta < \frac{\Gamma^{{\mathrm{symp}}}(M, L)}{\max \mathsf{f}_{\alpha}}$, the proof of the proposition is completed.
***Proof of Theorem \[theorementropy\]:***\
From Proposition \[propgrowth\] and Yomdin’s theorem (see ) it follows that if $\Gamma^{{\mathrm{symp}}}(M,L)>0$, then for every contact form $\alpha$ on $(\Sigma,\xi_{M})$ we have $$\label{entropyineq}
h_{{\mathrm{top}}}(\phi_{\alpha}) \geq \frac{\Gamma^{{\mathrm{symp}}}(M,L)}{\max(\mathsf{f}_{\alpha})}.$$ We then obtain Theorem \[theorementropy\] by combining with the inequality $$\Gamma^{{\mathrm{symp}}}(M,L) \geq \frac{\Gamma^{{\mathrm{alg}}}_S(M,L)}{\rho(S)}.$$ from Lemma \[prop:alg\_symp\].
Algebras in loop space homology {#sec:ring_module}
===============================
Let $V$ be a compact manifold and fix a point $q \in V$. We denote by $\Omega_q(V)$ the based loop space of $V$ with basepoint in $q$, which is the space of continuous maps from $[0,1]$ to $V$ that map $0$ and $1$ to $q$.
The concatenation of based loops gives $\Omega_q(V)$ the structure of an $H$-space (see [@Hatcher]). More precisely, the concatenation induces the so-called Pontrjagin product on the singular homology ${\mathrm{H}}_*(\Omega_{q}(V))$ of $\Omega_{q}(V)$ with $\mathbb{Z}_2$ coefficients. The Pontrjagin product $[a_1]\cdot[a_2]$ of two homology classes $[a_1],[a_2] \in {\mathrm{H}}_*(\Omega_{q}(V)$ is well-known to be associative. As it is distributive with respect to the vector space structure of ${\mathrm{H}}_*(\Omega_{q}(V))$, it makes ${\mathrm{H}}_*(\Omega_{q}(V))$ into a ring. Because the homology ${\mathrm{H}}_*(\Omega_{q}(V))$ is considered with coefficients in ${\mathbbm{Z}}_2$ it actually has the structure of an algebra.
Relation between the algebra structure of the singular homology of loop spaces and the algebra structures of the Floer homology of cotangent bundles
----------------------------------------------------------------------------------------------------------------------------------------------------
Given a manifold $V$ and $q \in V$ we denote by $L_q \subset T^* V$ the cotangent fibre over $q$. The singular homology ${\mathrm{H}}_*(\Omega_{q}(V))$ of the based loop space $\Omega_{q}(V)$ is isomorphic to the wrapped Floer homology ${\mathrm{HW}}(T^*V,L_q)$; see Viterbo [@Viterbo1999], Salamon-Weber [@SW] and Abbondandolo-Schwarz [@AS-iso] for different proofs.
The Floer homology ${\mathrm{HW}}(H_g,L_q)$ is isomorphic to the wrapped Floer homology ${\mathrm{HW}}(L_q)$ we use in this paper. The key point is that the Hamiltonian $H_g$ is quadratic in the fibres. This isomorphism is proven in [@Ritter2013], and it preserves the triangle product and the spectral value of homology classes.
Let $\Psi_{AS,q}: {\mathrm{H}}_*(\Omega_{q}(V)) \to {\mathrm{HW}}(T^*V,L_q)$ be the isomorphism constructed in [@AS-iso]. In [@AS-product] the authors proceed to study more properties of the map $\Psi_{AS,q}$. They show that $\Psi_{AS,q}$ is also algebra isomorphism if we consider ${\mathrm{H}}_*(\Omega_{q}(V))$ as an algebra with the Pontrjagin product and ${\mathrm{HW}}(H_g,L_q)$ as an algebra with the triangle product. Combining this with the isomorphism ${\mathrm{HW}}(H_g,L_q) \cong{\mathrm{HW}}(L_q)$ we obtain the following
\[ASring\] The singular homology ${\mathrm{H}}_*(\Omega_{q}(V))$ and the wrapped Floer homology ${\mathrm{HW}}(L_q)$ are isomorphic as algebras.
For simplicity we will still denote by $\Psi_{AS,q}$ the isomorphism between ${\mathrm{H}}_*(\Omega_{q}(V))$ and ${\mathrm{HW}}(L_q)$.
Topological operations {#sec:top_operations}
======================
Subcritical surgery {#subsec:subcritical}
-------------------
Here we study the Viterbo transfer maps under subcritical handle attachment in the situation that is sufficient for our purpose, that is we assume that the Lagrangians do not intersect the handle.
Let $W=(Y_W,\omega,\lambda)$ be a Liouville domain, $\Sigma = \partial W$, $\lambda|_{\Sigma} = \alpha$ and $\xi = \ker \alpha$. We recall some notions using the terminology of [@GeigesBook Section 2.5.2]. The form $d\alpha$ endows $\xi$ with a natural conformal symplectic bundle structure. Let $S$ be an isotropic submanifold of $(\Sigma,\xi)$. We write $TS^{\bot}$ for the sub-bundle of $\xi$ that is $d\alpha$-orthogonal to $TS$. Because $S$ is isotropic $TS \subset TS^{\bot}$. We can therefore write the normal bundle of $S$ in $\Sigma$ as $$T \Sigma /T S = T\Sigma/\xi \oplus \xi/TS^{\bot} \oplus TS^{\bot} / TS.$$ The *conformal symplectic normal bundle* $\mathrm{CSN(S)} = TS^{\bot} / TS$ has a natural conformal symplectic structure via $d\alpha$. If $S$ is a sphere, $T\Sigma/\xi \oplus \xi/TS^{\bot}$ has a trivialization. The following theorem is due to Weinstein.
[@Weinstein1990] Let $S^n$ be an isotropic sphere in $\Sigma$ with a trivialization of $\mathrm{CSN}(S)$. Then there is a Liouville domain $M$ with an exact embedding $W \subset M$, such that $\partial M$ is obtained from $\Sigma$ by surgery on $S$.
The Liouville domain $M$ is obtained by attaching an $(n+1)$-handle to $W$ and the Liouville vector field $X$ can by chosen such that there is exactly one point $p \in M\setminus W$ where $X$ vanishes. The integral lines of $X$ that are asymptotic to $p$ intersect $\Sigma$ in $S$ and $\partial M$ in the co-core sphere $B \subset \partial M$. (See [@Weinstein1990; @Cieliebak2001] or [@GeigesBook Chapter 6] for details.)
Let now $L^{'}_0$, $L^{'}_1$ be two asymptotically conical exact Lagrangians in $ W$ whose boundaries $\Lambda^{'}_0$ and $\Lambda^{'}_1$ in $\Sigma$ do not intersect $S$. Outside $S$ the integral lines of the Liouville vector field starting at $\partial W$ intersect $\partial M$ and so the completed Lagrangians $\widehat{L^{'}_{i}} \subset \widehat{M}$ intersect $\partial M$. Moreover, $L_{i} = \widehat{L^{'}_{i}} \cap M \subset M$ for $i=0,1$ are exact and conical in the complement of $W$. We say that $(M,L_0,L_1)$ is *obtained by surgery* from $(W,L^{'}_0,L^{'}_1).$
As described in section \[sec:Viterbo\] we get a Viterbo transfer map $${j}_{!}(L_0,L_1): \widetilde{{\mathrm{HW}}}(M,L_0 \to L_1) \rightarrow \widetilde{{\mathrm{HW}}}(W,L^{'}_0 \to L^{'}_1).$$
Assume that the isotropic sphere $S$ has the property that there is no Reeb chord from $\Lambda^{'}_0$ to $S$. If $S$ is subcritical, i.e. $\dim(S) < n-1$, this can be achieved by a generic choice of $S$.
The following proposition was proved by Cieliebak ([@Cieliebak2001]) for symplectic homology. The proof in our situation is analogous and even simpler. We give it here for the convenience of the reader.
\[Viterbo\_iso\] The Viterbo transfer map in the direct limit, $$\bar{j}_{!}(L_0,L_1): {\mathrm{HW}}(M,L_0,L_1) \to {\mathrm{HW}}(W,L^{'}_0, L^{'}_1),$$ is an isomorphism.
For the proof of Proposition \[Viterbo\_iso\] it is convenient to introduce the following weaker form of interleaving of f.d.s. Let $\sigma: [0,\infty) \rightarrow [0,\infty)$ be a monotone increasing function and $V$ a filtered directed system. Analogously to the notation in \[subsubsec:fds\] let $(V(\sigma), \pi(\sigma))$ be given by $V({\sigma})_t = V_{\sigma (t) t}$, $\pi(\sigma)_{s\rightarrow t}= \pi_{\sigma(s) s \rightarrow \sigma(t) t}$ and $\pi[\sigma]_t = \pi_{\sigma(t) t}$. If $f$ is a morphism from $(V,\pi)$ to another f.d.s. we write $f(\sigma)_t = f_{\sigma(t) t}$ for the induced morphism with domain $(V(\sigma), \pi(\sigma))$. Call two f.d.s. $(V,\pi_V)$ and $(W,\pi_W)$ *weakly interleaved* if there are morphisms $f:V\rightarrow W(\sigma_1)$ and $g:W \rightarrow V(\sigma_2)$ for monotone increasing functions $\sigma_1, \sigma_2 \geq 1$ such that $$f(\sigma_2)\circ g = \pi_{W}[\widetilde{\sigma}_1] \text{ and } g(\sigma_1)\circ f = \pi_{V}[\widetilde{\sigma}_2],$$ where $\widetilde{\sigma}_1$, and $\widetilde{\sigma}_2$ are suitably chosen. The fact that the map $\bar{j}_{!}(L_0,L_1)$ in Proposition \[Viterbo\_iso\] is an isomorphism will follow from a weak interleaving of the corresponding f.d.s., which is in general not an interleaving. This is the reason why we cannot directly prove lower bounds for $\Gamma^{{\mathrm{symp}}}(M,L_0 \to L_1)$ in terms of $\Gamma^{{\mathrm{symp}}}(W,L^{'}_0 \to L^{'}_1)$ and this was originally our motivation to introduce the algebraic growth of wrapped Floer homology.
*Proof of Proposition \[Viterbo\_iso\]:* Let $U = \widetilde{{\mathrm{HW}}}(M,L_0 \to L_1)$, and $V =\widetilde{{\mathrm{HW}}}(W,L^{'}_0 \to L^{'}_1)$. We will construct a filtered directed system $Q$ that is isomorphic to $V$ and weakly interleaved with $U$.
For convenience we may assume $\mathsf{K}(M,L_0 \to L_1) = 0$. Let $S \subset \partial W$ be the attaching sphere and $B \subset \partial M$ be the co-core sphere. For $a > 0$, choose a tubular neighbourhood $U_a \subset \partial W$ of $S$ such that there is no Reeb trajectory starting at $\Lambda^{'}_0$ that intersects $U_a$ at a time less than $a$, and such that $U_{b} \subset U_a$ if $a<b$. Denote the Liouville flow on $\widehat{M}$ by $\varphi_t$ and let $g: \partial M \setminus B \rightarrow (0,1]$ given by $g(x) = t$ where $t$ is the unique number such that $\varphi_{\log t}(x) \in \partial W$. Note that $g$ tends to $0$ as $x$ tends to $B$. Define $N_{a} := \{x \in \partial M \, |\, \varphi_{\log g(x)}(x) \in \partial W \setminus U_a \}$. Choose a family of smooth functions $f_a: \partial M \rightarrow (0,1]$, $a \in (0,\infty)$, with the property $$\begin{aligned}
f_a|_{N_a} = g, \text{ and} \text{ for all }x \in \partial M, \, f_a(x) \text{ is monotone decreasing in }a.
$$
Note that $W \subset M_{f_b} \subset M_{f_a}$, for $b>a$ and $\partial W \setminus U_a \subset \partial M_{f_a}$. Define $\sigma(a) = \frac{1}{\min_{\partial M} f_a}$. Define $Q_a = {\mathrm{HW}}^a(M_{f_a}, L_0 \to L_1)$, where by abuse of notation we write $L_i$ instead of $L_i \cap M_{f_a}$, $i= 0,1$. For $a < b$ define $\pi_{a\to b} : Q_a \to Q_b$ as the composition of the Viterbo map ${\mathrm{HW}}^a(M_{f_a}, L_0 \to L_1) \to {\mathrm{HW}}^a(M_{f_b}, L_0 \to L_1)$ and the persistence map ${\mathrm{HW}}^a(M_{f_b}, L_0 \to L_1) \to {\mathrm{HW}}^b(M_{f_b}, L_0 \to L_1)$. By the commutativity of the Viterbo map with persistence maps and by functoriality of the Viterbo map it follows that $\pi_{a\to c} = \pi_{b\to c} \circ \pi_{a \to b}$, for $a<b<c$, and hence $(Q,\pi)$ is a filtered directed system. Furthermore, $\phi:U \rightarrow Q$, with $\phi_a:{\mathrm{HW}}^a(M) \rightarrow {\mathrm{HW}}^a(M_{f_a})$ is a morphism of f.d.s. We define $\psi: Q \rightarrow U({\sigma})$ by the Viterbo transfer ${\mathrm{HW}}^a(M_{f_a}) \rightarrow {\mathrm{HW}}^a(M_{\min f_a}) = {\mathrm{HW}}^{\sigma(a)\cdot a}(M)$. It is clear that $(\phi, \psi)$ is a weak interleaving of $U$ and $Q$.
It remains to show that $Q$ and $V$ are isomorphic. Let $a > 0$. Assume that $L_0$ and $L_1$ are conical in the complement of $W_{\frac{1}{2}}$. Let $H_{\mu}$ be an admissible Hamiltonian with slope $\mu$ with respect to $W_{\frac{1}{2}}$. Consider a Hamiltonian $K_{\mu}$ such that $$\begin{aligned}
K_{\mu}(x) = &H_{\mu}(x), \text{ if } x \in W_{\frac{3}{4}}. \\
K_{\mu}(x) = &H_{\mu}(x), \text{ if } x = (r,y) \in (0, +\infty) \times \partial W\setminus U_a. \label{K2}\end{aligned}$$ It follows that $K_{\mu}(x) = 2\mu r + b$, for some $b\in {\mathbbm{R}}$, where $x$ is written in the coordinates $(r,y) \in (1, \infty) \times \partial W \setminus U_a$. Hence we can assume additionally that $$\begin{aligned}
K_{\mu}(x) = &2\mu r + b, \text{ where } x = (r,y) \in (1, \infty) \times \partial M_{f_a}. \label{K3} \end{aligned}$$ By definition of $U_a$, $\mathcal{A}_{H_{\mu}}$ and $\mathcal{A}_{K_{\mu}}$ have the same critical points, and so it follows from [@AbouzaidSeidel2010 Lemma 7.2] that we actually have ${\mathrm{HW}}(K_{\mu}) = {\mathrm{HW}}(H_{\mu})$. On the other hand $K_{\mu} - {\frac{1}{2}}\mu$ is admissible with respect to $M_{f_a}$ with slope $2\mu$. One concludes, reasoning as in Lemma \[M\_epsilon\], that $Q_a = {\mathrm{HW}}^{a}(M_{f_a}, L_0 \to L_1) \cong {\mathrm{HW}}^{\frac{1}{2}a}(W_{\frac{1}{2}},L^{'}_0 \to L^{'}_1)$ which is, by Lemma \[M\_epsilon\], isomorphic to ${\mathrm{HW}}^{a}(W,L^{'}_0 \to L^{'}_1) = V_a$. That this identification respects the persistence morphisms of $Q$ and $V$ is again deduced from the functoriality of the Viterbo maps and the fact that the Viterbo maps are themselves morphisms of filtered directed systems. Denote the isomorphism from $Q$ to $V$ by $\tau$. We have obtained a weak interleaving $(\tau \circ \phi, \psi \circ \tau^{-1})$. Moreover $\tau \circ \phi = {j}_{!}$ by construction.
Plumbing
--------
Let $Q_1$ and $Q_2$ be closed orientable $n$-dimensional manifolds. We let $D^*Q_i$ be the unit cotangent bundle of $Q_i$. We choose balls $B_i \subset Q_i$ in each $Q_i$. The plumbing $N$ of $D^*Q_1$ and $D^*Q_2$ is obtained by identifying $D^*B_1$ and $D^*B_2$ via a symplectomorphism that swaps the momentum and position coordinates of these manifolds; see [@AbouzaidSmith2012; @GeigesBook] for the details. There are obvious embeddings of $D^*(Q_1 \setminus B_1)$ and $D^*(Q_2 \setminus B_2)$ into $N$. It is shown in [@AbouzaidSmith2012 Section 4] that $N$ admits a Liouville structure which coincides with those of $D^*(Q_i \setminus B_i)$ on the image of these embeddings. This implies that for points $q_1 \in Q_1\setminus B_1$ the cotangent disc fibre $L_{q_1}$ over $q_1$ survives as a conical exact Lagrangian in the Liouville domain $N$.
This construction can be generalised in the following way. Let $\{Q_i \mid 1 \leq i \leq k\}$ be a finite collection of orientable $n$-dimensional manifolds. Let $\mathsf{T}$ be a tree with $k$ vertices and use a bijection to associate to each vertex a manifold $Q_i$. For each edge $\eta$ leaving the “vertex” $Q_i$ we choose an embedded open ball $B_i(\eta)$ in $Q_i$. We assume that these balls are chosen to be disjoint and do not cover $Q_i$. For all $i\neq j$ and every edge $\eta$ connecting $Q_i$ and $Q_j$ (there can be at most one such edge as $\mathsf{T}$ is a tree) we identify $D^*(B_i(\eta))$ and $D^*(B_j(\eta))$ by the recipe explained in the previous paragraph. The resulting manifold $N$ can be given a Liouville structure as explained in [@AbouzaidSmith2012 Section 4] and [@GeigesBook]. Let $\dot{Q}_1$ be the complement of the “edge balls” in $Q_1$, and $q_1 \in \dot{Q}_1$. In [@AbouzaidSmith2012 Section 4] the following result is proved.
[@AbouzaidSmith2012] \[thm:plumbing\] There exists an injective algebra homomorphism from the group algebra $\mathbb{Z}_2[\pi_1(Q_1)]$ to ${\mathrm{HW}}(N,L_{q_1})$.
In fact the injective algebra homomorphism obtained in [@AbouzaidSmith2012] is for the respective homologies with $\mathbb{Z}$ coefficients, and applying the Universal Coefficient Theorem one obtains the homomorphism mentioned above. Thus if $\pi_1(Q_1)$ grows exponentially then ${\mathrm{HW}}(N,L_{q_1})$ has exponential algebraic growth; see Section \[preliminaries\].
*Proof of Proposition \[prop:operations\]*: Part A) follows from Proposition \[Viterbo\_iso\] and Part B) follows from Theorem \[thm:plumbing\].
Construction of contact structures with positive entropy {#sec:constructions}
=========================================================
In this section we prove Theorem \[spheres\] and Theorem \[maincorollary\].
Preliminaries
-------------
Let $Q$ be a closed connected smooth manifold and $g$ a Riemannian metric on $Q$. Let $(D_{g}^*Q,\lambda_{geo}) \subset (T^*Q,\lambda_{geo})$ be the unit disk bundle with respect to the Riemannian metric $g$ where $\lambda_{geo}$ is the canonical Liouville form on $T^*Q$. By Theorem \[ASring\] of Abbondandolo and Schwarz the map $$\Psi_{AS,q_1} : {\mathrm{H}}_*(\Omega_{q_1}(Q))\to {\mathrm{HW}}(D_g^*Q , L_{q_1})$$ is an algebra isomorphism. It is well-known that there is an algebra isomorphism $$\Phi: \mathbb{Z}_2[\pi_1(Q,q_1)] \to {\mathrm{H}}_0(\Omega_{q_1}(Q)).$$ Composing these two maps we obtain an injective algebra homomorphism $$\widetilde{\Phi}: \mathbb{Z}_2[\pi_1(Q,q_1)] \to {\mathrm{HW}}(D_g^*Q , L_{q_1}).$$
For a finitely generated group $G$ and a finite set $\upsigma$ of generators of $G$, let $\widehat{\Gamma}_{\upsigma}(G)$ be the usual exponential growth of the group $G$ with respect to the set $\upsigma$; see [@delaHarpe Section VI.C]. To a finite set $\upsigma$ of generators of $\pi_1(Q,q_1)$, we associate the finite set $S \subset \mathbb{Z}_2[\pi_1(Q,q_1)]$ that is formed by the elements of $\upsigma$ and its inverses. It is immediate to see that $$\widehat{\Gamma}_{\upsigma}(\pi_1(Q,q_1)) = {\Gamma_S^{{\mathrm{alg}}}}(\mathbb{Z}_2[\pi_1(Q,q_1)]).$$ Using that $\widetilde{\Phi}$ is injective we obtain $$\widehat{\Gamma}_{\upsigma}(\pi_1(Q,q_1)) = {\Gamma_S^{{\mathrm{alg}}}}(\mathbb{Z}_2[\pi_1(Q,q_1)]) \leq {\Gamma_{\widetilde{\Phi}(S)}^{{\mathrm{alg}}}}( {\mathrm{HW}}(D_g^*Q , L_{q_1})),$$ We have shown the following
\[fundalgrowth\] It $\pi_1(Q,q_1)$ has exponential growth then there exists a finite set $S \subset {\mathrm{HW}}(D_g^*Q , L_{q_1})$ such that $ {\Gamma_S^{{\mathrm{alg}}}}(D_g^*Q , L_{q_1}) >0$.
Proof of statement (A) of Theorem \[spheres\] and statement $\clubsuit$ of Theorem \[maincorollary\]
------------------------------------------------------------------------------------------------------
*Proof of statement (A) of Theorem \[spheres\]*
Let $G$ be a finitely presented group such that
- ${\mathrm{H}}_1(G) = {\mathrm{H}}_2(G) =0$,
- $G$ has exponential growth,
- $G$ admits a presentation on which the number of relations does not exceed the number of generators.
Then, it follows from [@Kervaire], that for every $n\geq 4$ there exists a manifold $Q^n$ which is an integral homology sphere and which satisfies $\pi_1(Q^n) = G$. We denote by $\varrho(G)$ the minimal number of generators of $G$.
We denote by $D^*Q^n$ the unit disk bundle of $Q^n$, with respect to a Riemannian metric $g$ in $Q^n$, endowed with the canonical symplectic and Liouville forms. We choose a point $q\in Q^n$ and $g$ generically so that $q$ is not conjugate to itself. Let $S^*Q^n= \partial D^*Q^n $ be the unit cotangent bundle of $Q^n$. In order to prove our result we consider two distinct cases.
**Case 1: $n$ is odd and $\geq 5$.**\
In this case the Euler characteristic of $Q^n$ vanishes. Because $G$ grows exponentially, we know that ${\mathrm{HW}}_0(D^*Q^n,L_q)$ has exponential algebraic growth. Let $N^1$ be the plumbing of $D^*Q^n$ and $D^* S^n$ performed far from $L_q$. By Proposition \[prop:operations\], ${\mathrm{HW}}_0(N^1,L_q)$ has exponential algebraic growth.
It is a result of Milnor that the boundary of the plumbing of the unit disk bundles of two odd-dimensional homology spheres of dimension $\geq 3$ is a homology sphere; see [@Bredon Chapter VI - Section 18]. Applying this to the pair $D^*Q^n$ and $D^* S^n$ we conclude that $\partial N^1 $ is a homology sphere. Since $N^1$ retracts to the one point union of $Q$ and $S^n$ we know that the homology of $N^1$ is zero in every degree different from $0$ and $n$, where we have ${\mathrm{H}}_0(N^1) = \mathbb{Z}$ and ${\mathrm{H}}_n(N^1) = \mathbb{Z} \oplus \mathbb{Z}$.
**Case 2: $n$ is even and $\geq 4$.**\
In this case the Euler characteristic of $Q^n$ is $2$. We consider the plumbing associated to the E8 tree; see [@Bredon Chapter VI - Section 18]. To each vertex of the E8 tree we associate a disk bundle in the following way:
- to the leftmost vertex we associate $D^*Q^n$,
- to every other vertex we associate $D^* S^n$.
We let $N^1$ be the plumbing associated to the E8 tree determined by this choice of disk bundles at each vertex, and assume that the plumbing is done away from a cotangent fibre $L_q\subset D^*Q^n$ . It was shown by Milnor (see [@Bredon Chapter VI - Section 18] ) that $\partial N^1$ is a homology sphere. Since $N^1$ retracts to the wedge sum of $Q$ and seven copies of $S^n$ determined by the E8 tree, we know that the homology of $N^1$ is zero in every degree different from $0$ and $n$, where we have ${\mathrm{H}}_0(N^1) = \mathbb{Z}$ and ${\mathrm{H}}_n(N^1) = \oplus_{i=1}^8 \mathbb{Z} $. By Proposition \[prop:operations\], ${\mathrm{HW}}_0(N^1,L_q)$ has exponential algebraic growth.
**We now treat both cases simultaneously.** By attaching $2$-handles to $N^1$ away from $L_q$ we can obtain a simply connected Liouville domain $N^2$ such that ${\mathrm{HW}}(N^2,L_q)$ has exponential algebraic growth. We choose the framing of these handle attachments so that the first Chern class of $N^2$ vanishes.
The effect of the handle attachment on the homology of the boundary can be read from the surgery formula in [@Kosinsky Section X.1]. One concludes that the homology of $\partial N^2$ coincides with that of $\partial N_1$ except in degree $2$, and ${\mathrm{H}}_2(\partial N^2)$ is the direct sum of $\varrho(G)$ copies of $\mathbb{Z}$.
By Hurewicz’ Theorem there is a basis of ${\mathrm{H}}_2(\partial N^2)$ which is composed of embedded $S^2$. Since the first Chern class of $N^2$ vanishes, it follows from [@McLean2011 Lemma 2.19] that these $S^2$ can be made isotropic and disjoint from $L_q$ by an isotopy and that their symplectic normal bundle is trivial. We can thus perform the Weinstein handle attachment over these spheres. The resulting Liouville domain $N^3$ still contains the Lagrangian $L_q$ and it follows from Proposition \[prop:operations\] that ${\mathrm{HW}}(N^3,L_q)$ has exponential algebraic growth. By the surgery formula in [@Kosinsky Section X.1], the effect of these handle attachments on the homology of the boundary implies that ${\mathrm{H}}_2(\partial N^3)=0$ and that the homologies of $\partial N^3$ and $\partial N^2$ coincide in all other degrees. Therefore, $\partial N^3$ is a simply connected homology sphere. It follows from Whitehead’s Theorem for homology [@Hatcher Corollary 4.33] that $\partial N^3$ also has the homotopy groups of a sphere. Since the dimension of $\partial N^3$ is $>5$ the h-cobordism theorem tells us that $\partial N^3$ is homeomorphic to a sphere. Since the smooth spheres under connected sum form a finite group, we can take the connected sum of finitely many copies of $\partial N^3$ to get the sphere $\partial N^4$ with the standard smooth structure such that ${\mathrm{HW}}(N^4,L_q)$ has exponential algebraic growth. This proves statement (A) of Theorem \[spheres\].
*Proofs of statement $\clubsuit$ of Theorem \[maincorollary\]* Let $V$ be a $(2n-1)$-dimensional manifold where $n\geq 4$, and assume that there exists an exactly fillable contact structure $\xi$ on $V$. Denote by $M_V$ a Liouville domain whose boundary is $(V,\xi)$. Let $N^4$ be the Liouville domain constructed in the proof of statement (A) of Theorem \[spheres\]. By Proposition \[prop:operations\], the Liouville domain $N^5=N^4 \# M_V$ has an asymptotically conical exact Lagrangian $L$ such that ${\mathrm{HW}}(N^5,L)$ has exponential algebraic growth. The statement then follows from Theorem \[theorementropy\].
Proof of statement (B) of Theorem \[spheres\] and statement $\diamondsuit$ of Theorem \[maincorollary\]
-------------------------------------------------------------------------------------------------------
*Proof of statement (B) of Theorem \[spheres\]:*\
We will consider a carefully chosen $3$-manifold $Q$. Consider the Brieskorn manifolds of dimension $3$, $M(p,q,r) = \{(z_1,z_2,z_3) \in {\mathbbm{C}}^3 \, |\, {z_1}^p + {z_2}^q + {z_3}^r =0\} \cap S^5$. $M(p,q,r)$ is a ${\mathbbm{Z}}$-homology sphere if $p,q,r$ are relatively prime (see for example [@Saveliev2002]). It was shown by Milnor [@Milnor1975] that its fundamental group $\pi_1(M(p,q,r))$ is the commutator subgroup of the group $G = G(p,q,r) = \langle \gamma_1, \gamma_2, \gamma_3 \, | \, \gamma_1^p = \gamma_2^q = \gamma_3^r = \gamma_1\gamma_2\gamma_3 \rangle$ , see also [@Seade2006]. The groups $\Sigma = G/Z(G)$ are the triangle groups, where $Z(G)$ is the center of $G$. Consider the case $p=2$, $q=3$, $r=7$. A short computation shows that $G(2,3,7) = [G(2,3,7), G(2,3,7)] = \pi_1(M(2,3,7))$. One has $\widehat{\Gamma}(G(2,3,7)) \geq \widehat{\Gamma}(\Sigma(2,3,7))$, and the exponential growth of $\Sigma(2,3,7)$ is $\log(x)$, where $x \approx 1.17628$ is equal to Lehmer’s Salem number (see [@Hironaka2003] or [@Breuillard2014]). We take $Q=M(2,3,7)$. The integral homology of $D_g^*Q$ is the same as that of $Q$, which is $\mathbb{Z}$ in degrees $0$ and $3$ and vanishes in all other degrees. Moreover it is clear that $\pi_1(S^*Q)= \pi_1(Q\times S^2)= \pi_1(Q)$ is generated by the elements $\gamma_1$ and $\gamma_2$.
Let $N^1$ be the Liouville domain obtained by plumbing $D_g^*Q$ with the unit disk bundle $D^*S^3$ of $S^3$. We assume that the plumbing is performed away from the cotangent fibre $L_q$ over a point $q \in Q$. Therefore $L_q$ survives as a conical exact Lagrangian in $N^1$. By Proposition \[prop:operations\] we know that ${\mathrm{HW}}_*(N^1,L_{q})$ has exponential algebraic growth.
Since $N^1$ is the plumbing of $D_g^*Q$ and $D^*S^3$, and $Q$ and $S^3$ are both homology spheres we obtain that $\partial N^1$ is a homology sphere; see [@Bredon Chapter VI - Section 18]. Combining this with the fact that $N^1$ retracts to the one point union of $S^3$ and $Q$ we conclude that
- ${\mathrm{H}}_0(N_1) = \mathbb{Z}$, ${\mathrm{H}}_3(N_1) = \mathbb{Z} \oplus \mathbb{Z}$ and ${\mathrm{H}}_i(N_1) =0$ for $i\neq 0,3$,
- ${\mathrm{H}}_0( \partial N_1) = \mathbb{Z}$, ${\mathrm{H}}_5( \partial N_1) = \mathbb{Z}$, and ${\mathrm{H}}_i( \partial N_1) =0$ for $i\neq 0,5$.
Let now $\{\overline{\sigma}_1,\overline{\sigma}_{2}, \overline{\sigma}_3\}$ be generators of $\pi_1(\partial N_1)= \pi_1(Q)$ corresponding to $\gamma_1$, $\gamma_2$ and $\gamma_3$ respectively. By the h-principle for subcritical isotropic submanifolds of contact manifolds [@GeigesBook] we can isotope the curve $\overline{\sigma}_3$ to a curve ${\sigma}_3$ which is isotropic in $(S_g^*Q,\xi_{geo})$. We can also assume that ${\sigma}_3$ does not intersect $\Lambda_{q} := \partial L_{q}$. Since ${\sigma}_3$ is isotropic and has trivial normal bundle we can apply the Weinstein handle attachment [@Weinstein1990] and attach a 2-handle to $ N^1$ over ${\sigma}_3$, obtaining a new Liouville domain $N^2$. From the presentation of $\pi_1(Q)$ that we used, it is clear that $\partial N^2$ is simply connected, and so is $N^2$ by [@McLean2011 Lemma 2.9]. We choose the framing of the handle attachment so that $\partial N^2$ is spin. Using the Mayer-Vietoris sequence we obtain that ${\mathrm{H}}_0(\partial N^2) = \mathbb{Z}$, ${\mathrm{H}}_2(\partial N_2) = \mathbb{Z} $, and ${\mathrm{H}}_1(\partial N_2) =0$. By Smale’s classification of spin simply-connected five manifolds [@Smale] it follows that $\partial N_2$ is $S^3\times S^2$.
Since $N^2$ is obtained from $N^1$ via a subcritical handle attachment and the Lagrangian $L_q$ is far from the attaching locus of this handles, we know that $L_q$ survives as a conical exact Lagrangian in $N^2$. Moreover Proposition \[prop:operations\] implies that ${\mathrm{HW}}_*(N^2,L_{q})$ has exponential algebraic growth, and it follows from Theorem \[theorementropy\] that the contact manifold $\partial N_2$ has positive entropy.
*Proof of statement $\diamondsuit$ of Theorem \[maincorollary\]:* The statement is proved by a connected sum argument identical to the one in the proof of statement $\clubsuit$.
To guarantee the vanishing of the second Stiefel-Whitney class of $\partial N^2$ one must only guarantee the vanishing of the first Chern class of $N^2$. As observed in the proof of [@McLean2011 Lemma 2.10], one can choose the framing when performing the attachments of the 2-handles so as to guarantee the vanishing of the first Chern class of $N^2$.
Appendix {#appendix .unnumbered}
========
Construction of exact Lagrangian cobordisms. {#construction-of-exact-lagrangian-cobordisms. .unnumbered}
--------------------------------------------
Before proving the Lemma we recall that the symplectization of the contact form $\alpha$ on $(\Sigma,\xi)$ can also be given by $(\mathbb{R} \times \Sigma,e^s\alpha)$, where $s$ denotes the ${\mathbbm{R}}$-coordinate. It is straightforward to see that the diffeomorphism $\mathrm{F}: ((0,+\infty) \times \Sigma,r\alpha ) \to (\mathbb{R} \times \Sigma,e^s\alpha)$ given by $\mathrm{F}(r,p) = (\log r, p)$ is an exact symplectmorphism.
It follows that an exact Lagrangian submanifold $\mathcal{L}^-$ is conical in $ ((0,+\infty) \times \Sigma,r\alpha )$ outside $[1-2\epsilon,1-\epsilon] \times\Sigma$ if, and only if, $\widetilde{\mathcal{L}}^-:=\mathrm{F}(\mathcal{L}^-)$ is conical in $(\mathbb{R} \times \Sigma,e^s\alpha)$ outside $[\log (1-2\epsilon),\log(1-\epsilon)] \times \Sigma$. Analogously, an exact Lagrangian submanifold $\mathcal{L}^+$ is conical in $ ((0,+\infty) \times \Sigma,r\alpha )$ outside $[1+\epsilon,1+2\epsilon] \times\Sigma$ if, and only if, $\widetilde{\mathcal{L}}^+:=\mathrm{F}(\mathcal{L}^+)$ is conical in $(\mathbb{R} \times \Sigma,e^s\alpha)$ outside $[\log (1+\epsilon),\log(1+ 2\epsilon)] \times \Sigma$.
*Proof of Lemma \[lemmaBaptiste\]:*
We use the technique presented in [@EHK Lemma 6.3].
**Step 1:** We first apply the Legendrian neighbourhood Theorem [@MichorKriegl Proposition 43.18] to find a neighbourhood $\widetilde{U}(\Lambda_0)$ of the Legendrian of $\Lambda_0$ such that there exists a strict contactmorphism $\Upsilon: (\widetilde{U}(\Lambda_0), \alpha) \to (V(\widetilde{\Lambda}_0) \subset \mathbb{R}^{2n-1}, dz + \sum^{n-1}_{i=1}x_i dy_i) $ that satisfies $\Upsilon(\Lambda_0)= \widetilde{\Lambda}_0$, where $\widetilde{\Lambda}_0$ is the standard Legendrian unknot in $\mathbb{R}^{2n-1}$ (see [@EES Example 3.1]) and $V(\widetilde{\Lambda}_0)$ is a tubular neighbourhood of $\widetilde{\Lambda}_0$ in $\mathbb{R}^{2n-1}$. Given this identification, it suffices to establish the lemma for the case of $\widetilde{\Lambda}_0$ since it will follow if we can establish it for $U(\Lambda_0) \subset \widetilde{U}(\Lambda_0) $.
**Step 2:** Clearly, it suffices to establish the lemma for all $0<\epsilon<\frac{1}{e^{1000}}$. We thus fix $0<\epsilon<\frac{1}{e^{1000}}$ and a tubular neighbourhood $U(\widetilde{\Lambda}_0)$ of $\widetilde{\Lambda}_0$. To establish the lemma for $\widetilde{\Lambda}_0$ we assume that $\Lambda_1$ is a Legendrian sphere in $(\mathbb{R}^{2n-1},\alpha_{{\mathrm{can}}} = dz + \sum^{n-1}_{i=1}x_i dy_i)$ that is $\mu$-close to $\widetilde{\Lambda}_0$ in the $C^3$-sense, with $\mu>0$ so small that
- $\Lambda_1\subset U(\widetilde{\Lambda}_0)$,
- there exists a Legendrian isotopy $\theta: [-1,1] \times S^{n-1} \to \mathbb{R}^{2n-1}$ which is $\mu$-small in the $C^3$-topology and satisfies $\theta(\{-1\}\times S^{n-1}) = \widetilde{\Lambda}_0$ and $\theta(\{1\}\times S^{n-1}) = {\Lambda}_1$.
Moreover if $\mu>0$ is chosen sufficiently small we can also assume that
- the isotopy $\theta$ is constant in the first coordinate for $t \notin [\log (1-2\epsilon),\log(1-\epsilon)] $.
Extend $\theta$ to ${\mathbbm{R}}\times S^{n-1}$ by $\theta(t,p) =\theta(-1,p)$ for $t\leq -1$ and $\theta(t,p) =\theta(1,p)$ for $t\geq 1$.
We write $\theta(t,p)= (x(t,p),y(t,p),z(t,p))$ for coordinates $(x,y,z)\in \mathbb{R}^{n-1} \times \mathbb{R}^{n-1}\times \mathbb{R}$, set $F(t,p):= \alpha_{{\mathrm{can}}}(\partial_t \theta(t,p)) $, and define the cylinder $\Theta: \mathbb{R}\times S^{n-1}\to (\mathbb{R}\times \mathbb{R}^{2n-1},e^s\alpha_{{\mathrm{can}}})$ in the symplectization $(\mathbb{R}\times \mathbb{R}^{2n-1},e^s\alpha_{{\mathrm{can}}})$ of $\alpha_{{\mathrm{can}}}$ by $$\Theta(t,p) = (t,x(t,p),y(t,p),z(t,p)+F(t,p)).$$ It is clear that if $\mu>0$ is chosen sufficiently small then $\Theta$ will be an embedding, since it will be a small compact perturbation of the embedding $(t,x(t,p),y(t,p),z(t,p))$. Let $\Pi :\mathbb{R}\times \mathbb{R}^{2n-1} \to \mathbb{R}^{2n-1}$ be the projection of the symplectization to the contact manifold.
A direct computation shows that $$\Theta^*(e^s\alpha_{{\mathrm{can}}})= d(e^tF(t,p)).$$
**Step 3.** Step 2 implies that the cylinder $\widetilde{\mathcal{L}}^-= \Theta( \mathbb{R}\times S^{n-1})$ is an admissible exact Lagrangian submanifold of $(\mathbb{R}\times \mathbb{R}^{2n-1},e^s\alpha_{{\mathrm{can}}})$. By taking $\mu$ even smaller we can guarantee that
- $\widetilde{\mathcal{L}}^-$ is conical over $\Lambda_1$ in $[\log(1-\epsilon), +\infty) \times \mathbb{R}^{2n-1}$ and is conical over $\widetilde{\Lambda}_0$ in $(0,\log(1-2\epsilon)] \times \mathbb{R}^{2n-1}$,
- and the projection of $\widetilde{\mathcal{L}}^-$ to $ \mathbb{R}^{2n-1}$ is contained in $U(\widetilde{\Lambda}_0)$.
In order to construct $\widetilde{\mathcal{L}}^+$ we use the inverse isotopy $\theta^+(t,p) = \theta(-t,p)$, and apply the construction above. We have established statements a), b) and c) of Lemma \[lemmaBaptiste\].
To conclude d), notice that $f^-:= e^tF(t,p)$ has support in $[\log(1-2\epsilon),\log(1 -\epsilon)] \times S^{n-1}$ and if $\mu>0$ is chosen small enough then $f^-:= e^tF(t,p)$ will satisfy $| f^- |_{C^0} \leq \epsilon$ since the $C^0$-norm of $F(t,p)$ will be small. Applying the same argument to $\widetilde{\mathcal{L}}^+$ implies d).
**Step 4.** Statement e) is obtained by performing this construction for a smooth $1$-parameter family of Legendrian isotopies which starts with the isotopy $\theta$ and ends at the stationary isotopy from $\widetilde{\Lambda}_0$ to itself. As the construction above depends $C^\infty$-smoothly on the parameter we obtain a smooth $1$-parameter family of exact Lagrangian embeddings $\Psi: \mathbb{R}\times S^{n-1}\to (\mathbb{R}\times \mathbb{R}^{2n-1},e^s\alpha_{{\mathrm{can}}})$ which starts at $\mathcal{L}$ and ends at $\mathbb{R}\times \widetilde{\Lambda}_0$, and which is constant in $(\mathbb{R}\setminus [ \log (1-2\epsilon),\log(1+2\epsilon)]) \times S^{n-1}$. This is an exact Lagrangian isotopy from $\widetilde{\mathcal{L}} := \mathrm{F}(\mathcal{L})$ to $\mathbb{R}\times \widetilde{\Lambda}_0$ which is constant in $(\mathbb{R}\setminus [\log(1-2\epsilon),\log(1+2\epsilon)]) \times \widetilde{\Lambda}_0$. Statement e) then follows from [@Mclean2015 Lemma 5.6].
[^1]: In a recent work [@Dahinden] Dahinden extended the results in [@MacariniSchlenk2011] proving that on the unit cotangent bundles $(S^{*}Q, \xi)$ studied in [@MacariniSchlenk2011] every positive contactomorphism has positive topological entropy. It would be interested to investigate if Dahinden’s result is true for any contact manifold with positive entropy.
[^2]: This was explicitly observed in [@Mclean2015] although it is implicit in [@FrauenfelderSchlenk2006; @MacariniSchlenk2011].
[^3]: By [@AbouzaidSeidel2010 Lemma 8.1] any admissible Hamiltonian in $M_{1+\epsilon}$ can be perturbed to one in $\mathcal{H}_{{\mathrm{reg}}}(M_{1+\epsilon},L\to \widetilde{L}) \cap \mathcal{H}_{{\mathrm{reg}}}(M_{1+\epsilon},L)$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The last decade has witnessed an explosion in the development of models, theory and computational algorithms for “big data” analysis. In particular, distributed computing has served as a natural and dominating paradigm for statistical inference. However, the existing literature on parallel inference almost exclusively focuses on Euclidean data and parameters. While this assumption is valid for many applications, it is increasingly more common to encounter problems where the data or the parameters lie on a non-Euclidean space, like a manifold for example. Our work aims to fill a critical gap in the literature by generalizing parallel inference algorithms to optimization on manifolds. We show that our proposed algorithm is both communication efficient and carries theoretical convergence guarantees. In addition, we demonstrate the performance of our algorithm to the estimation of Fréchet means on simulated spherical data and the low-rank matrix completion problem over Grassmann manifolds applied to the Netflix prize data set.'
author:
- |
Bayan Saparbayeva\
Department of Applied and Computational Mathematics and Statistics\
Univeristy of Notre Dame\
Notre Dame, Indiana 46556, USA\
`[email protected] `\
\
Michael Minyi Zhang\
Department of Computer Science\
Princeton University\
Princeton, New Jersey 08540, USA\
`[email protected]`\
\
Lizhen Lin\
Department of Applied and Computational Mathematics and Statistics\
Univeristy of Notre Dame\
Notre Dame, Indiana 46556, USA\
`[email protected]`\
bibliography:
- 'main.bib'
title: Communication Efficient Parallel Algorithms for Optimization on Manifolds
---
Introduction {#sec-intro}
============
A natural representation for many statistical and machine learning problems is to assume the parameter of interest lies on a more general space than the Euclidean space. Typical examples of this situation include diffusion matrices in large scale diffusion tensor imaging (DTI) which are $3\times3$ positive definite matrices, now commonly used in neuroimaging for clinical trials [@dti-ref]. In computer vision, images are often preprocessed or reduced to a collection of subspaces [@subspacepaper; @facialsubpace] or, a digital image can also be represented by a set of $k$-landmarks, forming landmark based shapes [@kendall84]. One may also encounter data that are stored as orthonormal frames [@vecdata], surfaces[@surface], curves[@curve], and networks [@paperwitqeric].
In addition, parallel inference has become popular in overcoming the computational burden arising from the storage, processing and computation of big data, resulting in a vast literature in statistics and machine learning dedicated to this topic. The general scheme in the frequentist setting is to divide the data into subsets, obtain estimates from each subset which are combined to form an ultimate estimate for inference [@Duchi2012DualAF; @JMLR:v14:zhang13b; @Lee2015Communication]. In the Bayesian setting, the subset posterior distributions are first obtained in the dividing step, and these subset posterior measures or the MCMC samples from each subset posterior are then combined for final inference [@median-posterior; @median-selection; @NIPS2015_5986; @Neiswanger:2014; @consensus-mcmc; @nemeth2018merging]. Most of these methods are “embarrassingly parallel” which often do not require communication across different machines or subsets. Some communication efficient algorithms have also been proposed with prominent methods including [@JordanLeeYang] and [@Shamir].
Despite tremendous advancement in parallel inference, previous work largely focuses only on Euclidean data and parameter spaces. To better address challenges arising from inference of big non-Euclidean data or data with non-Euclidean parameters, there is a crucial need for developing valid and efficient inference methods including parallel or distributed inference and algorithms that can appropriately incorporate the underlying geometric structure.
For a majority of applications, the parameter spaces fall into the general category of *manifolds*, whose geometry is well-characterized. Although there is a recent literature on inference of manifold-valued data including methods based on Fréchet means or model based methods [@linclt; @rabivic03; @rabivic05; @rabibook; @sinicapaper] and even scalable methods for certain models [@recht2013parallel; @mackey2015distributed; @salehian2015efficient], there is still a vital lack of general parallel algorithms on manifolds. We aim to fill this critical gap by introducing our parallel inference strategy. The novelty of our paper is in the fact that is generalizable to a wide range of loss functions for manifold optimization problems and that we can parallelize the algorithm by splitting the data across processors. Furthermore, our theoretical development does not rely on previous results. In fact, generalizing Theorem 1 to the manifold setting requires totally different machineries from that of previous work.
Notably, our parallel optimization algorithm has several key features:
1. Our parallel algorithm efficiently exploits the geometric information of the data or parameters.
2. The algorithm minimizes expensive inter-processor communication.
3. The algorithm has theoretical guarantees in approximating the true optimizer, characterized in terms of convergence rates.
4. The algorithm has outstanding practical performance in simulation studies and real data examples.
Our paper is organized as follows: In Section \[sec:related\] we introduce related work to the topic of parallel inference. Next we present our proposed parallel optimization framework in Section \[sec-manifold\] and present theoretical convergence results for our parallel algorithm in Section \[sec:theory\]. In Section \[sec-simu\], we consider a simulation study of estimating the Fréchet means on the spheres and a real data example using the Netflix prize data set. The paper ends with a conclusion and discussion of future work in Section \[sec:conclusion\].
Related work {#sec:related}
============
In the typical “big data” scenario, it is usually the case that the entire data set cannot fit onto one machine. Hence, parallel inference algorithms with provably good theoretic convergence properties are crucial for this situation. In such a setting, we assume that we have $N=mn$ identically distributed observations $ \{x_{ij}: i=1, \ldots, n, j=1, \ldots, m\}$, which are i.i.d divided into $m$ subsets $ X_j=\{x_{ij}, i=1, ..., n\}, j=1, \ldots, m $ and stored in $m$ separate machines. While it is important to consider inference problems when the data are not i.i.d. distributed across processors, we will only consider the i.i.d. setting as a simplifying assumption for the theory.
For a loss function $\mathcal{L}:\Theta\times\mathcal{D}\rightarrow\mathbb{R},$ each machine $ j $ has access to a local loss function, $\mathcal{L}_j(\theta) = \frac{1}{n} \sum_{i=1}^n\mathcal{L}(\theta, x_{ij})$, where $\mathcal{D}$ is the data space. Then, the local loss functions are combined into a global loss function $ \mathcal{L}_N(\theta)=\frac{1}{m}\sum_{j=1}^m\mathcal{L}_j(\theta)$. For our intended optimization routine, we are actually looking for the minimizer of an expected loss function $\mathcal{L}^*(\theta)=\mathbb{E}_{x\in\mathcal{D}}\mathcal{L}(\theta, x) $. In the parallel setting, we cannot investigate $\mathcal{L}^*$ directly and we may only analyze it through $\mathcal{L}_N$. However, calculating the total loss function directly and exactly requires excessive inter-processor communication, which carries a huge computational burden as the number of processors increase. Thus, we must approximate the true parameter $ \theta^*=\arg\min_{\theta\in\Theta}\mathcal{L}^*(\theta) $ by an empirical risk minimizer $\hat{\theta}=\arg\min_{\theta\in\Theta}\mathcal{L}_N(\theta).$
In this work, we focus on generalizing a particular parallel inference framework, the Iterative Local Estimation Algorithm (ILEA) [@JordanLeeYang], to manifolds. This algorithm optimizes an approximate, surrogate loss function instead of the global loss function as a way to avoid processor communication. The idea of the surrogate function starts from the Taylor series expansion of $\mathcal{L}_N$ $$\mathcal{L}_N\big(\bar{\theta}+t(\theta-\bar{\theta})\big)= \mathcal{L}_N(\bar{\theta})+ t\langle\nabla\mathcal{L}_N(\bar{\theta}), \theta-\bar{\theta}\rangle + \sum_{s=2}^{\infty}\frac{t^s}{s!} \nabla^s\mathcal{L}_N(\bar{\theta})(\theta-\bar{\theta})^{\otimes s}.$$ The global high-order derivatives $\nabla^s\mathcal{L}_N(\bar{\theta})$ $(s\geq2)$ are replaced by local high-order derivatives $\nabla^s\mathcal{L}_1(\bar{\theta}) (s\geq2)$ from the first machine $$\tilde{\mathcal{L}}(\theta)= \mathcal{L}_N(\bar{\theta})+ \langle\nabla\mathcal{L}_N(\bar{\theta}), \theta-\bar{\theta}\rangle + \sum_{s=2}^{\infty}\frac{1}{s!} \nabla^s\mathcal{L}_1(\bar{\theta})(\theta-\bar{\theta})^{\otimes s}.$$ So the approximation error is $$\begin{split}
\tilde{\mathcal{L}}(\theta)-\mathcal{L}_N(\theta)&=\sum_{s=2}^{\infty}\frac{1}{s!}\big(\nabla^s\mathcal{L}_1(\bar{\theta})-\nabla^s\mathcal{L}_N(\bar{\theta})\big)(\theta-\bar{\theta})^{\otimes s}\\
&=\frac{1}{2}\Big\langle\theta-\bar{\theta},\big(\nabla^2\mathcal{L}_1(\bar{\theta})-\nabla^2\mathcal{L}_N(\bar{\theta})\big)\big(\theta-\bar{\theta}\big)\Big\rangle+O\big(\parallel\theta-\bar{\theta}\parallel^3\big)\\
&=O\Big( \frac{1}{\sqrt{n}} \parallel\theta-\bar{\theta}\parallel^2+\parallel\theta-\bar{\theta}\parallel^3 \Big).
\end{split}$$ The infinite sum $\sum_{s=2}^{\infty}\frac{1}{s!} \nabla^s\mathcal{L}_1(\bar{\theta})(\theta-\bar{\theta})^{\otimes s}$ in the $\tilde{\mathcal{L}}(\theta)$ can be replaced by $\mathcal{L}_1(\theta)-\mathcal{L}_1(\bar{\theta})-\langle\nabla\mathcal{L}_1(\bar{\theta}), \theta-\bar{\theta}\rangle$ $$\tilde{\mathcal{L}}(\theta)=\mathcal{L}_1(\theta)-\big(\mathcal{L}_1(\bar{\theta})-\mathcal{L}_N(\bar{\theta})\big)-\big\langle\nabla\mathcal{L}_1(\bar{\theta})-\nabla\mathcal{L}_N(\bar{\theta}), \theta-\bar{\theta}\big\rangle.$$ We can omit the additive constant $\big(\mathcal{L}_1(\bar{\theta})-\mathcal{L}_N(\bar{\theta})\big)+\langle\nabla\mathcal{L}_1(\bar{\theta})-\nabla\mathcal{L}_N(\bar{\theta}), \bar{\theta}\rangle$. Thus the surrogate loss function $\tilde{\mathcal{L}}(\theta)$ is defined as $$\tilde{\mathcal{L}}(\theta)=\mathcal{L}_1(\theta)-\langle\nabla\mathcal{L}_1(\bar{\theta})-\nabla\mathcal{L}_N(\bar{\theta}), \theta\rangle.$$ Thus, the surrogate minimizer $\tilde{\theta}=\arg\min_{\Theta}\tilde{\mathcal{L}}$ approximates the empirical risk minimizer $\hat{\theta}$.
[@JordanLeeYang] show that the consequent surrogate minimizers have a provably good convergence rate to $\hat{\theta}$ given the following regularity conditions:
1. The parameter space $\Theta$ is a compact and convex subset of $\mathbb{R}^d.$ Besides, $\theta^*\in{\rm int}(\Theta)$ and $R=\sup_{\theta\in\Theta}\parallel\theta-\theta^*\parallel>0,$
2. The Hessian matrix $I(\theta)=\nabla^2\mathcal{L}^*(\theta)$ is invertible at $\theta^*,$ that is there exist constants $(\mu_{-}, \mu_{+})$ such that $$\mu_{-}I_d\preceq I(\theta^*)\preceq\mu_{+}I_d,$$
3. For any $\delta>0,$ there exists $\epsilon>0,$ such that $$\inf \ \mathbb{P}\bigg\{\inf_{\parallel\theta-\theta^*\parallel\geq\delta}\big|\mathcal{L}(\theta)-\mathcal{L}(\theta^*)\big|\geq\epsilon\bigg\}=1,$$
4. For a ball around the true parameter $U(\rho)=\{\theta: \parallel\theta-\theta^*\parallel\leq\rho\}$ there exist constants $(G, L)$ and a function $\mathcal{K}(x)$ such that $$\begin{split}
\mathbb{E}\parallel\nabla\mathcal{L}(\theta)\parallel^{16}\leq G^{16} \quad \mathbb{E}\VERT\nabla^2\mathcal{L}(\theta)-I(\theta)\VERT\leq L^{16}, \\
\VERT\mathcal{L}(\theta, x)-\mathcal{L}(\theta', x)\VERT\leq\mathcal{K}(x)\parallel\theta-\theta'\parallel,
\end{split}$$ for all $\theta, \theta'\in U(\rho).$
which leads to the following theorem:
Suppose that the standard regularity conditions hold and initial estimator $\bar{\theta}$ lies in the neighborhood $U(\rho)$ of $\theta^*.$ Then the minimizer $\tilde{\theta}$ of the surrogate loss function $\tilde{\mathcal{L}}(\theta)$ satisfies $$\parallel\tilde{\theta}-\hat{\theta}\parallel\leq C_2(\parallel\bar{\theta}-\hat{\theta}\parallel+\parallel\hat{\theta}-\theta^*\parallel+\VERT\nabla^2\mathcal{L}_1(\theta^*)-\nabla^2\mathcal{L}_N(\theta^*)\VERT)\parallel\bar{\theta}-\hat{\theta}\parallel,$$ with probability at least $1-C_1mn^{-8},$ where the constants $C_1$ and $C_2$ are independent of $(m, n, N).$
Parallel optimizations on manifolds {#sec-manifold}
===================================
Our work aims to generalize the typical gradient descent optimization framework to manifold optimization. In particular, we will use the ILEA framework as our working example to generalize parallel optimization algorithms. Instead of working with $\mathbb{R}^d$, we have a $d$-dimensional manifold $M.$ We also consider a surrogate loss function $\tilde{\mathcal{L}}_j:\Theta\times\mathcal{Z}\rightarrow\mathbb{R},$ where $\Theta$ is a subset of the manifold $M$, that approximates the global loss function $\mathcal{L}_N.$ Here we choose to optimize $\tilde{\mathcal{L}}_j$ on the $j$th machine–that is, on different iterations we optimize on different machine for efficient exploration unlike from previous algorithm, where the surrogate function is always optimized on the first machine.
To generalize the idea of moving along a gradient on the manifold $ M $, we use the retraction map, which is not necessarily the exponential map that one would typically use in manifold gradient descent, but shares several important properties with the exponential map. Namely, a retraction on $M$ is a smooth mapping $\mathcal{R}: TM\rightarrow M$ with the following properties
1. $\mathcal{R}_{\theta}(0_{\theta})=\mathcal{R}(\theta, 0_{\theta})=\theta,$ where $\mathcal{R}_{\theta}$ is the restriction of $\mathcal{R}$ from $TM$ to the point $\theta$ and the tangent space $T_{\theta}M,$ $0_{\theta}$ denotes the zero vector on $T_{\theta}M,$
2. $D\mathcal{R}_{\theta}(0_{\theta})=D\mathcal{R}(\theta, 0_{\theta})={\rm id}_{T_{\theta}M},$ where ${\rm id}_{T_{\theta}M}$ denotes the identity mapping on $T_{\theta}M.$
We also demand that
1. For any $\theta_1, \theta_2\in M,$ curves $\mathcal{R}_{\theta_1}t\mathcal{R}_{\theta_1}^{-1}\theta_2$ and $\mathcal{R}_{\theta_2}s\mathcal{R}_{\theta_2}^{-1}\theta_1,$ where $s, t\in[0, 1],$ must coincide,
2. The triangle inequality holds, that is for any $\theta_1, \theta_2, \theta_3\in M$, it is the case that $d_{\mathcal{R}}(\theta_1, \theta_2)\leq d_{\mathcal{R}}(\theta_2, \theta_3)+d_{\mathcal{R}}(\theta_3, \theta_1)$ where $d_{\mathcal{R}}(\theta_1, \theta_2)$ is the length of the curve $\mathcal{R}_{\theta_1}t\mathcal{R}_{\theta_1}^{-1}\theta_2$ for $t\in[0, 1].$
Our construction starts with the Taylor’s formula for $\mathcal{L}_N$ on the manifold $M$ $$\mathcal{L}_N(\theta)=\mathcal{L}_N(\bar{\theta})+ \langle\nabla\mathcal{L}_N(\bar{\theta}), \log_{\bar{\theta}}\theta\rangle+ \sum_{s=2}^{\infty} \frac{1}{s!}\nabla^s\mathcal{L}_N(\bar{\theta})(\log_{\bar{\theta}}\theta)^{\otimes s}$$ Because we split the data across machines, evaluating the derivatives $\nabla^s\mathcal{L}_N(\bar{\theta})$ requires excessive processor communication. We want to reduce the amount of communication by replacing the *global high-order derivatives* $\nabla^s\mathcal{L}_N(\bar{\theta})$ ($s\geq2$) with the *high-order local derivatives* $\nabla^s\mathcal{L}_j(\bar{\theta}).$ This gives us the following surrogate to $\mathcal{L}_N$ $$\tilde{\mathcal{L}}_j(\theta)=\mathcal{L}_N(\bar{\theta})+\langle\nabla\mathcal{L}_N(\bar{\theta}), \log_{\bar{\theta}}\theta\rangle+ \sum_{s=2}^{\infty}\frac{1}{s!}\nabla^s\mathcal{L}_j(\bar{\theta})(\log_{\bar{\theta}}\theta)^{\otimes s}.$$ Then we have the following approximation error $$\begin{aligned}
\tilde{\mathcal{L}}_j(\theta)-\mathcal{L}_N(\theta)&=
\frac{1}{2}\langle\log_{\bar{\theta}}\theta, (\nabla^2\mathcal{L}_j(\bar{\theta})-\nabla^2\mathcal{L}_N(\bar{\theta}))\log_{\bar{\theta}}\theta\rangle+O\big(d_{g}(\bar{\theta}, \theta)^3\big) \\
&=O\bigg(\frac{1}{\sqrt{n}}d_{g}(\bar{\theta}, \theta)^2+d_{g}(\bar{\theta}, \theta)^3\bigg).\end{aligned}$$ We replace $\sum_{s=2}^{\infty}\frac{1}{s!}\nabla^s\mathcal{L}_j(\bar{\theta})(\log_{\bar{\theta}}\theta)^{\otimes s}$ with $\mathcal{L}_j(\theta)-\mathcal{L}_j(\bar{\theta})-\big\langle\nabla\mathcal{L}_j(\bar{\theta}), \log_{\bar{\theta}}\theta\big\rangle:$ $$\begin{aligned}
\tilde{\mathcal{L}}_j(\theta) &=\mathcal{L}_N(\bar{\theta})+ \langle\nabla\mathcal{L}_N(\bar{\theta}),\log_{\bar{\theta}}\theta\rangle+ \mathcal{L}_j(\theta)- \mathcal{L}_j(\bar{\theta})-\langle\nabla\mathcal{L}_j(\bar{\theta}), \log_{\bar{\theta}}\theta\rangle \\
&=\mathcal{L}_j(\theta)+(\mathcal{L}_N(\bar{\theta})-\mathcal{L}_j(\bar{\theta}))+\langle\nabla\mathcal{L}_N(\bar{\theta})-\nabla\mathcal{L}_j(\bar{\theta}),\log_{\bar{\theta}}\theta\rangle.\end{aligned}$$ Since we are not interested in the value of $\tilde{\mathcal{L}}_j$ but in its minimizer, we omit the additive constant $(\mathcal{L}_N(\bar{\theta})-\mathcal{L}_j(\bar{\theta}))$ and redefine $\tilde{\mathcal{L}}_j$ as $\tilde{\mathcal{L}}_j(\theta):=\mathcal{L}_j(\theta)-\langle\nabla\mathcal{L}_j(\bar{\theta})-\nabla\mathcal{L}_N(\bar{\theta}), \log_{\bar{\theta}}\theta\rangle.$ Then we can generalize the exponential map $\exp_{\bar{\theta}}$ and the inverse exponential map $\log_{\bar{\theta}}$ to the retraction map $\mathcal{R}_{\bar{\theta}}$ and the inverse retraction map $\mathcal{R}_{\bar{\theta}}^{-1},$ which is also called the lifting, and redefine $\tilde{\mathcal{L}}_j$ $$\tilde{\mathcal{L}}_j(\theta):=\mathcal{L}_j(\theta)-\langle\nabla\mathcal{L}_j(\bar{\theta})-\nabla\mathcal{L}_N(\bar{\theta}), \mathcal{R}_{\bar{\theta}}^{-1}\theta\rangle.$$ Therefore we have the following generalization of the Iterative Local Estimation Algorithm (ILEA) for the manifold $M$:
Initialize $\theta_0=\bar{\theta};$\
Return $\theta_T$
Convergence rates of the algorithm {#sec:theory}
==================================
To establish some theoretical convergence rates on our algorithm, we consequently have to impose some regularity conditions on the parameter space $\Theta,$ the loss function $\mathcal{L}$ and the population risk $\mathcal{L}^*$. We must establish these conditions specifically for manifolds instead of simply using the regularity conditions placed on Euclidean spaces. For example, in the manifold the Hessians $\nabla^2\mathcal{L}(\theta, x), \nabla^2\mathcal{L}(\theta', x)$ are defined in different tangent spaces meaning there cannot be any linear expressions of the second-order derivatives.
In the manifold for any $\xi\in T_{\theta'}M$ we can define the vector field as $\xi(\theta)=D(\mathcal{R}_{\theta}^{-1}\theta')\xi.$ We can also take the covariant derivative of $\xi(\theta)$ along the retraction $\mathcal{R}_{\theta'}t\mathcal{R}_{\theta'}\theta:$ $$\begin{gathered}
\label{covR}
\nabla_{D\big(R_{\theta'}^{-1}(R_{\theta'}t\mathcal{R}_{\theta'}\theta)\big)^{-1}\mathcal{R}_{\theta'}^{-1}\theta}\xi(\mathcal{R}_{\theta'}t\mathcal{R}_{\theta'}\theta)=\\
\nabla_{D\big(R_{\theta'}^{-1}(R_{\theta'}t\mathcal{R}_{\theta'}\theta)\big)^{-1}\mathcal{R}_{\theta'}^{-1}\theta}D\Big(\mathcal{R}_{\mathcal{R}_{\theta'}t\mathcal{R}_{\theta'}\theta}^{-1}\theta'\Big)\xi=\nabla D(t, \theta, \theta')\xi. \end{gathered}$$ The expression defines the linear map $\nabla D(t, \theta, \theta')$ from $T_{\theta'}M$ to $T_{\mathcal{R}_{\theta'}t\mathcal{R}_{\theta'}\theta}M$ and want to impose some conditions to this map. Finally, we impose the following regularity conditions on the parameter space $\Theta$, the loss function $\mathcal{L}$ and the population risk $\mathcal{L}^*$.
1. The parameter space $\Theta$ is a compact and $\mathcal{R}$-convex subset of $M,$ which means that for any $\theta_1, \theta_2\in\Theta$ curves $\mathcal{R}_{\theta_1}t\mathcal{R}_{\theta_1}\theta_2$ and $\exp_{\theta_1}t\log_{\theta_1}\theta_2$ must be within $\Theta$ for any $\theta_1, \theta_2\in M$ and also demand that there exists $L'\in\mathbb{R}$ such that $$d_{\mathcal{R}}(\theta_1, \theta_2)\leq L'd_g(\theta_1, \theta_2),$$ where $d_g(\theta_1, \theta_2)$ is the geodesic distance,
2. The matrix $I(\theta)=\nabla^2\mathcal{L}^*(\theta)$ is invertible at $\theta^*:$ $\exists$ constants $\mu_{-}, \mu_{+}\in\mathbb{R}$ such that $$\mu_{-}{\rm id}_{\theta^*}\preceq I(\theta^*)\preceq\mu_{+}{\rm id}_{\theta^*},$$
3. For any $\delta>0,$ there exists $\varepsilon>0$ such that $$%\lim_{n\to\infty}
\inf \ \mathbb{P}\Big\{\inf_{d_g(\theta^*, \theta)\geq\delta}\big|\mathcal{L}(\theta)-\mathcal{L}(\theta^*)\big|\geq\varepsilon\Big\}=1,$$
4. There exist constants $(G, L)$ and a function $\mathcal{K}(x)$ such that for all $\theta, \theta'\in U$ and $t\in[0, 1]$ $$\begin{aligned}
\mathbb{E}\parallel\nabla\mathcal{L}(\theta, \mathcal{D})\parallel^{16}\leq G^{16}, \qquad \mathbb{E}\big\VERT\nabla^2\mathcal{L}(\theta, \mathcal{D})-I(\theta)\big\VERT^{16}\leq L^{16}, \\
\parallel\nabla D(t, \theta, \theta')^*\nabla\mathcal{L}(\mathcal{R}_{\theta'}t\mathcal{R}_{\theta'}\theta, x)\parallel\leq \mathcal{K}(x)d_{\mathcal{R}}(\theta, \theta'), \\
\Big\VERT\big(D\mathcal{R}_{\theta}^{-1}\hat{\theta}\big)^*\nabla^2\mathcal{L}(\theta, x)\big(D\mathcal{R}_{\hat{\theta}}^{-1}\theta\big)^{-1}-\big(D\mathcal{R}_{\theta'}^{-1}\hat{\theta}\big)^*\nabla^2\mathcal{L}(\theta', x)\big(D\mathcal{R}_{\hat{\theta}}^{-1}\theta'\big)^{-1}\Big\VERT\leq\mathcal{K}(x)d_{\mathcal{R}_{\bar{\theta}}}(\theta, \theta'),\\
\Big\VERT\big(D\mathcal{R}_{\theta}^{-1}\hat{\theta}\big)^*\nabla^2\mathcal{L}(\theta, x)(D\mathcal{R}_{\theta}^{-1}\hat{\theta}\big)-\big(D\mathcal{R}_{\theta'}^{-1}\hat{\theta}\big)^*\nabla^2\mathcal{L}(\theta', x)\big(D\mathcal{R}_{\theta'}^{-1}\hat{\theta}\big)\Big\VERT\leq\mathcal{K}(x)d_{\mathcal{R}_{\bar{\theta}}}(\theta, \theta'),\end{aligned}$$ where $\VERT \VERT$ is a spectral norm of matrices, $ \VERT A\VERT=\sup\{\parallel Ax\parallel:x\in \mathbb{R}^{n}, \quad \parallel x\parallel=1\}.$ Moreover, $\mathcal{K}$ satisfies $\mathbb{E}\mathcal{K}\leq K^{16}$ for some constant $K>0.$
Given these conditions, we have the following theorem:
If the standard regularity conditions holds, the initial estimator $\bar{\theta}$ lies in the neighborhood $U$ of $\theta^*$ and $$\begin{split}
\Big\VERT\big(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta}\big)^*\big(\nabla^2\tilde{\mathcal{L}}_s(\theta^*)- I(\theta^*)\big)\big(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta}\big)\Big\VERT\leq\frac{\rho\mu_{-}R_{-}}{4},\\
\Big\VERT\big(D\mathcal{R}_{\theta}^{-1}\hat{\theta}\big)^*\nabla^2\tilde{\mathcal{L}}_s(\theta, x)(D\mathcal{R}_{\theta}^{-1}\hat{\theta}\big)-\big(D\mathcal{R}_{\theta'}^{-1}\hat{\theta}\big)^*\nabla^2\tilde{\mathcal{L}}_s(\theta', x)\big(D\mathcal{R}_{\theta'}^{-1}\hat{\theta}\big)\Big\VERT\leq\mathcal{K}(x)d_{\mathcal{R}_{\bar{\theta}}}(\theta, \theta'),
\end{split}$$ where $R_{-}=\frac{1}{\Big\VERT\big((D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})^*(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})\big)^{-1}\Big\VERT},$ then any minimizer $\tilde{\theta}$ of the surrogate loss function $\tilde{\mathcal{L}}_s(\theta)$ satisfies $$\begin{gathered}
d_{\mathcal{R}}(\tilde{\theta}, \hat{\theta})\leq C_2\bigg(1+d_{\mathcal{R}}(\bar{\theta}, \hat{\theta})+d_{\mathcal{R}}(\theta^*, \hat{\theta})+\\
C_3\Big\VERT\big(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta}\big)^*\big(\nabla^2\mathcal{L}_s(\theta^*)-\nabla^2\mathcal{L}_N(\theta^*)\big)\big(D\mathcal{R}_{\hat{\theta}}^{-1}\theta^*\big)^{-1}\Big\VERT\bigg)d_{\mathcal{R}}(\bar{\theta}, \hat{\theta}),
\end{gathered}$$ with probability at least $1-C_1mn^{-8},$ where constants $C_1, C_2$ and $C_3$ are independent of $(m, n, N).$
Simulation study and data analysis {#sec-simu}
==================================
To examine the quality of our parallel algorithm we first apply it to the estimation of Fréchet means on spheres, which has closed form expressions for the estimation of the extrinsic mean (true empirical minimizer). In addition, we apply our algorithm to Netflix movie-ranking data set as an example of optimization over Grassmannian manifolds in the low-rank matrix completion problem. In the following results, we demonstrate the utility of our algorithm both for high dimensional manifold-valued data (Section \[subsec:frechet\]) and Euclidean space data with non-Euclidean parameters (Section \[subsec:netflix\]). We wrote the code for our implementations in Python and carried out the parallelization of the code through MPI[^1][@Dalcin:2005].
Estimation of Fréchet means on manifolds {#subsec:frechet}
----------------------------------------
We first consider the estimation problem of Fréchet means [@frechet] on manifolds. In particular, the manifold under consideration is the sphere in which we wish to estimate both the extrinsic and intrinsic mean [@linclt]. Let $M$ be a general manifold and $\rho$ be a distance on $M$ which can be an intrinsic distance, by employing a Riemannian structure of $M$, or an extrinsic distance, via some embedding $J$ onto some Euclidean space. Also, let $x_1,\ldots, x_N$ be sample of point on the hypersphere $S^d$, the sample Fréchet mean of $x_1,\ldots, x_n$ is defined as $$\begin{aligned}
\hat\theta=\arg\min_{\theta\in M=S^d}\sum_{i=1}^N\rho^2(\theta, x_i),\end{aligned}$$ where $\rho$ is some distance on the sphere.
The extrinsic distance, for our spherical example, is defined to be $\rho(x,y)=\|J(x)-J(y)\|=\|x-y\|$ with $\|\cdot\|$ as the Euclidean distance and the embedding map $J(x)=x\in \mathbb R^{d+1}$ as the identity map. We call $\hat\theta$ the extrinsic Fréchet mean on the sphere. We choose this example in our simulation, as we know the true global optimizer which is given by $\bar{x}/\|\bar x\|$ where $\bar x$ is the standard sample mean of $x_1,\ldots, x_N$ in Euclidean distance. The intrinsic Fréchet mean, on the other hand, is defined to be where the distance $\rho$ is the geodesic distance (or the arc length). In this case we compare the estimator obtained from the parallel algorithm with the optimizer obtained from a gradient descent algorithm along the sphere applied to the entire data set. Despite that the spherical case may be an “easy” setting as it has a Betti number of zero, we chose this example so that we have ground truth with which to compare our results. We, in fact, perform favorably even when the dimensionality of the data is high as we increase the number of processors.
For this example, we simulate one million observations from a $100$-dimensional von Mises distribution projected onto the unit sphere with mean sampled randomly from $ N(0,I) $ and a precision of $2$. For the extrinsic mean example, the closed form expression of the sample mean acts as a “ground truth” to which we can compare our results. In both the extrinsic and intrinsic mean examples, we run $ 20 $ trials of our algorithm over 1, 2, 4, 6, 8 and 10 processors. For the extrinsic mean simulations we compare our results to the true global optimizer in terms of root mean squared error (RMSE) and for the intrinsic mean simulations we compare our distributed results to the single processor results, also in terms of RMSE.
![Extrinsic mean comparison (left) and intrinsic mean comparison (right) on spheres in $ S^{99} $[]{data-label="fig:parallel_comparison"}](extrinsic_mean_comparison_all_p.png "fig:"){width=".5\linewidth"}![Extrinsic mean comparison (left) and intrinsic mean comparison (right) on spheres in $ S^{99} $[]{data-label="fig:parallel_comparison"}](intrinsic_mean_comparison.png "fig:"){width=".5\linewidth"}
As we can see in Figure \[fig:parallel\_comparison\], even if we divide our observations to as many as 10 processors we still obtain favorable results for the estimation of the Fréchet mean in terms of RMSE to the ground truth for the extrinsic mean case and the single processor results for the intrinsic mean case. To visualize this comparison, we show in Figure \[fig:2d\_extrinsic\_mean\] an example of our method’s performance on two dimensional data so that we may see that our optimization results yield a very close estimate to the true global optimizer.
![Extrinsic mean results on $ S^{1} $, for one (left) and ten (right) processors[]{data-label="fig:2d_extrinsic_mean"}](extrinsic_mean_1.png "fig:"){width=".5\linewidth"}![Extrinsic mean results on $ S^{1} $, for one (left) and ten (right) processors[]{data-label="fig:2d_extrinsic_mean"}](extrinsic_mean_10.png "fig:"){width=".5\linewidth"}
Real data analysis: the Netflix example {#subsec:netflix}
---------------------------------------
Next, we consider an application of our algorithm to the Netflix movie rating dataset. This dataset of over a million entries, $ X \in \mathbb{R}^{M \times N} ,$ consists of $M = 17770$ movies and $N = 480189$ users, in which only a sparse subset of the users and movies have ratings. In order to build a better recommendation systems to users, we can frame the problem of predicting users’ ratings for movies as a low-rank matrix completion problem by learning the rank-$ r $ Grassmannian manifold $ U \in \mbox{Gr(M, r)} $ which optimizes for the set of observed entries $ (i,j) \in \Omega$ the loss function $$L(U) = \frac{1}{2} \sum_{(i,j) \in \Omega} \left( (UW)_{ij} - X_{ij} \right)^{2} + \frac{\lambda^{2}}{2} \sum_{(i,j) \notin \Omega}(UW)_{ij},$$ where $W$ is $r$-by-$N$ matrix. Each user $ k$ has the loss function $\mathcal{L}(U,k)=\frac{1}{2} \left| c_k\circ\left( Uw_k(U)-X_k \right) \right|^{2}$ , where $\circ$ is the Hadamard product, $(w_k)^{i}=W_{ik},$ and $$\begin{split}
(c_{k})^i=\begin{cases}
1, & {\rm if} \ \ \ (i, k)\in \Omega \\
\lambda, & {\rm if} \ \ \ (i, k)\notin \Omega
\end{cases},
\qquad
(X_{k})^i=\begin{cases}
X_{ik}, & {\rm if} \ \ \ (i, k)\in \Omega \\
0, & {\rm if} \ \ \ (i, k)\notin \Omega,
\end{cases}\\
w_k(U)=\big(U^T{\rm diag}(c_k\circ c_k)U\big)^{-1}U^T\big(c_k\circ c_k\circ X_k\big).
\end{split}$$ Which results in the following gradient $$\nabla\mathcal{L}(U, k)= \big(c_k\circ c_k\circ(Uw_k(U)-X_k)\big)w_k(U)^T=
{\rm diag}(c_k\circ c_k)(Uw_k(U)-X_k)w_k(U)^T.$$
We can assume that $N=pq,$ then for each local machine $\mathcal{M}_j,$ $j=1, ..., p,$ we have the local function $ \mathcal{L}_j(U)=\frac{1}{q}\sum_{k=(j-1)q+1}^{jq}\mathcal{L}(U,k)$. So the global function is $$\mathcal{L}_N(U)=\frac{1}{p}\sum_{j=1}^p\mathcal{L}_j(U)=\frac{1}{pq}\sum_{k=1}^{pq}\mathcal{L}(U, k)=\frac{1}{N}L(U).$$
For iterations $s=0, 1, ..., P-1$ we have $\nabla\mathcal{L}_j(U_s)=\sum_{k=(j-1)q+1}^{jq}\nabla\mathcal{L}(U_s, k)$. Therefore the global gradient is $\nabla\mathcal{L}_N(U_s)=\frac{1}{p}\sum_{j=1}^p\nabla\mathcal{L}_j(U_s)$. Instead of the logarithm map we will use the inverse retraction map $$\begin{array}{cccc}
{\rm R}_{[U]}^{-1}: & {\rm Gr}(m, r) & \rightarrow & T_{[U]}{\rm Gr}(m, r) \\
& [V] &\mapsto & V-U(U^TU)^{-1}U^TV.
\end{array}$$ Which gives us the following surrogate function $$\begin{aligned}
\tilde{\mathcal{L}}_s(V)&=\mathcal{L}_s(V)-\langle V-U_s(U_s^TU_s)^{-1}U_s^TV, \nabla\mathcal{L}_s(U_s)-\nabla\mathcal{L}_N(U_s)\rangle\\
&=
\mathcal{L}_s(V)-\langle V, \nabla\mathcal{L}_s(U_s)-\nabla\mathcal{L}_N(U_s)\rangle.
\end{aligned}$$ and its gradient $$\nabla\tilde{\mathcal{L}}_s(V)=\nabla\mathcal{L}_s(V)-(I_m-V(V^TV)^{-1}V^T)(\nabla\mathcal{L}_s(U_s)-\nabla\mathcal{L}_N(U_s)).$$ To optimize with respect to our loss function, we have to find $U_{s+1}=\arg\min\tilde{\mathcal{L}}_s$. To do this, we move according to the steepest descent by taking step size $ \lambda_0 $ in the direction $ \nabla\tilde{\mathcal{L}}_s(U_s) $ by taking the retraction, $ U_{s+1}= R_{[U_s]}\left(\lambda_0 \nabla\tilde{\mathcal{L}}_s(U_s) \right) $.[^2]
For our example we set the matrix rank to $ r=10 $ and the regularization parameter to $ \lambda=0.1 $ and divided the data randomly across $ 4 $ processors. Figure \[fig:netflix\] shows that we can perform distributed manifold gradient descent in this complicated problem and we can reach convergence fairly quickly (after about $ 1000 $ seconds).
![Test set RMSE of the Netflix example over time, evaluated on 10 trials.[]{data-label="fig:netflix"}](netflix_timing.png){width=".75\linewidth"}
Conclusion {#sec:conclusion}
==========
We propose in this paper a communication efficient parallel algorithm for general optimization problems on manifolds which is applicable to many different manifold spaces and loss functions. Moreover, our proposed algorithm can explore the geometry of the underlying space efficiently and perform well in simulation studies and practical examples all while having theoretical convergence guarantees.
In the age of “big data”, the need for distributable inference algorithms is crucial as we cannot reliably expect entire datasets to sit on a single processor anymore. Despite this, much of the previous work in parallel inference has only focused on data and parameters in Euclidean space. Realistically, much of the data that we are interested in is better modeled by manifolds and thus we need fast inference algorithms that are provably suitable for situations beyond the Euclidean setting. In future work, we aim to extend the situations under which parallel inference algorithms are generalizable to manifolds and demonstrate more critical problems (in neuroscience or computer vision, for example) in which parallel inference is a crucial solution.
### Acknowledgments {#acknowledgments .unnumbered}
Lizhen Lin acknowledges the support from NSF grants IIS 1663870, DMS Career 1654579, ARO grant W911NF1510440 and a DARPA grant N66001-17-1-4041. Bayan Saparbayeva was partially supported by DARPA N66001-17-1-4041. Michael Zhang is supported by NSF grant 1447721.
Proof to Theorem 2 {#proof-to-theorem-2 .unnumbered}
==================
For $j=1, ..., n,$ let $K_j=\sum_{i=1}^n\mathcal{K}(x_{ij})$ and $\delta_{\rho}=\min\Big\{\rho, \frac{\rho\mu_{-}R_{-}}{4K}\Big\}.$ Consider the following “good events”: $$\begin{aligned}
\begin{split}
\mathcal{E}_0 = & \bigg\{d_{\mathcal{R}}(\hat{\theta}, \theta^*)\leq\min\Big\{ \frac{\rho\mu_{-}}{8K}, \ \ \frac{(1-\rho)\mu_{-}\delta_{\rho}}{8\mu_{+}}, \ \ \sqrt{\frac{(1-\rho)\mu_{-}\delta_{\rho}}{16K}}\Big\}\bigg\} \\
\mathcal{E}_j = & \bigg\{K_j\leq2K, \Big\VERT\big(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta}\big)^*\big(\nabla^2\mathcal{L}_j(\theta^*)- I(\theta^*)\big)\big(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta}\big)\Big\VERT\leq\frac{\rho\mu_{-}R_{-}}{4}, \parallel\nabla\mathcal{L}_j(\theta^*)\parallel
\leq\frac{(1-\rho)\mu_{-}\delta_{\rho}}{4}\bigg\}.
\end{split}
\end{aligned}$$
Under standard regularity conditions, we have $$\mathbb{P}\bigg(\bigcup_{j=1}^m\mathcal{E}_j^c\bigg)\leq\big(c_1+c_2(\log2d)^{16}L^{16}+c_3G^{16}\big)\frac{m}{n^8},$$ where constants $c_1, c_2$ and $c_3$ are independent of $(n, m, N, d, G, L).$
Apply Lemma 6 in Zhang et al \[2013\] [@JMLR:v14:zhang13b], which was also generalized for the manifold.
Under the event $\mathcal{E}_j$ we have $$\parallel\theta_j-\theta^*\parallel\leq\frac{2\parallel\nabla\mathcal{L}_j(\theta^*)\parallel}{(1-\rho)\mu_{-}R_{-}R_{+}}, \qquad (1-\rho)\mu_{-}R_{-}R_{+}{\rm id}_{\theta}\preceq\nabla^2\mathcal{L}_j(\theta),$$ where $R_{+}=\frac{1}{\big\VERT(D\mathcal{R}_{\theta}^{-1}\hat{\theta})(D\mathcal{R}_{\theta}^{-1}\hat{\theta})^*\big\VERT}.$
We first prove that $\mathcal{L}_j$ is $(1-\rho)\mu_{-}R_{-}R_{+}$-strongly convex over the ball $U_{\delta_{\rho}}=\{\theta\in\Theta:d_{\mathcal{R}}(\theta, \theta^*)<\delta_{\rho}\}.$ Indeed $$\begin{gathered}
\VERT(D\mathcal{R}_{\theta}^{-1}\hat{\theta})^*\nabla^2\mathcal{L}_j(\theta)(D\mathcal{R}_{\theta}^{-1}\hat{\theta})-(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})^*\nabla^2\mathcal{L}^*(\theta^*)(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})\VERT\leq \\
\VERT(D\mathcal{R}_{\theta}^{-1}\hat{\theta})^*\nabla^2\mathcal{L}_j(\theta)(D\mathcal{R}_{\theta}^{-1}\hat{\theta})-(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})^*\nabla^2\mathcal{L}_j(\theta^*)(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})\VERT + \\
\qquad \VERT(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})^*\nabla^2\mathcal{L}_j(\theta^*)(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})-(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})^*I(\theta^*)(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})\VERT\leq \\
K_jd_{\mathcal{R}}(\theta, \theta^*)+\frac{\rho\mu_{-}R_{-}}{2}\leq
2Kd_{\mathcal{R}}(\theta, \theta^*)+\frac{\rho\mu_{-}R_{-}}{2}
\end{gathered}$$ Lets consider $$\begin{gathered}
(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})^*I(\theta^*)(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})- (D\mathcal{R}_{\theta}^{-1}\hat{\theta})^*\nabla^2\mathcal{L}_j(\theta)(D\mathcal{R}_{\theta}^{-1}\hat{\theta})\preceq\\
\VERT(D\mathcal{R}_{\theta}^{-1}\hat{\theta})^*\nabla^2\mathcal{L}_j(\theta)(D\mathcal{R}_{\theta}^{-1}\hat{\theta})-(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})^*I(\theta^*)(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})\VERT{\rm id}_{\hat{\theta}}\preceq\\
\Big(2Kd_{\mathcal{R}}(\theta, \theta^*)+\frac{\rho\mu_{-}R_{+}}{2}\Big){\rm id}_{\hat{\theta}}
\end{gathered}$$ From the first regularity condition we have $$\begin{gathered}
\mu_{-}(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})^*(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})-(D\mathcal{R}_{\theta}^{-1}\hat{\theta})^*\nabla^2\mathcal{L}_j(\theta)(D\mathcal{R}_{\theta}^{-1}\hat{\theta})\preceq \\
(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})^*I(\theta^*)(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})- (D\mathcal{R}_{\theta}^{-1}\hat{\theta})^*\nabla^2\mathcal{L}_j(\theta)(D\mathcal{R}_{\theta}^{-1}\hat{\theta}),
\end{gathered}$$ so $$\begin{gathered}
\mu_{-}R_{-}{\rm id}_{\hat{\theta}}-(D\mathcal{R}_{\theta}^{-1}\hat{\theta})^*\nabla^2\mathcal{L}_j(\theta)(D\mathcal{R}_{\theta}^{-1}\hat{\theta})\preceq \\
(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})^*I(\theta^*)(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})- (D\mathcal{R}_{\theta}^{-1}\hat{\theta})^*\nabla^2\mathcal{L}_j(\theta)(D\mathcal{R}_{\theta}^{-1}\hat{\theta}).
\end{gathered}$$ Then by our choice of $\delta_{\rho}\leq\frac{\rho\mu_{-}R_{-}}{4K}$ $$\begin{gathered}
(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})^*I(\theta^*)(D\mathcal{R}_{\theta^*}^{-1}\hat{\theta})- (D\mathcal{R}_{\theta}^{-1}\hat{\theta})^*\nabla^2\mathcal{L}_j(\theta)(D\mathcal{R}_{\theta}^{-1}\hat{\theta})\preceq\\
\Big(2Kd_{\mathcal{R}}(\theta, \theta^*)+\frac{\rho\mu_{-}R_{-}}{2}\Big){\rm id}_{\hat{\theta}}\preceq\rho\mu_{-}R_{-}{\rm id}_{\hat{\theta}}.
\end{gathered}$$ So finally we have $$\mu_{-}R_{-}{\rm id}_{\hat{\theta}}-(D\mathcal{R}_{\theta}^{-1}\hat{\theta})^*\nabla^2\mathcal{L}_j(\theta)(D\mathcal{R}_{\theta}^{-1}\hat{\theta})\preceq \rho\mu_{-}R_{-}{\rm id}_{\hat{\theta}}$$ and so $$(D\mathcal{R}_{\theta}^{-1}\hat{\theta})^*\nabla^2\mathcal{L}_j(\theta)(D\mathcal{R}_{\theta}^{-1}\hat{\theta})\succeq(1-\rho)\mu_{-}R_{-}{\rm id}_{\hat{\theta}},$$ and so $$\begin{gathered}
\nabla^2\mathcal{L}_j(\theta)\succeq(1-\rho)\mu_{-}R_{-}\big((D\mathcal{R}_{\theta}^{-1}\hat{\theta})^*\big)^{-1}(D\mathcal{R}_{\theta}^{-1}\hat{\theta})^{-1}\succeq (1-\rho)\mu_{-}R_{-}R_{+}{\rm id}_{\theta}.
\end{gathered}$$ It means that $\mathcal{L}_j$ is $(1-\rho)\mu_{-}R_{-}R_{+}$-strongly convex.
Thus by the definition of strong convexity $$\mathcal{L}_j(\theta')\geq\mathcal{L}_j(\theta^*)+\langle\nabla\mathcal{L}_1(\theta^*), \mathcal{R}_{\theta^*}^{-1}\theta'\rangle+\frac{(1-\rho)\mu_{-}R_{-}R_{+}}{2}d_{\mathcal{R}}(\theta', \theta^*)^2,$$ then $$\begin{aligned}
\begin{split}
d_{\mathcal{R}}(\theta_j, \theta^*)^2\leq&\frac{2}{(1-\rho)\mu_{-}R_{-}R_{+}}\big(\mathcal{L}_j(\theta_j)-\mathcal{L}_j(\theta^*)-\langle\nabla\mathcal{L}_j(\theta^*), \mathcal{R}_{\theta^*}^{-1}\theta_j\rangle\big)\leq\\
&\frac{2}{(1-\rho)\mu_{-}R_{-}R_{+}}\big(\mathcal{L}_j(\theta_j)-\mathcal{L}_j(\theta^*)+\parallel\nabla\mathcal{L}_j(\theta^*)\parallel d_{\mathcal{R}}(\theta_j, \theta^*)\big).
\end{split}
\end{aligned}$$ Dividing each side by $d_{\mathcal{R}}(\theta', \theta^*)$ $$d_{\mathcal{R}}(\theta_j, \theta^*)\leq \frac{2}{(1-\rho)\mu_{-}R_{-}R_{+}}\Big(\frac{\mathcal{L}_j(\theta_j)-\mathcal{L}_j(\theta^*)}{d_{\mathcal{R}}(\theta_j, \theta^*)}+\parallel\nabla\mathcal{L}_j(\theta^*)\parallel\Big).$$ Hence $$d_{\mathcal{R}}(\theta_j, \theta^*)< \frac{2\parallel\nabla\mathcal{L}_j(\theta^*)\parallel}{(1-\rho)\mu_{-}R_{-}R_{+}}.$$
Then from the previous lemma $$d_{\mathcal{R}}(\hat{\theta}, \theta^*)\leq\frac{2\parallel\nabla\mathcal{L}_N(\theta^*)\parallel}{(1-\rho)\mu_{-}R_{-}R_{+}}.$$
Under standard regularity conditions 2 and 4, there exist universal constants $c, c'$ such that for $\nu\in\{1, ..., 8\},$ $$\begin{aligned}
\begin{split}
\mathbb{E}\parallel\nabla\mathcal{L}_j(\theta^*)\parallel^{2\nu}&\leq\frac{cG^{2\nu}}{n^{nu}},\\
\mathbb{E}\big\VERT\nabla^2\mathcal{L}_j(\theta^*)-I(\theta^*)\big\VERT^{2\nu}&\leq\frac{c'(\log2d)^{\nu}L^{\nu}}{n^{\nu}}.
\end{split}\end{aligned}$$
Now we apply Markov’s inequality, Jensen’s inequality and the union bound to obtain that there exist constants $c_1, c_2, c_3$ independent of $(n, m, N, d, G, L)$ such that $$\mathbb{P}\bigg(\bigcup_{j=1}^m\mathcal{E}_j^c\bigg)\leq\big(c_1+c_2(\log2d)^{16}L^{16}+c_3G^{16}\big)\frac{m}{n^8}.$$
Suppose that standard regularity conditions hold. Then under event $\mathcal{E}_0\cap\mathcal{E}_s$ we have $$d_{\mathcal{R}}(\tilde{\theta}, \hat{\theta})\leq\frac{2\parallel\nabla\tilde{\mathcal{L}}_s(\hat{\theta})\parallel}{(1-\rho)\mu_{-}R_{-}R_{+}}.$$
A simple calculation yields $$\begin{gathered}
\nabla\tilde{\mathcal{L}}_s(\hat{\theta})=\nabla\mathcal{L}_s(\hat{\theta}) -(D\mathcal{R}_{\bar{\theta}}^{-1}\hat{\theta})^*(\nabla\mathcal{L}_s(\bar{\theta})-\nabla\mathcal{L}_N(\bar{\theta}))=\\
\nabla\mathcal{L}_s(\hat{\theta}) -(D\mathcal{R}_{\bar{\theta}}^{-1}\hat{\theta})^*\nabla\mathcal{L}_s(\bar{\theta})-\big(\nabla\mathcal{L}_N(\hat{\theta}) -(D\mathcal{R}_{\bar{\theta}}^{-1}\hat{\theta})^*\nabla\mathcal{L}_N(\bar{\theta})\big).\end{gathered}$$ Actually for any $\xi\in T_{\hat{\theta}}M$ we have the vector field $\xi(\theta)=D\big(\mathcal{R}_{\theta}^{-1}\hat{\theta}\big)\xi,$ then $$\xi(\mathcal{L}_s)(\theta)=\langle\nabla\mathcal{L}_s(\theta), \xi(\theta)\rangle=\langle D\big(\mathcal{R}_{\theta}^{-1}\hat{\theta}\big)^*\nabla\mathcal{L}_s(\theta), \xi\rangle.$$ Also along the retraction $\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta}$ we have the tangent vector field $$\eta(t)=D\big(\mathcal{R}_{\hat{\theta}}^{-1}(\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta})\big)^{-1}\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta}.$$ So along the curve $\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta}$ $$\begin{gathered}
\eta\xi(\mathcal{L}_s)(t)=
\langle\nabla^2\mathcal{L}_s(\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta})\eta(t), \xi(\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta})\rangle+
\langle\nabla\mathcal{L}_s(\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta}), \nabla_{\eta(t)}\xi(\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta})\rangle=\\
\langle D\big(\mathcal{R}_{\theta}^{-1}\hat{\theta}\big)^*\nabla^2\mathcal{L}_s(\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta})D\big(\mathcal{R}_{\hat{\theta}}^{-1}(\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta})\big)^{-1}\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta}+\nabla D(t, \bar{\theta}, \hat{\theta})\nabla\mathcal{L}_s(\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta}), \xi\rangle.\end{gathered}$$ Thus $$\begin{gathered}
\xi(\mathcal{L}_s)(\bar{\theta})-\xi(\mathcal{L}_s)(\hat{\theta})=\langle D\big(\mathcal{R}_{\bar{\theta}}^{-1}\hat{\theta}\big)^*\nabla\mathcal{L}_s(\bar{\theta})-\nabla\mathcal{L}_s(\hat{\theta}), \xi\rangle=\\
\bigg\langle\int_0^1\Big( D\big(\mathcal{R}_{\theta}^{-1}\hat{\theta}\big)^*\nabla^2\mathcal{L}_s(\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta})D\big(\mathcal{R}_{\hat{\theta}}^{-1}(\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta})\big)^{-1}\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta}+\nabla D(t, \bar{\theta}, \hat{\theta})\nabla\mathcal{L}_s(\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta})\Big)dt,\xi\bigg\rangle.\end{gathered}$$ Therefore we have $$\nabla\mathcal{L}_s(\hat{\theta}) -(D\mathcal{R}_{\bar{\theta}}\hat{\theta})^*\nabla\mathcal{L}_s(\bar{\theta})=-H_s\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta}-\int_0^1\nabla D(t, \bar{\theta}, \hat{\theta})^*\nabla\mathcal{L}_s(\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta}),$$ where $$H_s=\int_0^1(D\mathcal{R}_{\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta}}\hat{\theta})^*\nabla^2\mathcal{L}_s(\mathcal{R}_{\bar{\theta}}t\mathcal{R}_{\bar{\theta}}^{-1}\hat{\theta})\big(D\mathcal{R}_{\hat{\theta}}^{-1}(\mathcal{R}_{\hat{\theta}}t\mathcal{R}_{\hat{\theta}}^{-1}\bar{\theta})\big)^{-1}dt$$ satisfies $$\big\VERT H_s-(D\mathcal{R}_{\theta^*}\hat{\theta})^*\nabla^2\mathcal{L}_s(\theta^*)(D\mathcal{R}_{\hat{\theta}}\theta^*)^{-1}\big\VERT\leq2K\big(d_{\mathcal{R}}(\bar{\theta},\hat{\theta})+d_{\mathcal{R}}(\hat{\theta}, \theta^*)\big).$$ Therefore $$\begin{aligned}
\begin{split}
\VERT\nabla\tilde{\mathcal{L}}_s(\hat{\theta})\VERT\leq
&4Kd_{\mathcal{R}}(\hat{\theta}, \bar{\theta})+\big\VERT H_s-(D\mathcal{R}_{\theta^*}\hat{\theta})^*\nabla^2\mathcal{L}_s(\theta^*)(D\mathcal{R}_{\hat{\theta}}\theta^*)^{-1}\big\VERT d_{\mathcal{R}}(\hat{\theta},
\bar{\theta})+\\
&\quad
\Big(\frac{1}{m}\sum_{j=1}^m
\big\VERT H_j-(D\mathcal{R}_{\theta^*}\hat{\theta})^*\nabla^2\mathcal{L}_j(\theta^*)(D\mathcal{R}_{\hat{\theta}}\theta^*)^{-1}\big\VERT\Big)d_{\mathcal{R}}(\hat{\theta}, \bar{\theta})+\\
&\qquad \ \big\VERT(D\mathcal{R}_{\theta^*}\hat{\theta})^*\big(\nabla^2\mathcal{L}_s(\theta^*)-\nabla^2\mathcal{L}_N(\theta^*)\big)(D\mathcal{R}_{\hat{\theta}}\theta^*)^{-1}\big\VERT d_{\mathcal{R}}(\hat{\theta}, \bar{\theta})\leq \\
&\bigg(4K+\big\VERT H_s-(D\mathcal{R}_{\theta^*}\hat{\theta})^*\nabla^2\mathcal{L}_s(\theta^*)(D\mathcal{R}_{\hat{\theta}}\theta^*)^{-1}\big\VERT+\\
&\quad \Big(\frac{1}{m}\sum_{j=1}^m
\big\VERT H_j-(D\mathcal{R}_{\theta^*}\hat{\theta})^*\nabla^2\mathcal{L}_j(\theta^*)(D\mathcal{R}_{\hat{\theta}}\theta^*)^{-1}\big\VERT\Big)+\\
&\qquad \ \big\VERT(D\mathcal{R}_{\theta^*}\hat{\theta})^*\big(\nabla^2\mathcal{L}_s(\theta^*)-\nabla^2\mathcal{L}_N(\theta^*)\big)(D\mathcal{R}_{\hat{\theta}}\theta^*)^{-1}\big\VERT\bigg)d_{\mathcal{R}}(\hat{\theta}, \bar{\theta})\leq\\
&\bigg(4K+4Kd_{\mathcal{R}}(\theta_j, \hat{\theta})+4Kd_{\mathcal{R}}(\hat{\theta}, \bar{\theta})+\\
&\quad \big\VERT(D\mathcal{R}_{\theta^*}\hat{\theta})^*\big(\nabla^2\mathcal{L}_s(\theta^*)-\nabla^2\mathcal{L}_N(\theta^*)\big)(D\mathcal{R}_{\hat{\theta}}\theta^*)^{-1}\big\VERT\bigg)d_{\mathcal{R}}(\hat{\theta}, \bar{\theta}).
\end{split}\end{aligned}$$ Hence $$\begin{aligned}
\begin{split}
\big\VERT\nabla\tilde{\mathcal{L}}_s(\hat{\theta})\big\VERT\leq
&\bigg(4K+4Kd_{\mathcal{R}}(\bar{\theta}, \hat{\theta})+4Kd_{\mathcal{R}}(\hat{\theta}, \bar{\theta})+\\
&\quad \big\VERT(D\mathcal{R}_{\theta^*}\hat{\theta})^*\big(\nabla^2\mathcal{L}_s(\theta^*)-\nabla^2\mathcal{L}_N(\theta^*)\big)(D\mathcal{R}_{\hat{\theta}}\theta^*)^{-1}\big\VERT\bigg)d_{\mathcal{R}}(\hat{\theta}, \bar{\theta}).
\end{split}\end{aligned}$$
Then comes lemma 4.
[^1]: Our code is available at <https://github.com/michaelzhang01/parallel_manifold_opt>
[^2]: We select the step size parameter according to the modified Armijo algorithm seen in [@boumal2014thesis].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Although ultracool dwarfs (UCDs) are now known to generate and dissipate strong magnetic fields, a clear understanding of the underlying dynamo is still lacking. We have performed X-ray and radio observations of seven UCDs in a narrow range of spectral type (M6.5–M9.5) but spanning a wide range of projected rotational velocities ($\vsi \approx 3$–40 ). We have also analyzed unpublished archival observations of four additional objects. All of the newly-observed targets are detected in the X-ray, while only one is detected in the radio, with the remainder having sensitive upper limits. We present a database of UCDs with both radio and X-ray measurements and consider the data in light of the so-called Güdel-Benz relation (GBR) between magnetic activity in these bands. Some UCDs have very bright radio emission and faint X-ray emission compared to what would be expected for rapid rotators, while others show opposite behavior. We show that UCDs would still be radio-over-luminous relative to the GBR even if their X-ray emission were at standard rapid-rotator “saturation” levels. Recent results from Zeeman-Doppler imaging and geodynamo simulations suggest that rapidly-rotating UCDs may harbor a bistable dynamo that supports either a stronger, axisymmetric magnetic field or a weaker, non-axisymmetric field. We suggest that the data can be explained in a scenario in which strong-field objects obey the GBR while weak-field objects are radio-over-luminous and X-ray-under-luminous, possibly because of a population of gyrosynchrotron-emitting coronal electrons that is continuously replenished by low-energy reconnection events.'
author:
- 'P. K. G. Williams, B. A. Cook, and E. Berger'
bibliography:
- '../bib/pkgw.bib'
- 'extra.bib'
- 'pkgw.bib'
- 'extra.bib'
title: 'Trends in Ultracool Dwarf Magnetism. I. X-Ray Suppression and Radio Enhancement'
---
Introduction {#s.intro}
============
The magnetic activity of very low mass stars and brown dwarfs (collectively, “ultracool dwarfs” or UCDs; here, objects of spectral types M7 or later) raises several challenging astrophysical questions. UCD activity cannot be driven by a solar-type shell dynamo, in which the tachocline (the shearing interface layer between the radiative and convective zones) plays a crucial role; full convection sets in around spectral type M4, and so the tachocline simply does not exist in UCDs [@cb00]. Nonetheless, observations reveal persistent, strong (kG) magnetic fields at the bottom of the main sequence [@bbb+01; @b02; @b06b; @bbf+10; @had+06; @rb07; @opbh+09; @rb10; @mdp+10; @ahd+13]. Despite substantial theoretical work to understand the fully-convective dynamo , its properties, and those of the resulting magnetic fields, remain unclear. Further progress requires observational input based on studies of magnetic activity indicators.
X-ray and emission are two common indicators, and decades of study of solar-type stars have yielded several important relationships between these quantities and other stellar properties. One such relationship is between X-ray activity (, where the denominator is the bolometric stellar luminosity) and stellar rotation [@nhb+84], which follows a “saturation” pattern in which activity increases with rotation until it reaches $\LxLb \sim 10^{-3}$ [@v84; @pmm+03; @jjb+11; @wdmh11]. The same general pattern is observed in emission .
Radio emission is another powerful magnetic activity indicator, and there are also well-known correlations between the radio and soft X-ray (SXR) luminosities of magnetically active stellar systems. @dsl89 first noted a correlation betwen and in RSCVn systems. @gb93 analyzed a larger sample and found that for the most active F–M stars, $\Lx \sim
\sLr\times10^{15.5}$ Hz. This result was then extended to solar flares and other active binaries by @bg94. Over their whole dataset, spanning 10 orders of magnitude in radio spectral luminosity, $\Lx \propto \sLr^\alpha$ with $\alpha \sim 0.73$, a result now commonly known as the Güdel-Benz relation (GBR). Continuity over such a broad range strongly suggests both a common driver of emission in the two bands (despite the fact that the fundamental emission processes operate in very different conditions), as well as common physical processes across this range of emitters. In the standard interpretation, magnetic reconnection accelerates a population of nonthermal particles, leading to radio emission; these particles then deposit some of their energy in the chromosphere, where ablated material concentrates in coronal loops and emits thermally in the SXR band. This model is supported by the observed “Neupert effect” [@n68], in which $\text{d}\Lx/\text{d}t
\propto \Lr$, suggesting that the SXR emission tracks the total energy deposited by the particle acceleration process. This effect is well-established though far from universal in both the solar and stellar contexts .
Observations of magnetic activity tracers in UCDs paint a different picture. Despite the demonstrated existence of strong fields in at least some UCDs, their X-ray and emission both drop off precipitously [@smf+06; @bbf+10; @gmr+00; @whw+04]. The rotation/activity relation evolves significantly in UCDs, with evidence for trends in which X-ray and activity decrease as rotation increases [@bm95; @mb03; @wb09; @bbf+10], reminiscent of the much weaker “supersaturation” effect seen in active stars [@prs+96; @jjb+11]. But UCD radio emission remains stubbornly unchanged: in this regime, radio activity and radio surface flux ($L_\text{rad}/R_*^2$) increase with rotation, with no evidence of saturation [@mbr12].
These trends offer clues toward a deeper understanding of the fully convective dynamo and the structures of (sub)stellar magnetic fields and outer atmospheres. Progress toward this understanding, however, is hampered by the relatively small number of UCDs detected in the X-ray regime. In an attempt to improve this situation, we have observed numerous UCDs with the X-ray Observatory [@brr+05; @bgg+08; @bbg+08; @brpb+09; @bbf+10]. In this work, we report new observations of 7 UCDs of spectral type M7 with a wide range of rotational velocities, all of which were detected. Every source was also observed (non-simultaneously) with the upgraded Karl G. Jansky Very Large Array (VLA), yielding one detection. Our detections offer a striking counterpoint to previous observations of objects objects only a few spectral subtypes later, in which X-ray detections have been elusive . We also report the analysis of several unpublished measurements from archival data. We combine these results with data from our previous work and the literature to form a comprehensive database of UCD activity measurements.
In this work, we use this database to investigate the correlation between X-ray and radio emission in UCDs. We proceed here by describing the targets for which we present new results (§\[s.targets\]) and our data analysis (§§\[s.xray\], \[s.radio\]). We then discuss our results in the context of the full sample of UCDs with X-ray and radio observations (§\[s.trends\]). We offer a physical model that explains the observations (§\[s.disc\]). In @paper2 [hereafter ], we use the same database to investigate the correlation between UCD X-ray emission and rotation.
Throughout this work, we use the notation $[x] \equiv \log_{10} x$, with $x$ being measured in cgs units if it is a dimensional quantity, unless its units are specified otherwise.
Targets with New Results {#s.targets}
========================
Candidates for new radio and X-ray observations were selected by searching [dwarfarchives.org](http://dwarfarchives.org) for nearby UCDs of spectral type M7 that were visible to the VLA and had measurements of in the literature. We also searched the data archive for unpublished observations of late-type objects with measurements. Seven targets were observed with and the VLA, and we identified and analyzed four archival targets. The main characteristics of these objects, along with the appropriate references and 2MASS identifications [@the2mass], are provided in Table \[t.tinfo\], while their properties are discussed in greater detail in the Appendix. The Appendix also describes our method for computing bolometric luminosities.
Observations {#s.obs}
------------
The new observations were performed with /ACIS-S between 2011 December and 2013 February (proposal 13200167; observation IDs 13603–13609; PI: Berger), using the S3 backside-illuminated chip. All exposures were 20 ks, except for which was observed for 10 ks. Parameters of the observations are provided in Table \[t.xinfo\]. No grating was used, the data mode was , and the exposure mode was “timed” (TE).
The targets were also observed with the Karl G. Jansky Very Large Array (VLA) between 2012 February and 2012 May (project VLA/12A-089; PI: Berger). These observations were not simultaneous with the X-ray observations, and the relative timing between observations in the two bands was arbitrary. Each observing session lasted 1 hr, with a total correlated bandwidth of 2048 MHz divided into two basebands centered at 5000 and 7100 MHz, each containing 512 spectral channels. Bandpass, flux density scale, and complex gain calibrations were obtained in the usual way; the sources used are listed in Table \[t.rinfo\], along with other parameters of the observations. Approximately 65% of each session was spent integrating on the target source.
The archival data were obtained as follows. was observed with /ACIS-S on 2003 December 14 (proposal 05200058; observation ID 4476; PI: Garmire) for 24 ks. was observed in two parts (proposal 09200200; observation IDs 8903 and 10828; PI: Audard) on 2008 December 2 and 2008 December 4, for a total exposure time of 28 ks. was observed on 2009 September 15 (proposal 10200468; observation ID 9925; PI: Osten) with an exposure time of 37 ks. In all cases, the instrumental configuration was the same as in our new observations.
X-Ray Analysis {#s.xray}
==============
We analyzed the data in CIAO version 4.5 [@theciao] with CalDB version 4.5.5.1. We used the date of each observation and astrometric information from the Simbad database to predict positions for each target at the time of our observations. Based on Monte Carlo simulations, all of our predictions have uncertainties of $\lesssim2''$, and most of them have uncertainties of $\lesssim0.5''$. This is comparable to the astrometric precision of . We defined initial source apertures $2''$ in radius centered on our astrometric predictions of the source positions. As discussed below, two of our targets ( and ) are affected by pileup, leading us to use annular apertures instead. For the rest of the targets, the initial apertures needed no modification.
Following reprocessing to eliminate a substantial fraction of the background events, we estimated the mean residual background in each dataset by extracting events in an energy range of 0.3–7 keV in large, source-free regions near the target locations. In all cases, the expected number of background counts in the source aperture is $\lesssim$1.
Every source except the archival L8 dwarf was detected at $>$5$\sigma$ significance. In our new observations, the number of counts at the source location ranges between 6 (, with $0.17$ expected background events) and 252 (). With the high success rate in our new observations, we nearly double the number of UCDs with X-ray detections (cf. Table \[t.dbsrcs\]). The only target with nontrivial X-ray structure and annular source extraction apertures, the triple system , is shown in Figure \[f.g208img\]. Cutouts of the X-ray images around the predicted locations of the other sources (as well as radio images; see next section) are shown in Figure \[f.cutouts\].
During extraction we also checked for background flaring. The only dataset in which flares were seen was the observation of . We discuss our handling of this dataset separately below.
We searched for variability in the X-ray emission of the detected sources using a Bayesian blocks analysis [@s98; @snjc13]. This approach models the source flux as a series of independent, piecewise constant “blocks,” with overfitting being controlled by the use of a downward-sloping prior on the number of blocks ($N_b$). Our implementation of the algorithm uses the iterative approach described in @snjc13 with a Monte-Carlo-derived parametrization of the prior on $N_b$ that sets the probability of false detection of an extraneous block at 5%. This parametrization is given in Equation 21 of @snjc13 but is misstated; the correct equation is $$\text{ncp\_prior} = 4 - \ln \left(73.53 p_0 N^{-0.478}\right),$$ as may be verified with the example given in that work. Our implementation is in Python and derives from the MatLab code provided by @snjc13 and the Python module [@theastroml]. Compared to the latter module, our system adds support for time-tagged events and datasets with gaps in coverage (due to the background flaring in our case). It also fixes several minor bugs such as the mistaken equation above. Our implementation is publicly available[^1].
The Bayesian blocks analysis finds more than one block — that is, significant evidence of variability — in five sources. We plot the X-ray light curves resulting from this analysis in Figure \[f.gtilcs\], also showing the results of uniform binning for reference. The plots also indicate the “good time intervals” in which background flares were not an issue; they are continuous except for the observations of . Table \[t.xfluxes\] includes information on the flare durations and fluxes. Of the five sources with more than one block, four of them contain two blocks, suggesting partially-observed flares. The last, , has three blocks, with the pre- and post-flare fluxes agreeing at the 20% level. From visual inspection of Figure \[f.gtilcs\], one may conclude that both a flatter prior on $N_b$ (i.e., an assumption of a higher likelihood that sources are variable) and that a non-piecewise-constant emission model would lead to more faithful approximations of the data. Because the aim of our analysis is limited to identifying representative quiescent and (when appropriate) flaring X-ray fluxes, we do not explore these possible elaborations here.
We determined X-ray fluxes in the 0.2–2 keV band using either spectral modeling or a simple energy conversion factor (ECF). We used the ECF approach for all but the three brightest sources: , , and . After extracting spectra from the event data of these three sources, grouping into bins of $\ge$12 events, we used version 1 [@thesherpa] for the modeling, ignoring energies outside of the standard ACIS energy filter of 0.3–7 keV. We used the implementation of the @thesimplex simplex algorithm to optimize the modified $\chi^2$ statistic of @g86. Solar abundances were taken from @l03. Our X-ray flux results and the outcomes of the modeling are discussed below. We summarize the fluxes in Table \[t.xfluxes\], where the uncertainties are properly propagated and account for Poisson statistics as appropriate. Derived quantities for the UCDs, including X-ray luminosities, are presented in Table \[t.dbdata\].
In , a one-temperature, solar-abundance APEC [Astrophysical Plasma Emission Code; @theapec] model yields a satisfactory fit, achieving a reduced statistic $\chi_r^2=1.1$ with 18 degrees of freedom (DOF), although there are strong correlations among the residuals. The data and best-fit model are shown in Figure \[f.xspecs\].
The model-fitting procedure finds $kT = 0.78 \pm 0.04$ keV, but we caution generally against overinterpreting the parameters derived from our spectral fits. While the data we present cannot rule out a single-temperature solar-abundance model, the truth is likely more complicated. In particular, the value of $kT$ we report above is quite possibly an approximate average of multiple temperature components. Intriguingly, if we overfit the data by adopting a two-temperature model, we find temperatures of 0.3 and 1.2 keV, in good agreement with other results from mid-to-late M dwarfs . Similarly, if we use a single-temperature model with variable abundances, the fits tend to converge on an inverse first ionization potential (FIP) effect, with higher abundances found for elements with higher FIPs, as is commonly seen in higher-S/N spectra of similar objects . We emphasize, however, that these findings are not statistically significant.
Granting these caveats, the fluxes we derive from our spectral modeling are robust, because the essential shapes of the X-ray spectra are well-constrained by the data. For we find the X-ray flux in the 0.2–2 keV band to be $\fx = -13.29\pm0.04$. At our adopted distance, this corresponds to $\lx
= 27.1$ and $\lxlb = -3.1$, near the canonical “saturation” value of the X-ray/rotation activity relation observed in solar-type stars .
{#s.g208}
While and are clearly resolved in the image (Figure \[f.g208img\]), the tighter binary (A and B) is not. The following analysis considers only the blended emission of .
During our analysis of the dataset, we discovered significant background flaring activity. We used standard CIAO tools to extract a light curve of non-source regions on the source chip (S3), binning by 150 s in time and finding a non-flaring background rate of 0.32 s$^{-1}$ across the whole chip, which is consistent with typical nonflaring behavior. We flagged bins in which the measured background rate varied from this value by $>$4$\sigma$. After processing, the good exposure time was reduced from 23.8 to 12.1 ks. We note that the Bayesian blocks method can be applied to data with observational gaps without adjustment [@snjc13], making it well-suited for these data.
Both components of the system additionally achieved count rates sufficiently high ($>$0.1 s$^{-1}$) to make pileup a concern. We used the Portable, Interactive Multi-Mission Simulator () to estimate the pileup percentages during the observations. For both objects, pileup was estimated at the 5% level. To compensate for this we analyzed both components using annuli centered on the source positions, removing the centermost pixels where the count rate and hence pileup were greatest. For both sources, the annuli had inner (outer) radii of 0.5 (2.25) arcsec. This approach discards significant signal, but both sources were bright enough that we still retained strong detections sufficient for spectral modeling. The reported fluxes account for the reduced portion of the PSF being sampled. Filtering of background flares and removal of the central pixels reduced the number of counts detected at the position of from 1607 to 234; for , the numbers are 1072 and 170, respectively.
In , the Bayesian blocks analysis detects no significant variability, and a one-temperature, solar-abundance APEC model yields a satisfactory fit, achieving $\chi_r^2=1.2$ with 16 DOF. In this case as well there are strong correlations among the residuals. The best-fit temperature in the adopted model is $kT=0.30\pm0.02$ keV, subject to the same caveats mentioned above. We find $\fx = -13.05^{+0.04}_{-0.05}$ (0.2–2 keV). At our adopted distance, $\lx = 26.4$ and $\lxlb = -4.49$.
In , the Bayesian blocks analysis finds that the count rate in the final 9 ks of the observation is elevated by a factor of 3. Although flaring is often associated with spectral variability , there were few enough counts available that we chose to only model the mean spectrum of the source. We found that a one-temperature, solar-abundance model yielded a less satisfactory fit, with $\chi_r^2 = 2.0$ for 11 DOF. A two-temperature, solar-abundance model yields $\chi_r^2 = 0.8$ (9 DOF) with $kT_1=0.22\pm0.5$ keV and $kT_2=0.87\pm0.14$ keV. We find $\fx=-13.22^{+0.07}_{-0.08}$ (0.2–2 keV), $\lx = 26.2$, and $\lxlb=-4.48$. For comparison, a one-temperature fit with variable abundance finds $\chi^2_r
= 1.2$ (10 DOF), $kT=0.59^{+0.10}_{-0.15}$ keV, $Z/Z_\odot=0.08^{+0.05}_{-0.03}$, and $\fx=-13.35\pm0.35$. Despite their differences, the two models yield values of that agree within their uncertainties.
### Did ROSAT observe during a flare?
Previous X-ray observations of the system were performed as part of the *ROSAT* All-Sky Survey (RASS), with follow-up on the *ROSAT* High Resolution Imager (HRI). While the latter was capable of resolving the triple system into its two main components (resolution 2$''$; separation 7$''$), the former, with a resolution of 5–10$'$ [@therassbsc], was not. While @sl04 correctly identify the ROSAT source as the blend of all three components, we note that @lhm08 failed to highlight this in their summary of the X-ray properties of .
The RASS catalog luminosity for is $\lx=27.47$ in the 0.1–2.4 keV band. After correction to the ROSAT bandpass, this exceeds the sum of our resolved measurements by a factor of 6. This suggests that the RASS observations of may have occurred during a strong flare. The RASS exposure time during the observation was 730 s [@sl04], shorter than the typical timescale of such events. The later *ROSAT* HRI measurements find $\lx = 27.19$ with an integration time of 2849 seconds [@sl04], in better agreement with the results, though still exceeding them by a factor of 2. As X-ray flares may easily result in luminosity increases of an order of magnitude (cf. Figure \[f.lxspt\]), it would not take an unusually large flare to reconcile the two measurements. On the other hand, the fact that both *ROSAT* measurements exceed the combined flaring luminosity, despite their separation in time (8 months), suggests that perhaps the X-ray activity of this system has decreased since the time of the RASS, possibly due to a long-term magnetic activity cycle .
Other Detected Sources
----------------------
We determined X-ray fluxes for the other detected sources by assuming an energy conversion factor (ECF) of $(4.5\pm1)\times10^{-12}$ erg cm$^{-2}$ count$^{-1}$, where the uncertainty approximately accounts for the range of plasma temperatures and abundances commonly encountered. The applicable theoretical ECF reported by version 4.6a for a 0.5 keV APEC plasma with $Z/Z_\odot = 0.6$ is $4.4\times10^{-12}$ erg cm$^{-2}$ count$^{-1}$. Our adopted value agrees with those derived for , , and (ECF $= (4.3, 4.6,
4.3)\times10^{-12}$, respectively) as well as our previous observations of X-ray-emitting UCDs [@bgg+08; @bbg+08]. Spectral modeling of the other sources with larger numbers of events (, , ) yields results consistent with those reported here.
Undetected Source:
-------------------
Only one photon in the 0.3–7 keV range was detected at the predicted location of throughout both observations, compared to the expected background level of 0.7 counts in 28 ks. The resulting 95% confidence upper limit is 4.3 counts [@kbn91]. Using the above energy conversion factor, the time-averaged flux limit is $\fx < -15.2$. At our adopted distance of 5.0 pc, we can thus constrain the persistent emission of to be $\lx \lesssim 24.3$, or $\lxlb\lesssim-4.70$.
Radio Analysis {#s.radio}
==============
We calibrated the VLA using standard procedures in the CASA software system [@thecasa]. Radio-frequency interference was flagged automatically using the tool, which provides post-correlation [@odbb+10] and morphological [@ovdgr12] algorithms for identifying interference. At the time of analysis, did not include a set of tuning parameters suitable for the processing of VLA data, so these were developed manually.
We created deep Stokes I images of each field with 2048$\times$2048 pixels, each $1$$\times$$1$ arcsec$^2$, except the image of , for which a pixel scale of $1.5$$\times$$1.5$ arcsec$^2$ was used to include the nearby, bright blazar in the image. The imaging process used multi-frequency synthesis [@themfs] and CASA’s multi-frequency CLEAN algorithm with 1500 iterations. Two spectral Taylor series terms were used for each CLEAN component; this approach models both the flux and spectral index of each source. The reference frequency for each image is 6.05 GHz. Properties of the images are listed in Table \[t.rfluxes\]. Astrometric predictions of the source locations were computed as described above, and cutouts of the VLA images around the predicted source locations are show in Figure \[f.cutouts\]. The accuracy of VLA astrometry in our observing configuration is $1''$, comparable to that of our predictions. While our targets have relatively high proper motions, the time baseline between the pairs of VLA and observations is sufficiently small that the differences in the predicted positions are negligible.
We detect a radio source in the image of the field at position RA = 11:21:48.78, Dec = $-$13:13:09.4, coincident with our astrometric prediction of RA = 11:21:48.77, Dec = $-$13:13:09.5. We re-imaged this field, rephasing the data to place this position on a pixel center to obtain the most accurate source parameters from image-domain modeling. Fitting the rephased image with a point-source model yields a flux density of $63\pm7$ and a positional uncertainty of $0.4''$. @fwkk91 used the VLA to find the areal density of sources brighter than 50 at 5 GHz to be $6.4 \times
10^{-5}$ arcsec$^{-2}$, making the probability of a chance positional coincidence $2 \times 10^{-3}$ (taking our search area to be the synthesized beam size). We therefore identify this source with . The multi-frequency cleaning algorithm determines a spectral index of $\alpha =
-0.4 \pm 0.2$ ($S_\nu \propto \nu^\alpha$). We imaged the source in the Stokes V parameter and made no detection, with an image rms of 5.4 . Taking the Stokes V upper limit to be three times this value, we find that $|V|/I \lesssim 25\%$. We searched for flares and other forms of variability using a visibility-domain analysis of the Stokes I data as described in @wbz13. No significant indications of variability were seen.
There is a 50 (7$\sigma$) radio source in the image of the field at RA = 18:43:20.72, Dec = $+$40:40:33.01, which is 15$''$ distant from the astrometric prediction. We checked the image astrometry against two NVSS sources ( and ), finding agreement down to the $1.5''$ uncertainty in the survey’s astrometry. We conclude that this source is not . There are no sources within 1$'$ of our astrometric predictions in all of the other fields. In each of these cases, we place an upper limit on the target flux density of three times the image rms. The field of was visited twice; no source is detected in the individual visits or in a deep image formed by combining the two datasets. The results of our radio observations, including flux densities, are summarized in Table \[t.rfluxes\]. Derived parameters, including radio spectral luminosities, are presented in Table \[t.dbdata\].
Trends in Radio and X-Ray Emission {#s.trends}
==================================
We have combined our new measurements with data from the literature to compile a comprehensive database of UCDs with both radio and X-ray observations. In Table \[t.dbsrcs\], we list these objects and provide some of their properties. Different authors report X-ray luminosities that are integrated over varying energy regimes; for consistency we normalize all X-ray fluxes and luminosities to a common band of 0.2–2.0 keV. We used to compute the appropriate conversion factors, evaluating flux ratios in a range of different plasma temperatures in the APEC model. The resulting factors are listed in Table \[t.xbconvs\]. The conversion factors represent an approximate median for several temperatures in the range $kT=0.4$–$1.0$ keV and are stable to within $5\%$ for temperatures within this range. In Table \[t.dbdata\] we report all paired UCD radio and X-ray luminosities available, giving detailed references and using simultaneous measurements when available.
Although we focused on X-ray and radio fluxes when constructing our database, it contains many ancillary measurements such as distances, spectral types, photometry, and effective temperatures. It is compiled from simple textual tables that are maintained in the Git distributed version control system, taking inspiration from the architecture of the Open Exoplanet Catalogue [@r12barxiv]. Its design is intended to enable continuous refinement in a decentralized, collaborative manner. Further details will be presented in a future publication.
X-Ray Luminosity vs. Spectral Type {#s.lxspt}
----------------------------------
Low-mass stars of spectral types earlier than M6 obey an X-ray activity/rotation relationship with “saturation” at $\lxlb \sim -3$ [@pmm+03; @jjb+11; @wdmh11]. Because main-sequence late-type stars are generally rapid rotators [@ib08], most of them have X-ray emission around this saturation level, as demonstrated in Figure \[f.lxspt\]. The $<$M6 objects in this figure having $\lxlb \gtrsim -2$ are likely flares. All come from the work of @rgh06, who obtained 1080 M dwarf X-ray fluxes from the ROSAT All-Sky Survey, a fraction of which will inevitably have been measured during a flare. All of the $<$M6 objects with $\lxlb \lesssim -4.5$ and measured are slow rotators, and thus their low levels of X-ray emission are expected. There is a noticeable underdensity of sources with $\lxlb \sim -4$ or $\lx \sim 28$; it can also be seen clearly in the data of @pmm+03. This underdensity is reminiscent of the “Vaughan-Preston gap” [@vp80], a similar feature seen in the distribution of various chromospheric activity indicators observed in F and G stars, for which a variety of explanations have been offered, including changes in dynamo modes, evolutionary stages of rapid angular momentum loss, or two distinct waves of star formation in the solar neighborhood . Somewhat surprisingly, we are unable to locate in the literature any investigation of this feature; such an undertaking is beyond the scope of this work.
Figure \[f.lxspt\] shows that our observations lend further support to the conclusion that the standard X-ray “saturation” effect breaks down at spectral types $\gtrsim$M6 [@fgg03; @bbf+10], an effect that is also seen in [@gmr+00; @mb03]. Although the cause of this breakdown has been the subject of much study, the number of detected objects is small while there are many possibly-relevant physical effects: decreasing , increasing rotation, a disappearing radiative core, and an increasingly dipolar magnetic field [@mdp+08]. The role of rotation is of particular interest because of its known effect on X-ray activity and its strong correlation with spectral type; this issue is explored in .
The data show a similar breakdown in terms of , with $\lx \approx 29.3$ for spectral types $\lesssim$M4 but decreasing by 1 dex for each later spectral subtype. This empirical relationship indicates a breakdown of the relationship around where full convection sets in, at spectral types of $\gtrsim$M3.5–M4 [@cb00]. Furthermore, there are indications that it may be more appropriate to consider rather than in this regime: results from Zeeman-Doppler imaging studies suggest that mid-to-late M dwarfs harbor relatively weak, disordered magnetic fields similar to those of slowly-rotating solar-type stars [@mdp+10], and there is substantially less scatter in the slow-rotator region of the X-ray activity/rotation relationship when rather then is considered [@pmm+03].
The Güdel-Benz Relation
-----------------------
Our paired radio and X-ray observations allow us to consider ultracool dwarfs in the context of the GBR. This is particularly salient in the UCD regime because the first detection of radio emission from a brown dwarf implied a severe divergence from the GBR [@bbb+01], and subsequent observations have confirmed that this divergence is not uncommon . In Figure \[f.lxlr\], we plot our full database of UCDs with both radio and X-ray measurements as well as the original data of @bg94. Figure \[f.lxlrnof\] is similar but omits known flares. In these and several subsequent plots, we show the best linear fit to the @bg94 data from @bbf+10, $$\slr = 1.36 (\lx - 18.97).
\label{e.gbfit}$$ The scatter of the @bg94 data around this fit is 0.6 dex when is treated as an independent variable (i.e., the distance from the best-fit line is measured at fixed ). The scatter relative to the best-fit line (i.e., measured perpendicular to it) is 0.2 dex. In the @bg94 dMe data, $\slrlx \sim -15.5$ typically, a value we adopt as a reference when quantifying radio over-luminosity relative to the GBR.
As may be seen in Figure \[f.lxlr\], almost all of our new results could be consistent with the established GBR. Since the radio upper limits are 2 dex above the best-fit line. The lone new radio detection, , has $\slrlx = -13.9$ and lies 1.3 dex away from the best-fit line. It is thus an outlier but not nearly as extreme as, for example, , which has $\slrlx = -11.5$ and is 3.2 dex away from the fit in quiescence. We find no significant signs of variability in the emission of in either band, which suggests that the observed fluxes correspond to a quiescent rather than flaring state. , for which we have combined a new X-ray analysis with a radio measurement from the literature, has $\slrlx = -12.0$ and lies 2.7 dex away from the best-fit line in quiescence, making it a strong violator of the GBR.
UCD emission in both the radio and X-ray bands can show significant flaring during typical observational timescales , and the radio emission is additionally known to evolve over year timescales in some cases . The simultaneity of measurements is thus important to consider in the context of the GBR. In Table \[t.dbdata\], we annotate which radio/X-ray measurements arise from simultaneous observations, and we isolate these measurements in Figure \[f.lxlrsim\]. Simultaneous observations constitute about one third of the existing dataset, with many of the measurements coming from an observational campaign we have conducted over the past several years [@brr+05; @bbg+08; @bgg+08; @brpb+09; @bbf+10 Williams et al., in preparation]. These simultaneous measurements include several of the most extreme GBR violators, showing that strong divergence from the GBR is a genuine phenomenon and *not* merely due to flaring. There is only one simultaneously-observed UCD to be detected in the X-ray but not the radio: the L2+L3.5 ($\pm$1 subtype) binary [@aob+07]. Its detection by was marginal (4 counts), precluding a detailed analysis.
In Figure \[f.rxspt\] we plot as a function of spectral type. *There is clear evidence for new behavior at spectral types $\ge$M7*: while there are no measurements of $\slrlx > -13$ in objects earlier than M7, there are seven in later objects. The lower panel plots distance from the GBR linear fit (Equation \[e.gbfit\]) rather than the simple ratio . As can be seen, the choice of ordinate does not significantly affect the structure of the data, and we discuss trends in terms of because that quantity is closer to the observables.
In both panels of Figure \[f.rxspt\], divergence from the GBR seems to increase unconditionally with spectral type. This seeming trend is partially misleading because of radio sensitivity limitations. While there is new behavior at spectral types $\ge$M7, a population of $\ge$M7 objects consistent with the early- and mid-M dwarfs ($\slrlx\sim-15.5$) is not excluded. The new observations presented in this work, which are of nearby objects and were obtained with a highly sensitive radio telescope (the upgraded VLA), reach limits of $\lr\sim12.5$. As shown in Figure \[f.lxlr\], our data include the most sensitive upper limit on M dwarf radio emission available, $\slr < 11.8$ for . Nonetheless, in our sample $\lx \sim 26$, so that we are insensitive to $\slrlx \lesssim -13.5$. A source obeying the GBR with $\lx =
26$ and $d = 10$ pc would have a radio flux density of $S_\nu \sim 30$ nJy, accessible only to the proposed Square Kilometer Array [SKA; @theska]. Until the arrival of an SKA-class telescope, the only way to probe UCDs in this regime will be through observations of objects at distances of $\lesssim$3 pc, such as the recently-discovered L/T binary [@l13; @bsl13arxiv].
Another issue affecting Figure \[f.rxspt\] is the correlation between mass and rotational velocity in main-sequence late-type stars , which obscures the true physical process underlying the observed trend. We plot as a function of in Figure \[f.rxrot\]. Similar to Figure \[f.rxspt\], appears to increase with but the trend requires care in interpretation. Along with the limitations in sensitivity to small values of , some objects with low values of may be rapid rotators seen at low inclinations. In this context, it is striking that out of 13 objects with $\vsi > 20$ , seven have $\slrlx > -13.5$; that is, the fraction of radio-over-luminous objects is large. The lack of radio-over-luminous sources with small values of is consistent with the argument of @had+08 that viewing angle is an important factor affecting the observed radio emission: if increases with $v$ but is independent of $\sin i$, there should exist sources seen nearly pole-on with low values of and large radio excess. Such sources have not yet been observed. The presence of both lower and upper limits at high values of prevents a simple characterization of the trend in . A more full analysis should not only account for inclination effects, but also for source-to-source variation. Studies of the relationship between X-ray emission and rotation have shown that a useful parameter for doing so is the Rossby number, $\text{Ro} = P_\text{rot} / \tau_\text{conv}$, where $P_\text{rot}$ is the rotation period and $\tau_\text{conv}$ is the characteristic convective overturn time [@nhb+84; @pmm+03]. We pursue these matters in .
Finally, although we have discussed departure from the GBR in terms of “radio over-luminosity,” it could clearly be expressed in terms of “X-ray under-luminosity” as well, a framing well-motivated by the dropoff in seen in UCDs. This prompts us to consider a modification of the GBR in terms of the benchmark “saturation” level of X-ray emission, defining $\Lxsat = 10^{-3} \Lb$. In Figure \[f.lxlradj\] we plot the same radio data as in Figure \[f.lxlr\] but replace with . We omit flaring measurements and do not alter the data from @bg94. As should be expected from Figure \[f.lxspt\], the UCD measurements move closer to the canonical GBR. The data for objects $<$M6.5 straddle the canonical relation nicely, with a substantial subset of measurements that are radio *under*luminous. The UCDs, however, remain displaced to the left of the GBR; in other words, even if they emitted X-rays at the canonical “saturation” level, they would still be over-luminous in the radio. Any explanation of GBR divergence in UCDs must therefore not merely account for the suppression of X-ray emission relative to the stellar “saturation” trends, but also account for a comparative increase in radio emission. For instance, if the suppression of UCD X-ray emission is entirely due to less efficient heating of the corona leading to a temperature appreciably lower than the typical coronal value of $\approx$1 keV , a mechanism must still be proposed to explain the unexpectedly bright radio emission. We argue below that in the radio-over-luminous sources, no longer scales with .
Summary of Observed Trends
--------------------------
The X-ray emission of UCDs drops off rapidly with spectral type, with slightly different properties depending on whether or is considered (Figure \[f.lxspt\]). Some UCDs diverge strongly from the GBR, with values of exceeding the typical value of $10^{-15.5}$ by 4 orders of magnitude; others may be consistent with it, with the limited sensitivity of radio observations allowing us only to conclude that $\slrlx \lesssim -13.5$ (Figure \[f.lxlr\]). Although variability is an important consideration, extreme GBR divergence is seen in simultaneous radio and X-ray observations, confirming its reality (Figure \[f.lxlrsim\]). More extreme divergence from the GBR seems to become possible at later spectral types (Figure \[f.rxspt\]). Even if UCD X-ray activity () did not drop off rapidly with spectral type, but rather remained at the standard “saturation” level of $\lxlb = -3$, UCDs would still be radio-over-luminous compared to the GBR (Figure \[f.lxlradj\]).
The interpretation of these trends is complicated by the correlation between mass and rotational velocity in main-sequence late-type stars . Only objects with $\vsi \gtrsim 20$ are seen to diverge strongly from the GBR, and $\approx$50% of such objects do so (Figure \[f.rxrot\]). Although we discuss the role of rotation below, we defer detailed investigation of the relationship between rotation and magnetic activity in UCDs to .
Discussion {#s.disc}
==========
Our results underscore the wide range of magnetic phenomenology seen in the UCD regime. For instance, our new observations of the M7.5 dwarf yield a radio over-luminosity of $\lesssim$2.5 orders of magnitude. Meanwhile, simultaneous observations of the nearby M8.5 dwarf have revealed a radio over-luminosity of $\gtrsim$5 orders of magnitude [@bbg+08; @had+08]; this assessment is relative to its mean radio emission (0.5 mJy), not the bright, highly polarized pulses that it has also been seen to emit [2 mJy; @had+08]. Both objects are relatively rapid rotators, with $\vsi = 33$ and $50$ , respectively. (We emphasize that while inclination effects may cause a rapid rotator to appear as a slow rotator, they cannot cause a slow rotator to appear as a rapid rotator.) Both are nearby [9.7 and 5.7 pc; @cpbd+05; @rcl+03], and neither is known to have a companion . Despite these similarities, our new data show that while is least an order of magnitude brighter than in the X-ray, it is also at least an order of magnitude fainter in the radio.
@sab+12 considered the GBR in the UCD regime and proposed the existence of two populations: one comprising rapidly rotating, radio-bright, X-ray-dim UCDs; the other comprising slower-rotating, radio-dim, X-ray-bright objects. They additionally noted that only the radio-bright objects display bright, highly polarized radio pulses, while only the radio-dim objects produce X-ray flares, with the exception of which seems to flare in both bands. Our data show that rotational velocity is not strictly tied to this dichotomy. and are both rapid rotators (33 and 40 , respectively) that are nonetheless radio dim and X-ray bright. The late-M binary is also a rapid rotator [45 ; @mbi+11] that has relatively low divergence from the GBR, being radio over-luminous by 2.5 orders of magnitude in quiescence (Williams et al., in preparation).
We propose that the range of observed UCD behavior may be due to varying topology of their magnetic fields. In their study of the relationship between rotation and radio activity in UCDs, @mbr12 proposed such a connection; here, we extend it to include X-ray emission as well. The connection to magnetic topology is motivated by a series of studies of M dwarfs using Zeeman-Doppler imaging (ZDI) techniques [@dmp+08; @mdp+08; @mdp+10]. In ZDI, the magnetic field topology is reconstructed from measurements of the Zeeman effect in time-resolved optical spectropolarimetry [@thezdi][^2]. @mdp+10 find that some M dwarfs have strong, axisymmetric fields, while others have fields that are weak and non-axisymmetric. While early-M dwarfs inhabit the weak-field regime and mid-M dwarfs inhabit the strong-field regime, late-M dwarfs inhabit *both* regimes. This observational finding led @mdp+10 to propose two magnetic dynamo modes leading to differing magnetic topologies, with late-M dwarfs having a bistable dynamo that may inhabit either mode. This concept is supported by recent results from geodynamo simulations that show bimodal dynamo outcomes in rapid rotators [@mdsd11; @gmd+13].
We suggest that UCDs with strong, axisymmetric fields tend to have values of in line with the GBR, while the ones with weak, non-axisymmetric fields are radio over-luminous. Slow rotators ($\lesssim$20 ) have strong-field dynamos and thus stay close the GBR, as found by @sab+12. Rapid rotators may have dynamos in either mode, so some are strong GBR violators while others are not, explaining the results of @mbr12 and this work. The association of sources that stay near the GBR with strong fields provides continuity with the ZDI and X-ray observational results for mid-M dwarfs. The similar patterns seen in the ZDI topology and results are the primary motivation for our hypothesis.
We have described the difference between the proposed strong-field and weak-field sources in terms of , and argued that the weak-field sources are radio over-luminous and X-ray under-luminous. However, the weak-field sources also tend to have later spectral types, and as shown in Figure \[f.lxspt\] these sources have lower values of . Some component of the X-ray under-luminosity of the weak-field sources may be due to their cooler temperates rather than the proposed effects of field topology. To clarify this issue, we plot in Figure \[f.rbxb\] a modified Güdel-Benz relation in which we restrict the data to fully-convective dwarfs and normalize by , thereby rearranging the data points along lines of constant . It is noteworthy that even after accounting for changes in , the radio-bright sources are still X-ray underluminous compared to the radio-dim ones. Hypothetically, if the mean of the radio-bright objects were *much* lower than that of the radio-dim objects, the two groups might have indistinguishable values of , and the scatter in Figure \[f.rbxb\] might be narrow in the direction and broad in the direction. This is not the case. Instead, even after taking the changing into account, the radio-bright sources seem to require both smaller and larger . We argue below that for the radio-bright sources is in fact a more appropriate quantity to consider than , whereas for the radio-faint ones the opposite is true.
The obvious issue to consider in the bimodal dynamo scenario is how the two field topologies lead to such different values of . We start first with the weak-field, radio-bright mode. It is plausible that this mode extracts energy from stellar convective motions less efficiently than the strong-field mode, leading to a decrease in overall magnetic activity that manifests itself in decreased compared to the strong-field objects. Meanwhile, Figure \[f.lxlr\] shows that these objects tend to have similar values of $\slr \approx 13.5$. We conjecture that in this mode, tangled magnetic field structures lead to persistent small-scale reconnection events that are sufficient to maintain a population of gyrosynchrotron-emitting electrons filling a coronal region a few stellar radii ($R_*$) in size. This model has already been suggested for the radio-bright sources . Because the radio luminosity is energetically unimportant compared to or , the radio-emitting population could be maintained even at very low levels of magnetic activity. The observed spectral luminosity $\sLr\propto\nu^2 T_b R^2$, where $T_b$ is the brightness temperature and $R$ is the characteristic size of the emitting region. In the UCD regime, $R_*
\apx R_J$ with very weak dependence on mass. Supposing $R \apx R_*$, we find that in the radio-bright population $T_b \apx 10^{8.5{-}9.5}$ K. These results are consistent with values found in the dM1e binary YY Gem, for which $T_b =
1.1 \times 10^9$ K and $R \apx 2 R_*$ were derived from VLBI observations [@abg97]. Taking $T_b$ to be insensitive to mass, $\slr \approx 13.5$ should be a characteristic value for radio-bright UCDs observed at 8.5 GHz. @had+08 pointed out the relatively stable values of across the M spectral type and argued for a different emission mechanism (see below). Our interpretation differs in that we argue that emission from early-M dwarfs is instead due to magnetic reconnection in the standard chromospheric heating picture, leading to consistency with the GBR. In this scenario, the similar values of in the early-M dwarfs and the radio-bright UCDs are coincidental.
In the strong-field mode, we hypothesize that the coupling to internal stellar convective motions is stronger, resulting in comparatively higher levels of magnetic activity and . These are nonetheless lower than what is found in rapidly-rotating earlier-type stars, an effect often attributed to the outer layers of UCD atmospheres becoming appreciably neutral, reducing their coupling to the coronal magnetic field and thus their ability to inject energy into it through surface convective motions [@mbs+02]. The standard chromospheric evaporation model still applies, leading to values of compatible with the GBR. However, these objects are apparently unable to sustain a corona-filling population of gyrosynchrotron-emitting particles, perhaps because reconnection events are simply too rare. The rarer reconnection events would be more energetic than those seen in the weak-field sources, perhaps explaining the observation of @sab+12 that the radio-quiet sources are seen to flare in the X-ray while the radio-loud ones are not. This may also explain the conjecture of @rps10 that some UCDs exhibit large X-ray flares and low-level X-ray variability but, unlike active early-M dwarfs, not a continuous spectrum of flare energies.
This model posits that UCDs with stronger magnetic fields should be associated with higher values of . Figure \[f.bflx\] shows that this is qualitatively the case when the field strength is diagnosed with FeH spectroscopy [@rb06], a technique that has the advantage of having been applied to numerous cool dwarfs [@rb07; @rbb09; @rb10]. This technique measures $Bf$, the average unsigned magnetic field at the photosphere with $f\le1$ being a filling factor term. ZDI, being sensitive to the net signed field in each resolution element, measures strictly smaller values. Despite this distinction, the two measures are generally correlated, with ZDI values being 5–15% of FeH values depending on mass [@rb09b; @mdp+10]. A correlation similar to that in Figure \[f.bflx\] exists between $Bf$ and $\LhLb$ [@rb10]. Figure \[f.rbxb\] shows that the very radio-bright UCDs have $\lxlb \lesssim -4.5$; the five UCDs in Figure \[f.bflx\] meeting this criterion (, , , , ) have relatively low values of $Bf \lesssim 2.5$ kG. Three UCDs have $\lxlb \approx -4$ and $Bf < 2.5$ kG (, , ), consistent with the idea that both field topology and field strength are important parameters. All of the objects with $Bf > 2.5$ kG have $\lxlb > -4$.
A striking characteristic of the radio-bright UCDs is that several of them emit periodic, bright, highly polarized radio pulses [@had+06; @had+08; @bgg+08; @brpb+09]. These are characteristic of the electron cyclotron maser instability [ECMI; @theecm; @t06], a coherent process that results in radio emission at the cyclotron frequency $\omega_c = e B / m c$. This process is responsible for auroral radio bursts in the magnetic solar system planets, which have many similarities to those observed in the UCDs [@z98]. In particular, the periodic nature of the pulses is strongly suggestive of beamed emission tied to the stellar rotation. We suggest that, despite their generally tangled fields, the weak-field sources are able to maintain the large-scale converging magnetic structures needed to produce these auroral bursts. These structures need not dominate the ZDI-diagnosed field since their footpoints may cover only a small fraction of the photosphere; hence their possible existence is not inconsistent with the overall finding of a weak, non-axisymmetric field. (In @mdp+10, the ratio of the maximum and mean ZDI-diagnosed fields, $B_\text{max} / \langle B\rangle$, can exceed 5, and $B_\text{max}$ may be underestimated if the filling factor of the relevant field structure is low.) We have argued that the weak-field sources maintain a population of energetic electrons filling the coronal volume, providing a continual supply of particles capable of driving the bursts as well. The strong-field sources are likely also able to maintain the requisite converging fields but, lacking the corona-filling supply of particles, do not display the periodic bursts. As pointed out by @sab+12, however, the strong-field sources are slower rotators, and so the lack of observed bursts may be a selection effect because UCD radio observations are generally shorter in duration than the rotational period. @had+08 identify another possible selection effect, providing evidence that the sources with ECMI bursts are observed at high inclinations.
@had+06 [@had+08] further argue that the GBR violators are radio bright because the vast majority of their radio emission is due to the ECMI, including the non-burst emission that has generally been attributed to gyrosynchrotron processes . They suggest that depolarization and steady particle acceleration could cause ECMI emission to have the low variability and polarization that are associated with the non-burst emission of these sources. While this argument may apply to some UCDs, we disfavor its application to *all* of them, because it calls for an implausible physical configuration in sources such as . This L dwarf binary periodically emits rapid (100 s) pulses of 10 mJy and 100% circular polarization atop seemingly quiescent emission at 0.3 mJy with $\lesssim$15% circular polarization and a spectral index of $\alpha \apx
-0.7$ ($S_\nu \propto \nu^\alpha$) [@brpb+09]. Any variation besides the pulses is at the $\lesssim$1 mJy level. If the quiescent emission also originates in the ECMI, two emitting regions of identical field strengths would be required, both with the physical conditions suitable for ECMI cascade, but producing vastly different observed emission. It seems simpler to posit that both ECMI and gyrosynchrotron emission occur in this case. In other sources such as , both polarized and unpolarized emission of comparable flux densities and variability characteristics are observed, and the arguments of @had+06 [@had+08] are more persuasive.
In our scenario it is still not entirely clear what factors determine the overall level of magnetic activity, i.e. . There are substantial correlations among , spectral type (Figure \[f.lxspt\]), GBR deviation (Figure \[f.rbxb\]), and rotational velocity (). In the quest to understand these interconnections, two particular issues call for investigation. The first is to what extent the observed UCD “supersaturation”-like effect, an anticorrelation between and [@bbf+10], is a causal relationship, relating perhaps to evolution in the nature of the dynamo. The second is the question of which processes drive the observed dropoff in as a function of spectral type (Figure \[f.lxspt\]), relating perhaps to rotation, changes in internal structure, or increased photospheric neutrality [@mbs+02]. Because of the correlations among the variables in question, the role of rotation must be considered carefully in attempts to resolve these issues, and so we defer further analysis to .
Finally, we wish to emphasize that although we have discussed a model of two distinct dynamo modes, these modes are not necessarily mutually exclusive within a given source – it is plausible that one object could host multiple field-generating processes. With a sample of 29 sources and many non-detections, it would be premature to claim that there are two distinct subpopulations.
Summary and Conclusions {#s.conc}
=======================
We presented new X-ray () and radio (VLA) observations of seven ultracool dwarfs with spectral types between M6.5 and M9.5 and a wide range of . We have detected all of them in the X-ray band, nearly doubling the number of UCDs with X-ray detections. Despite the increased sensitivity of the upgraded VLA, only one of the sources was detected in the radio. Our results are thus broadly consistent with the Güdel-Benz relationship between radio and X-ray emission in stellar flare phenomena, though they still admit radio over-luminosities of a factor of $10^2$. This is in contrast to several spectacular recent results, in which some UCDs are seen to be radio-over-luminous by 5 orders of magnitude. Although UCDs are highly variable in both bands, making simultaneous observations important tools in the analysis of correlations such as the GBR, we have argued that the nonsimultaneity of our observations does not significantly affect their interpretation.
We have also assembled a comprehensive sample of UCDs with both radio and X-ray observations (Figure \[f.lxlr\]; Table \[t.dbdata\]), including all known measurements in which the observations were simultaneous. With the addition of our new measurements, there is strong evidence that UCDs display a wide range of behavior with regards to the GBR: some are strongly radio over-luminous, while others could be consistent with it. This range can be, and has been, interpreted as a bimodality in the UCD population [@mbr12; @sab+12], which has support from both ZDI observations and geodynamo simulations [@mdp+10; @gmd+13]. We have argued that one group of sources can maintain a population of gyrosynchrotron-emitting particles in the corona, setting an effective floor on , while the other group is less active than earlier-type stars but emits following the standard chromospheric evaporation model. Interpretation of the data, however, is made difficult by the many variables at play: $M_*$, , , age, metallicity, binarity, and long- and short-term variability. It is not clear that the population can be neatly divided into two distinct subgroups at this time.
Although the study of UCD magnetism continues to present many puzzles, we see several reasons to be optimistic for the future. Studies of progressively larger samples will help clarify trends and allow more robust examination of subsamples that control for variables such as mass, age, or rotation. The upgrade of the VLA presents a major opportunity in this regard, because it is becoming clear that radio observations present the best opportunity for exploring the magnetism of the coolest objects, whose faintness and rapid rotation make extremely difficult the application of techniques such as ZDI. Furthermore, new benchmark objects are being discovered that offer the chance for detailed study. These include the radio-brightest UCD, [@mbi+11 Williams et al., in preparation], and the coolest UCD yet detected in the radio, [T6.5; @rw12; @wbz13]. Wide-field infrared surveys and proper motion searches are discovering substantial numbers of UCDs, including extremely nearby examples such as the 2-pc L/T binary [@l13; @bsl13arxiv]. Finally, theoretical studies of the operation of dynamo action in fully convective bodies are progressing, with numerical geodynamo models being adapted to be able to simulate density contrasts closer to those seen in UCDs . Advances in all of these areas will greatly increase our understanding of the magnetism of brown dwarfs and, eventually, extrasolar planets.
We thank Scott Wolk and Katja Poppenhaeger for enlightening discussions. We also thank the anonymous referee for helpful, constructive comments. This work is supported in part by the National Science Foundation REU and Department of Defense ASSURE programs under NSF Grant no. 1262851 and by the Smithsonian Institution. E. B. and P. K. G. W. acknowledge support for this work from the National Science Foundation through Grant AST-1008361, and from the National Aeronautics and Space Administration through Chandra Award Number GO2-13007A issued by the Chandra X-ray Observatory Center, which is operated by the Smithsonian Astrophysical Observatory for and on behalf of the National Aeronautics Space Administration under contract NAS8-03060. The VLA is operated by the National Radio Astronomy Observatory, a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France, and NASA’s Astrophysics Data System.
Facilities: ,
Here we give our method for computing bolometric luminosities and discuss the properties of the sources for which we present new results.
Bolometric luminosities
-----------------------
We compute bolometric luminosities using bolometric corrections to observed absolute magnitudes. In particular, $$\lb = \frac{2}{5} (M_{\odot,\text{bol}} - M - \text{BC}) + [L_\odot],$$ where $M$ is the absolute magnitude of the star in some band, BC is the bolometric correction in that band, and $M_{\odot,\text{bol}} = 4.7554$ is the bolometric absolute magnitude of the Sun. For M dwarfs, we use the average of the bolometric luminosities calculated using $J$ and $K$ band magnitudes: $$\text{BC}_K = 2.43 + 0.0895\times (\text{SP})$$ and $$\text{BC}_J = 1.53 + 0.148\times(\text{SP}) - 0.0105\times(\text{SP})^2,$$ with SP = 0 (5) for spectral type M0 (M5) [@wgm99]. For L dwarfs, we use the $K$-band magnitude only: $$\text{BC}_K = \begin{cases}
2.37 + 0.075\times(\text{SP}), & 10 \leq \text{SP} < 14, \\
4.47 - 0.075\times(\text{SP}), & 14 \leq \text{SP} \leq 19,
\end{cases}$$ with SP used consistently as above (SP = 10 for L0) [@nty04].
Properties of Newly-Observed Targets
------------------------------------
### {#section-1}
(= GJ3622, LP731–58) is a nearby ($d \approx
4.5$ pc) M6.5 dwarf. It is a variable emitter with $\lhlb \approx
-4.4$ [@dlh86; @lbk10; @rb10] and has been suggested as a candidate UV Cet flare star [@dlh86]. $K$-band spectroscopy and modeling yield an estimated $\teff = 2772\pm25$ K and $[\text{Fe}/\text{H}] = -0.41\pm0.17$ [@racml12]. @rb10 used other spectroscopic measurements to determine $\vsi < 3\pm2$ and $Bf = 600\pm200$ G. ROSAT observations by @fgsb93 yielded an upper limit of $\lx < 26.42$ in the 0.1–2.4 keV band. @mbr12 observed in the radio and did not detect it, obtaining a limit of $S_\nu < 96$ ($\lrlb <
-8.20$) at 8.46 GHz. @mdp+10 used Zeeman-Doppler imaging to derive a low-resolution map of the stellar magnetic field structure. Although the S/N was low, their modeling indicates a dipolar field of 100 G with an inclination $i \approx 60\degr$. From variability in its radial velocity on the 10 level, @gw03 argue that is a binary system, but a companion has not yet been detected by other means.
### {#section-2}
(= GJ4281, LP760–3) is an M6.5 dwarf at a distance of 11 pc. @gl86 measured its EW to be 1.3 Å. @mb03 measure an EW of 4.4 Å ($\lhlb = -4.15$) and $\vsi =
7.0$ , consistent with a later determination of $\vsi < 12$ using a different method [@ljp+12]. A proper motion of $1200\pm130$ mas yr$^{-1}$ has been determined [@dhc05]. @jrj+09 determine $\teff = 2536$ K and $M = 0.093\pm0.005$ M$_\sun$ from photometric modeling. was undetected in a ROSAT survey of very-low-mass stars [@fgsb93], with $\lx
< 26.89$.
### {#section-3}
(= GJ3655, LP732–94) is an M8$+$L7.5 binary at a distance of 14 pc. The brown dwarf companion was detected by direct imaging with adaptive optics by @fcs03. Long-term monitoring has led to a detailed model of the binary orbit [@dli09; @kgb+10], with the total system mass being estimated as $0.144 \pm 0.013$ [@kgb+10]. The projected rotational velocities of the primary and secondary have been measured as $\vsi
= 15\pm1$ and $11\pm3$, respectively [@kgf+12]. @mb03 detected strong emission with EW = 29.4 Å, computing $\lhlb = -3.70$. Comparable results were obtained by @scb+07, who found even stronger emission with $\lhlb = -3.40$, making it one of the most -luminous UCDs known.
### {#section-4}
(= V492 Lyr, GJ4073, LP229–30) is an M8 dwarf at a distance of 14 pc. In an investigation of candidate cataclysmic variable systems, @lhlc99 confirmed its dMe nature and measured $\lhlb \approx -4.1$. Subsequent activity observations by @scb+07 find $\lhlb \approx -4.3$. High-resolution spectra obtained by @rb10 reveal a rotational velocity of $\vsi = 5.0\pm3.2$ , $\lhlb = -4.1$, and $Bf = 1200\pm800$ G. This is consistent with the results of @dmm+12, who find $\vsi < 12$ . Previous radio observations with the VLA yielded an upper limit of $S_\nu <
48$ at 8.46 GHz [@b06b]. was undetected in the ROSAT survey of @fgsb93 with an upper limit of $\lx < 26.83$. In a search for brown dwarf companions, no objects were found within limits of $\Delta J <
9.6$, $\Delta \theta > 10''$ [@mz04].
### {#section-5}
(= NLTT3868) is an M9 dwarf at a distance of 11 pc. In the survey of @scb+07, it was found that $\lhlb \approx -4.5$. @rb10 found a rotational velocity of $\vsi = 13\pm2$ as well as $\lhlb =
-4.50$ and $Bf = 1400\pm200$ G. @dmm+12 obtain consistent results, with $\vsi < 12$ . Radio observations with the VLA have yielded an upper limit of $S_\nu < 33$ at 8.46 GHz [@b06b].
### {#section-6}
(= DENIS-PJ115542.9$-$222458) is an M7.5 dwarf at a distance of 9.7 pc. Despite the fact that it has been suspected to be a very nearby object for nearly 40 years [@thelhs], it is poorly-studied. @rb10 found a rotational velocity of $\vsi = 33\pm3$ and $\lhlb = -4.58$. Radio observations with the VLA have yielded an upper limit of $S_\nu < 90$ at 8.46 GHz [@mbr12].
### {#section-7}
(= LSPMJ1521$+$5053) is an M7.5 dwarf at a distance of 16.1 pc. @scb+07 measured $\lhlb \approx -4.9$. @rb10 measured $\vsi =
40\pm4$ and $\lhlb = -4.88$. Radio observations with the VLA have yielded an upper limit of $S_\nu < 39$ at 8.46 GHz [@b06b]. @scc+05 found no evidence for any companions to between 0.1–15 arcsec with a separation-dependent contrast of $\Delta H \lesssim
12$ mag.
Properties of Targets with New Analysis
---------------------------------------
### {#section-8}
The system (= GJ1245ABC) is a well-studied, nearby ($4.6$ pc), cool triple [@mhf+88]. (= GJ1245AC, LHS3494AB, LSPMJ1953$+$4424W) is separated from (= GJ1245B, LHS3495, LSPMJ1953$+$4424E) by 7$''$, and is itself a tighter binary of 1$''$ separation. Both components of the overall system are flare stars [@rcv80] and it has long been a target of activity studies . The blended pair has a radio flux density $<$192 at 5 GHz [@bbfk09], $\lxlb \approx -3.78$ [@sfg95], $\lhlb \approx -4.3$ [@mb03], and $\vsi = 17.4 \pm
1.4$ [@dfpm98]. Resolved measurements of are uncommon, with @lhm08 recently providing the first spectral type estimate for B (M8.5); it is the only component of the system that fits our definition of being an ultracool dwarf. @mb03 were able to measure in A ($22.5 \pm 2$ ) and ($6.8 \pm 1.9$ ), but not B.
### {#section-9}
(= 2MUCD10158; hereafter ), an L8 dwarf at a distance of 5 pc, was identified as an extremely cool object by @mdb+99, with $\teff \approx 1500$ K [@rb08]. Its emission is extremely faint, $\lhlb < -8.28$ [@rb08]. Multiple measurements of its yield either 40 or 60 , depending on the analysis method used, likely due to its unusually late type [@mb03; @zomb+06; @rb08]. Brown dwarf companions with separations of 7–165$''$ are unlikely [@cmp+11]. Optical monitoring suggests short- and long-term aperiodic variations [@k13].
### {#section-10}
(= LSPMJ0027$+$2219, NLTT1470) is an M8+M9 binary at a distance of 10 pc. @gmr+00 found moderate activity in the blended system, with $\lhlb = -4.52$. @fbd+05 used adaptive-optics imaging to reveal the binarity of the system, finding a separation of 0.1$''$ with $\Delta K' = 0.26 \pm 0.05$ mag. A measurement of the trigonometric parallax, $\pi = 75.82 \pm 1.62$ mas, has only recently become available [@gc09]. @rb10 found $\lhlb = -4.53$, in very good agreement with @gmr+00, but were unable to assess the magnetic field strength. @pbol+07 discovered radio emission from the system with a flux density at 8.5 GHz of $365 \pm 16$ . Both of the components are rapid rotators, although their projected rotational velocities differ by 15%: $\vsi = 55 \pm 2$ and $83 \pm 3$ for the A and B components, respectively [@kgf+12]. The total mass of the system is $0.120^{+0.008}_{-0.007}$ and it is also likely to be young, at $140\pm30$ Myr [@dlb+10].
[^1]: The implementation is versioned using Git and is currently available at <https://github.com/pkgw/pwpy/blob/master/scilib/xbblocks.py>. The version used in this work is that included in commit . The design of Git ensures that this commit identifier uniquely specifies the exact content and complete revision history of the code in question.
[^2]: While the Zeeman effect provides a direct measurement of the field, it is important to note that the quantity being measured is the net *signed* field in each resolution element. FeH spectroscopy, discussed below, probes the magnitude but not the topology of the *unsigned* field.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
---
[**A lower estimate for the modified Steiner functional**]{}
[George K. Savvidy]{}
Physics Department,University of Crete,71409 Iraklion,Greece
Institut für Theoretische Physik,D-60325 Frankfurt,Germany
e-mail: [email protected]
and
[Rolf Schneider]{}
Mathimatisches Institut, Universitat Freiburg, Albertstr.23b, 7800 Freiburg, Germany
e-mail: [email protected]
[**Absrtact**]{}
We prove inequality (1) for the modified Steiner functional A(M), which extends the notion of the integral of mean curvature for convex surfaces. We also establish an expression for A(M) in terms of an integral over all hyperplanes intersecting the polyhedral surface M.
In the articles [@Amb], [@Sav1], [@Sav2] the authors suggest a new version of string theory, which can be considered as a natural extension of the Feynman–Kac integral over paths to an integral over surfaces. Both amplitudes coinside in the case, when the surface degenarates into a single partical world line.
The string has been conjectured to describe a wide variety of physical phenomena, including strong interaction, the three dimensional Ising model, and unified models incorporating gravity. The Feynman integral for the string is just the partition function for the randomly fluctuating surfaces, and in this statistical approach the surface is associated with a connected polyhedral surface embedded in euclidean space.
To prove the convergence of the partition function for this new string, the authors of [@Sav1], [@Sav2] require a lower estimate for the action $A(M)$ on which the theory is based. The purpose of the present note is to prove the inequality $$\label{1}
A(M) > 2\pi \Delta,$$ where $A(M)$ is the modified Steiner functional as introduced in [@Amb], [@Sav1], [@Sav2] and $\Delta$ is the diameter of the polyhedral surface $M$ in ${\sf R}^{d}$. We also establish an expression for A(M) in terms of an integral over all hyperplanes intersecting the polyhedral surface M.
[**2.Proof of the inequality**]{}
We recall the definition of $A(M)$.
[**Definition.**]{}Let $M$ be an embedded connected closed polyhedral surface in euclidean space ${\sf R}^{d}$ $(d \ge 3)$. Let ${\cal F}_{1}(M)$ be the set of edges of $M$. For $e \in {\cal F}_{1}(M)$ we denote by $L(e)$ the length of $e$ and by $\alpha(e)$, where $0 <
\alpha(e) < \pi$, the angle between the two faces of $M$ incident with $e$. Then the modified Steiner functional is defined by $$A(M) := \sum_{e \in {\cal F}_{1}(M)} L(e)[\pi - \alpha(e)].$$
[**Theorem.**]{} [*If $\Delta$ denotes the diameter of M, then*]{} $$A(M) > 2\pi \Delta.$$
[*Proof.*]{} We first consider a simple closed polygon $P$ in ${\sf R}^{d}$. For a vertex $v$ of $P$, we denote by $\alpha(v)$, where $0 < \alpha(v) < \pi$, the angle between the two edges of $P$ incident with $v$. The (absolute) total curvature of $P$ is defined by $$\kappa(P) := \sum_{v}[\pi - \alpha(v)],$$ where the sum extends over the vertices of $P$. It is known that $$\label{2}
\kappa(P) \ge 2\pi$$ (Fenchel’s inequality for polygons; see, e.g., [@Mil]).
In the proof of inequality (\[1\]) we shall use some integral geometry, in particular the space ${\cal E}^{d}_{d-1}$ of hyperplanes in ${\sf R}^{d}$ with its (suitably normalized) rigid motion invariant measure $\mu_{d-1}$; see, e.g., [@Sch]. According to [@Sch], (1.9), the measure $\mu_{d-1}$ can be represented as follows. For a nonnegative measurable function $f$ on ${\cal E}^{d}_{d-1}$ we have $$\int \limits_{{\cal E}^{d}_{d-1}}fd\mu_{d-1} =
\int \limits_{S^{d-1}}\int \limits_{-\infty}
^{\infty}f(H_{u,\tau}) d\tau d\sigma(u).$$ Here $S^{d-1} := \{u \in {\sf R}^{d}: \|u\| = 1\}$ is the unit sphere of ${\sf R}^{d}$, $$H_{u,\tau} := \{x \in {\sf R}^{d}: \langle x,u \rangle =
\tau\}, \quad u \in S^{d-1}, \enspace
\tau \in {\sf R},$$ is a general hyperplane with unit normal vector $u$, and $\sigma$ is the spherical Lebesgue measure on $S^{d-1}$, normalized to total measure 1. By $\langle \cdot, \cdot \rangle$ we denote the scalar product in ${\sf R}^{d}$.
The hyperplane $H \in {\cal E}^{d}_{d-1}$ is said to intersect the polyhedral surface $M$ [*in general position*]{} if $H \cap M \not= \emptyset$ and $H$ does not contain a vertex of $M$. In that case, the intersection of $H$ with an edge of $M$ is either empty or a point, and the intersection of $H$ with a face of $M$ is either empty or a segment. It follows that the intersection $H \cap M$ is the union of finitely many simple closed polygons $P_{1}(H),\dots,
P_{k}(H)$, and from inequality (\[2\]) (applied in $H$ instead of ${\sf R}^{d}$) we have $$\kappa(H \cap M) := \kappa(P_{1}(H)) + \dots + \kappa(P_{k}(H)) \ge 2\pi.$$ It follows that $$\label{3}
I := \int \limits_{{\cal E}^{d}_{d-1}} \kappa(H \cap M) d\mu_{d-1}(H)
\nonumber
\ge 2\pi \mu_{d-1}(\{H \in {\cal E}^{d}_{d-1} : H \cap M \not= \emptyset\}),$$ since the set of all hyperplanes intersecting $M$, but not in general position, has $\mu_{d-1}$- measure zero. Let $S$ be a segment connecting two points of $M$ with maximal distance, so that the length of $S$ is equal to the diameter $\Delta$ of $M$. Let $s$ be a unit vector parallel to $S$. Then $$\begin{aligned}
& & \mu_{d-1}(\{H \in {\cal E}^{d}_{d-1}: H \cap M \not= \emptyset\})
= \int \limits_{S^{d-1}} \int
\limits_{-\infty}^{\infty}{\bf 1}_{\{H_{u,\tau}\cap M \not=
\emptyset\}} d\tau d\sigma(u) \\
& & > \int \limits_{S^{d-1}}
\int \limits_{-\infty}^{\infty}{\bf 1}_{\{H_{u,\tau} \cap S \not=
\emptyset\}} d\tau d\sigma(u)
= \int \limits_{S^{d-1}}\Delta | \langle u,s \rangle| d\sigma(u) \\
& & = c_{1}\Delta.\end{aligned}$$ Here ${\bf 1}_X$ denotes the indicator function of $X$. By $c_{1},\dots,c_{6}$ we denote constants depending only on the dimension $d$. We have proved that $$\label{4}
I > c_{2}\Delta.$$
On the other hand, if the hyperplane $H$ intersects $M$ in general position, we can write $$\label{5}
\kappa(H \cap M) = \sum_{e \in {\cal F}_{1}(M)} [\pi - \beta(M,e,H)],$$ where $\beta(M,e,H)$ is defined as follows. If $H$ meets the edge $e$ (and hence the relative interior of $e$) and if $F_{1},F_{2}$ are the two faces of $M$ incident with $e$, then $\beta(M,e,H) \in (0,\pi)$ is the angle between the segments $H \cap F_{1}$ and $H \cap F_{2}$ at the point $H \cap e$. If $H$ does not meet $e$, we put $\beta(M,e,H) = \pi$. We can now write $$\label{6}
I = \sum_{e \in {\cal F}_{1}(M)} \;
\int \limits_{{\cal E}^{d}_{d-1}}[\pi - \beta(M,e,H)]
d\mu_{d-1}(H).$$ Let $e \in {\cal F}_{1}(M)$ be a fixed edge. We have $$\begin{aligned}
\label{7}
& & \int \limits_{{\cal E}^{d}_{d-1}}
[\pi - \beta(M,e,H)]d\mu_{d-1}(H) \nonumber \\
& = & \int \limits_{S^{d-1}}
\int \limits_{-\infty}^{\infty}[\pi - \beta(M,e,H_{u,\tau})]
d\tau d\sigma(u) \nonumber \\
& = & \int \limits_{S^{d-1}}
\left[\pi - \beta(M,e,H_{u,\langle x,u \rangle}\right]L(e) |
\langle u,w(e) \rangle| d\sigma(u),\end{aligned}$$ where $x$ is some fixed point in the relative interior of the edge $e$ and $w(e)$ denotes a unit vector parallel to the edge $e$.
Let $F_{1}, F_{2}$ be the two faces of $M$ incident with $e$ and let $\alpha(e) \in (0,\pi)$ be the angle between $F_{1}$ and $F_{2}$, as defined initially. We assert that $$\label{8}
\int \limits_{S^{d-1}}\left[\pi - \beta(M,e,H_{u,\langle x,u \rangle})\right]
| \langle u,w(e)
\rangle | d \sigma(u) = c_{5}[\pi - \alpha(e)].$$ For the proof we may assume, without loss of generality, that $x$ is the origin of ${\sf
R}^{d}$. The integral in (\[8\]) can be written in the form $$\label{9}
J := \int \limits_{{\cal L}^{d}_{d-1}}f(H)d\nu_{d-1}(H),$$ where ${\cal L}^{d}_{d-1}$ denotes the space of $(d-1)$-dimensional linear subspaces of ${\sf R}
^{d}$ and $\nu_{d-1}$ is its normalized invariant measure; the function $f$ is defined by $$f(H) = [\pi - \beta(M,e,H)] | \langle u_{H},w(e) \rangle |,$$ where $u_{H}$ is a unit normal vector of $H$. The edge $e$ and the two adjacent faces $F_{1},
F_{2}$ of $M$ lie in a 3-dimensional linear subspace $A$ of ${\sf R}^{d}$. For $\nu_{d-1}$- almost all $H \in {\cal L}^{d}_{d-1}$, the intersection $A \cap H$ is a 2-dimensional linear subspace. In that case, the angle $\beta(M,e,H)$ depends only on $M$ and this subspace, so that we can write $\beta(M,e,H) = \beta(M,e,A \cap H)$. Moreover, $$|\langle u_{H},w(e) \rangle | =
| \langle u_{H},u_{A \cap H} \rangle \langle u_{A \cap H},
w(e) \rangle |,$$ where $u_{A \cap H}$ is a unit normal vector of $A \cap H$ in $A$.
Using a general formula of integral geometry, one can write the integral (\[9\]) in the form $$\int \limits_{{\cal L}^{d}_{d-1}}f(H)d\nu_{d-1}(H)
= c_{3} \int \limits_{{\cal L}^{A}_{2}}
\int \limits_{{\cal L}^{L}_{d-1}}f(H)[H,A]^{2}
d\nu_{d-1}^{L}(H)d\nu_{2}^{A}(L).$$ Here ${\cal L}^{A}_{2}$ denotes the space of 2-dimensional linear subspaces of $A$ and $\nu_{2}
^{A}$ is the normalized invariant measure on this space. For fixed $L \in {\cal L}^{A}_{2}$, ${\cal L}^{L}_{d-1}$ denotes the space of hyperplanes containing $L$, and $\nu_{d-1}^{L}$ is the invariant measure on this space. $[H,A]$ is a certain function depending only on the relative position of $H$ and $A$; it is invariant under simultaneous rotations of $H$ and $A$. The identity above is equivalent to a special case of formula (14.40) in Santaló [@San], but written in the style of [@Sch]. Applying this to our present situation, we obtain $$\begin{aligned}
J & = & c_{3} \int \limits_{{\cal L}^{A}_{2}}
\int \limits_{{\cal L}^{L}_{d-1}}[\pi - \beta
(M,e,H)] \,| \langle u_{H}, u_{A \cap H}\rangle
\langle u_{A \cap H},w(e) \rangle | \,
[H,A]^{2}d\nu_{d-1}^{L}(H)d\nu_{2}^{A}(L) \\
& = & c_{3} \int \limits_{{\cal L}^{A}_{2}}
[\pi - \beta(M,e,L)] \,| \langle u_{L},w(e) \rangle |
\int \limits_{{\cal L}^{L}_{d-1}} [H,A]^{2} \,| \langle u_{H},u_{L}\rangle | \,
d \nu_{d-1}^{L}(H)d\nu_{2}^{A}(L) \\
& = & c_{4} \int \limits_{{\cal L}^{A}_{2}}[\pi - \beta(M,e,L)] \,
| \langle u_{L},w(e) \rangle | \,
d\nu_{2}^{A}(L).\end{aligned}$$ The final integral is a mean value over 2-dimensional linear subspaces in a 3-dimensional euclidean space. Its value can be obtained from the more general Theorem 3.2.1 in [@Sch]. In this way we arrive at $$J = c_{5}[\pi - \alpha(e)],$$ which proves (\[8\]).
Taking (\[3\]), (\[6\]), (\[7\]), (\[8\]) together, we deduce that $$\label{10}
A(M) > c_{6}\Delta.$$ In order to find the optimal constant $c_{6}$ for which (\[10\]) holds generally, we consider the boundary $M_{\epsilon}$ of a triangular prism with height $\Delta$ and base a regular triangle with edge length $\epsilon$. For $\epsilon \rightarrow 0$, the diameter of $M_{\epsilon}$ tends to $\Delta$ and $A(M_{\epsilon})$ tends to $2\pi \Delta$. It follows that $c_{6} \le 2\pi$. On the other hand, from the way inequality (\[10\]) was obtained it is easy to see that $A(M) >
A(M_{\epsilon})$, if $\epsilon > 0$ is sufficiently small. Thus $c_{6} = 2\pi$ is the optimal constant.
[**3. Concluding Remark**]{}
We want to stress that the represantaton (7) is very useful for studing more complicated models [@Amb], [@Sav1], [@Sav2] and various phenomena.
This wark was supported in part by Alexander von Humboldt Foundation.
[89]{} R.V. Ambartzumian, G.K. Savvidy, K.G. Savvidy and G.S. Sukiasian, Alternative model of random surfaces. [*Phys. Lett.*]{} B [**275**]{} (1992), 99 - 102. J.W. Milnor, On the total curvature of knots. [*Ann. Math.*]{} [**52**]{} (1950), 248 - 257. L.A. Santaló, [*Integral geometry and geometric probability.*]{} Addison-Wesley, Reading, Mass. 1976. G.K. Savvidy, K.G. Savvidy.[*Int.J.Mod.Phys.*]{} A [**8** ]{} (1993) 3993; hep–th 9208041. G.K. Savvidy, K.G. Savvidy. [*Mod.Phys.Lett.*]{} A [**8** ]{} (1993) 2963; hep–th 9301001. R. Schneider und W. Weil, [*Integralgeometrie.*]{} Teubner, Stuttgart 1992.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We show that the evolutionary track of a low-mass red giant should make an extended zigzag on the Hertzsprung-Russel diagram just after the bump luminosity, if fast internal rotation and enhanced extra mixing in the radiative zone bring the temperature gradient close to the adiabatic one. This can explain both the location and peculiar surface chemical composition of Li-rich K giants studied by [@kea11]. We also discuss a striking resemblance between the photometric and composition peculiarities of these stars and giant components of RS CVn binaries. We demonstrate that the observationally constrained values of the temperature gradient in the Li-rich K giants agree with the required rate of extra mixing only if the turbulence which is believed to be responsible for this extra mixing is highly anisotropic, with its associated transport coefficients in the horizontal direction strongly dominating over those in the vertical direction.'
author:
- 'Pavel A. Denissenkov'
title: 'A NEW TWIST IN THE EVOLUTION OF LOW-MASS STARS'
---
Introduction
============
During their first ascent along the red giant branch (RGB), low-mass stars are known to experience extra mixing in their convectively stable radiative zones, between the hydrogen burning shell (HBS) and the bottom of the convective envelope (BCE). This occurs above the bump luminosity (@grea00), after the HBS has crossed and erased a chemical composition discontinuity left behind by the BCE at the end of the first dredge-up (FDU). As a result, the atmospheric $^{12}$C/$^{13}$C ratio and, in metal-poor stars, carbon abundance resume their post-FDU declines with increasing luminosity. Extra mixing is needed to explain the observed low carbon isotopic ratios in evolved stars with initial masses up to $2.2\,M_\odot$ (@chl10). Another abundance anomaly that has also been associated with the start of RGB extra mixing at the bump luminosity is the lithium enrichment that has been detected in a small fraction (1–2%) of K giants (@chb00). To produce $^7$Li via the $^7$Be mechanism (@cf71), extra mixing has to be nearly three orders of magnitude faster in these stars than in other low-mass red giants above the bump luminosity (@sb99 [@dw00]).
The problem of the origin of Li enrichment in K giants has recently been addressed again by @kea11 [hereafter KRL]. They observed 2000 K giants with determined Hipparcos parallaxes and found that 15 of them are Li-rich. However, the most interesting new result is that many of these Li-rich K giants are located well below the bump luminosity, close to the red horizontal-branch clump region of the Hertzsprung-Russell diagram (HRD), where the low-mass stars should arrive much later, after they will have experienced the He-core flash. The newly discovered Li-rich K giants also exhibit very low $^{12}$C/$^{13}$C ratios, approaching the equilibrium value for the CN cycle in four of them. Analysing these data, KRL have proposed that the lithium enrichment in these K giants occurred during the He-core flash rather than at the bump luminosity.
In this Letter, we present an alternative explanation; namely, we assume that these stars have already reached the bump luminosity and, therefore, that RGB extra mixing is now operating in their radiative zones. However, for some reason, presumably because of their very rapid internal rotation, this mixing is so efficient that not only can it trigger the $^7$Be mechanism, but its driving force or associated heat transport can also modify the radiative zone’s thermal structure, resulting in a temperature gradient that is closer to the adiabatic one. In this case, a bump-luminosity star in which such enhanced extra mixing has just started should make an extended zigzag on the HRD towards much lower luminosities, comparable to those of red clump stars, before resuming its ascent along the RGB (@dea06 [@dp08b]). This hypothesis is supported by a striking resemblance between the photometric and chemical composition peculiarities of the primary (red-giant) components of the RS CVn binaries and the Li-rich K-giants studied by KRL.
Photometric Peculiarities Caused by Enhanced Extra Mixing {#sec:photopec}
=========================================================
The RS CVn binaries are close stellar systems in which the primary components are low-mass red giants. In many cases, the RS CVn giant components have synchronized their spin and orbital periods and, as a result of their fast rotation caused by the tidal spinning up, they show signatures of chromospheric activity (@fea02 [@mea04; @fh05]). [@mea04], and [@dea06] have noted a remarkable photometric peculiarity intrinsic to the RS CVn giants and other K-giants in tidally locked binaries — most of them reside on the lower RGB, below the bump luminosity.
[@dp08b] have shown that, after having reached the bump luminosity, a low-mass star will make a long excursion towards much lower luminosities if the temperature gradient $\nabla\equiv (d\ln T/d\ln P)$ in its radiative zone increases and takes a value between the radiative and adiabatic temperature gradients, $\nabla_{\rm rad}<\nabla <\nabla_{\rm ad}$. Such a deviation of $\nabla$ from $\nabla_{\rm rad}$ can be explained in two ways. On the one hand, a rotational distortion of level surfaces in the radiative zone leads to an increase of $\nabla =\nabla_{\rm rad}$ to $\nabla\approx (1+2\varepsilon)\nabla_{\rm rad}$, where $\varepsilon\approx (\Omega^2r^3/3GM_r) < \varepsilon_{\rm crit}$ is the ratio of the equatorial centrifugal acceleration to gravity divided by 3, and $\varepsilon_{\rm crit} = 0.24$ is its critical value at which the Roche equipotential surface breaks up at the equator (@dvb03b). It is this effect that was proposed by [@dea06] to be responsible for the photometric peculiarity of red giants in tidally locked binaries. On the other hand, fast differential rotation and its related hydrodynamic or magnetohydrodynamic instabilities can be driving mechanisms for extra mixing in stellar radiative zones (e.g., @z92 [@s99]). Besides chemical elements, this extra mixing should also transport heat. The efficiency of the latter process in the standard mixing length theory is estimated by the quantity $\Gamma\approx\gamma (D_{\rm mix}/K)$, where $D_{\rm mix}$ is the rate (diffusion coefficient) of extra mixing, $K$ is the thermal diffusivity, and $\gamma$ is the factor determined by the geometry of the fluid elements. In this case, the temperature gradient lies between $\nabla_{\rm rad}$ and $\nabla_{\rm ad}$, and its value is given by $\nabla\approx (1-f)\nabla_{\rm rad} + f\nabla_{\rm ad}$, where $f = 6\Gamma^2/(1+\Gamma +6\Gamma^2)$ (@m95).
The giant components of the RS CVn binaries selected by us from the paper of [@mea04] for comparison and the Li-rich K giants discovered by KRL both have nearly solar metallicities and masses close to $2\,M_\odot$. Therefore, we compare their positions on the HRD (the red and blue star symbols in Fig. \[fig:f1\]) with the evolution of two stellar models calculated[^1] for the solar heavy-element and helium mass fractions, $Z=0.019$ and $Y=0.28$, and initial masses $1.7\,M_\odot$ and $2.2\,M_\odot$ (the black and green curves in the same figure). The neighbouring open symbols along the evolutionary tracks are separated by one million years. Their concentrations on the green curve at low $T_{\rm eff}$ are increased at the bump luminosity, $\log_{10}(L/L_\odot)\approx 2.25$, and at the red clump, $\log_{10}(L/L_\odot)\approx 1.7$. Before the star has evolved beyond the bump luminosity, its radiative zone contains a strong gradient of the mean molecular weight, $\nabla_\mu\equiv d\ln\mu/d\ln P > 0$, that is believed to prevent any extra mixing. The blue star symbols are all located below the bump luminosity, in a region of the HRD that the standard evolutionary track (the green curve) of a low-mass red giant with a chemically uniform radiative zone can reach only after the He-core flash. This explains why KRL have inferred that both the production of Li and the reduction of the carbon isotopic ratio in their Li-rich K-giants occurred during the He-core flash.
However, the $^7$Be mechanism of surface Li enrichment requires strongly enhanced extra mixing with a rate (diffusion coefficient) increased by nearly three orders of magnitude compared to that of canonical extra mixing operating in the majority of RGB stars above the bump luminosity (@dvb03a). We surmise that such mixing or its associated fast rotation should modify $\nabla$ in the radiative zone bringing its value closer to $\nabla_{\rm ad}$. The red curve originating from $\log_{10}(L/L_\odot)\approx 2.25$ in Fig. \[fig:f1\] shows a bifurcation of the evolution of the $2.2\,M_\odot$ star after the bump luminosity caused by our assumption that the start of enhanced extra mixing with $D_{\rm mix} = 2\times 10^{11}$cm$^2$s$^{-1}$ in this star has initiated a convective heat transport in its radiative zone with the efficiency $\Gamma = 0.75$. It is seen that the star spends quite a long time at luminosities close to those of the Li-rich K-giants. We have assumed that the enhanced extra mixing has the same maximum depth, $r_{\rm mix} = 0.05\,R_\odot$, as the canonical one (@dp08a). Before the HBS, advancing in mass outwards, has crossed the chemical composition discontinuity left behind by the BCE at the end of the FDU, the mixing is allowed to operate only in the chemically uniform zone between the BCE and the current location of the discontinuity at $r>r_{\rm mix}$. Note that it is the increase of $\nabla$, not the effect of mixing, that forces the star to make the extended zigzag. Fig. \[fig:f1\] also shows that the region occupied by the majority of tidally locked giants in close binaries, including the RS CVn giant components (red star symbols), is reached by the evolutionary U-turn of the $1.7\,M_\odot$ star (red curve) when $\Gamma = 1$.
Chemical Composition Peculiarities Caused by Enhanced Extra Mixing {#sec:chempec}
==================================================================
The parameter values $D_{\rm mix} = 2\times 10^{11}$cm$^2$s$^{-1}$, $r_{\rm mix} = 0.05\,R_\odot$, and $\Gamma = 0.75$ lead to both Li enrichment[^2] and reduction of the $^{12}$C/$^{13}$C ratio consistent with those reported by KRL for the Li-rich K giants (solid red curves in Fig. \[fig:f2\] and Fig. \[fig:f3\]). Furthermore, an increase of the efficiency of convective heat transport to $\Gamma = 1$ shifts the HRD location of the Li enrichment and, especially, that of the $^{12}$C/$^{13}$C sharp decline to the edges of their corresponding observational domains (dashed red curves in the same figures). Therefore, if our hypothesis is correct, $\Gamma$ should be close to 0.75 and not exceed 1 by much.
A similar value of $\Gamma$ is needed to explain the photometric peculiarity of red giants in tidally locked binaries, including the RS CVn giant components (Fig. \[fig:f1\]). [@dea06] have predicted that some of these stars, those that have already reached the bump luminosity and are now making extended zigzags towards lower luminosities as a result of their fast rotation, should have carbon isotopic ratios smaller than the standard post-FDU value of $^{12}$C/$^{13}\mbox{C}\approx 25$. Recently, two such stars have actually been found (@tea10 [@bea10]); the chromospherically active RS CVn-type stars $\lambda$And (HD222107) and 29 Draconis (HD160538) are located below the bump luminosity but have $^{12}$C/$^{13}\mbox{C}=14$ and 16, respectively.
When comparing the Li-rich K giants with the RS CVn binaries, it is important to bear in mind that a hunt after the former is biased towards finding the stars that have already reached the bump luminosity and are now experiencing enhanced extra mixing, while a search for the latter selects both the pre- and post-bump luminosity stars. Therefore, it is not surprising that the third RS CVn-type star, 33 Piscium, studied by the same group does not show signatures of extra mixing (@bea11). It would be interesting to measure carbon isotopic ratios in the following RS CVn stars: HD19754, HD182776, HD202134, HD204128, and HD205249. They have relatively high (for red giants) projected rotational velocities ($v\sin i > 7$kms$^{-1}$) and probably strongly increased atmospheric abundances of Na and Al (@mea04), which might be a manifestation of enhanced extra mixing (@dvb03a).
For comparison, green curves in Fig. \[fig:f2\] and Fig. \[fig:f3\] show the evolutionary changes of the Li abundance and $^{12}$C/$^{13}$C ratio at the surface of the $2.2\,M_\odot$ star produced by the FDU (before the bump luminosity) and thermohaline convection driven by $^3$He burning that reduces $\mu$ in the tail of the HBS after the bump luminosity (@chz07). The rate of the latter is approximated by $D_\mu = 2\pi^2a^2[\nabla_\mu/(\nabla_{\rm rad}-\nabla_{\rm ad})]K$ (@d10a) with the observationally constrained value of the aspect ratio, $a\equiv l/d=10$, of a fluid element, where $l$ and $d$ are its length and diameter. Note that direct numerical simulations of the $^3$He-driven thermohaline convection in a bump-luminosity red giant give an estimate of $a\approx 1$, which does not support the hypothesis that it is the principal mechanism of RGB extra mixing (@d10a [@dm11; @tea11]).
Discussion {#sec:disc}
==========
For the majority of low-mass red giants, that are experiencing extra mixing above the bump luminosity, the observed evolutionary changes of their surface chemical composition are reproduced reasonably well either with the above-mentioned thermohaline diffusion coefficient $D_\mu$ that uses $a=10$, or with the diffusion coefficient $D_{\rm mix}\approx 0.01K$–$0.1K$ and appropriately chosen mixing depth, e.g. $r_{\rm mix} = 0.05\,R_\odot$ (@dp08a [@dp08b]). For the radiative zone of our $2.2\,M_\odot$ bump-luminosity model, these diffusion coefficients are plotted in Fig. \[fig:f4\] (the green and black curves). Their corresponding efficiencies of heat transport are negligibly small. Therefore, in the majority of cases, the (canonical) RGB extra mixing should not lead to noticeable changes in the evolution of these stars on the HRD.
On the contrary, the value of $D_{\rm mix} = 2\times 10^{11}$ cm$^2$s$^{-1}$, that is required to understand the origin of Li-rich K giants in our models, results in $\Gamma = \gamma(D_{\rm mix}/K)\gg 1$, at least in the inner half of the radiative zone (compare the red and blue curves in Fig. \[fig:f4\]), therefore such (enhanced) extra mixing has to be accompanied by an efficient heat transport. We have shown that the increase of $\Gamma$ to a value between 0.75 and 1 also helps to explain the photometric peculiarities intrinsic to both the majority of red giants in tidally locked binaries, including the RS CVn stars, and the Li-rich K giants studied by KRL. However, post-bump luminosity stars with $\Gamma\gg 1$ would make zigzags that are too lengthy, in conflict with observations. This disagreement can be resolved if, following [@z92], we assume that the enhanced extra mixing is produced by highly anisotropic turbulence whose associated chemical and heat transport in the horizontal direction strongly dominates over those in the vertical direction, $D_{\rm h}\gg D_{\rm mix}$. [@dp08b] have demonstrated that, in this case, $\Gamma = D_{\rm mix}/[2a^2(K+D_{\rm h})]$ and $f = 3\Gamma(D_{\rm mix}/K)/[1+\Gamma + 3\Gamma(D_{\rm mix}/K)]$, while the expression $\nabla = (1-f)\nabla_{\rm rad} + f\nabla_{\rm ad}$ still holds. In the standard mixing length theory, $\Gamma = 0.75$ and $\Gamma = 1$ correspond to $f= 0.66$ and $f=0.75$, respectively. For the anisotropic turbulent mixing, we obtain $\Gamma \ll 1$, assuming that $D_{\rm h}\gg D_{\rm mix}\gg K$ and $a\geq 1$. At the same time, we can still get the observationally supported value of $f\approx 0.75$, provided that $\Gamma(D_{\rm mix}/K)\approx 1$, or $2(D_{\rm h}/K)\approx (D_{\rm mix}/K)^2$. This means that, for $D_{\rm mix}\approx 10^2K$ (Fig. \[fig:f4\]), we need $D_{\rm h}\approx 5\times 10^3D_{\rm mix}$ to keep $f$ close to 0.75. Interestingly, [@d10b] has used a similar ratio of $D_{\rm h}$ to $D_{\rm mix}$ for the rotation-driven turbulent diffusion to make his model of magnetic braking of solar rotation consistent with observational data. Moreover, the same order of magnitude ratios $D_{\rm h}/D_{\rm mix}$ were obtained for radiative zones of bump luminosity red giant models by [@pea06] who took into account self-consistently the transport of angular momentum by rotation-driven meridional circulation and shear turbulence.
Given that $\nabla_{\rm rad}\approx 0.2$ and $\nabla_{\rm ad}\approx 0.4$ in the radiative zone of a bump-luminosity red giant, the maximum deviation of $\nabla$ from $\nabla_{\rm rad}$ that can be achieved at the break-up rotation corresponds to $f\approx 2\varepsilon_{\rm crit}=0.48$, which is too small to reproduce the extended zigzags presumably made by the Li-rich K giants and RS CVn giant components. However, this conclusion is based only on the analysis of the modification of the temperature gradient by the rotational distortion. It ignores the fact that effects of rotation are also incorporated in other equations of stellar structure (@dvb03b). With all the effects taken into account, [@dea06] were actually able to construct the evolutionary track of a $1.7\,M_\odot$ star whose post-bump luminosity zigzag reached the HRD domain occupied by the RS CVn binaries.
Comparing the green and red curves in Fig. \[fig:f4\], we infer that the $^3$He-driven thermohaline convection could be responsible for the observed Li enrichment only if the aspect ratio of its fluid elements were as large as $a\approx 300$. At present, we consider this highly improbable. The more likely interpretation from our point of view is that the Li enrichment in the stars studied by KRL and their location on the lower RGB below the bump luminosity are both caused by their fast internal rotation and its associated turbulent mixing and heat transport. This hypothesis is supported by the fact that surface rotation velocities of field stars with $M\ga 1.6\,M_\odot$ remain constant as they evolve on and away from the main sequence (@ws97). It is likely that only the most rapidly rotating stars, of which a handful have $v\sin i > 200$ kms$^{-1}$, become Li-rich and make extended zigzags after the bump luminosity. Our hypothesis does not exclude the possibility of Li enrichment above the bump luminosity in a red giant that has been spun up as a result of its engulfing an orbiting giant planet (@dw00).
The only possible scenario in which a substantial mixing of H-rich material occurs during the He-core flash is the “hydrogen injection flash” (@mea11). The He ignition starts off-center, which leads to the formation of a He convective shell. For this scenario to work, convection from the He-burning shell must penetrate into the HBS. In the MESA stellar evolution code, such penetration, also known as convective overshooting, is modeled by the exponentially decaying diffusion coefficient $D_{\rm OV} = D_0\exp[-2|r-r_0|/(fH_P)]$, where $H_P$ is the pressure scale height, and $D_0$ is a convective diffusion coefficient estimated using a mixing-length theory at the radius $r_0$ near the convective boundary (@hea97). Our computations of the He-core flash in the $1.7\,M_\odot$ star show that H can be injected into the He convective shell only when $f\ga 0.15$. This exceeds the observationally constrained value of $f$ by nearly an order of magnitude (@nea10, and references therein). Besides, given the discussed similarities between the Li-rich K giants studied by KRL and the RS CVn giant components and the fact that the latter stars are definitely on the lower RGB, because otherwise they would have undergone a common-envelope event with their close binary companions, we believe that these stars are in the same evolutionary phase. It should also be noted that the inclusion of convective overshooting in main-sequence stars would shift the maximum initial mass of the solar-composition stars that experience the He-core flash from $M\approx 2.2\,M_\odot$ to $M\approx 1.8\,M_\odot$.
Our hypothesis could eventually be tested by asteroseismology because high-precision photometry capable of measuring g-mode oscillations in red giants allows “to distinguish unambiguously between hydrogen-shell-burning stars (period spacing mostly $\sim$50 seconds) and those that are also burning helium (period spacing $\sim$100 to 300 seconds)” (@beddea11).
I am thankful to Falk Herwig and Don VandenBerg for useful comments and for supporting my work through their Discovery Grants from Natural Sciences and Engineering Research Council of Canada.
Barisevicius, G., Tautvaisienė, G., Berdyugina, S., Chornyi, Y., & Ilyin, I. 2010, BaltA, 19, 157
Barisevicius, G., Tautvaisienė, G., Berdyugina, S., Chornyi, Y., & Ilyin, I. 2010, BaltA, 20, 53
Bedding, S., Mosser, B., Huber, D., et al. 2011, Nature, 471, 608
Cameron, A. G. W., & Fowler, W. A. 1971, ApJ, 164, 111
Charbonnel, C., & Balachandran, S. C. 2000, A&A, 359, 563
Charbonnel, C., & Zahn, J.-P. 2007, A&A, 467, L15
Charbonnel, C., & Lagarde, N. 2010, A&A, 522, A10
Denissenkov, P. A. 2010a, ApJ, 723, 563
Denissenkov, P. A. 2010b, ApJ, 719, 28
Denissenkov, P. A., Chaboyer, B., & Li, K. 2006, ApJ, 641, 1087
Denissenkov, P. A., & Weiss, A. 2000, A&A, 358, L49
Denissenkov, P. A., & VandenBerg, D. A. 2003a, ApJ, 593, 509
Denissenkov, P. A., & VandenBerg, D. A. 2003b, ApJ, 598, 1246
Denissenkov, P. A., & Pinsonneault, M. 2008a, ApJ, 679, 1541
Denissenkov, P. A., & Pinsonneault, M. 2008b, ApJ, 684, 626
Denissenkov, P. A., & Merryfiled, W. J. 2011, ApJ, 727, L8
Fekel, F. C., Henry, G. W., Eaton, J. A., Sperauskas, J., & Hall, D. S. 2002, AJ, 124, 1064
Fekel, F. C., & Henry, G. W. 2005, AJ, 129, 1669
Gratton, R. G., Sneden, C., Carretta, E., & Bragaglia, A. 2000, A&A, 354, 169
Herwig, F., Blöcker, T., Schönberner, D., & El Eid, M. 1997, A&A, 324, L81
Kumar, Y. B., Reddy, B. E., & Lambert, D. L. 2011, ApJ, 730, L12 (KRL)
Maeder, A. 1995, A&A, 299, 84
Mocák, M., Siess, L., & Müller, E. 2011, A&A, 533, 53
Morel, T., Micela, G., Favata, F., & Katz, D. 2004, A&A, 426, 1007
Noels, A., Montalban, J., Miglio, A., Godart, M., & Ventura, P. 2010, Ap&SS, 328, 227
Palacios, A., Charbonnel, C., Talon, S., & Siess, L. 2006, A&A, 453, 261
Paxton, B., Bildsten, L., Dotter, A., Herwig, F., Lessafre, P., & Timmes, F. 2011, ApJS, 192, 3
Sackmann, I.-J., & Boothroyd, A. I. 1999, ApJ, 510, 217
Spruit, H. C. 1999, A&A, 349, 189
Tautvaisienė, G., Barisevicius, G., Berdyugina, S., Chornyi, Y., & Ilyin, I. 2010, BaltA, 19, 95
Traxler, A., Garaud, T., & Stellmach, S. 2011, ApJ, 728, L29
Wolff, S.-C., & Simon, T. 1997, PASP, 109, 759
Zahn, J.-P. 1992, A&A, 256, 115
[^1]: All the stellar evolution computations for this paper have been done with the MESA code (@pea11) that is freely available at [http://mesa.sourceforge.net]{}.
[^2]: For the Li abundance, we use the notation $\varepsilon(^7\mbox{Li})\equiv\log_{10}
[n(^7\mbox{Li})/n(\mbox{H})] + 12$, where $n$ is the number density of the corresponding nucleus.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this paper a class of single machine scheduling problems is considered. It is assumed that job processing times and due dates can be uncertain and they are specified in the form of discrete scenario set. A probability distribution in the scenario set is known. In order to choose a schedule some risk criteria such as the value at risk (VaR) an conditional value at risk (CVaR) are used. Various positive and negative complexity results are provided for basic single machine scheduling problems. In this paper new complexity results are shown and some known complexity results are strengthen.'
author:
- |
Adam Kasperski$^\dag$ Pawe[ł]{} Zieli[ń]{}ski$^\ddag$\
[*$^\dag$Department of Operations Research, Faculty of Computer Science and Management,*]{}\
[*Wroc[ł]{}aw University of Science and Technology, Wroc[ł]{}aw, Poland*]{}\
[*$^\ddag$Department of Computer Science, Faculty of Fundamental Problems of Technology,*]{}\
[*Wroc[ł]{}aw University of Science and Technology, Wroc[ł]{}aw, Poland*]{}
title: 'Risk averse single machine scheduling - complexity and approximation'
---
Introduction
============
Scheduling under risk and uncertainty has attracted considerable attention in recent literature. In practical applications of scheduling models the exact values of input parameters, such as job processing times or due dates, are often unknown in advance. Hence, a solution must be computed, before the true realization of the input data reveals. Typically, a *scenario set* $\mathcal{U}$ is a part of the input, which contains all possible realizations of the problem parameters, called *scenarios*. If the probability distribution in $\mathcal{U}$ is unknown, then *robust optimization* framework can be applied and solution performance in a worst case is optimized. First robust scheduling problems have been discussed in [@DK95; @KY97; @YK93]. Two uncertainty representations, namely a *discrete* and *interval* ones were considered. In the former, scenario set $\mathcal{U}$ contains a finite number of distinct scenarios. In the latter, for each uncertain parameter an interval of its possible values is specified and $\mathcal{U}$ is the Cartesian product of these intervals. In order to compute a solution the minmax and minmax regret criteria can be applied. Minmax (regret) scheduling problems have various complexity properties, depending on the cost function and the uncertainty representation (see, e.g., [@AV00; @AL06; @K05; @AAK11; @DR16]). For a survey of minmax (regret) scheduling problems we refer the reader to [@KZ14s].
The robust scheduling models have well known drawbacks. Minimizing the maximum cost can lead to very conservative solutions. The reason is that the probability of occurrence of the worst scenario may be very small and the information connected with the remaining scenarios is ignored while computing a solution. One method of overcoming this drawback was given in [@KZ16d], where the OWA criterion, proposed in [@YA88], was applied to compute an optimal schedule. In this approach, a set of weights is specified by the decision maker, which reflect his attitude towards a risk. The OWA operator contains the maximum, average and Hurwicz criteria as special cases. However, it does not take into account a probabilistic information, which may be available for scenario set $\mathcal{U}$.
In the case, when a probability distribution in $\mathcal{U}$ is known, the stochastic scheduling models are considered. The parameters of scheduling problem are then random variables with known probability distributions. Under this assumption, the expected solution performance is typically optimized (see, e.g., [@MSU99; @PI08; @SU05; @SSU16]). However, this criterion assumes that the decision maker is risk neutral and leads to solutions that guarantee an optimal long run performance. Such a solution may be questionable, for example, if it is implemented only once (see, e.g., [@KY97]). In this case, the decision maker attitude towards a risk should be taken into account.
In [@KPU02] a criterion called *conditional value at risk* (CVaR) was applied to a stochastic portfolio selection problem. Using this criterion, the decision maker provides a parameter $\alpha\in [0,1)$, which reflects his attitude towards a risk. When $\alpha=0$, then CVaR becomes the expectation. However, for greater value of $\alpha$, more attention is paid to the worst outcomes, which fits into the robust optimization framework. The conditional value at risk is closely connected with the *value at risk* (VaR) criterion (see, e.g., [@P00]), which is just the $\alpha$-quantile of a random outcome. Both risk criteria have attracted considerable attention in stochastic optimization (see, e.g., [@NST15; @ZSZD17; @NIK10; @O12]). This paper is motivated by the recent papers [@SSL14] and [@NBN17], in which the following stochastic scheduling models were discussed. We are given a scheduling problem with discrete scenario set $\mathcal{U}$. Each scenario $\xi_i\in \mathcal{U}$ is a realization of the problem parameters (for example, processing times and due dates), which can occur with a known positive probability ${\rm Pr}[\xi_i]$. The cost of a given schedule is a discrete random variable with the probability distribution induced by the probability distribution in $\mathcal{U}$. The VaR and CVaR criteria, with a fixed level $\alpha$, are used to compute a best solution.
In [@SSL14] and [@NBN17] solution methods based on mixed integer programming models were proposed to minimize VaR and CVaR in scheduling problems with the total weighted tardiness criterion. The aim of this paper is to analyze the models discussed in [@SSL14] and [@NBN17] from the complexity point of view. We will consider the class of single machine scheduling problems with basic cost functions, such as the maximum tardiness, the total flow time, the total tardiness and the number of late jobs. We will discuss also the weighted versions of these cost functions. We provide a picture of computational complexity for all these problems by proving some positive and negative complexity results. Since VaR and CVaR generalize the maximum criterion, we can use some results known from robust minmax scheduling. The complexity results for minmax versions of single machine scheduling problems under discrete scenario set were obtained in [@AAK11; @AC08; @DK95; @MNO13]. In this paper we will show that some of these results can be strengthen.
This paper is organized as follows. In Section \[sec1\] we recall the definitions of the VaR and CVaR criteria and show their properties, which will be used later on. In Section \[sec2\] the problems discussed in this paper are defined. In Section \[sec3\] some general relationships between the problems with various risk criteria are shown. Finally, Sections \[sec4\] and \[sec5\] contain some new negative and positive complexity results for the the considered problems. These results are summarized in the tables presented in Section \[sec2\].
The risk criteria {#sec1}
=================
Let $\rm Y$ be a random variable. We will consider the following risk criteria [@P00; @RU00]:
- *Value at Risk* ($\alpha$-quantile of $\mathrm{Y}$): $${\bf VaR}_\alpha[\rm Y]=\inf\{t: {\rm Pr}[Y\leq t] \geq \alpha\}, \alpha\in(0,1],$$
- *Conditional value at risk*: $${\bf CVaR}_\alpha[\rm Y]=\inf\{\gamma+\frac{1}{1-\alpha}{\bf E}[\rm Y-\gamma]^+: \gamma \in {\mathbb{R}}\}, \alpha\in [0,1),$$
where $[x]^+=\max\{0,x\}$. Assume that $\mathrm{Y}$ is a discrete random variable taking nonnegative values $b_1,\dots, b_K$. Then $ {\bf VaR}_\alpha[\rm Y]$ and ${\bf CVaR}_\alpha[\rm Y]$ can be computed by using the following programs, respectively (see, e.g., [@NBN17; @O12; @RU00]): $$\begin{aligned}
\text{(a)}\;\;\;\;\;& \min \theta & \text{(b)}\;\;\;\;\;&\min \gamma + \frac{1}{1-\alpha}\sum_{i\in [K]} {\rm Pr}[\mathrm{Y}=b_k] u_k\nonumber\\
\text{s.t. }& b_k-\theta\leq M \beta_k,\;\;\; k\in [K] &\text{s.t. }& \gamma + u_k \geq \displaystyle b_k,\;\;\; k\in [K] \label{exVC1}\\
&\sum_{k\in [K]} {\rm Pr}[\mathrm{Y}=b_k] \beta_k \leq 1-\alpha && u_k\geq 0,\;\;\; k\in [K]\nonumber\\
& \beta_k\in \{0,1\},\;\;\; k\in [K] && \nonumber \end{aligned}$$ where $M\geq \max\{b_1,\dots, b_K\}$ and $[K]=\{1,\dots,K\}$. Notice that (\[exVC1\])b is a linear programming problem. In the following, we will use the following dual to (\[exVC1\])b: $$\label{cvarmod1}
\begin{array}{lll}
\displaystyle \max & \displaystyle \sum_{k\in [K]} b_k r_k \\
\text{s.t.} & \displaystyle\sum_{k\in [K]} r_k=1\\
&0 \leq r_k\leq \frac{{\rm Pr}[\mathrm{Y}=b_k] }{1-\alpha}, & k\in [K]
\end{array}$$ The equality constraint in (\[cvarmod1\]) follows from the fact that all $b_k$, $k\in [K]$, are nonnegative. Substituting $r_k=q_k/(1-\alpha)$ into (\[cvarmod1\]), we get the following equivalent formulation for ${\bf CVaR}_\alpha[\rm Y]$: $$\label{cvarmod2}
\begin{array}{lll}
\displaystyle \max & \displaystyle \frac{1}{1-\alpha} \sum_{k\in [K]} b_k q_k\\
\text{s.t.} & \displaystyle\sum_{k\in [K]} q_k=1-\alpha\\
&0 \leq q_k\leq {\rm Pr}[\mathrm{Y}=b_k], & k\in [K]
\end{array}$$ Program (\[cvarmod2\]) can be solved by using a greedy method, which is illustrated in Figure \[fig0\]. Namely, we fix the optimal values of $q_k$ by greedily distributing the amount $1-\alpha$ among the largest values of $b_i$. It is easy to see that ${\bf CVaR}_{0}[{\rm Y}]={\bf E}[{\rm Y}]=\sum_{k\in [K]} b_k {\rm Pr}[\mathrm{Y}=b_k]$. On the other hand, ${\bf CVaR}_{1-\epsilon}[{\rm Y}]= {\bf VaR}_1[{\rm Y}]={\bf Max}[{\rm Y}]=\max_{k\in [K]} b_k$ for sufficiently small $\epsilon>0$ and any probability distribution.
![A computation of ${\bf CVaR}_{0.5}[{\rm Y}]$ for $\rm Y$ taking the values of 13, 22, 29, 33, and 36 with the probabilities 0.3, 0.1, 0.2, 0.1, 0.3, respectively. The value of ${\bf CVaR}_{0.5}[{\rm Y}]$ is the grey area divided by $1-\alpha=0.5$.[]{data-label="fig0"}](fig0.eps){width="6cm"}
We now show several properties of the risk measures which will be used later on in this paper.
\[lem01\] Let $\mathrm{Y}$ be a discrete random variable which can take $K$ nonnegative values $b_1,\dots, b_K$. The following inequalities hold for each $\alpha\in [0,1)$: $${\bf E}[\mathrm{Y}]\leq {\bf CVaR}_\alpha [\mathrm{Y}] \leq
\min\left \{\frac{1}{{\Pr}_{\min}},\frac{1}{1-\alpha} \right\} {\bf E}[\mathrm{Y}],$$ where ${\rm Pr}_{\min}=\min_{k\in [K]} {\rm Pr}[\mathrm{Y}=b_k]$.
Fix $\alpha\in [0,1)$. The inequality ${\bf E}[\mathrm{Y}]\leq {\bf CVaR}_\alpha [\mathrm{Y}]$ follows directly from the definition of the expected value and the conditional value at risk. We now prove the second inequality. Let $r^*_1, \dots r^*_k$ be the optimal values in (\[cvarmod1\]). Then the inequality $${\bf CVaR}_\alpha [\mathrm{Y}]=\sum_{k\in [K]} r^*_k b_k \leq \sum_{k\in [K]} \frac{{\rm Pr}[\mathrm{Y}=b_k]}{(1-\alpha)} b_k= \frac{1}{1-\alpha}{\bf E}[\mathrm{Y}]$$ holds. Since the value of ${\bf CVaR}_\alpha [\mathrm{Y}]$ is a convex combination of $b_1,\ldots,b_k$ (see (\[cvarmod1\])), we have ${\bf CVaR}_\alpha [\mathrm{Y}] \leq {\bf Max}[\mathrm{Y}]=b_{\max}\leq \sum_{k\in [K]} \frac{{\rm Pr}[\mathrm{Y}=b_k]}{{\rm Pr}_{\min}} b_k=\frac{1}{{\rm Pr}_{\min}}{\bf E}[\mathrm{Y}],$ and the lemma follows.
\[lemowa1\] Let $\mathrm{X}$ and $\mathrm{Y}$ be two discrete random variables taking nonnegative values $a_1,\dots, a_K$, and $b_1,\dots, b_K$, respectively, with ${\rm Pr}[\mathrm{X}=a_i]={\rm Pr}[\mathrm{Y}=b_i]$ and $a_i\leq \gamma b_i$ for each $i\in [K]$ and some fixed $\gamma \geq 0$. Then ${\bf CVaR}_\alpha[\mathrm{X}]\leq \gamma {\bf CVaR}_\alpha[\mathrm{Y}]$ for each $\alpha\in [0,1)$ and ${\bf VaR}_\alpha[\mathrm{X}]\leq \gamma {\bf VaR}_\alpha[\mathrm{Y}]$ for each $\alpha\in (0,1]$.
Let us compute ${\bf CVaR}_\alpha[\mathrm{X}]$ by using (\[cvarmod1\]) and denote by $r^*_k$, $k\in [K]$, the optimal values in (\[cvarmod1\]). Then ${\bf CVaR}_\alpha[\mathrm{X}]=\sum_{k\in [K]} r^*_k a_k\leq \gamma \sum_{k\in [K]} r^*_k b_k\leq \gamma {\bf CVaR}_\alpha[\mathrm{Y}]$. Let us compute ${\bf VaR}_{\alpha}[\mathrm{Y}]$ by solving the problem (\[exVC1\])a. Let $\theta^*$, $\beta^*_k$, $k\in [K]$, be an optimal solution to (\[exVC1\])a. Since $\gamma\geq 0$, the constraint $\gamma b_k -\gamma \theta^* \leq \gamma M \beta^*_k$ holds for each $k\in [K]$. By $a_k\leq \gamma b_k$ for each $k\in [K]$, we get $a_k-\gamma \theta^* \leq M' \beta^*_k,$ where $M'=\gamma M \geq \max\{a_1,\dots, a_K\}$, $k\in [K]$. In consequence, $$\label{exVC12}
\begin{array}{llll}
& \displaystyle a_k-\gamma \theta^*\leq M' \beta^*_k & k\in [K]\\
&\displaystyle \sum_{k\in [K]} {\rm Pr}[\mathrm{X}=a_k] \cdot \beta^*_k \leq 1-\alpha\\
\end{array}$$ and ${\bf VaR}_{\alpha}[\mathrm{X}]\leq \gamma \theta^*=\gamma{\bf VaR}_{\alpha}[\mathrm{Y}]$.
Problem formulations {#sec2}
====================
We are given a set $J$ of $n$ jobs, which can be partially ordered by some precedence constraints. Namely, $i\rightarrow j$ means that job $j$ cannot start before job $i$ is completed. For each job $j\in J$ a nonnegative processing time $p_j$, a nonnegative due date $d_j$ and a nonnegative weight $w_j$ can be specified. A schedule $\pi$ is a feasible (i.e. preserving the precedence constraints) permutation of the jobs and $\Pi$ is the set of all feasible schedules. We will use $C_j(\pi)$ to denote the completion time of job $j$ in schedule $\pi$. Obeying the standard notation, we will use $T_j(\pi)=[C_j(\pi)-d_j]^+$ to define the tardiness of $j$ in $\pi$, and $U_j(\pi)=1$ if $C_j(\pi)>d_j$ (job $j$ is late in $\pi$) and $U_j(\pi)=0$ (job $j$ is on-time in $\pi$), otherwise. In the deterministic case we seek a schedule $\pi\in \Pi$ that minimizes a given cost function $f(\pi)$. The basic cost functions are the *total flow time* $\sum_{j\in J} C_j(\pi)$, the *total tardiness* $\sum_{j\in J} T_j(\pi)$, the *maximum tardiness* $\max_{j\in J} T_j(\pi)$ and the *total number of late jobs* $\sum_{j\in J} U_j(\pi)$. We can also consider the weighted versions of these functions. Scheduling problems $\mathcal{P}$ will be denoted by means of the standard Graham’s notation (see, e.g., [@B07]).
In this paper we assume that job processing times and due dates can be uncertain. The uncertainty is modeled by a discrete *scenario set* $\mathcal{U}=\{\xi_1,\xi_2,\dots,\xi_K\}$. Each realization of the parameters $\xi \in \mathcal{U}$ is called a *scenario*. For each scenario $\xi\in \mathcal{U}$ a probability ${\rm Pr}[\xi]$ of its occurrence is known (without loss of generality we can assume ${\rm Pr}[\xi]>0$). We will use $p_j(\xi)$ and $d_j(\xi)$ to denote the processing time and due date of job $j$ under scenario $\xi\in \mathcal{U}$, respectively. We will denote by $C_j(\pi, \xi)$, $T_j(\pi, \xi)$ and $U_j(\pi, \xi)$ the completion time, tardiness and unit penalty of job $\pi$, respectively, under scenario $\xi\in \mathcal{U}$. Also, $f(\pi, \xi)$ stands for the cost of schedule $\pi$ under scenario $\xi\in \mathcal{U}$. Given a feasible schedule $\pi \in \Pi$, we denote by ${\rm F}(\pi)$ a random cost of $\pi$. Notice that ${\rm F}(\pi)$ is a discrete random variable with the probability distribution induced by the probability distribution in $\mathcal{U}$.
For a fixed value of $\alpha$, we can compute a performance measure of $\pi$, namely the expected cost $\textbf{E}[{\rm F}(\pi)]$, the maximum cost $\textbf{Max}[{\rm F}(\pi)]$, the value at risk $\textbf{VaR}_\alpha[{\rm F}(\pi)]$ and the conditional value at risk $\textbf{CVaR}_\alpha[{\rm F}(\pi)]$. A sample problem $1||\sum C_j$ with 4 jobs and 5 processing time scenarios is shown in Figure \[fig1\]. Let $\pi=(1,2,3,4)$. It is easily seen that $\textbf{E}[{\rm F}(\pi)]=26$, $\textbf{VaR}_{0.5}[{\rm F}(\pi)]=29$, $\textbf{CVaR}_{0.5}[{\rm F}(\pi)]=34$ and $\textbf{Max}[{\rm F}(\pi)]=36$.
![A sample scheduling problem $1||\sum C_j$ with 5 processing time scenarios.[]{data-label="fig1"}](fig1.eps){width="12cm"}
In this paper we will study the problems $\textsc{Min-VaR}_\alpha~\mathcal{P}$, $\textsc{Min-CVaR}_\alpha~\mathcal{P}$, $\textsc{Min-Exp}$ $\mathcal{P}$, and $\textsc{Min-Max}~\mathcal{P}$, in which we minimize the corresponding performance measure for a fixed $\alpha$ and a specific single machine scheduling problem $\mathcal{P}$, under a given scenario set $\mathcal{U}$. Notice that the robust $\textsc{Min-Max}~\mathcal{P}$ problem is a special case of both $\textsc{Min-VaR}_\alpha~\mathcal{P}$ and $\textsc{Min-CVaR}_\alpha~\mathcal{P}$. Also, $\textsc{Min-Exp}$ $\mathcal{P}$ is a special case of $\textsc{Min-CVaR}_\alpha~\mathcal{P}$.
In the next sections we provide a number of new positive and negative complexity and approximation results for basic single machine scheduling problems $\mathcal{P}$. Tables \[tabs1\]-\[tabs3\] summarize the known and new results. In Table \[tabs1\], the negative results for uncertain due dates and deterministic processing times are shown. In Table \[tabs2\], the negative results for uncertain processing times and deterministic due dates are presented. Finally, in Table \[tabs3\], some positive results are shown.
$\mathcal{P}$ $\textsc{Min-Exp}~\mathcal{P}$ $\textsc{Min-VaR}_\alpha~\mathcal{P}$ $\textsc{Min-CVaR}_\alpha~\mathcal{P}$ $\textsc{Min-Max}~\mathcal{P}$
-------------------- ----------------------------------------------- --------------------------------------- ---------------------------------------- -------------------------------------- -- -- -- -- -- --
$1|p_j=1|T_{\max}$ str. NP-hard str. NP-hard str. NP-hard poly. sol. [@KZ16d]
not appr. within not at all appr. not appr. within
$\frac{7}{6}-\epsilon$, $\epsilon>0$ [@KZ16d] for any $\alpha\in (0,1)$ $\frac{7}{6}-\epsilon$, $\epsilon>0$
for any $\alpha\in (0,1)$
$1|p_j=1|\sum U_j$ poly sol. str. NP-hard str. NP-hard str. NP-hard
(assignment) not at all appr. for any $\alpha\in (0,1)$ not appr. for any
for any $\alpha\in (0,1)$ constant $\gamma>1$
$1||\sum U_j$ NP-hard as above as above as above
$1|p_j=1|\sum T_j$ poly sol. str. NP-hard str. NP-hard str. NP-hard
(assignment) not at all appr. for any $\alpha\in (0,1)$ not appr. within
for any $\alpha\in (0,1)$ $\frac{5}{4}-\epsilon$, $\epsilon>0$
$1||\sum T_j$ str. NP-hard as above as above as above
: Complexity results for uncertain due dates (processing times are deterministic).[]{data-label="tabs1"}
$\mathcal{P}$ $\textsc{Min-Exp}~\mathcal{P}$ $\textsc{Min-VaR}_\alpha~\mathcal{P}$ $\textsc{Min-CVaR}_\alpha~\mathcal{P}$ $\textsc{Min-Max}~\mathcal{P}$
--------------- -------------------------------- --------------------------------------- ---------------------------------------- ------------------------------------------------------ -- -- -- -- -- --
$1||\sum C_j$ poly sol. str. NP-hard str. NP-hard str. NP-hard
not appr. within for any $\alpha\in (0,1)$ not appr. within
$\frac{6}{5}-\epsilon$, $\epsilon>0$ $\frac{6}{5}-\epsilon$, $\epsilon>0$ [@KY97; @MNO13]
$1||\sum U_j$ open str. NP-hard str. NP-hard str. NP-hard
for any $\alpha\in (0,1)$ for any $\alpha\in (0,1)$
$1||\sum T_j$ NP-hard [@LA77] str. NP-hard str. NP-hard str. NP-hard
not appr. within for any $\alpha\in [0,1)$ not appr. within
$\frac{6}{5}-\epsilon$, $\epsilon>0$ $\frac{6}{5}-\epsilon$, $\epsilon>0$
: Complexity results for uncertain processing times (the due dates are deterministic).[]{data-label="tabs2"}
$\mathcal{P}$ $\textsc{Min-Exp}~\mathcal{P}$ $\textsc{Min-VaR}_\alpha~\mathcal{P}$ $\textsc{Min-CVaR}_\alpha~\mathcal{P}$ $\textsc{Min-Max}~\mathcal{P}$
------------------------- -------------------------------- --------------------------------------- -------------------------------------------------------------------------- -------------------------------- -- -- -- -- -- --
$1|prec|\max w_j T_j$ $O(f_{\max}^K Kn^2)$ $O(f_{\max}^K Kn^2)$ $O(f_{\max}^K Kn^2)$ $O(Kn^2)$ [@KZ16d]
FPTAS FPTAS FPTAS
for const. $K$ for const. $K$ for const. $K$
$1|prec|\sum w_j C_j$ as the determ. appr. within 2 appr. within 2 appr. within 2 [@MNO13]
problem for const. $K$
$1|prec^*|\sum w_j C_j$ poly sol. appr. within 2 appr. within appr. within 2 [@MNO13]
for const. $K$ $\min\{\frac{1}{1-\alpha},2\}$
$1|p_j=1|\sum w_j U_j$ poly sol. - appr. within appr. within $K$
$\min\{\frac{1}{{\rm Pr}_{\min}},\frac{1}{1-\alpha}\}$
$1||\sum w_j U_j$ appr. within - appr. within appr. within
determ. proc. times $4+\epsilon$, $\epsilon>0$ $\min\{\frac{4+\epsilon}{{\rm Pr}_{\min}},\frac{4+\epsilon}{1-\alpha}\}$ $(4+\epsilon)K$, $\epsilon>0$
$1|p_j=1|\sum w_j T_j$ poly sol. - appr. within appr. within $K$
$\min\{\frac{1}{{\rm Pr}_{\min}},\frac{1}{1-\alpha}\}$
$1||\sum w_j T_j$ appr. within - appr. within appr. within
determ. proc. times $4+\epsilon$, $\epsilon>0$ $\min\{\frac{4+\epsilon}{{\rm Pr}_{\min}},\frac{4+\epsilon}{1-\alpha}\}$ $(4+\epsilon)K$, $\epsilon>0$
: Positive complexity results.[]{data-label="tabs3"}
[$f_{\max}$ is an upper bound on the cost of any schedule under any scenario; $prec^*$ is a polynomially solvable structure of the precedence constraints; ${\rm Pr}_{\min}=\min_{k\in [K]} {\rm Pr}[\xi_k]$.]{}
Some general properties {#sec3}
=======================
In this section we will show some general relationships between the problems with various performance criteria. These properties will be used later to establish some positive and negative complexity results for particular problems.
The following statements are true:
1. If $\textsc{Min-Exp}~\mathcal{P}$ is approximable within $\sigma>1$ (for $\sigma=1$ it is polynomially solvable), then $\textsc{Min-CVaR}_\alpha~\mathcal{P}$ is approximable within $\sigma\rho$, where $\rho=
\min\{\frac{1}{{\rm Pr}_{\min}},\frac{1}{1-\alpha}\}$, for each constant $\alpha\in [0,1)$.
2. If $\textsc{Min-Exp}~\mathcal{P}$ with $K$-scenarios is NP-hard and hard to approximate within $\rho>1$, then $\textsc{Min-CVaR}_\alpha~\mathcal{P}$ with $K+1$ scenarios is also NP-hard and hard to approximate within $\rho$ for each constant $\alpha\in [0,1)$.
\[thmexp\]
We first prove assertion 1. Let $\pi^*$ minimize the expected cost and $\pi'$ minimize the conditional value at risk for a fixed $\alpha \in [0,1)$. We will denote by $\hat{\pi}$ a $ \sigma$-approximation schedule for $\textsc{Min-Exp}~\mathcal{P}$. Using Lemma \[lem01\] we get $${\bf CVaR}_\alpha [{\rm F}(\hat{\pi})]\leq
\rho {\bf E}[{\rm F}(\hat{\pi})]\leq \sigma\rho {\bf E}[{\rm F}(\pi^*)]
\leq \sigma\rho {\bf E}[{\rm F}(\pi')]\leq \sigma\rho {\bf CVaR}_\alpha [{\rm F}(\pi')],$$ and the assertion follows.
![Illustration of the proof of Theorem \[thmexp\].[]{data-label="fig2a"}](fig3.eps){width="6cm"}
In order to prove assertion 2, consider an instance of $\textsc{Min-Exp}~\mathcal{P}$ with $\mathcal{U}=\{\xi_1,\dots,\xi_K\}$. Fix $\alpha\in (0,1)$ (the statement trivially holds for $\alpha=0$) and add one additional scenario $\xi'$ under which the cost of each schedule is 0 (for example, all job processing times are 0 under $\xi'$). We fix ${\rm Pr}'[\xi']=\alpha$ and ${\rm Pr}'[\xi_i]={\rm Pr}[\xi_i]\cdot(1-\alpha)$ for each $i\in [K]$. Denote by ${\rm F}'(\pi)$ the random cost of $\pi$ under the new scenario set $\mathcal{U}'$. For each schedule $\pi$ we get (see Figure \[fig2a\]): $${\bf CVaR}_{\alpha}[{\rm F}'(\pi)]=\frac{1}{1-\alpha}\sum_{i\in [K]} {\rm Pr}'[\xi_i] f(\pi, \xi_i)=\sum_{i\in [K]}{\rm Pr}[\xi_i] f(\pi, \xi_i)={\bf E}[{\rm F}(\pi)].$$ Hence there is a cost preserving reduction from $\textsc{Min-Exp}~\mathcal{P}$ with $K$ scenarios to $\textsc{Min-CVaR}_\alpha~\mathcal{P}$ with $K+1$ scenarios and the theorem follows.
\[thm01\] Assume that $w_j=1$ for each job $j\in J$ in problem $\mathcal{P}$. If <span style="font-variant:small-caps;">Min-Max</span> $\mathcal{P}$ with $K\geq 2$ scenarios is NP-hard and hard to approximate within $\rho>1$, then
1. $\textsc{Min-VaR}_\alpha~\mathcal{P}$ with $K+1$ scenarios is NP-hard and hard to approximate within $\rho>1$ for each constant $\alpha\in (0,1)$.
2. $\textsc{Min-CVaR}_\alpha~\mathcal{P}$ with $K+1$ scenarios is NP-hard for each constant $\alpha\in (0,1)$.
Choose an instance of the $\textsc{Min-Max}~\mathcal{P}$ problem with $\mathcal{U}=\{\xi_1,\dots, \xi_K\}$, $K\geq 2$. Fix $\alpha\in (0,1)$ and create $\mathcal{U}'$ by adding to $\mathcal{U}$ a dummy scenario $\xi'$ such that the cost of each schedule under $\xi'$ equals $M$ and $M\geq f(\pi, \xi_i)$ for each $i\in [K]$ and each $\pi \in \Pi$. It is enough to fix $p_j(\xi')=p_{\max}$ and $d_j(\xi')=d_{\min}$ for each job $j\in J$, where $p_{\max}=\max_{j\in J, i\in [K]} p_j(\xi_i)$ is the maximum job processing time and $d_{\min}=\min_{j\in J, i\in [K]} d_j(\xi_i)$ is the minimum due date over all scenarios. For each of the two assertions, we define an appropriate probability distribution in $\mathcal{U}'$. We will use ${\rm F}'(\pi)$ to denote the random cost of $\pi$ under $\mathcal{U}'$.
In order to prove the statement 1, we fix ${\rm Pr}[\xi']=1-\alpha$ and ${\rm Pr}[\xi_i]=\frac{\alpha}{K}$ for each $i\in [K]$ (see Figure \[fig2\]a). The equality ${\bf VaR}_\alpha[{\rm F}'(\pi)]={\bf Max}[{\rm F}(\pi)]$ holds. Hence, there is a cost preserving reduction from $\textsc{Min-Max}~\mathcal{P}$ with $K$ scenarios to $\textsc{Min-VaR}_\alpha~\mathcal{P}$ with $K+1$ scenarios and the statement follows. To prove the statement 2, we fix ${\rm Pr}[\xi']=\gamma$ and ${\rm Pr}[\xi_i]=\beta$ for each $i\in [K]$, where $\gamma$ and $\beta$ satisfy the following system of equations (see Figure \[fig2\]b):
$$\left\{
\begin{array}{ll}
\beta+\gamma=1-\alpha \\
K\beta+\gamma=1
\end{array}\right.$$ In consequence $\beta=\frac{\alpha}{K-1}$ and $\gamma=1-\frac{K\alpha}{K-1}$. Observe that $\beta>1$ as $\alpha\in (0,1)$.
![Illustration of the proof of Theorem \[thm01\].[]{data-label="fig2"}](fig2.eps){width="14cm"}
For each schedule $\pi$ we get $${\bf CVaR}_{\alpha}[{\rm F}'(\pi)]=\frac{1}{1-\alpha}(\beta \cdot {\bf Max}[\mathrm{F}(\pi)] + \gamma M),$$ where $\beta, \gamma$ and $M$ are numbers depending on $K$ and $\alpha$. Hence $\textsc{Min-Max}~\mathcal{P}$ and the corresponding instance of $\textsc{Min-CVaR}_\alpha~\mathcal{P}$ have the same optimal solutions and the theorem follows.
Negative complexity results {#sec4}
===========================
In this section we will prove some negative complexity results for basic single machine scheduling problems. These results are summarized in Tables \[tabs1\] and \[tabs2\].
Uncertain due dates
-------------------
We first address the problem of minimizing the value at risk criterion. The following theorem characterizes the complexity of some basic problems:
\[thm1\] For each $\alpha\in (0,1)$, $\textsc{Min-Var}_\alpha~\mathcal{P}$ is strongly NP-hard and not at all approximable, when $\mathcal{P} \in \{1|p_j=1|T_{\max},\,
1|p_j=1|\sum T_j, \,1|p_j=1| \sum U_j\}$.
Consider an instance of the following strongly NP-hard <span style="font-variant:small-caps;">Min 3-Sat</span> problem [@KM94; @AZ02]. We are given boolean variables $x_1,\dots, x_n$, a collection of clauses $\mathcal{C}_1,\dots \mathcal{C}_m$, where each clause is a disjunction of at most $3$ literals (variables or their negations) and we ask if there is an assignment to the variables which satisfies at most $L < m$ clauses. We can ask equivalently, if there is an assignment to the variables for which at least $l=m-L$ clauses are not satisfied.
Given an instance of <span style="font-variant:small-caps;">Min 3-Sat</span>, we create two jobs $J_{x_i}$ and $J_{\overline{x}_i}$ for each variable $x_i$, $i\in [n]$. A *due date scenario* $\xi_i$ corresponds to clause $\mathcal{C}_i=(l_1 \vee l_2 \vee l_3)$ and is formed as follows. For each $q=1,2,3$, if $l_q=x_j$, then the due date of $J_{x_j}$ is $2j-1$ and the due date of $J_{\overline{x}_j}$ is $2j$; if $l_q=\overline{x}_j$, then the due date of $J_{x_j}$ is $2j$ and the due date of $J_{\overline{x}_j}$ is $2j-1$; if neither $x_j$ nor $\overline{x}_j$ appears in $\mathcal{C}_i$, then the due dates of $J_{x_j}$ and $J_{\overline{x}_j}$ are set to $2j$. An example is shown in Table \[tab1\].
$\xi_1$ $\xi_2$ $\xi_3$ $\xi_4$ $\xi_5$
---------------------- --------- --------- --------- --------- ---------
$J_{x_1}$ 1 2 2 1 1
$J_{\overline{x}_1}$ 2 2 1 2 2
$J_{x_2}$ 4 4 3 3 4
$J_{\overline{x}_2}$ 3 3 4 4 4
$J_{x_3}$ 6 6 6 5 5
$J_{\overline{x}_3}$ 5 5 6 6 6
$J_{x_4}$ 8 7 8 8 8
$J_{\overline{x}_4}$ 8 8 7 8 7
: The set of jobs and the due date scenarios for the formula $(x_1\vee \overline{x}_2 \vee \overline{x}_3)\wedge (\overline{x}_2 \vee \overline{x}_3 \vee x_4) \wedge (\overline{x}_1 \vee x_2 \vee \overline{x}_4) \wedge (x_1 \vee x_2 \vee x_3) \wedge (x_1 \vee x_3 \vee \overline{x}_4)$.[]{data-label="tab1"}
Let us define a subset of the schedules $\Pi'\subseteq \Pi$ such that each schedule $\pi\in \Pi'$ is of the form $\pi=(J_1,J_1',J_2,J_2',\dots,J_n,J_n')$, where $J_j,J_j'\in\{J_{x_j},J_{\overline{x}_j}\}$ for $j\in [n]$. Observe that $\Pi'$ contains exactly $2^n$ schedules and each such a schedule corresponds to the assignment to the variables such that $x_j=0$ if $J_{x_j}$ is processed before $J_{\overline{x}_j}$ and $x_j=1$ otherwise. Note that this correspondence is one-to-one. In the following we assume that $f(\pi, \xi_i)$ is the maximum tardiness, or the total tardiness, or the sum of unit penalties in $\pi$ under $\xi_i$. The reasoning will be the same for each of these cost functions. If $\pi \notin \Pi'$, then $f(\pi, \xi_i)>0$ for each scenario $\xi_i$. Indeed, suppose that $\pi \notin \Pi'$ and let $J_j$ ($J_j')$ be the last job in $\pi$ which is not placed properly, i.e. $J_j,(J_j')\notin\{J_{x_j},J_{\overline{x}_j}\}$. Then $J_j$ ($J_j'$) is late under all scenarios. On the other hand, if $\pi \in \Pi'$, then the number of scenarios under which no job is late is equal to the number of unsatisfiable clauses for the assignment corresponding to $\pi$. Fix $\alpha\in (0,1)$. We will add to $\mathcal{U}$ one additional scenario $\xi'$ and define a probability distribution in $\mathcal{U}$, depending on the fixed $\alpha$, so that the answer to <span style="font-variant:small-caps;">Min 3-Sat</span> is yes if and only if there is schedule $\pi$ for which $\mathbf{VaR}_{\alpha}[\mathrm{F}(\pi)]\leq 0$. This will prove the stated result. We consider two cases:
1. $l/m\geq \alpha$. We create dummy scenario $\xi'$ under which the due date of all jobs is equal to 0. The probability of this scenario is equal to $\frac{l-\alpha m}{l}$. The probability of each of the remaining scenarios is equal to $\frac{1}{m}(1-\frac{l-\alpha m}{l})=\frac{\alpha}{l}$. Assume that the answer to <span style="font-variant:small-caps;">Min 3-Sat</span> is yes. So, there is an assignment to the variables which satisfies at most $m-l$ clauses. By the above construction, there is a schedule $\pi\in \Pi'$ whose cost is positive under at most $m-l$ scenarios plus the dummy one. It holds $$\mathrm{Pr}[\mathrm{F}(\pi)>0]\leq \frac{l-\alpha m}{l} + (m-l)\frac{\alpha}{l}=1-\alpha.$$ Hence $\mathrm{Pr}[\mathrm{F}(\pi)\leq 0]\geq \alpha$ and $\mathbf{VaR}_{\alpha}[\mathrm{F}(\pi)]\leq 0$. Assume that the answer to <span style="font-variant:small-caps;">Min 3-Sat</span> is no. Then, for every schedule $\pi$ there are more than $m-l$ scenarios under which the cost of $\pi$ is positive plus the dummy one. Hence $\mathrm{Pr}[\mathrm{F}(\pi)>0]> (1-\alpha)$ and $\mathrm{Pr}[\mathrm{F}(\pi)\leq 0]<\alpha$. In consequence, $\mathbf{VaR}_{\alpha}[\mathrm{F}(\pi)]>0$.
2. $l/m<\alpha$. We create dummy scenario $\xi'$ under which the due date of each job equals $2n$. The probability of the dummy scenario is $\frac{m\alpha-l}{m-l}$. The probability of each of the remaining scenarios is equal to $\frac{1}{m}(1-\frac{m\alpha-l}{m-l})=\frac{1-\alpha}{m-l}$. Assume that the answer to <span style="font-variant:small-caps;">Min 3-Sat</span> is yes. So, there is an assignment to the variables which satisfies at most $m-l$ clauses. By the construction, there is a schedule $\pi$ whose cost is positive under at most $m-l$ scenarios. Hence $$\mathrm{Pr}[\mathrm{F}(\pi)\leq 0]=1-\mathrm{Pr}[\mathrm{F}(\pi)>0]\geq 1-(m-l)\frac{1-\alpha}{m-l}=\alpha$$ and $\mathbf{VaR}_{\alpha}[\mathrm{F}(\pi)]\leq 0$. Assume that the answer to <span style="font-variant:small-caps;">Min 3-Sat</span> is no. Then for each assignment more than $m-l$ clauses are satisfied. By the construction, for every schedule $\pi$ there are more than $m-l$ scenarios under which the cost $\pi$ is positive. Therefore $\mathrm{Pr}[\mathrm{F}(\pi)>0]>(m-l)\frac{1-\alpha}{m-l}=(1-\alpha)$ and $\mathrm{Pr}[\mathrm{F}(\pi)\leq 0]<\alpha$, so $\mathbf{VaR}_{\alpha}[\mathrm{F}(\pi)]>0$.
It follows from Theorem \[thm1\] that the problem discussed in [@NBN17] is strongly NP-hard and not at all approximable even in the very restrictive case in which all job processing times are equal to 1. It was shown in [@KZ16d] that $\textsc{Min-Exp}~1|p_j=1|T_{\max}$ is strongly NP-hard and hard to approximate within $7/6-\epsilon$ for any $\epsilon>0$. Hence, we immediately get from Theorem \[thmexp\] that for each constant $\alpha\in [0,1)$, $\textsc{Min-CVaR}_\alpha~1|p_j=1|T_{\max}$ is strongly NP-hard and hard to approximate within $7/6-\epsilon$ for any $\epsilon>0$.
We consider now the problem with the total tardiness criterion. The deterministic $1||\sum T_j$ problem is known to be NP-hard [@LA77]. However, $1|p_j=1|\sum T_j$ is polynomially solvable(see, e.g., [@B07]). The following result characterizes the complexity of the minmax version of this problem:
<span style="font-variant:small-caps;">Min-Max</span> $1|p_j=1|\sum T_j$ is strongly NP-hard and not approximable within $\frac{5}{4}-\epsilon$ for any $\epsilon>0$. \[tmm\]
We will show a reduction from the strongly NP-complete <span style="font-variant:small-caps;">3-Sat</span> problem, in which we are given boolean variables $x_1,\dots, x_n$, a collection of clauses $\mathcal{C}_1,\dots \mathcal{C}_m$, where each clause is a disjunction of at most $3$ literals (variables or their negations) and we ask if there is an assignment to the variables which satisfies all clauses (see, e.g., [@GJ79]). Given an instance of <span style="font-variant:small-caps;">3-Sat</span>, we create two jobs $J_{x_j}$ and $J_{\overline{x}_j}$ for each variable $x_j$, $j\in [n]$, $|J|=2n$. A due date scenario $\xi_i$ corresponding to clause $\mathcal{C}_i=(l_1 \vee l_2 \vee l_3)$ is created in the same way as in the proof of Theorem \[thm1\]. Additionally, for each variable $x_j$ we create scenario $\xi_j'$ under which the due dates of $J_{x_j}$ and $J_{\overline{x}_j}$ are $2(j-1)+\frac{1}{2}$ and the due dates of the remaining jobs are set to $2n$ (see Table \[tab1a\]). We first show that the answer to <span style="font-variant:small-caps;">3-Sat</span> is yes if and only if there is a schedule $\pi$ such that $\max_{\xi\in \mathcal{U}} \sum_{j\in J} T_j(\pi, \xi)\leq 2$.
$\xi_1$ $\xi_2$ $\xi_3$ $\xi_4$ $\xi_5$ $\xi'_1$ $\xi'_2$ $\xi'_3$ $\xi'_4$
---------------------- --------- --------- --------- --------- --------- --------------- ----------------- ----------------- ----------------- --
$J_{x_1}$ 1 2 2 1 1 $\frac{1}{2}$ 8 8 8
$J_{\overline{x}_1}$ 2 2 1 2 2 $\frac{1}{2}$ 8 8 8
$J_{x_2}$ 4 4 3 3 4 8 $2+\frac{1}{2}$ 8 8
$J_{\overline{x}_2}$ 3 3 4 4 4 8 $2+\frac{1}{2}$ 8 8
$J_{x_3}$ 6 6 6 5 5 8 8 $4+\frac{1}{2}$ 8
$J_{\overline{x}_3}$ 5 5 6 6 6 8 8 $4+\frac{1}{2}$ 8
$J_{x_4}$ 8 7 8 8 8 8 8 8 $6+\frac{1}{2}$
$J_{\overline{x}_4}$ 8 8 7 8 7 8 8 8 $6+\frac{1}{2}$
: The set of jobs and the due date scenarios for the formula $(x_1\vee \overline{x}_2 \vee \overline{x}_3)\wedge (\overline{x}_2 \vee \overline{x}_3 \vee x_4) \wedge (\overline{x}_1 \vee x_2 \vee \overline{x}_4) \wedge (x_1 \vee x_2 \vee x_3) \wedge (x_1 \vee x_3 \vee \overline{x}_4)$. Schedule $\pi=(J_{x_1},J_{\overline{x}_1}, J_{\overline{x}_2}, J_{x_2}, J_{x_3}, J_{\overline{x}_3}, J_{\overline{x}_4}, J_{x_4}))$ corresponds to a truth assignment.[]{data-label="tab1a"}
Assume that the answer to <span style="font-variant:small-caps;">3-Sat</span> is yes. Consider schedule $\pi=(J_1,J'_1,J_2,J'_2,\dots,J_n, J'_n)$, where $J_j,J'_j\in\{J_{x_j}, J_{\overline{x}_j}\}$. Furthermore $J_{x_j}$ is processed before $J_{\overline{x}_j}$ if and only if $x_j=1$. Since in every clause at least one literal is true, at most two jobs in $\pi$ are late under each scenario $\xi_i\in \mathcal{U}$. The tardiness of each job in $\pi$ under any $\xi_i\in \mathcal{U}$ is at most 1. Furthermore, the total tardiness in $\pi$ under any $\xi'_j$ is exactly 2. In consequence, $\max_{\xi\in \mathcal{U}} \sum_{j\in J} T_j(\pi, \xi)\leq 2$.
Assume that there is a schedule $\pi$, such that $\max_{\xi\in \mathcal{U}} \sum_{j\in J} T_j(\pi, \xi)\leq 2$. We claim that $\pi=(J_1,J'_1,J_2,J'_2,\dots,J_n, J'_n)$, where $J_j,J'_j\in\{J_{x_j}, J_{\overline{x}_j}\}$. Suppose that this is not the case, and let $J_k$ ($J'_k$) be the last job in $\pi$ which is not placed properly. The completion time of $J_k$ ($J'_k$) is at least $2k+1$. So, its tardiness under $\xi'_k$ is at least $2k+1-(2k-2+\frac{1}{2})=2.5$. Let $x_j=1$ if and only if $J_{x_j}$ is processed before $J_{\overline{x}_j}$ in $\pi$. Since only two jobs can be late under any $\xi_i$, this assignment satisfies all clauses and the answer to <span style="font-variant:small-caps;">3-Sat</span> is yes.
In order to prove the lower approximation bound, it is enough to observe that if the answer to <span style="font-variant:small-caps;">3-Sat</span> is no, then each schedule has the total tardiness 3 under some scenario $\xi_i$ or 2.5 under some scenario $\xi'_j$, which gives a gap at least $\frac{5}{4}$.
From the fact that $1||\sum T_j$ is weakly NP-hard (see [@DL90]), we get immediately that more general $\textsc{Min-Exp}~1||\sum T_j$ problem is weakly NP-hard as well. The next theorem strengthens this result.
\[thmsT\] $\textsc{Min-Exp}~1||\sum T_j$ is strongly NP-hard.
We will show a polynomial time reduction from the deterministic $1||\sum w_j T_j$ problem, which is known to be strongly NP-hard [@LA77]. Consider an instance of $1||\sum w_j T_j$. Let $W=\sum_{j\in J} w_j>0$ and $P=\sum_{j\in J} p_j$. We build an instance of $\textsc{Min-Exp}~1||\sum T_j$ with the same set of jobs $J$ and job processing times $p_j$, $j\in J$. We create $K=|J|=n$ due date scenarios as follows. Under scenario $\xi_j$, $j\in [n]$, job $j$ has due date equal to $d_j$ and all the remaining jobs have due dates equal to $P$. We also fix ${\Pr}[\xi_i]=w_i/W$, $i\in [n]$. For any schedule $\pi$, we get ${\bf E}[{\rm F}(\pi)]=\sum_{i\in [K]} {\rm Pr}[\xi_i] \sum_{j\in J} T_j(\pi, \xi_i)=\frac{1}{W}\sum_{i\in [n]} w_i \sum_{j\in J} T_j(\pi, \xi_i)$. By the construction, we get $\sum_{j\in J} T_j(\pi, \xi_i)=[C_i(\pi)-d_i]^+$, so ${\bf E}[{\rm F}(\pi)]=\frac{1}{W}\sum_{i\in [n]} w_i [C_i(\pi)-d_i]^+$. In consequence $1||\sum w_j T_j$ and $\textsc{Min-Exp}~1||\sum T_j$ have the same optimal solutions and the theorem follows.
It was shown in [@AAK11] that <span style="font-variant:small-caps;">Min-Max</span> $1|p_j=1|\sum U_j$ with uncertain due dates is strongly NP-hard. The following theorem strengthens this result:
<span style="font-variant:small-caps;">Min-Max</span> $1|p_j=1|\sum U_j$ is not approximable within any constant factor unless P=NP. \[tnapr\]
Consider the following <span style="font-variant:small-caps;">Min-Max 0-1 Selection</span> problem. We are given a set of items $E=\{e_1,e_2,\dots,e_n\}$ and an integer $q\in [n]$. For each item $e_j$, $j\in [n]$, there is a cost $c_j(\xi_i)\in\{0,1\}$ under scenario $\xi_i$, $i\in [K]$. We seek a selection $X\subseteq E$ of exactly $q$ items, $|X|=q$, which minimizes the maximum cost over all scenarios, i.e. the value of $\max_{i\in [K]} \sum_{e_i\in X} c_j(\xi_i)$. This problem was discussed in [@KKZ13], where it was shown that it is not approximable within any constant factor $\gamma\geq 1$. We will show that there is a cost preserving reduction from <span style="font-variant:small-caps;">Min-Max 0-1 Selection</span> to the considered scheduling problem, which will imply the stated result.
Given an instance of <span style="font-variant:small-caps;">Min-Max 0-1 Selection</span>, we build the corresponding instance of <span style="font-variant:small-caps;">Min-Max</span> $1|p_j=1|\sum U_j$ as follows. We create a set of jobs $J=E$, $|J|=n$, with deterministic unit processing times. For each $i\in [K]$, if $c_j(\xi_i)=1$ then $d_j(\xi'_i)=n-q$, and if $c_j(\xi_i)=0$, then $d_j(\xi'_i)=n$. So, we create $K$ due date scenarios that correspond to the cost scenarios of <span style="font-variant:small-caps;">Min-Max 0-1 Selection</span>.
Suppose that there is a solution $X$ to <span style="font-variant:small-caps;">Min-Max 0-1 Selection</span> such that $\sum_{e_i\in X} c_j(\xi_i)\leq C$ for each $i\in [K]$. Hence $X$ contains at most $C$ items, $C\leq q$, with the cost equal to 1 under each scenario. In the corresponding schedule $\pi$, we first process $n-q$ jobs from $J\setminus X$ and then the jobs in $X$ in any order. It is easily seen that there are at most $C$ late jobs in $\pi$ under each scenario $\xi'_i$, hence the maximum cost of schedule $\pi$ is at most $C$. Conversely, let $\pi$ be a schedule in which there are at most $C$ late jobs under each scenario $\xi'_i$. Clearly $C\leq q$ since the first $n-q$ jobs in $\pi$ must be on-time in all scenarios. Let us form solution $X$ by choosing the items corresponding to the last $q$ jobs in $\pi$. Among these jobs at most $C$ are late under each scenario, hence the cost of $X$ is at most $C$ under each scenario $\xi_i$.
Thus, by Theorem \[thm01\], $\textsc{Min-CVaR}_\alpha~1|p_j=1|\sum U_j$ is strongly NP-hard for any $\alpha\in (0,1)$ (notice that $p_{\max}=1$ in the proof of Theorem \[thm01\] and in the new scenario set $\mathcal{U}'$ still only due dates are uncertain).
\[thmsU\] $\textsc{Min-Exp}~1||\sum U_j$ is NP-hard.
Choose the deterministic $1||\sum w_j U_j$ problem, which is known to be NP-hard [@KR74]. The reduction from this problem to $\textsc{Min-Exp}~1||\sum U_j$ is the same as the one in the proof of Theorem \[thmsT\].
It is worth noting that in the proof of Theorem \[thmsU\] we require an arbitrary probability distribution in the scenario set and we have shown that the problem is only weakly NP-hard. Its complexity for uniform probability distribution is open.
Uncertain processing times
--------------------------
In this section we characterize the complexity of the problems under consideration when only processing times are uncertain. It has been shown in [@KY97] that $\textsc{Min-Max}~1||\sum C_j$ is strongly NP-hard. Furthermore, this problem is also hard to approximate within $\frac{6}{5}-\epsilon$ for any $\epsilon>0$ [@MNO13]. Using Theorem \[thm01\], we can immediately conclude that the same negative result holds for $\textsc{Min-VaR}_{\alpha}~1||\sum C_j$ for any $\alpha\in (0,1]$. Also, strong NP-hardness of the min-max problem implies that $\textsc{Min-CVaR}_{\alpha}~1||\sum C_j$ is strongly NP-hard for each fixed $\alpha \in (0,1)$. Observe that the boundary case $\alpha=0$ (i.e. $\textsc{Min-Exp}~1||\sum C_j$) is polynomially solvable, as it easily reduces to the deterministic $1||\sum C_j$ problem. Since $1||\sum T_j$ is a special case of $1||\sum C_j$, with $d_j=0$ for each $j\in J$, the same negative results are true for the problem with the total tardiness criterion. Observe, however that $\textsc{Min-Exp}~1||\sum T_j$ is also NP-hard, since the deterministic $1||\sum T_j$ problem is known to be weakly NP-hard [@DL90].
It has been shown in [@AC08] that <span style="font-variant:small-caps;">Min-Max</span> $1||\sum U_j$ with uncertain processing times and deterministic due dates is NP-hard. The following theorem strengthens this result:
\[thmsUp\] <span style="font-variant:small-caps;">Min-Max</span> $1||\sum U_j$ is strongly NP-hard. This assertion remains true even when all the jobs have a common deterministic due date.
We show a polynomial time reduction from the <span style="font-variant:small-caps;">3-Sat</span> problem (see the proof of Theorem \[tmm\]). Given an instance of <span style="font-variant:small-caps;">3-Sat</span>, we create an instance of <span style="font-variant:small-caps;">Min-Max</span> $1||\sum U_j$ in the following way. For each variable $x_i$ we create two jobs $J_{x_i}$ and $J_{\overline{x}_i}$, so $J$ contains $2n$ jobs. The due dates of all these jobs are the same under each scenario and equal 2. For each clause $C_j=(l_1,l_2,l_3)$ we construct processing time scenario $\xi_i$, under which the jobs $J_{\overline{l}_1}, J_{\overline{l}_2}, J_{\overline{l}_3}$ have processing time equal to 1 and all the remaining jobs have processing times equal to 0. Then, for each pair of jobs $J_{x_i}, J_{\overline{x}_i}$ we construct scenario $\xi'_i$ under which the processing times of $J_{x_i}, J_{\overline{x}_i}$ are 2 and all the remaining jobs have processing times equal to 0. A sample reduction is shown in Table \[tab1b\]. We will show that the answer to <span style="font-variant:small-caps;">3-Sat</span> is yes if and only if there is a schedule $\pi$ such that $\max_{\xi \in \mathcal{U}} \sum_{j\in J} U(\pi,\xi)\leq n$.
$\xi_1$ $\xi_2$ $\xi_3$ $\xi_4$ $\xi_5$ $\xi'_1$ $\xi'_2$ $\xi'_3$ $\xi'_4$ $d_i$
---------------------- --------- --------- --------- --------- --------- ---------- ---------- ---------- ---------- -------
$J_{x_1}$ 0 0 1 0 0 2 0 0 0 2
$J_{\overline{x}_1}$ 1 0 0 1 1 2 0 0 0 2
$J_{x_2}$ 1 1 0 0 0 0 2 0 0 2
$J_{\overline{x}_2}$ 0 0 1 1 0 0 2 0 0 2
$J_{x_3}$ 1 1 0 0 0 0 0 2 0 2
$J_{\overline{x}_3}$ 0 0 0 1 1 0 0 2 0 2
$J_{x_4}$ 0 0 1 0 1 0 0 0 2 2
$J_{\overline{x}_4}$ 0 1 0 0 0 0 0 0 2 2
: Processing time scenarios for the formula $(x_1\vee \overline{x}_2 \vee \overline{x}_3)\wedge (\overline{x}_2 \vee \overline{x}_3 \vee x_4) \wedge (\overline{x}_1 \vee x_2 \vee \overline{x}_4) \wedge (x_1 \vee x_2 \vee x_3) \wedge (x_1 \vee x_3 \vee \overline{x}_4)$. Schedule $\pi=(J_{x_1},J_{\overline{x}_2},J_{x_3},J_{\overline{x}_4} | J_{\overline{x}_1}, J_{x_2}, J_{\overline{x}_3}, J_{x_4})$ corresponds to a satisfying truth assignment.[]{data-label="tab1b"}
Assume that the answer to <span style="font-variant:small-caps;">3-Sat</span> is yes. Then there exists a truth assignment to the variables which satisfies all the clauses. Let us form schedule $\pi$ by processing first the jobs corresponding to true literals in any order and processing then the remaining jobs in any order. From the construction of the scenario set it follows that the completion time of the $n$th job in $\pi$ under each scenario is not greater than 2. In consequence, at most $n$ jobs in $\pi$ are late under each scenario and $\max_{\xi \in \mathcal{U}} \sum_{j\in J} U(\pi,\xi)\leq n$.
Assume that there is a schedule $\pi$ such that $\sum_{j\in J} U(\pi,\xi)\leq n$ for each $\xi \in \mathcal{U}$, which means that at most $n$ jobs in $\pi$ are late under each scenario. Observe first that $J_{x_i}$ and $J_{\overline{x}_i}$ cannot appear among the first $n$ jobs in $\pi$ for any $i\in [n]$; otherwise more than $n$ jobs would be late in $\pi$ under $\xi'_i$. Hence the first $n$ jobs in $\pi$ correspond to a truth assignment to the variables $x_1,\dots,x_n$, i.e. when $J_l$ is among the first $n$ jobs, then the literal $l$ is true. Since $f(\pi,\xi_i)\leq n$, the completion time of the $n$-th job in $\pi$ under $\xi_i$ is not greater than 2. We conclude that at most two jobs among the first $n$ job have processing time equal to 1 under $\xi_i$, so there are at most two false literals for each clause and the answer to <span style="font-variant:small-caps;">3-Sat</span> is yes.
We thus get from Theorem \[thmsUp\] that $\textsc{Min-VaR}_{\alpha}~1||\sum U_j$ is strongly NP-hard for any $\alpha\in (0,1)$ and $\textsc{Min-CVaR}_{\alpha}~1||\sum U_j$ is strongly NP-hard for any $\alpha\in (0,1]$. The boundary case with $\alpha=0$ (i.e. <span style="font-variant:small-caps;">Min-Exp</span> $1||\sum U_j$ with uncertain processing times) is an interesting open problem.
Positive complexity results {#sec5}
===========================
In this section we establish some positive complexity results. Namely, we provide several polynomial and approximation algorithms for particular problems. A summary of the results can be found in Table \[tabs3\].
Problems with uncertain due dates
---------------------------------
Consider the $\textsc{Min-Exp}~1|p_j=1|\sum w_jU_j$ problem with uncertain due dates. We introduce variables $x_{ij}\in \{0,1\}$, $i\in [n]$, $j\in [n]$, where $x_{ij}=1$ if $j\in [n]$ is the $i$th job in the schedule constructed. The variables satisfy the assignment constraints, i.e. $\sum_{i\in [n]} x_{ij}=1$ for each $j\in [n]$ and $\sum_{j\in [n]} x_{ij}=1$ for each $i\in [n]$. If $x_{ij}=1$, then the completion time of job $j$ equals $i$. Define $c_{ijk}=w_j$ if $i>d_j(\xi_k)$ and $c_{ijk}=0$ otherwise, for each $i,j\in [n]$ and $k\in [K]$. If the variables $x_{ij}$ describe $\pi$, then $${\bf E}[{\rm F}(\pi)]=\sum_{k\in [K]} \sum_{i\in [n]}\sum_{j\in [n]}{\rm Pr}[\xi_k] c_{ijk}x_{ij}=\sum_{i\in [n]}\sum_{j\in [n]}c^*_{ij} x_{ij},$$ where $c^*_{ij}=\sum_{k\in [K]} {\rm Pr}[\xi_k] c_{ijk}$. Hence the problem is equivalent to the <span style="font-variant:small-caps;">Minimum Assignment</span> with the cost matrix $c^*_{ij}$. The same result holds for $\textsc{Min-Exp}~1|p_j=1|\sum w_j T_j$. It is enough to define $c_{ijk}=w_j[i-d_j(\xi_k)]^+$ for $i,j\in [n]$, $k\in [K]$. We thus get the following results:
$\textsc{Min-Exp}~\mathcal{P}$ is polynomially solvable, when $\mathcal{P}\in\{1|p_j=1|\sum w_j U_j,\, 1|p_j=1|\sum w_j T_j\}$. \[texppol\]
From Theorems \[texppol\] and \[thmexp\], we immediately get the following approximation result:
$\textsc{Min-CVaR}_{\alpha}~\mathcal{P}$ is approximable within $\rho=\min\{\frac{1}{{\rm Pr}_{\min}},\frac{1}{1-\alpha}\}$, when $\mathcal{P}\in\{1|p_j=1|\sum w_j U_j, \,1|p_j=1|\sum w_j T_j\}$.
Since $\textsc{Min-Max}~1|p_j=1|\sum w_jU_j$ and $\textsc{Min-Max}~1|p_j=1|\sum w_jT_j$ are special cases of the min-max version of <span style="font-variant:small-caps;">Minimum Assignment</span>, which is approximable within $K$ (see, e.g., [@ABV09]), both problems are approximable within $K$ as well.
We now study the $\textsc{Min-Exp}~1|| \sum w_j T_j$ problem with uncertain due dates and deterministic processing times. This problem is strongly NP-hard since $1||\sum w_j T_j$ is strongly NP-hard. The expected cost of $\pi$ can be rewritten as ${\bf E}[{\rm F}(\pi)]=\sum_{j\in J} \sum_{i\in [K]} {\rm Pr}[\xi_i] [C_j(\pi)-d_j(\xi_i)]^+$. We thus get a single machine scheduling problem $1||\sum f_j$ with job-dependent cost functions of form $f_j(C_j(\pi))= \sum_{i\in [K]} {\rm Pr}[\xi_i] [C_j(\pi)-d_j(\xi_i)]^+$, $j\in J$. Note also that these functions are nonnegative and nondecreasing with respect to $C_j(\pi)$. The same analysis can be done for the $\textsc{Min-Exp}~1|| \sum w_j U_j$ problem with uncertain due dates and deterministic processing times. Hence and from [@CMSV17], where a $(4+\epsilon)$-approximation algorithm, for any $\epsilon>0$, for this class of problems was provided, we get the following result (see also Theorem \[thmexp\]):
\[tpdexp\] If $\mathcal{P}\in \{1||\sum w_j U_j,\, 1||\sum w_j T_j\}$, then $\textsc{Min-Exp}$ $\mathcal{P}$ is approximable within $4+\epsilon$ and $\textsc{Min-CVaR}_\alpha~\mathcal{P}$ is approximable within $\min\{\frac{4+\epsilon}{{\rm Pr}_{\min}},\frac{4+\epsilon}{1-\alpha}\}$, $\epsilon>0$, for any $\epsilon>0$ and each constant $\alpha\in [0,1)$
When the probability distribution in $\mathcal{U}$ is uniform, then the approximation ratio in Theorem \[thmexp\] can be improved to $\min\{(4+\epsilon)K, \frac{4+\epsilon}{1-\alpha}\}$. Since <span style="font-variant:small-caps;">Min-Max</span> $\mathcal{P}$ is a special case of $\textsc{Min-CVar}_{\alpha}~\mathcal{P}$ with uniform probability distribution and $\alpha$ sufficiently large, we get that $\textsc{Min-Max}$ $\mathcal{P}$, $\mathcal{P}\in \{1||\sum w_j U_j,\, 1||\sum w_j T_j\}$, is approximable within $(4+\epsilon)K$ for any $\epsilon>0$.
The total weighted flow time criterion
--------------------------------------
In this section we focus on the problems with the total weighted flow time criterion. We start by recalling a well known property (see, e.g., [@MNO13]), which states that every such a problem with uncertain processing times and deterministic weights can be transformed into an equivalent problem with uncertain weights and deterministic processing times. This transformation goes as follows. For each processing time scenario $\xi_i$, $i\in [K]$, we invert the role of processing times and weights obtaining the weight scenario $\xi'_i$. Formally, $p_j=w_j$ and $w_j(\xi'_i)=p_j(\xi_i)$ for each $i\in [K]$. The new scenario set $\mathcal{U}'$ contains scenario $\xi'_i$ with ${\rm Pr}[\xi_i']={\rm Pr}[\xi_i]$ for each $i\in [K]$. We also invert the precedence constraints, i.e. if $i\rightarrow j$ in the original problem, then $j\rightarrow i$ in the new one. Given a feasible schedule $\pi=(\pi(1),\dots,\pi(n))$, let $\pi'=(\pi(n),\dots,\pi(1))$ be the corresponding inverted schedule. Of course, schedule $\pi'$ is feasible for the inverted precedence constraints. It is easy to verify that $f(\pi, \xi_i)=f(\pi',\xi'_i)$ for each $i\in [K]$. In consequence ${\bf CVaR}_\alpha[{\rm F}(\pi)]={\bf CVar}_\alpha[{\rm F'}(\pi')]$ and ${\bf VaR}_\alpha[{\rm F}(\pi)]={\bf VaR}_\alpha[{\rm F'}(\pi')]$, where ${\rm F'}(\pi')$ is the random cost of $\pi'$ for scenario set $\mathcal{U}'$. Hence, the original problem with uncertain processing times and the new one with uncertain weights have the optimal solutions with the same performance measure.
From now on we make the assumption that the jobs have deterministic processing times $p_j$, $j\in J$ and $w_j(\xi_i)$ is the weight of job $j$ under scenario $\xi_i$, $i\in [K]$. The value of ${\bf CVaR}_\alpha[{\rm F}(\pi)]$, for a fixed schedule $\pi$, can be computed by solving the following optimization problem (see the formulation (\[exVC1\])b): $$\label{exWC1}
\begin{array}{llll}
\min & \displaystyle \gamma + \frac{1}{1-\alpha}\sum_{i\in [K]} {\rm Pr}[\xi_k] u_k\\
\text{s.t.}& \gamma + u_k \geq \displaystyle \sum_{j\in J}w_j(\xi_k) C_j(\pi) & k\in [K]\\
& u_k\geq 0 & k\in [K]
\end{array}$$
Let $\delta_{ij}\in \{0,1\}$, $i,j\in [n]$, be binary variables such that $\delta_{ij}=1$ if job $i$ is processed before job $j$ in a schedule constructed. The vectors of all feasible job completion times $(C_1,\dots, C_n)$ can be described by the following system of constraints [@PO80]: $$\label{cCC2}
\begin {array}{llll}
VC: & C_j=p_j+\sum_{i\in J\setminus\{j\}} \delta_{ij} p_i & j\in J\\
&\delta_{ij}+\delta_{ji}=1 & i,j\in J, i\neq j \\
&\delta_{ij}+\delta_{jk}+\delta_{ki} \geq 1 & i,j,k \in J\\
&\delta_{ij}=1 & i\rightarrow j\\
&\delta_{ij}\in \{0,1\}& i,j \in J
\end{array}$$ Let us denote by $VC'$ the relaxation of $VC$, in which the constraints $\delta_{ij}\in \{0,1\}$ are relaxed with $0\leq \delta_{ij}\leq 1$. It has been proved in [@SH96b; @HA97] that each vector $(C_1,\dots, C_n)$ that satisfies $VC'$ also satisfies the following inequalities: $$\label{Schin}
\sum_{j\in I} p_jC_j\geq \frac{1}{2}\left((\sum_{j\in I} p_j)^2+\sum_{j\in I} p_j^2\right) \text{ for all }
I \subseteq J.$$ The formulations (\[cCC2\]) and (\[exWC1\]) lead to the following mixed integer programming model for $\textsc{Min-CVaR}_{\alpha}~1|prec|\sum w_j C_j$ with uncertain weights: $$\label{exWC3}
\begin{array}{llll}
\min & \displaystyle \gamma + \frac{1}{1-\alpha}\sum_{i\in [K]} {\rm Pr}[\xi_k] u_k\\
\text{s.t.}& \gamma + u_k \geq \sum_{j\in J}w_j(\xi_k) C_j & k\in [K]\\
& \text{Constraints VC} \\
& u_k\geq 0 & k\in [K]
\end{array}$$
We now solve the relaxation of (\[exWC3\]), in which $VC$ is replaced with $VC'$. Let $(C_1^*, \dots, C_n^*)$ be the relaxed optimal job completion times and $z^*$ be the optimal value of the relaxation. Consider discrete random variable $\mathrm{Y}$, which takes the value $\sum_{j\in J} w_j(\xi_i)C^*_j$ with probability ${\rm Pr}[\xi_i]$, $i\in [K]$. The equality $z^*={\bf CVaR}_\alpha[\mathrm{Y}]$ holds. We relabel the jobs so that $C^*_1\leq C^*_2\leq \cdots\leq\ C_n^*$ and form schedule $\pi=(1,2,\dots,n)$ in nondecreasing order of $C^*_j$. Since the vector $(C_j^*)$ satisfies $VC'$ it must also satisfy (\[Schin\]). Hence, setting $I=\{1,\dots,j\}$, we get $$\sum_{i=1}^j p_iC^*_i\geq \frac{1}{2}\left((\sum_{i=1}^j p_i)^2+\sum_{i=1}^j p_i^2\right)\geq \frac{1}{2}\left((\sum_{i=1}^j p_i)^2\right).$$ Since $C^*_j\geq C_i^*$ for each $i\in \{1\dots j\}$, we get $C^*_j\sum_{i=1}^j p_i\geq \sum_{i=1}^j p_iC^*_i \geq \frac{1}{2}(\sum_{i=1}^j p_i)^2$ and, finally $C_j=\sum_{i=1}^j p_j \leq 2 C^*_j$ for each $j\in J$ – this reasoning is the same as in [@SH96b].
For each scenario $\xi_i\in \mathcal{U}$, the inequality $f(\pi,\xi_i)=\sum_{j\in J} w_j(\xi_i)C_j \leq 2 \sum_{j\in J} w_j(\xi_i)C^*_j $ holds. By Lemma \[lemowa1\], we have ${\bf CVaR}_\alpha[{\rm F}(\pi)]\leq 2\cdot {\bf CVaR}_\alpha[Y]=2z^*$. Since $z^*$ is a lower bound on the value of an optimal solution, $\pi$ is a 2-approximate schedule. Let us summarize the obtained result.
\[thmcappr1\] $\textsc{Min-CVaR}_{\alpha}~1|prec|\sum w_j C_j$ is approximable within 2 for each $\alpha\in [0,1)$.
This result can be refined when the deterministic $1|prec|\sum w_j C_j$ problem is polynomially solvable (for example, when the precedence constraints form an sp-graph, see, e.g., [@B07]). In this case $\textsc{Min-Exp}~1|prec|\sum w_j C_j$ is polynomially solvable, and we can also apply Theorem \[thmexp\], which leads to the following result:
\[thmcappr2\] If $1|prec|\sum w_j C_j$ is polynomially solvable, then $\textsc{Min-CVaR}_{\alpha}~1|prec|\sum w_j C_j$ is approximable within $\min\{\frac{1}{1-\alpha}, 2\}$ for each $\alpha\in [0,1)$.
Observe that $\frac{1}{1-\alpha}<2$ for each $\alpha<0.5$. Let us consider $\textsc{Min-VaR}_{\alpha}~1|prec|\sum w_j C_j$ problem. The value of ${\bf VaR}_\alpha[{\rm F}(\pi)]$, for a fixed schedule $\pi$, can be computed by solving the following MIP problem (see (\[exVC1\])a): $$\label{exVC123}
\begin{array}{llll}
\min & \theta\\
\text{s.t.} & \displaystyle \sum_{j\in J} w_j(\xi_k)C_j(\pi)-\theta\leq M_k \beta_k & k\in [K]\\
&\displaystyle \sum_{k\in [K]} {\rm Pr}[\xi_i] \beta_k \leq 1-\alpha\\
& \beta_k\in \{0,1\} & k\in [K]
\end{array}$$ where $M_k$ is an upper bound on the schedule cost under scenario $\xi_k$, $k\in [K]$. Using the formulation (\[cCC2\]) together with (\[exVC1\]), we can get a mixed integer programming formulation for $\textsc{Min-VaR}_{\alpha}~1|prec|\sum w_j C_j$. By replacing the constraints $VC$ with relaxed $VC'$ in the constructed model, we get a mixed integer problem with $K$ binary variables. This problem can be solved in polynomial time when $K$ is a constant. The same analysis as for $\textsc{Min-CVaR}_{\alpha}~1|prec|\sum w_j C_j$ (we also use Lemma \[lemowa1\]) leads to the following result:
\[thmcappr3\] If the number of scenarios is constant, then $\textsc{Min-VaR}_{\alpha}~1|prec|\sum w_j C_j$ is approximable within 2 for each $\alpha\in (0,1]$.
The bottleneck objective
------------------------
In this section we address a class of single machine scheduling problems with a bottleneck objective, i.e. in which $f(\pi)=\max_{j\in J} f_j(C_j(\pi))$, where $f_j(t)$ is the cost of completing job $j$ at time $t$. An important and well known example is $1|prec| \max w_j T_j$, in which the maximum weighted tardiness is minimized. This problem can be solved in $O(n^2)$ time by Lawler’s algorithm [@LA73]. We will use the fact that the minmax versions of the bottleneck problems are polynomially solvable for a wide class of cost functions [@BFSS16; @KZ16d]. In particular, the minmax version of $1|prec| \max w_j T_j$ with uncertain processing times and uncertain due dates can be solved in $O(Kn^2)$ time by using the algorithm constructed in [@KZ16d]. In the following, we will assume that $f(\pi, \xi)=\max_{j\in J} w_j T_j(\pi, \xi)$ for a given scenario $\xi\in \mathcal{U}$. We also assume that job processing times and due dates are nonnegative integers under all scenarios and job weights are positive integers. In consequence, the value of $f(\pi, \xi)$ is a nonnegative integer for each $\xi$.
Let $f_{\max}$ be an upper bound on the schedule cost over all scenarios. Let $h:{\mathbb{Q}}_{+}^K\rightarrow {\mathbb{Q}}_{+}$ be a nondecreasing function with respect to ${\mathbb{Q}}_{+}^K$. Suppose that $h$ can be evaluated in $g(K)$ time for a given vector $\pmb{t}=(t_1,\dots,t_K)\in {\mathbb{Z}}_{+}^K$. Consider the corresponding scheduling problem $\mathcal{PS}$, in which we seek a feasible schedule $\pi\in \Pi$ minimizing $H(\pi)=h(f(\pi,\xi_1),\dots, f(\pi, \xi_K))$. We can find such a schedule by solving a number of the following *auxiliary problems*: given a vector $\pmb{t}\in{\mathbb{Z}}_{+}^K$, check if $\Pi(\pmb{t})=\{\pi \in \Pi\,:\,
f(\pi,\xi_i)\leq t_i, i\in [K]\}$ is nonempty, and if so, return any schedule $\pi_{\pmb{t}}\in \Pi(\pmb{t})$. From the monotonicity of the function $h$, it follows that for each $\pi \in \Pi(\pmb{t})$ the inequality $h(f(\pi,\xi_1),\ldots, f(\pi, \xi_K))\leq
h(\pmb{t})$ is true. Thus, in order to solve the problem $\mathcal{PS}$, it suffices to enumerate all possible vectors $\pmb{t}=(t_1,\dots,t_K)$, where $t_i\in \{0,1,\dots,f_{\max}\}$, $i\in [K]$, and compute $\pi_{\pmb{t}}\in \Pi(\pmb{t})$ if $\Pi(\pmb{t})$ is nonempty. A schedule $\pi_{\pmb{t}}$ with the minimum value of $H(\pi_{\pmb{t}})$ is returned.
The crucial step in this method is solving the auxiliary problem. We now show that this can be done in polynomial time for the bottleneck problem with the maximum weighted tardiness criterion. Given any $\pmb{t}\in{\mathbb{Z}}_{+}^K$, we first form scenario set $\mathcal{U}'$ by specifying the following parameters for each $\xi_i \in \mathcal{U}$ and $j\in J$: $$p_j(\xi_i')=p_j(\xi_i), \; w'_j=1,\;
d_j(\xi_i')=\max\{C\geq 0\,:\,w_j(C-d_j(\xi_i))\leq t_i\}=t_i/w_j+d_j(\xi_i).$$ The scenario set $\mathcal{U}'$ can be built in $O(Kn)$ time. We then solve the minmax problem with scenario set $\mathcal{U}'$, which can be done in $O(Kn^2)$ time by using the algorithm constructed in [@KZ16d]. If the maximum cost of the schedule $\pi$ returned is 0, then $\pi_{\pmb{t}}=\pi$; otherwise $\Pi(\pmb{t})$ is empty. Since all the risk criteria considered in this paper are nondecreasing functions with respect to schedule costs over scenarios (see Lemma \[lemowa1\] for $\gamma=1$) and $g(K)$ is negligible in comparison with $Kn^2$, we get the following result:
$ \textsc{Min-Exp}~\mathcal{P}$, $\textsc{Min-VaR}_{\alpha}~\mathcal{P}$ and $\textsc{Min-CVaR}_{\alpha}~\mathcal{P}$ are solvable in $O(f_{\max}^K (Kn^2))$ time, when $\mathcal{P}$ is $1|prec|\max w_j T_j$.
The above running time is pseudopolynomial if $K$ is constant. Notice that the special cases, when $\mathcal{P}$ is $1|prec, p_j=1| T_{\max}$ are solvable in $O(Kn^{K+2})$ time, which is polynomial if $K$ is constant (as we can fix $f_{\max}=n$).
We now show that the problems admit an FPTAS if $K$ is a constant and $h(\gamma \pmb{t})\leq \gamma h( \pmb{t})$, for any $\pmb{t}\in{\mathbb{Q}}_{+}^K$, $\gamma\geq 0$. First we partition the interval $[0,f_{\max}]$ into geometrically increasing subintervals: $[0,1)\cup \bigcup_{\ell \in [\eta]}[(1+\epsilon)^{\ell-1}, (1+\epsilon)^{\ell})$, where $\eta=\lceil \log_{1+\epsilon} f_{\max} \rceil$ and $\epsilon\in (0,1)$. Then we enumerate all possible vectors $\pmb{t}=(t_1,\dots,t_K)$, where $t_i\in \{0,1\}\cup \bigcup_{\ell \in [\eta]}\{(1+\epsilon)^{\ell}\}$, $i\in [K]$, and find $\pi_{\pmb{t}}\in \Pi(\pmb{t})$ if $\Pi(\pmb{t})\not=\emptyset$. Finally, we output a schedule $\pi_{\hat{\pmb{t}}}$ that minimizes value of $H(\pi_{\pmb{t}})$ over the nonempty subsets of schedules. Obviously, the running time is $O(( \log_{1+\epsilon} f_{\max})^K (Kn^2+g(K)))=
O((\epsilon^{-1} \log f_{\max})^K (Kn^2+g(K)))$. Let $\pi^*$ be an optimal schedule. Fix $\ell_i\in \{0,1,\ldots,\eta\}$ for each $i\in [K]$, such that $(1+\epsilon)^{\ell_i-1}\leq f(\pi^*,\xi_i)<(1+\epsilon)^{\ell_i}$, where we assume that $(1+\epsilon)^{\ell_i-1}=0$ for $\ell_i=0$. This clearly forces $\Pi((1+\epsilon)^{\ell_1},\dots,(1+\epsilon)^{\ell_K})\not=\emptyset$. Moreover, $(1+\epsilon)^{\ell_i}\leq (1+\epsilon) f(\pi^*,\xi_i)$ for $\ell_i$, $i\in [K]$. By the definition of $\pi_{\hat{\pmb{t}}}$, we get $H(\pi_{\hat{\pmb{t}}})\leq h((1+\epsilon)^{\ell_1},\dots,(1+\epsilon)^{\ell_K})$. Since $h$ is a nondecreasing function and $h(\gamma \pmb{t})\leq \gamma h( \pmb{t})$, $h((1+\epsilon)^{\ell_1},\dots,(1+\epsilon)^{\ell_K})\leq
(1+\epsilon)h(f(\pi^*,\xi_1),\dots,f(\pi^*,\xi_K))$. Hence, $H(\pi_{\hat{\pmb{t}}})\leq (1+\epsilon)H(\pi^*)$. By Lemma \[lemowa1\], the risk criteria satisfy the additional assumption on the function $h(\pmb{t})$. This leads to the following theorem:
$ \textsc{Min-Exp}~\mathcal{P}$, $\textsc{Min-VaR}_{\alpha}~\mathcal{P}$ and $\textsc{Min-CVaR}_{\alpha}~\mathcal{P}$ admit an FPTAS, when $\mathcal{P}$ is $1|prec|\max w_j T_j$ and the number of scenarios is constant.
Conclusions and open problems
=============================
In this paper we have discussed a wide class of single machine scheduling problems with uncertain job processing times and due dates. This uncertainty is modeled by a discrete scenario set with a known probability distribution. In order to compute a solution we have applied the risk criteria, namely, the value at risk and conditional value at risk. The expectation and the maximum criteria are special cases of the risk measures. We have provided a number of negative and positive complexity results for problems with basic cost functions. Moreover, we have sharpened some negative ones obtained in [@AAK11; @AC08]. The picture of the complexity is presented in Tables \[tabs1\]-\[tabs3\]. Obviously, the negative results obtained remain true for more general cases, for instance, for the problems with more than one machine.
There is still a number of interesting open problems on the models discussed. The negative results for uncertain due dates assume that the number of due dates scenarios is a part of input. The complexity status of the problems when the number of due date scenarios is fixed (in particular, equals 2) is open. For uncertain processing times, an interesting open problem is $\textsc{Min-Exp}~1||\sum U_j$ (see Table \[tabs2\]). There is still a gap between the positive and negative results, in particular, we conjecture that the negative results for $\textsc{Min-Var}~\mathcal{P}$ for uncertain processing times (see Table \[tabs2\]) can be strengthen. Now they are just the same as for the $\textsc{Min-Max}~\mathcal{P}$.
### Acknowledgements {#acknowledgements .unnumbered}
This work was supported by the National Center for Science (Narodowe Centrum Nauki), grant 2017/25/B/ST6/00486.
[10]{}
H. Aissi, M. A. Aloulou, and M. Y. Kovalyov. Minimizing the number of late jobs on a single machine under due date uncertainty. , 14:351–360, 2011.
H. Aissi, C. Bazgan, and D. Vanderpooten. Min-max and min-max regret versions of combinatorial optimization problems: a survey. , 197:427–438, 2009.
M. A. Aloulou and F. D. Croce. Complexity of single machine scheduling problems under scenario-based uncertainty. , 36:338–342, 2008.
S. Atakan, K. Bulbul, and N. Noyan. Minimizng value-at-risk in single machine scheduling. , 248:25–73, 2017.
I. Averbakh. Minmax regret solutions for minimax optimization problems with uncertainty. , 27:57–65, 2000.
A. Avidor and U. Zwick. Approximating [MIN $k$-SAT]{}. , 2518:465–475, 2002.
N. Brauner, F. Gerd, S. Yakov, and S. Dzmitry. Lawler’s minmax cost algorithm: optimality conditions and uncertainty. , 19:401–408, 2016.
P. Brucker. . Springer Verlag, Heidelberg, 5th edition, 2007.
Z. Chang, S. Song, Y. Zhang, J.-Y. Ding, and R. Chiong. Distributionally robust single machine scheduling with risk aversion. , 256:261–274, 2017.
M. Cheung, J. Mestre, D. B. Shmoys, and J. Verschae. A [P]{}rimal-[D]{}ual [A]{}pproximation [A]{}lgorithm for [M]{}in-[S]{}um [S]{}ingle-[M]{}achine [S]{}cheduling [P]{}roblems. , 31:825–838, 2017.
R. L. Daniels and P. Kouvelis. Robust scheduling to hedge against processing time uncertainty in single-stage production. , 41:363–376, 1995.
M. Drwal and R. Rischke. Complexity of interval minmax regret scheduling on parallel identical machines with total completion time criterion. , 44:354–358, 2016.
J. Du and J. Y.-T. Leung. Minimizing [Total]{} [Tardiness]{} on [One]{} [Machine]{} is [NP]{}-hard. , 15:483–495, 1990.
M. R. Garey and D. S. Johnson. . W. H. Freeman and Company, 1979.
L. A. Hall, A. S. Schulz, D. B. Shmoys, and J. Wein. Scheduling to minimize average completion time: off-line and on-line approximation problems. , 22:513–544, 1997.
R. M. Karp. Reducibility [A]{}mong [C]{}ombinatorial [P]{}roblems. In [*Complexity of Computer Computations*]{}, pages 85–103, 1972.
A. Kasperski. Minimizing maximal regret in the single machine sequencing problem with maximum lateness criterion. , 33(4):431–436, 2005.
A. Kasperski, A. Kurpisz, and P. Zieli[ń]{}ski. Approximating the min-max (regret) selecting items problem. , 113:23–29, 2013.
A. Kasperski and P. Zieli[ń]{}ski. Minmax (regret) sequencing problems. In F. Werner and Y. Sotskov, editors, [*Sequencing and scheduling with inaccurate data*]{}, chapter 8, pages 159–210. Nova Science Publishers, 2014.
A. Kasperski and P. Zieli[ń]{}ski. Single machine scheduling problems with uncertain parameters and the [OWA]{} criterion. , 19:177–190, 2016.
R. Kohli, R. Krishnamurti, and P. Mirchandani. The minimum satisfiability problem. , 7:275–283, 1994.
P. Kouvelis and G. Yu. . Kluwer Academic Publishers, 1997.
P. Krokhmal, J. Palmquist, and S. P. Uryasev. Portfolio optimization with conditional value-at-risk objective and constraints. , 4:43–68, 2002.
E. L. Lawler. Optimal sequencing of a single machine subject to precedence constraints. , 19:544–546, 1973.
E. L. Lawler. A pseudopolynomial algorithm for sequencing jobs to minimize total tardiness. , 1:331–342, 1977.
V. Lebedev and I. Averbakh. Complexity of minimizing the total flow time with interval data and minmax regret criterion. , 154:2167–2177, 2006.
M. Mastrolilli, N. Mutsanas, and O. Svensson. Single machine scheduling with scenarios. , 477:57–66, 2013.
R. H. M[ö]{}hring, A. S. Schulz, and M. Uetz. Approximation in stochastic scheduling: the power of [LP]{}-based priority policies. , 46:924–942, 1999.
K. Natarajan, D. Shi, and K.-C. Toh. A probabilistic model for minmax regret in combinatorial optimization. , 62:160–181, 2014.
E. Nikolova. Approximation algorithms for offline risk-averse combinatorial optimization. In [*Proceedings of APPROX’10*]{}, 2010.
W. Ogryczak. obust [D]{}ecisions under [R]{}isk for [I]{}mprecise [P]{}robabilities. In Y. Ermoliev, M. Makowski, and K. Marti, editors, [*[M]{}anaging [S]{}afety of [H]{}eterogeneous [S]{}ystems*]{}, pages 51–66. Springer-Verlag, 2012.
G. C. Pflug. Some remarks on the [V]{}alue-at-[R]{}isk and the [C]{}onditional [V]{}alue-at-[R]{}isk. In S. P. Uryasev, editor, [*Probabilistic [C]{}onstrained [O]{}ptimization: [M]{}ethodology and [A]{}pplications*]{}, pages 272–281. Kluwer Academic Publishers, 2000.
M. Pinedo. . Springer, 2008.
C. N. Potts. An algorithm for the single machine sequencing problem with precedence constraints. , 13:78–87, 1980.
R. T. Rockafellar and S. P. Uryasev. Optimization of conditional value-at-risk. , 2:21–41, 2000.
S. Sarin, H. Sherali, and L. Liao. Minimizing conditional-value-at-risk for stochastic scheduling problems. , 17:5–15, 2014.
A. S. Schulz. Scheduling to minimize total weighted completion time: Performance guarantees of [LP-Based]{} heuristics and lower bounds. In [*IPCO*]{}, pages 301–315, 1996.
M. Skutella, M. Sviridenko, and M. Uetz. Unrelated [M]{}achine [S]{}cheduling with [S]{}tochastic [P]{}rocessing [T]{}imes. , 41:851–864, 2016.
M. Skutella and M. Uetz. Stochastic [M]{}achine [S]{}cheduling with [P]{}recedence [C]{}onstraints. , 34:788–802, 2005.
R. R. Yager. On ordered weighted averaging aggregation operators in multi-criteria decision making. , 18:183–190, 1988.
G. Yu and P. Kouvelis. Complexity results for a class of min-max problems with robust optimization applications. In P. M. Pardalos, editor, [*Complexity in Numerical Optimization*]{}. World Scientyfic, 1993.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'It has been recently suggested the existence of a neutral bound state of two neutrons and a $\Lambda$ hyperon, $^3_\Lambda n$. We point out that using either simple separable potentials or a full-fledged calculation with realistic baryon-baryon interactions derived from the constituent quark cluster model there is no possibility for the existence of such a $\Lambda nn$ bound state. For this purpose, we performed a full Faddeev calculation of the $\Lambda nn$ system in the $(I,J^P)=(1,1/2^+)$ channel using the interactions derived from the constituent quark cluster model which describe well the two-body $NN$ and $NY$ data and the $\Lambda np$ hypertriton.'
author:
- 'H. Garcilazo'
- 'A. Valcarce'
title: '**Nonexistence of a $\Lambda nn$ bound state**'
---
In a recent Rapid Communication by the experimental HypHI Collaboration [@HYPHI] it has been suggested the existence of a neutral bound state of two neutrons and a $\Lambda$ hyperon, $^3_\Lambda n$. They analyze the experimental data obtained from the reaction $^6$Li +$^{12}$C at 2A GeV to study the invariant mass distribution of $d+\pi^-$ and $t+\pi^-$. The signal observed in the invariant mass distributions of $d+\pi^-$ and $t+\pi^-$ final states was attributed to a strangeness-changing weak process corresponding to the two- and three-body decays of an unknown bound state of two neutrons associated with a $\Lambda$, $^3_\Lambda n$, via $^3_\Lambda n \to t + \pi^-$ and $^3_\Lambda n \to t^* + \pi^- \to d + n + \pi^-$.
This is an intriguing conclusion since one would naively expect the $\Lambda nn$ system to be unbound. In the $\Lambda nn$ system the two neutrons interact in the $^1S_0$ partial wave while in the $\Lambda np$ system they interact in the $^3S_1$ partial wave. Thus, since the $NN$ interaction in the $^1S_0$ channel is weaker than the $^3S_1$ channel, and the $\Lambda np$ system is bound by only 0.13 MeV, one may have anticipated that the $\Lambda nn$ system should be unbound. The unbinding of the $\Lambda nn$ system was first demonstrated by Dalitz and Downs [@Dal58] using a variational approach.
In a previous, by now somewhat older, paper [@GAJPG] we concluded the nonexistence of $\Lambda nn$ bound states by solving the Faddeev equations with separable potentials whose parameters were adjusted to reproduce the $\Lambda n$ scattering length and effective range of the two-body channels as obtained from four different versions of the Niemegen model [@NIEM1; @NIEM2; @NIEM3; @NIEM4] as well as the corresponding $NN$ spin-singlet and spin-triplet low-energy parameters. This leads to integral equations in one continuous variable.
As pointed out in Ref. [@GAJPG], if a system can have at most one bound state then the simplest way to determine if it is bound or not is by looking at the Fredholm determinant $D_F(E)$ at zero energy. If there are no interactions then $D_F(0)=1$, if the system is attractive then $D_F(0)<1$, and if a bound state exists then $D_F(0)<0$. We found in Ref. [@GAJPG] that $D_F(0)$ lies between 0.46 and 0.59 for the different models constructed by the Niemegen group so that the system is quite far from being bound.
Of course, it can be argued that the use of simple separable potentials is not a realistic assumption. Besides, since our previous work the knowledge of the strangeness –1 two-baryon system has improved and the models to study these systems are more tightly constrained. Therefore, we have now reexamined the $\Lambda nn$ system within a realistic baryon-baryon formalism obtained from the quark model. The baryon-baryon interactions involved in the study of the coupled $\Sigma NN - \Lambda NN$ system are obtained from the constituent quark cluster model [@Val05; @Gar05]. In this model baryons are described as clusters of three interacting massive (constituent) quarks, the mass coming from the spontaneous breaking of chiral symmetry. The first ingredient of the quark-quark interaction is a confining potential. Perturbative aspects of QCD are taken into account by means of a one-gluon potential. Spontaneous breaking of chiral symmetry gives rise to boson exchanges between quarks. In particular, there appear pseudoscalar boson exchanges and their corresponding scalar partners [@GFV07; @GAXXX]. Explicit expressions of all the interacting potentials and a more detailed discussion of the model can be found in Refs. [@Gar05; @GFV07].
In Refs. [@GFV07; @GAXXX] we established the formalism to study the $\Lambda NN$ system at threshold using the baryon-baryon interactions obtained from the constituent quark cluster model which leads to integral equations in the two continuous variables $p$ and $q$, where $p$ is the relative momentum of the pair and $q$ is the relative momentum of the third particle with respect to the pair. In order to solve these equations the two-body $t-$matrices are expanded in terms of Legendre polynomials leading to integral equations in only one continuous variable coupling the various Legendre components required for convergence.
$(\ell_\Sigma s_\Sigma j_\Sigma i_\Sigma\lambda_\Sigma J_\Sigma)_N$ $(\ell_\Lambda,s_\Lambda j_\Lambda i_\Lambda\lambda_\Lambda J_\Lambda)_N$ $(\ell_N,s_N j_N i_N \lambda_N J_N)_\Sigma$ $(\ell_N s_N j_N i_N \lambda_N J_N)_\Lambda$
--------------------------------------------------------------------- --------------------------------------------------------------------------- --------------------------------------------- ----------------------------------------------
(000$\frac{1}{2}$0$\frac{1}{2}$),(011$\frac{1}{2}$0$\frac{1}{2}$), (000$\frac{1}{2}$0$\frac{1}{2}$),(011$\frac{1}{2}$0$\frac{1}{2}$), (00010$\frac{1}{2}$),(01100$\frac{1}{2}$), (00010$\frac{1}{2}$),
(211$\frac{1}{2}$0$\frac{1}{2}$),(011$\frac{1}{2}$2$\frac{3}{2}$), (211$\frac{1}{2}$0$\frac{1}{2}$),(011$\frac{1}{2}$2$\frac{3}{2}$), (21100$\frac{1}{2}$),(01102$\frac{3}{2}$),
(211$\frac{1}{2}$2$\frac{3}{2}$),(000$\frac{3}{2}$0$\frac{1}{2}$), (211$\frac{1}{2}$2$\frac{3}{2}$) (21102$\frac{3}{2}$)
(011$\frac{3}{2}$0$\frac{1}{2}$),(211$\frac{3}{2}$0$\frac{1}{2}$),
(011$\frac{3}{2}$2$\frac{3}{2}$),(211$\frac{3}{2}$2$\frac{3}{2}$)
: Two-body $\Sigma N$ channels with a nucleon as spectator $(\ell_\Sigma s_\Sigma j_\Sigma i_\Sigma\lambda_\Sigma J_\Sigma)_N$, two-body $\Lambda N$ channels with a nucleon as spectator $(\ell_\Lambda,s_\Lambda j_\Lambda i_\Lambda\lambda_\Lambda J_\Lambda)_N$, two-body $NN$ channels with a $\Sigma$ as spectator $(\ell_N,s_N j_N i_N \lambda_N J_N)_\Sigma$, and two-body $NN$ channels with a $\Lambda$ as spectator $(\ell_N s_N j_N i_N \lambda_N J_N)_\Lambda$ that contribute to the $(I,J^P)=(1,1/2^+)$ state. $\ell$, $s$, $j$, and $i$, are, respectively, the orbital angular momentum, spin, total angular momentum, and isospin of a pair, while $\lambda$ and $J$ are the orbital angular momentum of the third particle with respect to the pair and the result of coupling $\lambda$ with the spin of the third particle.
\[t1\]
This model takes into account the coupling $N\Lambda-N\Sigma$ as well as the tensor force responsible for the coupling between S and D waves. In particular, for the $\Lambda NN$ channel $(I,J^P)=(1,1/2^+)$ which corresponds to the conjectured $\Lambda nn$ bound state there is a total of 21 coupled channels contributing to the state. We give in Table \[t1\] the quantum numbers of these contributing channels.
In Ref. [@GAXXX] we showed that if one increases the triplet $N\Lambda$ interaction by increasing the triplet scattering length then the $\Lambda NN$ state with $(I,J^P)=(0,3/2^+)$ becomes bound and since that state does not exist we are allowed to set an upper limit of 1.58 fm for the $\Lambda N$ spin triplet scattering length. Since, in addition, the fit of the hyperon-nucleon cross sections is worsened [@GFV07] when the spin-triplet scattering length is smaller than 1.41 fm we concluded that $1.41 \le a_{1/2,1} \le 1.58$ fm. By requiring that the hypertriton binding energy had the experimental value $B=0.13\pm 0.05$ MeV we obtained for the $\Lambda N$ spin-singlet scattering length the limits $2.37 \le a_{1/2,0} \le 2.48$ fm.
Thus, we constructed twelve different models corresponding to different choices of the spin-singlet and spin-triplet $\Lambda N$ scattering lengths which describe equally well all the available experimental data. We solved the three-body problem taking full account of the $\Lambda NN - \Sigma NN$ coupling as well as the effect of the D waves. We present in Table \[t2\] the Fredholm determinant at zero energy of the $(I,J^P)=(1,1/2^+)$ state for these models. The realistic quark model interactions predict a Fredholm determinant at zero energy ranging between 0.38 and 0.42, close to the interval 0.46–0.59 obtained from the separable potentials of the Niemegen group. As one can see, in all cases the Fredholm determinant at zero energy is positive and far from zero, excluding the possibility for binding in this system. From the results of Table \[t2\] and from the energy dependence of the Fredholm determinant shown in Fig. 2 of Ref. [@GAXXX] one can infer that the $(I,J^P)=(1,1/2^+)$ state is unbound by at least 5–10 MeV, which is a large energy in comparison with the 0.13 MeV binding energy of the hypertriton.
$a_{1/2,1}=1.41$ $a_{1/2,1}=1.46$ $a_{1/2,1}=1.52$ $a_{1/2,1}=1.58$
------------------ ------------------ ------------------ ------------------ ------------------
$a_{1/2,0}=2.33$ 0.42 0.41 0.40 0.38
$a_{1/2,0}=2.39$ 0.42 0.41 0.39 0.38
$a_{1/2,0}=2.48$ 0.42 0.41 0.40 0.38
: Fredholm determinant at zero energy $D_F(0)$ for several hyperon-nucleon interactions characterized by $\Lambda N$ scattering lengths $a_{1/2,0}$ and $a_{1/2,1}$ (in fm).
\[t2\]
To summarize, we have shown that using either simple separable potentials or a full-fledged calculation with realistic baryon-baryon interactions derived from the constituent quark cluster model there is no possibility for the existence of a $\Lambda nn$ bound state. Thus, the signal observed in the invariant mass distributions of $d+\pi^-$ and $t+\pi^-$ final states in the analysis of the experimental data obtained from the reaction $^6$Li +$^{12}$C at 2A GeV and adduced to the existence of a neutral bound state of two neutrons and a $\Lambda$ hyperon must be due to a different effect.
This work has been partially funded by COFAA-IPN (México), by Ministerio de Educación y Ciencia and EU FEDER under Contract No. FPA2010-21750-C02-02 and by the Spanish Consolider-Ingenio 2010 Program CPAN (CSD2007-00042).
[99]{}
C. Rappold [*et al.*]{} (HypHI Collaboration), Phys. Rev. C [**88**]{}, 041001R (2013).
R. H. Dalitz and B. W. Downs, Phys. Rev. [**110**]{}, 958 (1958); Phys. Rev. [**111**]{}, 967 (1958); Phys. Rev. [**114**]{}, 593 (1959).
H. Garcilazo, J. Phys. G [**13**]{}, L63 (1987).
M. M. Nagels, T. A. Rijken, and J. J. de Swart, Ann. Phys., (N.Y.) [**79**]{}, 338 (1973).
M. M. Nagels, T. A. Rijken, and J. J. de Swart, [*Few Particle Problems in the Nuclear Interaction*]{} edited by I. Slaus [*et al.*]{} (Amsterdam, North Holland) pp. 42-45.
M. M. Nagels, T. A. Rijken, and J. J. deSwart, Phys. Rev. D [**15**]{}, 2547 (1977).
M. M. Nagels, T. A. Rijken, and J. J. de Swart, Phys. Rev. D [**20**]{}, 1633 (1979).
A. Valcarce, H. Garcilazo, F. Fernández, and P. González, Rep. Prog. Phys. [**68**]{}, 965 (2005).
A. Valcarce, H. Garcilazo, and J. Vijande, Phys. Rev. C [**72**]{}, 025206 (2005); J. Vijande, F. Fernández, and A. Valcarce, J. Phys. G [**31**]{}, 481 (2005).
H. Garcilazo, T. Fernández-Caramés, and A. Valcarce, Phys. Rev. C [**75**]{}, 034002 (2007).
H. Garcilazo, A. Valcarce, and T. Fernández-Caramés, Phys. Rev. C [**76**]{}, 034001 (2007).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We present measurements on the single molecule magnet Fe$_8$ in the presence of pulsed microwave radiation at 118 GHz. The spin dynamics is studied via time resolved magnetization experiments using a Hall probe magnetometer. We investigate the relaxation behavior of magnetization after the microwave pulse. The analysis of the experimental data is performed in terms of different contributions to the magnetization after-pulse relaxation. We find that the phonon bottleneck with a characteristic relaxation time of $\sim 10-100$ ms strongly affects the magnetization dynamics. In addition, the spatial effect of spin diffusion is evidenced by using samples of different sizes and different ways of the sample’s irradiation with microwaves.'
author:
- 'K. Petukhov,$^{1,2}$ S. Bahr,$^1$ W. Wernsdorfer,$^1$ A.-L. Barra$^2$ and V.Mosser$^3$'
title: 'Magnetization dynamics in the single-molecule magnet Fe$_8$ under pulsed microwave irradiation'
---
Introduction
============
Single molecule magnets (SMMs) have attracted much interest in recent years because of their unique magnetic properties. Having a regular structure, a well defined spin ground state and magnetic anisotropy they exhibit quantum phenomena even at macroscopic scales.[@novak:1995; @friedman:1996; @thomas:1996] Features such as quantum tunneling between spin states, interference between tunneling paths or blocking of the spin orientation at very low temperature show the quantum nature of SMMs. [@barco:2004; @sorace:2003; @wernsdorfer:science1999; @wernsdorfer:2002a; @wernsdorfer:2000] In addition SMMs are supposed to be good candidates for data storage or quantum computing. [@leuenberger:2001]
Recent works in the field of SMMs focused on spin dynamics and interactions with millimeter-wave radiation. The aima are to control the spin orientation in the sample and to selectively induce transitions between spin states. The crucial point for any application of SMMs is the knowledge of the spin relaxation time and the spin decoherence time. Therefore, various experiments have been performed in studying spin dynamics in SMMs in the presence of microwaves. Most measurements are based on standard electron paramagnetic resonance (EPR) techniques [@zipse:2003] or on optical spectroscopy [@mukhin:2001], while others are based on magnetization measurements of the sample. In measuring the absorption of the microwaves via the decrease of magnetization we can obtain information about both the magnetization of the sample and EPR-like spectra. This technique also allows the precise control over the excitation of the sample and makes it possible to quantify the nonresonant heating.[@petukhov:2005] The magnetization sensor can be either a Hall magnetometer [@sorace:2003; @bal:epl2005], a micrometer sized superconducting quantum inference device (SQUID) [@wernsdorfer:2004], a standard SQUID [@cage:2005] or an inductive pickup loop [@bal:2004]. Differences in these techniques lie mainly in the rapidity and sensitivity of the measurement, in the possibility of applying magnetic fields, and in the compatibility with microwaves.
In this paper we study the spin dynamics of the single-molecule magnet $Fe_{8}O_{2}(OH)_{12}(tacn)_{6}$, hereafter called Fe$_8$. This molecule contains eight Fe(III) ions with spins s=5/2. These spins are strongly superexchange coupled forming a spin ground state $S=10$ and the spin dynamics can be described assuming the giant spin model by an effective Hamiltonian [@barra:1996] $$\mathcal{H} = - DS_z^2 + E(S_x^2 -S_y^2) + \mathcal{O}(4) -
g\mu_B\vec{S}\cdot\vec{H}$$
![\[barrier\] Spin states and energy barrier of the Fe$_8$ system. For an excited spin there are various processes relevant for spin dynamics.](Figure0_small.eps){width="3.0in"}
$\vec{H}$ is the applied magnetic field, $\mathcal{O}(4)$ contains fourth order terms of spin operators and $g\approx 2$ represents the gyromagnetic factor. The anisotropy parameters $D=0.275$ K and $E=0.046$ K have been determined by various experimental methods. [@wernsdorfer:science1999] Classical EPR techniques, frequency domain magnetic resonance spectroscopy and neutron spectroscopy are complementary methods and give similar results. [@barra:1996; @mukhin:2001; @caciuffo:1998; @park:2002] The nondiagonal terms in the Hamiltonian are responsible for the tunneling processes between spin states, whereas $D$ defines the anisotropy barrier of approximately 25 K as can be seen in Fig. \[barrier\].
In terms of the spin dynamics the giant spin model reveals various relaxation processes that are important for the evolution of the spin system in the time domain. As sketched in Fig. \[barrier\] the main parameters of the spin system are the spin relaxation time $\tau_{1}$ (time scale $\sim 10^{-7}~s$), the excitation over the barrier by a thermally activated multistep Orbach process with time constant $\tau_{orbach}$ (time scale $\sim 10^{-8}~s \times e^{\Delta E/k_{\rm B}T}$) where $\Delta E$ is the barrier height) and the tunnel probability between degenerated states with time constant $\tau_{tunnel}$ (time scale $\sim 10^{4}~s$ for the ground state tunneling).[@barra:1996; @sangregorio:1997] The use of a large crystal of the single molecule magnet Fe$_8$ and, in consequence, the interactions between molecules make it necessary to introduce spin-phonon and spin-spin interactions. Effects such as spin decoherence (typical time scale $\tau_ {2}\sim 10^{-9}~s$), phonon bottleneck (typical time scale $\tau_{ph}\sim 10^{0}~s$) or spin diffusion ($\tau_{diff}$) have to be taken into account for a complete description of the spin dynamics.[@Chiorescu:PRL2000; @abragam]
In this paper, a series of measurements on the SMM Fe$_8$ is presented investigating the relaxation of magnetization on millisecond and microsecond scales. In Section II we describe the experimental setup and the various experimental conditions. In Section III we present the experimental data that will be discussed in Section IV. Finally in Section V, we give some concluding remarks.
Experimental techniques
=======================
General setup
-------------
The measurements are performed using a commercial 16 T superconducting solenoid and a cryostat at low temperatures in the range of 1.4 K to 10 K with temperature stability better than 0.05 K. The magnetization of the Fe$_8$ sample is measured by a Hall magnetometer. The Hall bars were patterned by Thales Research and Technology (Palaiseau), using photolithography and dry etching, in a delta-doped AlGaAs/InGaAs/GaAs pseudomorphic heterostructure grown by Picogiga International using molecular beam epitaxy (MBE). A two-dimensional electron gas is induced in the 13 nm thick $In_{0.15}Ga_{0.85}As$ well by the inclusion of a Si delta-doping layer in the graded $Al_xGa_{1-x}As$ barrier. All layers, apart from the quantum well, are fully depleted of electrons and holes. The two-dimensional electron gas density $n_s$ is about $8.9 \times 10^{11}~cm^{-2}$ in the quantum well, corresponding to a sensitivity of about 700 $\Omega$/T, essentially constant under $-100~^\circ$C. The sample is placed on top of the 10 $\mu$m $\times$ 10 $\mu$m Hall junction with its easy axis approximately parallel to the magnetic field of the solenoid. The three samples used in our experiments (150 $\times$ 100 $\times$ 30 $\mu$m$^3$, 160 $\times$ 180 $\times$ 100 $\mu$m$^3$ and 680 $\times$ 570 $\times$ 170 $\mu$m$^3$) are exposed to microwave radiation. Microwaves are generated by a continuous wave (cw), mechanically tunable Gunn oscillator with a nominal output power of 30 mW and a frequency range of 110 GHz to 119 GHz. Pulses are generated using a SPST fast PIN diode switch (switching time less of than 3 ns) triggered by a commercial pulse generator. An oversized circular waveguide of 10 mm diameter leads the microwaves into the cryostat. In some of our experiments we use transition parts from oversized circular-to-rectangular WR6 waveguides. In other experiments we use only a cylindrical cone as an end piece of the circular waveguide that is right in front of the irradiated sample. The different configurations of the coupling of the microwaves to the sample that are investigated and compared are explained in the next paragraph. The cryostat is filled with exchange gas that thermodynamically couples the sample to the bath and allows a rather fast heat exchange. As the signal of the Hall magnetometer is in the range of a few microvolts a low-noise preamplifier is used in order to increase the signal-to-noise ratio in the rather long coaxial cables. Finally, the signal acquisition is done by a fast digital oscilloscope having a bandwidth of 1 GHz and 10 G/s sample rate and is done by taking an average over typically 32 frames.
Coupling of the microwaves to the sample
----------------------------------------
![\[wguides\] Different configurations of irradiating the sample with microwaves. We used several waveguides in order to change the electromagnetic environment of the sample, i.e. the coupling of the microwaves to the sample.](Figure1_small.eps){width="3.0in"}
### Conical waveguide
The simplest method of coupling microwaves to the sample is provided by a conical waveguide focusing microwaves from the oversized circular waveguide to the sample. In our experiments a cone with an opening of 4 mm in diameter is used. This setup is sketched in Fig. \[wguides\]a and will be denoted hereafter as *setup I*. The conical waveguide has the advantage that it conserves the polarization of the passing light, thus allowing experiments depending on the angle of polarization or experiments with circularly polarized microwaves. Due to the large dimensions of the waveguide compared to the wavelength the propagation of the microwaves can be considered as quasioptical and the attenuation of microwave power is rather small.
### Rectangular waveguide
Compared to the conical end piece of the waveguide a rectangular waveguide can focus the microwaves even better. The opening of a typical D-band WR6 waveguide is 1.7 mm $\times$ 0.8 mm, thus the cross section is nine times less than that in the case of the conical waveguide. However, the attenuation using the rectangular waveguide is fairly large, especially because of the transition part between oversized cylindrical and rectangular waveguides. At the end of the rectangular waveguide the field distribution is well defined. This feature allows us to irradiate the crystal in different ways. It can be put either in the central region of the waveguide (Fig. \[wguides\]b, *setup II*), or at the edge of the waveguide, in order to partially irradiate the crystal (Fig. \[wguides\]c, *setup III*). When irradiated partially, the magnetization of the crystal is measured with a Hall sensor at the nonirradiated side of the crystal. This method allows us to point out the importance of the inhomogeneous distribution of magnetization in the sample and spin diffusion processes that take place in large samples.
### Microwave resonator
In some of our experiments a cylindrical cavity made of copper with detachable end plates is used, and the sample is placed on top of one of the end plates (Fig. \[wguides\]d, *setup IV*). The inner diameter of the resonator is 10.10 mm and the height is 5.5 mm. A standard WR6 waveguide is coupled to the sidewall of the cavity by a coupling hole. The microwaves from the waveguide enter into the cavity at half height and can excite various modes obeying the selection rule $TE_{odd,*,*}$ and $TM_{even,*,*}$ (Table \[modes\]). All the modes have zero tangential electrical field on the end plates. The magnetic field has one or more maxima on the end plates according to the mode and the direction of the magnetic field is always radial. The sample is mounted on one of the end plates in such a way that the magnetic field in the cavity and the easy axes of the sample are parallel.
Resonator Mode Frequency \[GHz\]
---------------- -------------------
$TM_{413}$ 108.631
$TE_{114}$ 110.471
$TM_{014}$ 111.436
$TM_{033}$ 114.137
$TE_{314}$ 116.097
$TM_{214}$ 119.410
$TE_{124}$ 120.176
$TM_{024}$ 120.932
: Possible modes in a perfect cylindrical cavity in the range of 108 GHz to 121 GHz and for the dimension of the cavity specified in the text. Due to perturbations inside the cavity the calculated frequencies might differ from the theoretical ones.[]{data-label="modes"}
By using a resonator we expect the amplitude of ac magnetic field to increase by a few orders of magnitude and thus allow better excitation of the sample even with very short pulses ($<$ 10 $\mu$s). The Q-factor for resonant modes is numerically calculated to be in the range of $10^2$ to $10^3$ depending on the mode. Therefore, the resonant modes are expected to have a full width at half maximum in the order of $0.1$ to $1$ GHz. Consequently we expect each mode to exist in a rather broad frequency band. The modes should be present in the resonator even when the microwave frequency does not exactly match the calculated resonance frequency.
In the experiments two slightly different end plates are used. In the first case an *unprotected* Hall bar with a sample on top is directly glued on top of the end plate. The position of the sample is about 2 mm off center of the end plate. In respect to the size of the cavity the Hall sensor and the sample represent a perturbation of the resonator that might be non-negligible. In consequence the frequencies for the different modes might slightly and inhomogeneously shift according to the calculated frequencies. Nevertheless the density of modes between 109 GHz and 120 GHz should remain rather high.
In the second case in order to perturb the cavity as weak as possible we *protected* the Hall bar with a copper foil. The position of the sample in this case is about 1 mm off center of the end plate. A small hole is drilled into the end plate and the Hall sensor is placed into the hole and is finally coated with a thin copper foil (thickness of 10 $\mu$m). Thus only the sample placed on the copper foil is directly exposed to the electromagnetic field inside the cavity whereas the Hall sensor and the cables are outside the resonator. In this setup the Hall sensor is protected from the microwave radiation by the thin copper foil, however the sensitivity in measuring the sample’s magnetization is expected to be weaker as in the unprotected case.
Measurements
============
Magnetometry combined with microwaves
-------------------------------------
When the sample of SMM placed in the magnetic field $B$ is exposed to the cw microwaves, the magnetization curves show resonant absorption dips, similar to EPR spectroscopy spectra. The absorption of microwave radiation takes place at certain field values at a given frequency, when the microwave frequency matches the energy difference between two neighboring energy states, thus the allowed transitions are $\Delta m_s=\pm
1$. The populating of the upper levels (see Fig. \[barrier\]) reduces the net sample’s magnetization $M$, the change of which $\Delta M$ can be detected via Hall voltage measurements. If the applied magnetic field is ramped while the microwaves are applied, the obtained magnetization spectra clearly show a series of nearly evenly spaced absorption dips, which can be easily attributed to the appropriate transitions, as shown in Fig. \[CW\_stuff\]a by the thick solid curve. This curve is placed on the top of “pure” magnetization curves, i.e. the curves measured without microwaves, depicted by the thin solid curves in Fig. \[CW\_stuff\]a. These reference curves were measured at different cryostat temperatures in the range from 2 K (top curve) to 20 K (bottom curve) with 1 K incremental step. As can be seen from Fig. \[CW\_stuff\]a, the 2 K magnetization data measured without microwaves and the magnetization data measured in the presence of microwaves do not match: the latter curve lies much below the first curve, as the temperature during the microwave experiment would be higher compared to that of the pure magnetization experiment. The difference between the two curves is denoted by the magnetization difference $\Delta M$, which is a good measure of the amount of microwave radiation absorbed by the spin system.
 Magnetization of Fe$_8$ as a function of magnetic field. The curves are normalized to the saturation magnetization value $M_S$. The thin solid curves represent the magnetization measured without microwaves at temperatures from 2 K to 20 K in 1 K steps. The thick solid curve represents magnetization curve in presence of cw microwaves of frequency $f_{CW}=118$ GHz measured at 2 K. The thick arrow represents the magnetization difference $\Delta M$ between data taken with and without microwaves at 2 K. (b) Spin temperature $T_{\rm S}$ vs applied field $B$, calculated using the mapping procedure described in the text. The dashed curves depict off-resonance absorption.](Figure2_small.eps){width="3.0in"}
Spin temperature
----------------
Although the difference of magnetization $\Delta M$ can be used to qualify the amount of absorbed microwave photons, it is rather inconvenient to speak in terms of relative units of $\Delta M$. Another more significant complication in the use of $\Delta M$ for quantitative characterizations concerns the loss of sensitivity of $\Delta M$ close to zero field. As the magnetic field $B$ goes to zero, the magnetization also goes to zero, and hence the sensitivity of detection of absorption peaks goes to zero as well. Therefore, we need to perform a transformation of the magnetization to a physical quantity which does not depend on the magnetic field $B$.
Such a quantity called *spin temperature* was explicitly introduced in our earlier paper [@petukhov:2005] as a perfect measure of the amount of microwave radiation absorbed by an SMM spin system. The concept of spin temperature can be easily understood from Fig. \[CW\_stuff\]a. We can map the magnetization curve (magnetization spectrum) obtained under the use of microwaves onto underlying reference magnetization curves, measured at different cryostat temperatures without microwaves. For each magnetization point of the absorption spectra one finds, at the corresponding field $B$, the temperature $T_{\rm S}$ that gives the same magnetization measured without microwave radiation \[Fig. \[CW\_stuff\]a\]. The temperatures in between the reference magnetization curves are obtained with a linear interpolation. A typical result of such a mapping is depicted in Fig. \[CW\_stuff\]b. $T_{\rm S}$ can be called the spin temperature because the irradiation time is much longer than the lifetimes of the energy levels of the spin system which were found to be around 10$^{-7}$ seconds [@Wernsdorfer:EPL2000]. The phonon relaxation time $\tau_{ph}$ from the crystal to the heat bath (cryostat) is much longer, typically between milliseconds and seconds [@Chiorescu:PRL2000]. The spin and phonon systems of the crystal are therefore in equilibrium.
Figure \[CW\_stuff\]b shows spin temperature data calculated for the magnetization measurements at cryostat temperatures of 2 K and 10 K, performed at frequency of cw microwaves of $f_{CW}=118$ GHz. From Fig. \[CW\_stuff\]b we can conclude that the obtained spin temperatures $T_{\rm S}$ are much larger than the cryostat temperature $T$. This is associated with a strong heating of the spin system. This effect is more prominent at lower $T$: the baseline of 2 K $T_{\rm S}$ spectrum is around 7 K, while 10 K $T_{\rm S}$ spectrum’s background is very close to the nominal cryostat temperature $T=10$ K, as depicted by the dashed curves in Fig. \[CW\_stuff\]b. We also see that both backgrounds are not flat, but are magnetic field dependent. The presence of the spectra’s nonflat background is due the presence of off-resonance absorption of microwaves, which takes place between the resonant absorption peaks. The nonresonant (or background) absorption is modulated in the following way: it has larger contribution where the resonant absorption has larger spectral weight, i.e. higher peaks of $T_{\rm S}$. Interestingly, the off-resonance absorption was also evidenced in the EPR spectra of SMMs, but its origin remains undiscussed [@park:2002; @Hill:PRB02]. On the top of the nonresonant background one can see a perfect EPR-like absorption spectra, and the $T_{\rm S}$ peak positions exactly match the magnetic field values, corresponding to $|\Delta m_S|=1$ transitions (see Fig. \[CW\_stuff\]).
Pulsed microwave measurements
-----------------------------
Another way to perform $\Delta M$ measurements in order to calculate the spin temperature $T_{\rm S}$ is to utilize a pulsed microwave (PW) radiation [@petukhov:2005]. This method gives direct information about $\Delta M$ at a given magnetic field value $B$, at a given temperature $T$, and at a given microwave power, a measure of which is a pulse length $t$. This advanced method also drastically reduces the heating of the sample with microwaves, since the repetition rate of microwave pulses (typically 200 ms in our experiments) is much larger than the pulse length values, typically $t \leq 10$ ms. Restoration of the after-pulse magnetization to the equilibrium value $M_0$ normally takes less than 100 ms, as can be seen in Fig. \[OSC\].
![\[OSC\]Typical oscillogram of a pulsed experiment. The magnetization was measured as a function of time for a microwave pulse length of $t=10$ ms at temperature $T$ = 10 K.](Figure3_small.eps){width="3.4in"}
The scheme of possible pulsed experiment is depicted in Fig. \[OSC\]. The top part of Fig. \[OSC\] schematically shows microwave pulse of duration $t=10$ ms, and the bottom part of the figure shows time-resolved development of magnetization data collected during such pulsed experiment. The magnetization before and at the end of the pulse has values $M_{0}$ and $M_{\rm MW}$, respectively. The difference between the unperturbed magnetization value $M_{0}$ and magnetization at the final edge of the pulse $M_{\rm MW}$ (i.e. the height of the magnetization response) $\Delta M= M_{\rm MW}-M_{0}$ is identical to the magnetization difference $\Delta M$ defined for the cw microwaves case, as graphically explained by Fig. \[CW\_stuff\]a. Thus, having a set of reference magnetization curves, shown by thin curves in Fig. \[CW\_stuff\]a, and magnetization difference $\Delta M$ defined from the PW measurements, as shown in Fig. \[OSC\], one can perform spin temperature $T_{\rm S}$ calculations. Such calculations for Fe$_8$ have been performed in our previous work for the PW configuration and it has been shown that obtained spin temperatures $T_{\rm S}$ are much closer to the cryostat temperature $T$ than that for cw experiments [@petukhov:2005]. The linewidths and shapes of PW $T_{\rm S}$ spectra depend on the pulse length, but in general the peak positions of cw and PW configurations are identical. In contrast to the cw experiments, the PW method can successfully resolve absorption peaks near zero field [@petukhov:2005].
Unlike cw measurements, PW magnetization profiles contain information not only about $\Delta M$ but also about the magnetization dynamics. Let us assume that the applied magnetic field $B$ and the microwave frequency $f_{PW}$ do match the resonance condition, i.e., the in-resonance microwave pulse is applied. Since the sample’s magnetization is connected to the spin state level occupancy of SMMs, the magnetization dynamics should be connected to the level’s lifetime. The spin-spin relaxation time $\tau_2$ is usually much shorter than the spin-phonon relaxation time $\tau_1$; if $\tau_1$ obtained through magnetization dynamics is short enough, it can determine the upper limit of $\tau_2$. Finally, since there is an increase of the sample’s temperature due to the heating with microwaves, the phonon relaxation time $\tau_{ph}$ from the crystal to the heat bath (cryostat) can be systematically studied from the magnetization “cooling” after long enough pulses in PW experiments. Thus, the detailed consideration of the time-resolved magnetization profile, depicted in Fig. \[OSC\], might provide information about the $\tau_1$, $\tau_2$, and $\tau_{ph}$ relaxation times.
Let us consider the magnetization behavior during a PW experiment in detail in Fig. \[OSC\]. At the beginning of the pulse, the magnetization rapidly decreases (region \[1\]) and starts to saturate (region \[2\]) until the end of the pulse. We need to note that a complete saturation is observed only for rather long pulses of several seconds. After the microwave pulse is switched off, the magnetization restores back to the equilibrium value $M_{0}$. At the beginning of its restoration the magnetization increases rapidly (region \[3\]), several millisecond later magnetization increase changes to the slower behavior (region \[4\]), until it levels out at $M_{0}$. This slow restoration lasts long, up to a hundred of milliseconds, but we were able to follow it completely, since the typical repetition time of pulses was 200 ms. This brings us to the conclusion, that region \[4\] might comprise the information about the cooling of the sample after the microwave pulse, i.e. the phonon relaxation time $\tau_{ph}$ from the crystal to the heat bath (cryostat). Exactly this relaxation time is typically of the order of magnitude of several tens of milliseconds up to seconds [@Chiorescu:PRL2000]. The fast-running beginning of region \[3\] could contain the longitudinal relaxation time $\tau_1$ (typically $\sim
10^{-7}$ seconds [@Wernsdorfer:EPL2000]). There was another interesting observation in region \[3\] of the magnetization curve in Fig. \[OSC\]: during some of our experiments we have observed that right after the pulse was switched off the magnetization continued to decrease for some time and only then started to increase to the equilibrium value. Similar behavior of magnetization after microwave pulses in Fe$_8$ was observed in the recent work of Bal *et al.* [@bal:epl2005]. Below we will explicitly investigate such an *overshooting* of magnetization after the microwave pulse. In principle, regions \[1\] and \[2\] are very similar to the regions \[3\] and \[4\], correspondingly. The problem with the use of this part of magnetization evolution curve is that regions \[1\] and \[2\] are limited by the pulse duration, and therefore it seems to be problematic to estimate relaxation times from this part of the data, especially the long-lasting $\tau_{ph}$. In this paper, we pay attention to the time-resolved behavior of the magnetization after the pulse in a PW experiment, i.e. to regions \[3\] and \[4\] as determined in Fig. \[OSC\].
The model
---------
![\[fits\]Fit of magnetization restoration data (open circles) with a single exponential relaxation (solid line). The inset shows the fit with two exponents: fast relaxation with relaxation time $\tau_{short}$ and cooling with relaxation time $\tau_{long}$. Note that the depicted magnetization data curve contains only 2% of the experimental points taken during the single experiment, i.e. only every 50th point is shown.](Figure4_small.eps){width="3.0in"}
First, in order to analyze the behavior of the magnetization after the microwave pulse, we have tried to fit the magnetization data in regions \[3\] and \[4\] by a single exponential relaxation. We have found that in many cases a single exponential description was unsatisfactory, as shown in Fig. \[fits\]. This is not surprising in the framework of consideration concerning different relaxation times given above. Indeed, the $\tau_1$ relaxation time, which can be found right after the pulse is much shorter than the cooling $\tau_{ph}$ relaxation time, which can be a major contribution in the tail of magnetization restoration. Therefore, we have separately considered two different regions of the magnetization restoration curve, as depicted in the inset of Fig. \[fits\]. Firstly, we have assumed that the magnetization data, obtained right after the pulse was switched off (typically, within the time frame of 10-20% of the pulse length), could contain the information about the spin-lattice $\tau_{1}$ relaxation time. This data can be described by a fast exponential relaxation, and we will denote the corresponding relaxation time by $\tau_{short}$, as shown in the inset of Fig \[fits\]. Another valuable contribution to the overall magnetization restoration comes from the cooling of the specimen after the microwave pulse, such a process can be described by the long-lasting relaxation process with relaxation time $\tau_{long}$, taken later after the pulse was switched off (typically, 3 to 4 pulse length values later after the pulse edge until the end of the magnetization data), as depicted in the inset of Fig \[fits\]. If the slow relaxation $\tau_{long}$ is responsible for the sample’s cooling, it can only be sensitive to the sample size and its thermal coupling to the bath, with both parameters unchanged during an experimental set. Therefore, we expect this contribution to be temperature and pulse length independent. Nevertheless, the slow relaxation $\tau_{long}$ contribution (i.e. sample thermalization) can become dominating over the fast relaxation $\tau_{short}$ on increase of the temperature and/or for very long pulses, since the $\tau_{short}$ drastically shortens under such conditions and can be unresolved. In this case, a single exponential relaxation could be suitable for the magnetization data description and it could give solely the relaxation time $\tau_{long}$. In our experiments we avoid such a situation and we carefully adjust the experimental condition to have two clearly distinguishable regions \[3\] and \[4\], where the uncontroversial analysis by means of $\tau_{short}$ and $\tau_{long}$ can be performed. This analysis was applied in the following magnetization relaxation measurements.
Relaxation of magnetization
---------------------------
We have studied the relaxation of magnetization employing different sample irradiation configurations, described in the “Experimental setup” part. In all the cases, the applied magnetic field was set to 0.2 T and the frequency of microwaves during pulses was 118 GHz. Thus, below we describe the studies of the magnetization dynamics of the first transition from the ground state $m_S=-10$ to the first excited state $m_S= -9$, since the given magnetic field and frequency values match the resonance condition for Fe$_8$ placed into magnetic field along its easy axis [@petukhov:2005]. While the temperature and the pulse length were changed during the PW experiments, shown below, the repetition time of microwave pulses was always set to 200 ms.
### Measurements with conical waveguide
![\[Tdeptau\]Temperature dependence of relaxation time $\tau_{short}$ measured for a small sample with a conical waveguide. The microwave pulses of repetition time of 200 ms and of different lengths were applied: 0.05 ms (dotted curve), 0.1 ms (solid curve), 0.5 ms (dashed curve), 1 ms (dashed-dotted curve), 5 ms (bold solid curve), and 10 ms (bold dashed curve). Color online.](Figure5_small.eps){width="3.4in"}
Measurements with a conical waveguide are performed on a tiny sample placed in *setup I* configuration, as shown in Fig. \[wguides\]a. The volume of the sample is 150 $\times$ 100 $\times$ 30 $\mu$m$^{3}$. We have investigated the $\tau_{short}$ and $\tau_{long}$ relaxation times as a function of temperature $T$ for different values of pulse length. The results obtained from such PW experiments for the short relaxation time $\tau_{short}$ are depicted in Fig. \[Tdeptau\].
The fast relaxation shows the relaxation time $\tau_{short}$ of the order of magnitude of 1-1.5 ms. The temperature dependence of $\tau_{short}$ is not strong, but is clearly pronounced: on warming from 2 K to approximately 4 K the relaxation time $\tau_{short}$ decreases, reaches its minimum near 4 K, and then smoothly increases on further warming (see Fig. \[Tdeptau\]). The decrease of relaxation time $\tau_{short}$ on warming from 2 K to approximately 4 K is much steeper for shorter pulses, while the $\tau_{short}$ behavior above 4 K is similar for all the pulse length values.
![\[relax\]Generic plot depicting temperature dependence of fast relaxation time $\tau_{short}$ and slow relaxation time $\tau_{long}$ for different sample sizes and different sample’s irradiation configurations: *setup I* (solid curves), *setup II* for a big and a small sample (big and small solid circles, respectively), *setup III* (open triangles), and *setup IV* (solid rhombus with dotted curve). Note that the Latin numbers stand for corresponding setup notation.](Figure6_small.eps){width="3.0in"}
For the same sample and experimental configuration *setup I*, the slow relaxation time $\tau_{long}$ is approximately one order of magnitude greater, that is, its value lies around 20 ms. This long-lasting relaxation time is temperature independent within 5% in the measured temperature range, as shown in Fig. \[relax\], where both $\tau_{short}$ and $\tau_{long}$ obtained from the 10 ms data using *setup I* configuration are shown by solid curves for comparison. This behavior of $\tau_{long}$ perfectly fits into the above given explanation of $\tau_{long}$-process as a cooling of the sample after the microwave pulse. Such a cooling rate is only defined by the sample thermal coupling to the bath, and therefore, it is permanent for a given experimental setup.
Figure \[relax\] shows the generic plot of relaxation times $\tau_{short}$ and $\tau_{long}$, obtained from all the experimental configurations at 10 ms pulse length over 200 ms pulse repetition time. Such a rather long pulse length was chosen mainly due to the following reasons. At first, at rather long pulses the contribution of the $\tau_{long}$-process, i.e. the sample thermalization after the pulse, becomes valuable, making better separation of $\tau_{short}$ and $\tau_{long}$ data intervals possible, and thus both fast and long relaxations can be estimated and compared for the same system under the same experimental condition. Secondly, at long pulses, a high signal-to-noise ratio enables better fitting, even for the limited regions of experimental data. Finally, as will be shown below, at such long pulses there are no *overshooting* phenomena observed for both small and big samples at all the temperatures applied; with the presence of overshooting, the analysis of magnetization data in terms of relaxation exponents becomes controversial.
### Measurements with a rectangular waveguide
The irradiation of the sample with a piece of the rectangular waveguide WR6 is advantageous in the sense that the distribution of electromagnetic field is known at the waveguide cut edge. Another advantage of the use of the WR6 waveguide is that the area of its opening (1.36 mm$^2$) is approximately 9 times smaller than the area of the 4 mm opening of the conical waveguide, so we could expect a higher density of microwaves exposed to the sample. On measurements employing rectangular waveguide, *setup II* configuration of irradiating the sample with microwave radiation is used. In this configuration, the sample is placed in the geometrical center of the waveguide opening, as schematically shown in Fig. \[wguides\]b. *Setup I* configuration provides the point of maximal magnetic field for the propagating TE$_{10}$ mode in a rectangular waveguide [@poole].
In our magnetization study employing *setup II* configuration, we have used two samples of different volumes: a big sample with a volume 680$\times$570$\times$170 $\mu$m$^3$ and a smaller sample with a volume of 160$\times$180$\times$100 $\mu$m$^3$. We will refer these samples hereafter as *big* and *small*, correspondingly. The measurements were performed upon irradiation with pulsed microwaves of 118 GHz and at applied magnetic field of 0.2 T, these conditions correspond to the first transition from the ground state -10 $\rightarrow$-9 for Fe$_8$ system along the easy axis. We have found that the relaxation time $\tau_{short}$ for the *small sample* behaves very similar to the relaxation time $\tau_{short}$ of the sample used with *setup I*. Indeed, this relaxation time decreases during the temperature increase from 2 K to 4 K, where it reaches its minimal value of around 1.5 ms, as shown by the small solid circles in Fig. \[relax\]. On the consequent increase of the temperature above 4 K, we see the increase of $\tau_{short}$ with temperature $T$. The $T$-dependence of the fast relaxation parameter $\tau_{short}$ for the *big* sample is qualitatively similar to that of the *small* sample, the corresponding data is depicted with big solid circles in Fig. \[relax\]. Here, the relaxation time also reveals a minimum at around 4 K, but the absolute values of $\tau_{short}$ for the *big* sample are approximately four times higher, and both dependences can be perfectly scaled one onto another. The $\tau_{long}$ values for both the *small* and *big* samples are shown in Fig. \[relax\] by the small and big solid circles, respectively; both these relaxation times lie around 25-30 ms and they are temperature independent, similar for the earlier described configuration *setup I*. This is not surprising, since in both configurations of coupling of the samples to the electromagnetic (EM) radiation, the surrounding of the sample was the same: we have used the same sample holder in both cases, and the same amount of exchange gas was contained in the sample volume chamber for the sample’s thermalization. This was not the case when the sample was placed onto the interior wall of the massive copper cavity resonator.
![\[cavity\]Magnetization loops measured at several cw microwave frequencies using a cavity. Note, that the curves are equally scaled.](Figure7_small.eps){width="3.0in"}
### Measurements with a cavity
![\[ST\]Spin temperature calculated during a pulse of length of 0.1 ms at nominal cryostat temperature $T=2$ K for the different ways of irradiating sample with microwaves. Color online.](Figure8_small.eps){width="3.4in"}
In order to increase the amplitude of the EM radiation exposed to the sample, we have used the cylindrical copper cavity resonator, construction details of which are given above. The corresponding configuration is denoted as *setup IV* and it is shown in Fig. \[wguides\]c. The use of a microwave cavity assumes the use of different modes compatible with the cavity geometry and the microwave frequency. The general problem of the cavity usage is that the modes of the cavity are coupled to the sample in a different way, i.e. the electromagnetic environment of the sample inside the cavity is strongly dependent on the mode and the frequency. This leads not only to different amplitudes of exposed (and therefore, absorbed) microwaves, but also to the irregularities of the absorption spectra. Figure \[cavity\] shows the microwave absorption spectra of magnetization obtained at several frequencies, separated just 1.5 GHz from each other; the spectra are obtained by the use of cw microwaves with *setup IV*. The sample used for the studies with the cavity has dimensions of 680$\times$570$\times$170 $\mu$m$^3$. Unfortunately, when a smaller sample is used in the cavity-employed configuration *setup IV*, the sensitivity is reduced drastically. This is due to the fact that the sample’s change of magnetization is sensed by the Hall bar separated by a copper foil, as described above.
It can be seen that the modes differ not only in the amplitude of absorption peaks, but some of them are also highly distorted (peaks instead of dips, amplitude-phase mixing, asymmetry for opposite field directions). For the magnetization relaxation measurements we choose the frequency of 118 GHz, since the mode working at this frequency shows the largest amplitudes of absorption peaks in the positive magnetic-field direction; so, all the measurements presented below are taken at the frequency of 118 GHz.
In general, we find that there is no increase of the EM-field amplitude exposed on the sample in comparison with the case, when the sample is irradiated with microwaves without a cavity. The best way to quantitatively characterize the amount of EM radiation (photons) absorbed by the sample is to consider the spin-temperature $T_{\rm S}$ growth due to the exposure of a microwave pulse. Using the mapping procedure described above, we convert the magnetization data obtained at a pulse length of 0.1 ms into $T_{\rm S}$ for the different configurations of coupling of the sample to the microwaves. The plot, representing spin temperature $T_{\rm S}$ calculated when no cavity is used (*setup I* and *setup II*) and in cavity-employed configuration (*setup IV*), is depicted in Fig. \[ST\]. The plot presented in Fig. \[ST\] shows no evidence of enhancement of absorption of microwaves, when the resonant cavity is used, although the best performing mode is chosen for this spin temperature comparison. Instead, both waveguide-employed configurations clearly show a better performance. The same sample is used for *setup II* and *setup IV* configurations, while the smaller sample is used in configuration *setup I*; details of the sample’s size are given above.
The use of a cavity also significantly extends the relaxation of magnetization after the microwave pulse. This slowing of magnetization restoration is also clearly visible from the spin-temperature dynamics, shown in Fig. \[ST\], as compared to the experiment without a cavity *setup II* on the same sample. We have summarized the relaxation parameters when the cavity was used on the generic plot shown in Fig. \[relax\] and compared these relaxation time values to the parameters, obtained from the magnetization restoration during pulsed microwave measurements on the same sample without cavity. The relaxation time data obtained during cavity-employed experiment are denoted as “$\tau$, *IV*” on Fig. \[relax\]. The obtained relaxation time $\tau$ is around 100-200 ms, which is one order of magnitude larger than the slow relaxation time $\tau_2$, obtained in no-cavity experiments. We cannot unambiguously attribute the obtained relaxation time $\tau$ to the fast relaxation $\tau_{short}$, although the data for its calculation were taken right after the pulse, as what was done for the $\tau_{short}$ definition in no-cavity setups. We think that the calculated relaxation time $\tau$ rather corresponds to the mixture of $\tau_{short}$ and cooling of the sample thermally coupled to the massive copper cavity, i.e. the relaxation time $\tau_{long}$.
### Background absorption
 Relaxation time $\tau_{short}$ and $\tau_{long}$ as a function of applied magnetic field in PW measurements of pulse length of 10 ms and frequency $f=118$ GHz. The measurements were performed using *setup I* at several temperatures. (b) Relaxation time $\tau_{short}$ and $\tau_{long}$ as a function of applied magnetic field at temperature $T=2$ K and frequency $f=118$ GHz. The measurements were performed using *setup I* at several pulse length values: 0.5 ms (solid curve), 1 ms (dotted-dashed curve), 2 ms (dotted curve), 5 ms (bold solid curve), and 10 ms (dashed curve). The repetition time of the pulses was 200 ms. Color online.](Figure9_small.eps){width="3.4in"}
All the measurements, described above are performed at a resonance condition corresponding to the transition from the ground state, i.e. -10 $\rightarrow$ -9. At 118 GHz, a frequency which is used for current studies, the appropriate applied magnetic field is always set to 0.2 T, and thus the resonant condition is fulfilled for the Fe$_8$ system along the easy axis. Thus, only the resonant absorption is detected. Nevertheless, as we have mentioned above, there is also a significant off-resonant, or background, absorption.
In Fig. \[backgrnd\] we present the relaxation time $\tau_{short}$ and $\tau_{long}$ as a function of the applied magnetic field. The magnetic field is set in discrete steps from zero field to 1.5 T with an increment of 0.05 T. These relaxation times are calculated from the PW measurements performed using configuration *setup I* at several temperatures \[Fig. \[backgrnd\]a\] and at several pulse length values \[Fig. \[backgrnd\]b\].
Both figures show that the long relaxation time $\tau_{long}$ remains pulse length and field independent within the noise bandwidth and equals to approximately 25 ms; there is no resonance structure evidenced in $\tau_{long}$ field-dependence. This is consistent with our consideration of the slow relaxation as a cooling of the system. Note that above approximately 0.5 T the magnetization deviation amplitudes are reduced for short pulse values and the corresponding $\tau_{long}$ curves become very noisy in Fig. \[backgrnd\]b.
$\tau_{short}$ follows the resonance behavior and clearly pronounced resonant dips can be seen in both figures. We see that off-resonance and in-resonance relaxation time $\tau_{short}$ values lie within the same order of magnitude (around 1-3 ms), while $\tau_{long}$ is one order of magnitude larger. For the pulse length value of 0.5 ms \[Fig. \[backgrnd\]b\], the difference between the in-resonance value at $H=0.2$ T and the off-resonance value of $H=0.38$ T is a factor of 2, while for the pulses with a duration of 10 ms this factor is reduced to 1.2.
Magnetization overshooting
--------------------------
![\[over3pulse\]Typical plots of PW measurements of magnetization as a function of time performed on *big* (solid circles) and on *small* (open circles) samples. The data plotted are taken at 5 ms (a), 0.5 ms (b), and 50 $\mu$s (c) right after the microwave pulse was switched off. Color online.](Figure10_small.eps){width="3.4in"}
![\[over4freq\]Position of the magnetization overshooting minima as a function of pulse length for the big (solid symbols) and for the small (open symbols) samples. The measurements were done at temperatures of 1.8 K (squares), 2 K (circles), 2.5 K (triangles), 3 K (diamonds), and 4 K (stars). Color online.](Figure11_small.eps){width="3.4in"}
As we have mentioned above, in some of our PW experiments we observed an *overshooting* of the magnetization after the microwave pulse, when the magnetization continued to decrease even after the pulse is switched off. This phenomenon, for example, can be clearly seen in the magnetization data recalculated into spin temperature for *setup II* and *setup IV*, as shown in Fig. \[ST\]. A similar effect is also evidenced in the work of Bal *et al.* [@bal:epl2005]. Such an *overshooting*, however, is not observed in magnetization measurements employing *setup I*, where we have performed pulsed microwave experiments with pulses of the length of 10 ms down to 1 $\mu$s. Note that two different samples are used with *setup II*, and the volume of the small sample used in this study is approximately 6 times larger than the volume of the crystal, used in the measurement with *setup I*.
In measurements employing rectangular waveguide, two configurations *setup II* and *setup III* are possible. In the former configuration, the sample is placed in the geometrical center of the waveguide opening; in the latter configuration, the sample is placed at the midpoint of the shortest wall of the waveguide, as schematically shown in Fig. \[wguides\]. Both configurations provide the points of maximal magnetic field for the propagating TE$_{10}$ mode in a rectangular waveguide [@poole]. In *setup II* configurations, a sample of any size can be used, while only a sample of large enough volume can be used in *setup III*, because a tiny sample cannot be properly placed at the edge of the waveguide for partial irradiation with microwaves.
In order to find the nature of the overshooting of magnetization restoration after the microwave pulse, we employ *setup III* configuration, which allows partial irradiation of the sample with microwaves. We construct a sample holder, where *setup II* and *setup III* can be used simultaneously, i.e. two samples can be exposed to the microwaves at the same time. The change of each sample’s magnetization can be sensed by an individual Hall bar mounted underneath the sample.
For measurements we use two different size samples of Fe$_8$: one sample, hereafter referred to as *small*, had dimensions of 160$\times$180$\times$100 $\mu$m; and another sample, hereafter referred to as *big*, had dimensions of 680$\times$570$\times$170 $\mu$m. The *small* sample was placed in *setup II* and was entirely irradiated with microwave radiation. The *big* sample was placed in configuration *setup III* and only a part of it was exposed to the microwaves; the Hall bar was placed under the “dark” part of the big sample. Both samples were mounted with their easy axes parallel to the direction of the applied magnetic field set to the value of 0.2 T, which corresponds to the first transition -10 $\rightarrow$ -9 at the frequency of 118 GHz.
The typical oscillograms of pulsed microwave measurements performed at temperature $T=2$ K and pulse durations of 5 ms, 0.5 ms, and 50 $\mu$s for the *big* and the *small* samples are presented in Fig. \[over3pulse\]; and the data plotted are taken right after the microwave pulse was switched off. As seen in Fig. \[over3pulse\]a, after rather long pulses of duration of 5 ms, no magnetization overshooting is observed for both the *big* and *small* samples. At ten times shorter pulse length of 0.5 ms, the magnetization restoration of the *small* sample reveals no overshooting feature, while the magnetization of the *big* sample continues to decrease, reaches the minimum at approximately 0.6 ms after the microwave pulse is switched off, and only then increases and saturates to the equilibrium value \[see Fig. \[over3pulse\]b\]. When the microwaves are applied within the pulses of length of 50 $\mu$s, the magnetization data of the *big* sample show even more overshooting: the minimum of the magnetization curve is observed approximately 1.2 ms after the pulse edge, see Fig. \[over3pulse\]c. At the same time, the *small* sample magnetization data also show an appearance of overshooting having its minimum at 0.2 ms after the pulse edge, as can be evidenced from Fig. \[over3pulse\]c.
By performing a series of similar PW experiments in a broader range of microwave pulse length values and at several helium temperatures, we obtain a generic plot, depicted in Fig \[over4freq\]. Here, the position of magnetization minima, i.e. the overshooting time, is plotted as a function of the applied microwave pulse length values at the resonance condition of the transition -10$\rightarrow$-9 (118 GHz, 0.2 T). The measurements are done consequently on the *big* and on the *small* sample at same temperature and pulse length values. From Fig. \[over4freq\] it can be clearly seen, that the *big* sample shows well pronounced overshooting already at pulse lengths of 1 ms and above, while shorter pulses of the length of approximately 100 $\mu$s are needed in order to observe measurable overshooting of the magnetization of the *small* sample. Another interesting finding, which can be concluded from the dependencies shown in Fig \[over4freq\] is that the overshooting time strongly decreases with the temperature increase. This temperature dependence is very intense: by increasing the temperature from 1.8 K to 2.5 K, the overshooting time is reduced twice in its value. At high enough temperatures the overshooting feature completely disappears for both samples. The pulse length dependence of the overshooting time can also be easily understood in terms of its strong temperature dependence, since the PW configuration of experiments, as well as that of cw, leads to the significant heating of the system, as was shown previously [@petukhov:2005]. Heating with microwaves perfectly explains why less overshooting is observed at longer pulses than at shorter pulses.
We have also estimated the relaxation time $\tau_{short}$ for the *big* sample used in *setup III* configuration. The corresponding data are plotted by the big open triangles in Fig. \[relax\]. We observe, that the relaxation time $\tau_{short}$ is very similar to $\tau_{short}$ values obtained in other cavity-free configurations. In particular, the profile of the relaxation time $\tau_{short}$ as a function of temperature for the *big* sample measured with *setup III* is similar to that of the same sample measured with *setup II* (big solid circles in Fig. \[relax\]). The corresponding absolute values are very close and the difference between the two curves of 3 ms can be explained by the partial irradiation of the sample with microwaves in *setup III*. Thus, only a fraction of molecules contributes to the magnetization change, and we can consider that the “effective” size of the sample is smaller.
Discussion
==========
The magnetization dynamics measurements presented in this work intend to define some characteristic relaxation times, which should be taken into account when the spin dynamics of Fe$_{8}$ SMM is considered. In particular, we have investigated magnetization recovery right after the microwave pulse, where the spin-phonon relaxation time $\tau_1$ can contribute to the magnetization relaxation. We have found that the after-pulse fast relaxation $\tau_{short}$ is typically on the order of magnitude of several milliseconds, as can be seen in Fig. \[relax\]. It was found that the lower limit for $\tau_{short}$ is $1.4 \cdot
10^{-3}$ s, which is orders of magnitude larger than the longitudinal relaxation time $\tau_1$, which is expected to be $\sim 10^{-7}$ s [@Wernsdorfer:EPL2000]. Such an obvious discrepancy shows that $\tau_1$ contribution to the $\tau_{short}$-process is not major at the conditions of the performed experiments. Indeed, the temperature behavior of $\tau_{short}$ is also incompatible with the expected temperature behavior of $\tau_{1}$, which should decrease with temperature growth.
One of the dominant contributions to the $\tau_{short}$ relaxation can be the phonon-bottleneck effect, which can screen out the shorter relaxations, such as $\tau_1$. Within the model described above, we have performed the magnetization data treatment by means of the long relaxation time $\tau_{long}$, which is believed to be characteristic for the cooling of the specimen after the microwave pulse. The values of $\tau_{long}$ were found to be an order of magnitude higher than $\tau_{short}$, typically around 30-50 ms. We have also not evidenced any temperature (see Fig. \[relax\]) or magnetic-field (see Fig. \[backgrnd\]a) dependence of $\tau_{long}$. It can be noticed from Fig. \[relax\] that $\tau_{long}$ has a pronounced sample size dependence: for the larger sample $\tau_{long}$ shown by the big solid circles lies above the $\tau_{long}$ data for the smaller sample, depicted by the small solid circles. Another interesting observation is that $\tau_{long}$ has a prominent power dependence: as can be concluded from Fig. \[backgrnd\]b, the longer pulses provide large $\tau_{long}$ than shorter pulses. There is nearly a factor of 2 difference between the $\tau_{long}$ data obtained after a pulse with durations of 0.5 ms and 10 ms. Thus, we attribute the obtained $\tau_{long}$ relaxation time to the phonon relaxation time from the crystal to the heat bath $\tau_{ph}$. Our values are in good agreement with previously published literature values [@Chiorescu:PRL2000].
Another observation, which can support the idea that $\tau_{ph}$ is admixed to the $\tau_{1}$ data is that values of $\tau_{1}$ obtained at nonresonant and resonant conditions are rather similar, as shown in Fig. \[backgrnd\]. Although the modulation due to the resonant absorption can be clearly seen from the data, the $\tau_1$ values obtained in resonance and out of resonance differ only by a factor of two. Also, from the plot in Fig. \[backgrnd\]b it can be seen that $\tau_{short}$ and $\tau_{long}$ relaxation times experience fairly similar power dependence: an extension of both relaxation times is observed for longer pulses. This pulse length (or power) dependence of $\tau_{short}$ can be a plain evidence that a process of cooling of the crystal contributes to $\tau_{short}$ too.
We have also found another pertaining to time process, which builds the overall profile of the magnetization restoration curve, as sketched in Fig. \[OSC\]. As it was shown in comparative experiment on small and big samples (*setup II* and *setup III*), under certain physical conditions an overshooting can be observed in magnetization dynamics. The generic plot depicted in Fig. \[over4freq\] shows the mapping of the occurrence of the overshooting. It is shown in Fig. \[over4freq\] that the sample’s size and the sample’s temperature are two factors responsible for the phenomenon of overshooting in the following way: the larger the sample and the lower the temperature, the more prominent the overshooting. It perfectly explains why no overshooting is evidenced in our previous work [@petukhov:2005] and in investigations employing *setup I* in this work: we have used a very small sample, the volume of which was approximately 6 times less than the volume of the *small* sample used to make the plot in Fig. \[over4freq\].
Such a spatial effect, which also depends on the sample’s spin temperature, can be described by the sample’s thermal spin equilibration. In terms of spin language, this process is known as *spin diffusion*: for the system of identical spins, where the level population at one part in the sample is different from those at other points, the spin flip process will act to make the population difference uniform throughout the specimen.[@abragam] Thus, spin diffusion creates a uniform spin temperature throughout the sample. The presence of spin diffusion is a sequence of the fact that we measure an array of magnetic molecules, and the spin interactions between them are presented. Here one can see that the term “single” in the SMM notation is, to a certain extent, an idealization. In the strict sense, the spin diffusion is completely inevitable until one single molecule is measured.
As can be seen from Fig. \[over4freq\], the overshooting time is comparable to the pulse length. For rather big samples it can be in the order of magnitude of several milliseconds, which is already comparable to the values of $\tau_{short}$. Therefore, experimental conditions should be chosen carefully for such pulsed microwave experiments. Ideally, one should employ a smallest possible sample; then even shorter microwave pulses can be utilized than those depicted in Fig. \[over4freq\]. Nevertheless, to perform microsecond and submicrosecond pulsed microwave measurements one needs a higher power of microwaves.
As can be seen from Fig. \[ST\], the use of a microwave resonator cannot serve to reach this goal (*setup IV* in Fig. \[ST\]). The problem is that microwaves can only be guided to the cavity by a standard rectangular waveguide, for which the electromagnetic-field distribution of propagating mode is known and the effective magnetic coupling via coupling hole is possible at the position of the magnetic-field antinode. Employing such a configuration, we produce unavoidable losses due to the transition from the oversized circular waveguide to the WR6 rectangular waveguide and thus reduce the overall performance of the use of a cavity. For the same reason the configuration *setup II* is less advantageous as configuration *setup I*, as shown in Fig. \[ST\]: the use of a circular-to-rectangular transition leads to high losses. Therefore, there is no gain in the use of better focusing lower-cross-section rectangular waveguide.
Conclusions
===========
We have presented the magnetization dynamics experiments employing magnetization measurements combined with pulsed microwave absorption measurements. The analysis of the magnetization dynamics is performed in terms of characteristic exponents, which describe the fast and slow components of magnetization relaxation. These exponents are physically connected to different contributions to the overall magnetization dynamics. We have found that the spin-phonon relaxation time $\tau_1$ is screened out by other longer-lasting relaxations. The phonon-bottleneck effect is probably the major contribution to the magnetization relaxation, giving a slow relaxation. We have found that the phonon relaxation time $\tau_{ph}$ is around 30 ms in our experiments, which is comparable to other studies.[@Chiorescu:PRL2000] We have also evidenced the effect of spin diffusion inside the specimen, which should be taken into consideration, when after-pulse magnetization dynamics is analyzed.
Finally, we can propose that more advanced microwave experiments are needed to resolve the spin-phonon relaxation time $\tau_1$, such as the “pump and probe” technique employing two frequencies of pulsed microwaves. But in all cases special care should be taken concerning the sample’s coupling to the microwaves and to the phonon bath.
Acknowledgments
===============
We thank R. Sessoli and L. Sorace for helpful discussions. The samples for the investigations were kindly provided by A. Cornia. This paper is partially financed by EC-RTN-QUEMOLNA Contract No. MRTN-CT-2003-504880.
[10]{}
M. Novak and R. Sessoli, in: L. Gunther and B. Barbara (eds.): [*Quantum Tunneling of Magnetization-QTM’94*]{}, Vol. 301 (2005) of [*NATO ASI Series E: Applied Sciences*]{}. London: Kluwer Academic Publishers, pp. 171–188.
Jonathan R. Friedman, M. P. Sarachik, J. Tejada, and R. Ziolo, Phys. Rev. Lett. [**76**]{}, 3830 (1996).
L. Thomas, F. Lionti, R. Ballou, D. Gatteschi, R. Sessoli, and B. Barbara, Nature (London) [**383**]{}, 145 (1996).
E. del Barco, A. D. Kent, E. C. Yang, and D. N. Hendrickson, Phys. Rev. Lett. [**93**]{}, 157202 (2004).
L. Sorace, W. Wernsdorfer, C. Thirion, A.-L. Barra, M. Pacchioni, D. Mailly, and B. Barbara, Phys. Rev. B [**68**]{}, 220407(R) (2003).
W. Wernsdorfer and R. Sessoli, Science [**284**]{}, 133 (1999).
W. Wernsdorfer, M. Soler, G. Christou, and D. N. Hendrickson, J. Appl. Phys. [**91**]{}, 7164 (2002).
W. Wernsdorfer, R. Sessoli, A. Caneschi, D. Gatteschi, A. Cornia, and D. Mailly, J. Appl. Phys. [**87**]{}, 5481 (2000).
M. N. Leuenberger and D. Loss, Nature (London) [**410**]{}, 789 (2001).
D. Zipse, J. M. North, N. S. Dalal, S. Hill, and R. S. Edwards, Phys. Rev. B [**68**]{}, 184408 (2003).
A. Mukhin, B. Gorshunov, M. Dressel, C. Sangregorio, and D. Gatteschi, Phys. Rev. B [**63**]{}, 214411 (2001).
K. Petukhov, W. Wernsdorfer, A.-L. Barra, and V. Mosser, Phys. Rev. B [**72**]{}, 052401 (2005).
M. Bal, J. R. Friedman, Y. Suzuki, E. M. Rumberger, D. N. Hendrickson, N. Avraham, Y. Myasoedov, H. Shtrikman, and E. Zeldov, Europhys. Lett. [**71**]{}, 110 (2005).
W. Wernsdorfer, A. Müller, D. Mailly, and B. Barbara, Europhys. Lett. [**66**]{}, 861 (2004);
W. Wernsdorfer, D. Mailly, G. A. Timco, and R. E. P. Winpenny, Phys. Rev. B [**72**]{}, 060409 (2005).
B. Cage, S. E. Russek, D. Zipse, J. M. North, and N. S. Dalal, Appl. Phys. Lett. [**87**]{}, 082501 (2005).
M. Bal, Jonathan R. Friedman, Yoko Suzuki, K. M. Mertes, E. M. Rumberger, D. N. Hendrickson, Y. Myasoedov, H. Shtrikman, N. Avraham, and E. Zeldov, Phys. Rev. B [**70**]{}, 100408(R) (2004).
A.-L. Barra, P. Debrunner, D. Gatteschi, Ch. E. Schulz, and R. Sessoli, Europhys. Lett. [**35**]{}, 133 (1996).
R. Caciuffo, G. Amoretti, A. Murani, R. Sessoli, A. Caneschi, and D. Gatteschi, Phys. Rev. Lett. [**81**]{}, 4744 (1998).
K. Park, M. A. Novotny, N. S. Dalal, S. Hill, and P. A. Rikvold, Phys. Rev. B [**66**]{}, 144409 (2002).
C. Sangregorio, T. Ohm, C. Paulsen, R. Sessoli, and D. Gatteschi, Phys. Rev. Lett. [**78**]{}, 4645 (1997).
A. Abragam and B. Bleaney, [*Electron Paramagnetic Resonance of Transition Ions*]{} (Clarendon Press, Oxford, 1970).
I. Chiorescu, W. Wernsdorfer, A. M${\rm \ddot u}$ller, H. B${\rm
\ddot o}$gge, and B. Barbara, Phys. Rev. Lett. [**84**]{}, 3454 (2000).
W. Wernsdorfer, A. Caneschi, R. Sessoli, D. Gatteschi, A. Cornia, V. Villar, and C. Paulsen, EuroPhys. Lett. [**50**]{}, 552 (2000).
S. Hill, S. Maccagnano, Kyungwha Park, R. M. Achey, J. M. North, and N. S. Dalal, Phys. Rev. B [**65**]{}, 224410 (2002).
Ch. P. Poole, [*Electron Spin Resonance: A Comprehensive Treatise on Experimental Techniques*]{} (Wiley, New York, 1983).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'A. Aguado-Barahona'
- 'R. Barrena'
- 'A. Streblyanska'
- 'A. Ferragamo'
- 'J.A. Rubiño-Martín'
- '[D. Tramonte]{}'
- '[H. Lietzen]{}'
bibliography:
- 'LP15\_year2.bib'
title: 'Optical validation and characterization of [*Planck* ]{}PSZ2 sources at the Canary Islands observatories. II. Second year of LP15 observations'
---
[The second legacy catalogue of Planck Sunyaev-Zeldovich (SZ) sources, hereafter PSZ2, provides the largest galaxy cluster sample selected by means of their SZ signature in a full sky survey. In order to fully characterise this PSZ2 sample for cosmological studies, all the members should be validated and the physical properties of the clusters, including mass and redshift, should be derived. However, at the time of its publication roughly 21 per cent of the 1653 PSZ2 members had no known counterpart at other wavelengths. ]{} [ Here, we present the second and last year of observations of our optical follow-up programme 128-MULTIPLE-16/15B (hereafter LP15), which has been developed with the aim of validating all the unidentified PSZ2 sources in the northern sky, with declination above $-15\degree$, and with no correspondence in the first Planck catalogue PSZ1. The description of the programme and the first year of observations were presented in Streblyanska et al. (2019). ]{} [ The LP15 programme was awarded 44 observing nights, spread over two years in the Isaac Newton Telescope (INT), the Telescopio Nazionale Galileo (TNG) and the Gran Telescopio Canarias (GTC), all at Roque de los Muchachos Observatory (La Palma). Following the same methodology described in Streblyanska et al. (2019), at the end of the LP15 programme we performed deep optical imaging for more than 200 sources with the INT, and spectroscopy for almost 100 sources with the TNG and GTC. We adopted a robust confirmation criteria based on velocity dispersion and richness estimations in order to carry out the final classification of the new galaxy clusters as the optical counterparts of the PSZ2 detections. ]{} [Here, we present the observations of the second year of LP15, as well as the final results of the programme. The full LP15 sample comprises 190 previously unidentified PSZ2 sources. Of those, 106 objects were studied in Streblyanska et al. (2019), while the remaining sample (except for 6 candidates) has been completed in the second year and it is discussed here. In addition to the LP15 sample, in this paper we have studied 42 additional PSZ2 objects, which were originally validated as real clusters due to their matching with a WISE or PSZ1 counterpart, but they had no measured spectroscopic redshift. In total, we have confirmed the optical counterpart for 81 PSZ2 sources after the full LP15 programme, 55 of them with new spectroscopic information. Out of those 81 sources, 40 clusters are presented in this paper. After the LP15 observational programme the purity of the PSZ2 catalogue has increased from $76.7\,\%$ originally to $86.2\,\%$. In addition, we study the possible reasons of having false detection, and we report a clear correlation between the number of unconfirmed sources and galactic thermal dust emission. ]{}
Introduction
============
Galaxy Clusters (GCs) are a very powerful tool to test the cosmological model [@Allen11]. They have been used for cosmology studies since Zwicky’s discovery of dark matter in the Coma Cluster [@Zwicky33], and thanks to the recent all-sky surveys appearing during the last decade, their study has increased significantly. In particular, the [*Planck* ]{}[^1] satellite [@Planck14I] provided for the first time the possibility for the detection of GCs on a full-sky survey by means of the Sunyaev-Zeldovich (SZ) effect [@SZ72], which is Cosmic Microwave Background (CMB) spectral distortion. As CMB photons go through a GC, they can interact with the hot intra-cluster medium (ICM) electrons via inverse Compton scattering. As a result of this interaction, the CMB photons gain energy and the overall CMB spectrum is shifted towards higher frequencies, producing a characteristic spectral dependence which can be used for their detection.
In order to use the GCs surveys to constrain cosmological parameters , it is important to obtain unbiased measurements of the cluster mass and redshift. In particular, due to the fact that the surface brightness of the SZ effect does not depend on redshift, follow-up campaigns at other wavelengths are needed in order to complement the SZ information. The detailed characterisation of [*Planck* ]{}SZ survey cluster sample is needed to improve cosmological constraints from [*Planck* ]{}survey alone, and also to obtain constraints from future large galaxy cluster surveys, e.g. SRG/eROSITA cluster survey [@eRosita], whose spacecraft was successfully launched on July 13th and the first light of SRG/eROSITA telescope is planned to be obtained very soon.
The PSZ2 catalogue [@PSZ2] is the second [*Planck* ]{}catalogue of Sunyaev-Zeldovich sources derived from the full 29 months mission data. This catalogue is based on the results from three cluster detection codes (MMF1, MMF3 and PwS) described in detail in [@PSZ1; @PSZ2]. The PSZ2 was validated using external X-ray, optical, SZ and near-IR data and contained, at the time of its publication, 1653 detections of which 1203 were confirmed as actual galaxy clusters, 1094 of them including redshifts. The PSZ2 catalogue contains all objects found by at least one of the three detection algorithms with $S/N \geq
4.5$ for the SZ detection.
A first validation process was performed in [@PSZ2]. It began using a cross-match with the PSZ1, continuing the search for possible counterparts in the MCXC catalogue [@Piffaretti11] which is based on the *ROSAT* All Sky Survey [RASS, @Voges99; @Voges00] and on the serendipitous *ROSAT* and Einstein cluster catalogues, in the Sloan Digital Sky Survey [SDSS @York00], in the redMaPPer catalogue [@Rykoff14], in NED[^2] in similar follow-ups [@PcolintXXVI; @paper1] as the one presented in this work, in the AllWISE mid-infrared source catalogue [@Cutri13] as well as in SZ catalogues such as the catalogues obtain by the South Pole Telescope [SPT, @Bleem15], by the Atacama Cosmology Telescope [ACT, @Hasselfield13] and by direct follow-up with the Arc-minute Micro-kelvin Interferometre [AMI, @Perrott15].
This paper is the last of a series of papers where optical characterisation of SZ sources was performed using the Canary Islands Observatories. @paper1 [@paper2] study PSZ1 sources thanks to the observational programme [ITP13-08]{}. Later, a second long-time programme was granted [128-MULTIPLE-16/15B]{} (hereafter [LP15]{} ) to study the PSZ2 sources. @paper4 [hereafter Paper I] and this paper are the result of this last programme. The main motivation of these follow-up campaigns is to identify and confirm optical cluster counterparts of unknown sources. We perform photometric and spectroscopic observations in order to study the optical richness and estimate velocity dispersion. We leave the mass estimates for future works.
This paper is structured as follows. Sections \[sec:observations\] and \[sec:criteria\] describe the observational programme carried out for this work, including the instrumentation setups, the observational approach as well as the criteria used to validate candidates as optical counterparts for SZ sources. Section \[sec:2ndyear\] presents the observations for the 2nd year of this programme and details some special cases. In Section \[sec:wise\] we update information about already known counterpart validated using ALLWISE [@Cutri13]. Section \[sec:fullpsz2\] summarises the whole programme and gives the final results for this work. Section \[sec:conclusions\] presents the conclusions.
We adopt $\Lambda$CDM cosmology with $\Omega_{\rm m} = 0.3075$, $\Omega_{\Lambda} = 0.691$ and H$_0 = 67.74$kms$^{-1}$Mpc$^{-1}$.
LP15 optical follow-up campaign {#sec:observations}
===============================
Sample definition and observational strategy
--------------------------------------------
The [LP15]{} programme, the sample definition and the observing strategy are described in detail in paper I. Here, we briefly summarise its basic characteristics and present the second year of observations. As a reminder, the main motivation of the programme is to carry out a systematic follow-up of the complete set of PSZ2 cluster candidates in the northern sky, with no confirmed counterparts at the moment of the catalogue publication. For LP15, we consider only sources located at declination $> -15\degree$. As a reference, the full PSZ2 catalogue contains a total set of 1003 sources in that region.
The [LP15]{} sample is defined by all those sources in PSZ2 with declination above $-15\degree$, which also have [validation = -1]{} (i.e., no known counterpart at the time of the publication of the catalogue), and [PSZ1 = -1]{} (i.e. no matching detection in the PSZ1). This corresponds to 190 targets in total, 106 of which were already discussed in paper I.
During the second year of the programme, covering the semesters 2016B and 2017A, we continued with the observations of the remaining [LP15]{} sample. As for the previous year, the second year of observations were carried out at the Roque de los Muchachos Observatory (ORM) located at La Palma island (Spain) between August 2016 and August 2017. The three telescopes used in this work are: a) the 2.5m Isaac Newton Telescope (INT) operated by the Isaac Newton Group of Telescopes; b) the 3.6m Italian Telescopio Nazionale Galileo (TNG) operated by the Galileo Galilei Foundation of the INAF (Istituto Nazionale di Astrofisica); and c) the 10.4m Gran Telescopio Canarias (GTC) operated by the Instituto de Astrofísica de Canarias (IAC). More information about their instruments and technical features were presented in Table 1 of Paper I, as well as information about the observing nights and the number of candidates observed.
Our follow-up programme is structured as follows. The first step is to select a target for deep-imaging observations and if galaxy over-densities are found, confirm the cluster using spectroscopy. The photometric observations are carried out in the INT using three broadband Sloan filters ($g^\prime$, $r^\prime$, $i^\prime$). Based on colour combinations of these filters, we are able to estimate photometric redshifts [@paper1] of possible members of the cluster up to $z\sim 0.8$. In the second step, candidates are definitely confirmed using multi-object spectroscopy (MOS) either in the TNG or in GTC, depending on the photometric estimation of the redshift of the candidate. Due to the greater collecting area of the GTC, we use it for distant clusters ($z>0.35$), while the TNG for the closest ones ($z<0.35$). The final step is to validate the candidates using the confirmation criteria explained in Sect. \[sec:criteria\].
Imaging and spectroscopic observation and data reduction {#sec:imagespecdata}
--------------------------------------------------------
The complete technical description of the telescopes and instrument set-ups used during [LP15]{} programme and information on corresponding data reduction is detailed in Paper I. Here, we briefly summarise the information of the imaging and spectroscopic observations and data reduction performed during the second year of the programme.
Optical observations were carried out during multiple runs between August 2016 and August 2017. Deep images were obtained using the Wide-Field Camera (WFC) installed in the 2.5m in INT. Spectroscopic observations were obtained using the multi-object spectrographs DOLORES@TNG and OSIRIS@GTC. The spectroscopic data from DOLORES was retrieved during multiple runs between November 2016 and August 2017 and the data from OSIRIS was acquired in service mode between September 2016 and August 2017.
All images and spectra were reduced using [IRAF]{}[^3] standard routines. The astrometry was performed using the [images.imcoords]{} task and the USNO B1.0 catalogue [@Monet03] as reference obtaining a final accuracy of [*rms*]{} $\sim 0.2\arcsec$ across the full field of view. The photometric calibration is referred to SDSS photometry. In the case of fields with SDSS coverage, we performed the direct cross-correlation with the SDSS photometric data. Using [SExtractor]{} [@Bertin96] programme in single-image mode, we created individual source catalogues for each band. All catalogues were then merged in the final one with a search radius of $1\arcsec$.
Standard spectroscopic data reduction included sky subtraction, extraction of spectra, cosmic rays rejection, and wavelength calibration (using specific arcs). Our final reduced spectra have typical $S/N \sim 5$ per pixel (at around $6000\,\AA$) for galaxies with magnitudes $r^\prime \sim 20.5$ and $21.7$ observed with the TNG and GTC, respectively.
We estimated radial velocities using task [RVSAO]{}[^4]. Figure \[fig1\] shows two examples of spectra for low and high-z galaxies obtained at the TNG and GTC. Thanks to the capabilities of the MOS observations, we were able to retrieve typically 20 members per cluster with which to estimate the mean redshift and velocity dispersion of the systems. The cluster redshift is taken to be the mean value of the galaxy members retrieved. The galaxies are considered members only if they showed radial velocities of $\pm 2500$kms$^{-1}$ in rest frame with respect to the mean velocity of the system. Then, we follow an iterative method considering galaxies as members if their radial velocity is less than 2.5 times the velocity dispersion away from the cluster mean velocity. We follow this procedure in order to minimise the contamination of interlopers. The velocity dispersion and mass of the clusters will be published in a future work.
The broad band images used to carry out this work has been already included in the Virtual Observatory (VO) collection for public access. In the near future, the photometric and spectroscopic catalogues will be also available through this platform.
{width="\columnwidth"} {width="\columnwidth"}
Cluster identification and validation criteria {#sec:criteria}
==============================================
Here we describe the methodology used to validate a cluster candidate as the optical counterpart of a PSZ2 target. This procedure is similar to the one adopted in paper I, and it is an extension of the methodology applied in @paper1 [@paper2]. Compared to other methods in the literature, we have improved the validation criteria by including visual inspection and comparison between the RGB images, and also the Compton $y$-maps [@Planck15XXII], making photometric redshift estimates and analysing the red sequence [hereafter RS @Gladders00] using colour-magnitude diagrams. We have also performed a richness study considering galaxy counts in clusters. For approximately 30$\%$ of our sample we have performed spectroscopic confirmation by estimating the velocity dispersion of the candidates.
Photometric analysis {#sec:cri_photo}
--------------------
The first step in our validation process is to visually inspect the deep RGB images and compare them with the Compton $y$-map of the SZ source. Having as reference the nominal [*Planck* ]{}pointing coordinates, we look for galaxy over-densities around $5{'}$ radius, which is 2.5 times the [*Planck* ]{}mean position error [@PSZ2]. However, for a few cases, when a possible counterpart is found further than $5{'}$, we perform a more detailed analysis. It must be clear that in this first step we identify galaxy over-densities in the field of view but not everyone of them is associated with the SZ source. The next steps are performed in order to study this possible association.
We use colour-magnitude diagrams to identify likely clusters members and fit the RS to make an estimate of the photometric redshift using the methodology explained in [@paper2], Sect. 3 and in [@paper1], Sect. 4.2. To do so, we use $(g^\prime - r^\prime)$ and $(r^\prime -i^\prime)$ colours and consider galaxies within $\pm 0.05$ from the Brightest Cluster Galaxy (BCG) colour.
The SZ clusters of the [*Planck* ]{}catalogue are expected to be massive structures ($\ga 5 \times 10^{14}$M$_\odot$) [see @PSZ1] and therefore, they should present rich galaxy cluster populations. In order to discard the low mass systems from the real massive counterparts, we define a richness parameter ($R_{0}$) and its significance above the background level $\sigma_{R}$. The detailed description of our richness calculations is given in Paper I. Briefly, for a given cluster, we count likely cluster members (assumed as galaxies within the RS $\pm 0.15$, in colour) showing $r'$-magnitudes in the range $[m_{r^{\prime}}^{\star}-1, m_{r^{\prime}}^{\star}+1.5]$ within the 1Mpc from the cluster centre. For this task, we consider as a BCG the most luminous galaxy of the identified likely cluster members. In order to obtain the background subtracted richness and decontaminate from the galaxy field contribution, we compute galaxy counts outside the 1Mpc region and using the same restrictions in colour and magnitude as explained before, thereby creating a field galaxy sample or local background for each cluster ($R_{f}$). This field sample is scaled to the 1Mpc area and subtracted from the cluster counts to obtain a statistical estimate of the number of galaxies for each cluster, the so-called richness ($R_{cor} \equiv R_{0}-R_{f}$). We based our confirmation on the significance above the background level $\sigma_{R}$ which is computed as $R_{cor} / \sqrt{R_{f}}$. For a better understanding see Fig. \[figrich\] which shows galaxy counts as a function of redshift in the field of the confirmed cluster PSZ2 G032.31$+$66.07. The initial value found for this cluster was $R_{0} = 20.0$, and for the field at the redshift of the cluster was $R_{f}
= 6.4$, yielding to $R_{cor} = 13.6$ and $\sigma_{R} = 5.4$.
![Galaxy counts as a function of redshift in the field of the spectroscopically confirmed cluster PSZ2 G032.31$+$66.07. The galaxy counts for this particular cluster and its 1-$\sigma$ error bars are shown in green. The blue line represents the galaxy counts outside 1Mpc region from the optical centre of the cluster and the dashed blue line represents 1$\sigma$ uncertainty above the latter.[]{data-label="figrich"}](Figures/fieldrichness115.jpeg){width="\columnwidth"}
Spectroscopic analysis {#sec:cri_spec}
----------------------
We performed spectroscopic observations for approximately 30$\%$ of our sample, and include public data from the SDSS spectroscopic archive. We built radial velocity catalogues for each cluster and obtained the velocity dispersion in order to distinguish between poor and massive systems.
[*Planck* ]{}systems must have masses $M_{500} \gtrsim 10^{14}$M$_\odot$ at low redshift ($z < 0.2$) and $M_{500} \gtrsim 2 \times 10^{14}$M$_\odot$ at higher redshift [see @vanderBurg16 Fig.4]. Using the scaling relations M$_{500} -
\sigma_{v}$ from [@Munari13] and $M_{500}$–$M_{200}$ from [@Komatsu11], those values of mass lead to a velocity dispersion of $\sigma_{v} >
500$kms$^{-1}$ for $z < 0.2$ and $\sigma_{v} > 650$kms$^{-1}$ for $z
\geq 0.2$. We adopt here the same criteria used in previous validation studies of [*Planck* ]{}clusters [@paper1; @paper2; @paper4]. The velocity dispersion of the candidates will be presented in a future work.
Confirmation criteria adopted in this work {#sec:cri_cri}
------------------------------------------
Table \[tab:cri\] summarises the set of criteria adopted in this work in order to confirm or reject a cluster candidate as the optical counterpart of the SZ signal. These are the very same criteria used in Paper I, and provide a classification of the candidates according to four possible values of a [Flag]{}. Values of [Flag]{}$=1$ or 2 correspond to validated clusters; [Flag]{}$=3$ corresponds to clusters located along the line of sight of the [*Planck* ]{}signal but possibly not associated with the SZ emission; and ND refer to non-detection. For those cases where we have enough spectroscopic information to provide an estimate of $\sigma_{v}$, if that value is found to be above the corresponding threshold, we validate the candidate with [Flag]{} 1. However, if $\sigma_{v}$ is below the threshold, we assume that the system has a low mass and it is probably not linked to the SZ emission, being the candidates classified as [ Flag]{} 3. In the case that no spectroscopic information is available, or if we cannot estimate of the velocity dispersion due to an insufficient number of galaxy members (less than $5$ members), we validate the candidates using the richness estimate. Systems showing a $\sigma_{R} > 1.5$ are validated photometrically, but waiting for a definitive spectroscopic confirmation. These systems are classified with [Flag]{} 2. Clusters with [Flag]{} 3 represent very poor systems ($\sigma_{R} < 1.5$) with no spectroscopic information. The ND (non-detection) flag is used for those SZ candidates where we found no galaxy over-density in the optical images. We also consider the criterion that a [*Planck* ]{}cluster must be placed within 5${'}$ radius from the nominal pointing because it represents 2.5 times the mean position error with respect to the SZ peak emission. Nevertheless, this criterion can be modulated due to the wide range of uncertainties in the position error in the PSZ2 catalogue, and the different shapes of the $y$-maps. The cases that do not match the validation criteria but are positively confirmed will be discussed in Sect. \[sec:notes\].
LP15 sample: 2nd year of observations {#sec:2ndyear}
=====================================
Year Observed val spec [Flag 1]{} [Flag 2]{} [Flag 3]{} ND
-------- ---------- ----- ------ ------------ ------------ ------------ ----
1 106 41 34 31 10 8 57
2 78 40 22 18 22 6 32
TOTAL: 184 81 56 49 32 14 89
Table \[tab:LP15summary\] summarises the basic information of programme LP15 after the two years of observations, concerning the characterisation of the [LP15]{} sample. The results for year one of the programme were already discussed in Paper I. Here, table \[tab:LP15\] presents the results for 78 PSZ2 galaxy cluster candidates studied in this optical follow-up during the second and last year. The table is organised as follows. Columns 1, 2 and 3 are the official ID number, the [*Planck* ]{}Name and the SZ signal-to-noise ratio, respectively, as they appear in the PSZ2 catalogue. Columns 4 and 5 are the J2000 coordinates of the BCG when present, otherwise geometrical centre of the cluster is provided. Column 6 is the distance between [*Planck* ]{}and the optical centre reported in this work. Columns 7 and 8 present the spectroscopic information when available: the mean spectroscopic redshift of the cluster and/or the BCG, and the number of spectroscopic members retrieved. Columns 9, 10 and 11 provide the photometric information: the photometric redshift, the estimation of the richness and the value $\sigma_{R}$ as explained in Sect. \[sec:cri\_photo\]. Column 12 lists the cluster classification following the [Flag]{} system described in Sect. \[sec:cri\_cri\]. Finally, column 13 is reserved for special comments or notes about the individual candidates.
Following the confirmation criteria explained in Sect. \[sec:cri\_cri\], we find that 37 of our candidates have a single optical counterpart, and one additional is classified as double detection. We classify a source as double detection when we find two or more over-densities around the SZ emission peak that might contribute to this emission. The ones validated with a single optical counterpart are classified as follows: 17 as [Flag]{} 1 and 22 as [Flag]{} 2. In addition, we find 32 non-detection, flagged as ND, and six systems not associated with the corresponding SZ source ([Flag]{} 3). This means a total of 38 PSZ2 sources remaining unconfirmed.
We have partially focused our work using SDSS DR12 data to confirm PSZ2 clusters classified by [@Streblyanska18] as ‘potentially associated’ with the SZ emission. We have obtained the redshift and the velocity dispersion for six of the photometrically confirmed clusters and we have re-confirmed six clusters using our own deep INT imaging data classifying them as [Flag]{} 1 and 2, respectively. From the ‘potentially associated’ sub-sample of [@Streblyanska18] we have confirmed five as [Flag]{} 1. Here, we invalidate the PSZ2 G328.96$+$71.97, confirmed by [@Streblyanska18]. New SDSS DR14 data reveals that the counterpart proposed by the authors is part of a larger system whose BCG is 34${\overset{'}{.}}$6 away from the [*Planck* ]{}SZ pointing. This system will be discussed in detail in Section \[sec:notes\].
![Spatial distribution of the optical counterparts centres with respect to the nominal centre in the [*Planck* ]{}PSZ2 catalogue. Green dots, blue triangles and red crosses correspond to clusters classified with [Flag]{} 1, [Flag]{} 2 and [Flag]{} 3, respectively. Dashed and dotted lines show the regions enclosing 68$\%$ and 95$\%$ of the confirmed clusters, flagged as ‘1’ and ‘2’, respectively[]{data-label="Figoff"}](Figures/off.jpeg){width="\columnwidth"}
Figure \[Figoff\] represents the spatial distribution of the optical counterpart centre with respect to the nominal centre in the [*Planck* ]{}PSZ2 catalogue for the clusters in our sample flagged as ‘1’,‘2’ and ‘3’. The median of the offsets of the clusters validated is $3{\overset{'}{.}}10$ which is in agreement with the median of the position error computed from the PSZ2 catalogue ($2{\overset{'}{.}}43$). The mean values are also in agreement ($3.3$ and $2.6$) but they are more sensitive to cases which their position error is too high. We find 25 (35) clusters within $4{\overset{'}{.}}46$ ($6{\overset{'}{.}}38$) representing the 68$\%$ (95$\%$) of the confirmed clusters.
![Cluster optical centre offsets relative to their [*Planck* ]{}SZ position as a function of cluster redshift. The dashed horizontal line at $5{'}$ shows the maximum offset expected for a [*Planck* ]{}SZ detection. The dotted line corresponds to the angle subtended by 1Mpc in projection at the corresponding redshift. Symbols used are the same as in Fig. \[Figoff\].[]{data-label="Figdz"}](Figures/Dz.jpeg){width="\columnwidth"}
Figure \[Figdz\] shows the cluster optical centre offsets relative to their [*Planck* ]{}SZ position as a function of cluster redshift. Although the limit in distance is $5{'}$ (see Sect. \[sec:criteria\]), there are some confirmed sources that exceed that value due to different reasons such as the high position error in the PSZ2 catalogue or elongated $y$-map contour around the SZ emission peak (see Sect. \[sec:notes\]).
![Comparison between photometric and spectroscopic redshift estimations. Dashed line represents the 1:1 relation while dotted lines show the photometric redshift error $\delta z/(1+z) = 0.026$.[]{data-label="Figzz"}](Figures/zz_sdss.jpeg){width="\columnwidth"}
Figure \[Figzz\] presents a comparison between photometric and spectroscopic redshift estimations. Photometric redshifts were obtained as explained in Sect. \[sec:observations\]. We show every observation carried out during the second year of the [LP15]{}, including the ones explained in Sect. \[sec:wise\]. This study yields a mean photometric redshift error of $\delta z/(1+z) = 0.026$.
Finally, we note that there are six objects in the [LP15]{} sample that have not been observed during the programme. One of them, PSZ2 G186.50$-$13.45, was already validated in [@Streblyanska18], with a photometric redshift of $z_{\rm phot}=0.25$. According to our validation criterion, this case would correspond to a [Flag]{}$=2$.
A dedicated proposal has been submitted and approved to complete the observations for the remaining five objects: PSZ2 G023.05+20.52, PSZ2 G092.34+14.22, PSZ2 G206.55-43.22, PSZ2 G210.37-37.00 and PSZ2 G247.14+25.88. These observations will be conducted in July 2019.
------------- ---------------------- -------- -------------- ---------------- -------------- --------------------------------------- ---------------- ---------------- --------------- ------------------ ---------- ---------------------------------------------------------------------
ID$^{1}$ Planck Name SZ SNR R. A. Decl. Dist.(${'}$) $<z_{\rm spec}>$ ; $z_{\rm spec,BCG}$ $N_{\rm spec}$ $z_{\rm phot}$ $R_{\rm cor}$ $\sigma_{\rm R}$ [Flag]{} Notes$^{2}$
115$^{a,b}$ PSZ2 G032.31$+$66.07 5.14 14 37 23.35 $+$24 24 21.70 3.10 0.610 ;$-$ 38 0.62$\pm$0.05 13.6$\pm$3.7 5.4 1 (1), sub-structured
194 PSZ2 G048.47$+$34.86 5.74 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
242 PSZ2 G058.31$+$41.96 4.54 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
421-A$^{c}$ PSZ2 G092.69$+$59.92 4.90 14 26 03.78 $+$51 14 18.50 3.85 0.462 ; 0.4568 25 0.50$\pm$0.05 11.6$\pm$3.4 4.3 1 (1,2)
421-B$^{c}$ 14 26 13.10 $+$51 11 53.17 4.42 0.844 ;$-$ 5 $-$ $-$ $-$ 3 (3)
424$^{b}$ PSZ2 G093.41$-$16.26 4.59 22 24 07.25 $+$37 58 30.46 3.10 $-$ $-$ 0.24$\pm$0.03 40.6$\pm$6.4 7.9 2 WHL J222407.2$+$375831
432$^{a,b}$ PSZ2 G094.31$-$11.31 4.72 22 12 56.10 $+$42 35 46.34 1.08 0.204 ;$-$ 27 0.24$\pm$0.03 $-$ $-$ 1
500 PSZ2 G104.52$+$39.39 4.60 15 58 38.88 $+$70 27 24.20 5.62 $-$ $-$ 0.30$\pm$0.04 16.8$\pm$4.1 7.8 2
511 PSZ2 G105.94$-$16.14 4.62 $-$ $-$ $-$ $-$ $-$ $-$ ND
545 PSZ2 G112.54$+$59.53 5.37 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
546$^{c}$ PSZ2 G112.69$+$33.37 4.63 16 19 49.39 $+$79 06 24.49 4.78 0.521 ; 0.5194 15 0.51$\pm$0.03 $-$ $-$ 1 (1),WHL J161949.3$+$790624
592 PSZ2 G120.75$+$25.39 4.69 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
600 PSZ2 G122.81$+$24.74 4.60 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
613 PSZ2 G125.25$+$33.33 5.38 11 41 11.26 $+$83 27 38.91 1.80 $-$ $-$ 0.20$\pm$0.03 21.0$\pm$4.6 4.1 2
616 PSZ2 G125.41$+$27.95 4.76 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
620$^{b}$ PSZ2 G125.84$-$18.72 5.30 01 06 55.65 $+$44 04 25.72 1.81 0.189 ;$-$ 46 0.19$\pm$0.01 $-$ $-$ 1 WHL J010709.2$+$440918
624$^{b}$ PSZ2 G126.36$-$19.11 5.01 01 09 19.57 $+$43 37 40.41 0.60 0.203 ; 0.2007 22 0.22$\pm$0.01 $-$ $-$ 1 WHL J010919.5$+$433741
627 PSZ2 G126.62$-$53.42 4.55 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
628$^{b}$ PSZ2 G126.72$-$21.03 4.68 01 10 27.91 $+$41 40 57.27 0.84 0.196 ;$-$ 9 0.22$\pm$0.03 $-$ $-$ 1 WHL J011025.0$+$414119
640 PSZ2 G129.99$-$22.42 4.55 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
644-A$^{b}$ PSZ2 G130.64$+$37.16 4.80 10 47 45.54 $+$77 59 56.67 2.82 0.473 ; 0.4722 14 0.44$\pm$0.03 16.7$\pm$4.1 5.0 1 WHL J104745.5$+$775957
644-B$^{b}$ 10 46 29.31 $+$78 07 44.06 6.07 $-$ 0.24$\pm$0.02 49.3$\pm$7.0 11.9 2 (1), WHL J104629.2$+$780744
646$^{b}$ PSZ2 G131.15$-$14.72 5.37 01 38 42.22 $+$47 22 35.27 1.71 $-$ $-$ 0.22$\pm$0.03 50.2$\pm$7.1 11.3 2 WHL J013846.1$+$472236
647 PSZ2 G131.19$+$14.48 4.80 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
648 PSZ2 G131.27$-$25.82 4.50 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
667$^{c}$ PSZ2 G136.02$-$47.15 4.64 01 28 23.61 $+$14 41 13.60 7.09 0.466 ; 0.4648 8 0.50$\pm$0.03 13.1$\pm$3.6 4.2 3 WHL J012823.6$+$144114
700 PSZ2 G143.90$+$25.06 4.91 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
712$^{a}$ PSZ2 G146.10$-$55.55 4.67 01 42 46.53 $+$04 59 48.72 4.71 0.429 ;$-$ 1 0.42$\pm$0.04 10.8$\pm$3.3 3.0 2
713 PSZ2 G146.13$+$40.97 4.90 09 40 17.10 $+$66 24 02.56 7.18 0.342 ; 0.3379 4 0.34$\pm$0.04 7.0$\pm$2.6 3.5 2
717 PSZ2 G146.88$+$17.13 6.13 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
720 PSZ2 G147.17$+$42.67 4.92 09 50 01.152 $+$64 55 29.52 1.34 $-$; 0.4401 1 0.46$\pm$0.04 7.4$\pm$2.7 3.1 2 (4)
727 PSZ2 G149.73$+$24.49 4.52 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
732$^{a,b}$ PSZ2 G150.64$-$14.21 4.68 03 17 04.20 $+$40 41 33.22 3.08 $-$ $-$ 0.22$\pm$0.02 26.1$\pm$5.1 8.2 2 WHL J031704.2$+$404133
739$^{a,c}$ PSZ2 G152.40$+$75.00 4.70 12 13 19.17 $+$39 46 26.84 5.06 0.455 ;$-$ 9 0.42$\pm$0.03 31.3$\pm$5.6 9.0 1 (1),WHL J121319.2$+$394627
740 PSZ2 G152.47$+$42.11 4.81 09 29 52.64 $+$61 39 40.00 0.24 0.900 ;$-$ 6 $-$ $-$ $-$ 3
746 PSZ2 G153.68$+$36.96 5.07 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
747 PSZ2 G153.80$+$33.79 4.52 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
754 PSZ2 G156.24$+$22.32 4.79 06 45 02.14 $+$59 27 13.30 0.75 $-$ $-$ 0.30$\pm$0.05 $-$ $-$ 2
769 PSZ2 G160.94$+$44.85 4.98 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
780 PSZ2 G163.89$+$11.55 4.78 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
788 PSZ2 G165.39$+$09.22 5.60 05 48 09.37 $+$46 04 41.41 4.39 $-$ $-$ 0.70$\pm$0.10 15.1$\pm$3.9 5.4 2 Clearly sub-structured, second centre at 05:48:10.56 $+$46:07:18.38
789 PSZ2 G165.41$+$25.93 4.51 07 23 27.93 $+$52 07 32.70 4.83 $-$ $-$ 0.67$\pm$0.04 4.3$\pm$2.1 1.8 2
797 PSZ2 G166.56$-$17.69 4.76 04 04 53.39 $+$28 18 31.85 3.44 $-$ $-$ 0.70$\pm$0.07 3.3$\pm$1.8 1.4 ND
799 PSZ2 G167.43$-$53.67 4.65 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
812$^{a,c}$ PSZ2 G171.48$+$16.17 4.75 06 38 00.94 $+$43 50 57.20 0.75 0.385 ; 0.3881 25 0.40$\pm$0.05 15.9$\pm$4.0 2.7 1 WHL J063743.6$+$434859
820 PSZ2 G173.76$+$22.92 5.80 07 17 26.66 $+$44 05 00.28 1.63 0.063 ; 0.0652 2 0.06$\pm$0.02 $-$ $-$ 2 (5)
831$^{a}$ PSZ2 G177.03$+$32.64 4.93 08 13 08.56 $+$43 13 53.07 3.56 0.511 ;$-$ 9 $-$ $-$ $-$ 1 (1)
835 PSZ2 G179.33$-$22.22 5.02 $-$ $-$ $-$ $-$ $-$ $-$ $-$ $-$ ND
836$^{a,b}$ PSZ2 G179.45$-$43.92 4.54 03 19 18.34 $+$02 05 35.60 2.58 0.397 ; 0.4005 23 0.34$\pm$0.03 19.1$\pm$4.4 5.8 1 WHL J031918.3$+$020535
849 PSZ2 G183.32$-$31.51 4.56 04 05 20.11 $+$07 51 26.07 2.31 $-$ $-$ 0.55$\pm$0.10 1.8$\pm$1.3 0.9 3
------------- ---------------------- -------- -------------- ---------------- -------------- --------------------------------------- ---------------- ---------------- --------------- ------------------ ---------- ---------------------------------------------------------------------
: List of 78 PSZ2 cluster candidates analysed in this paper.[]{data-label="tab:LP15"}
[@p[1cm]{} c c c c c c c c c c c p[4.5cm]{} @]{} &\
ID$^{1}$& Planck Name & SZ SNR & R. A. & Decl.& Dist.(${'}$) & $<z_{\rm spec}>$ ; $z_{\rm spec,BCG}$&$N_{\rm spec}$&$z_{\rm phot}$&$R_{\rm cor}$&$\sigma_{\rm R}$&[Flag]{} &\
852 & PSZ2 G183.92$+$16.36 & 4.97 & 07 01 30.22 & $+$32 54 51.20 & 6.61 & 0.091 ; 0.0914 & 18 & 0.15$\pm$0.03 & 30.3$\pm$5.5 & 4.4 & 1 & ABELL 567\
859 & PSZ2 G185.68$+$09.82 & 5.18 & 06 37 14.93 & $+$28 38 02.80 & 1.47 & 0.390 ; 0.3897 & 39 & 0.40$\pm$0.05 & $-$ & $-$ & 1 &\
860 & PSZ2 G185.72$-$32.23 & 5.12 & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & ND &\
878 & PSZ2 G191.57$+$58.88 & 5.17 & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & ND &\
887 & PSZ2 G193.90$+$09.41 & 5.06 & 06 51 11.80 & $+$21 08 10.16 & 0.73 & 0.194 ; 0.1936 & 29 & 0.16$\pm$0.03 & 23.5$\pm$4.8 & 3.0 & 1 &\
912 & PSZ2 G201.20$-$42.83 & 4.70 & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & ND &\
916$^{a,c,d}$ & PSZ2 G202.61$-$26.26 & 4.87 & 04 59 50.17 & $-$03 16 47.52 & 5.52 & $-$ & $-$ & 0.23$\pm$0.03 & 8.5$\pm$2.9 & 4.6 & 3 & WHL J045950.2-031647\
917$^{a,c}$ & PSZ2 G202.66$+$66.98 & 4.63 & 11 07 30.90 & $+$28 51 01.20 & 4.68 & 0.482 ; 0.4814 & 20 & 0.48$\pm$0.04 & 11.9$\pm$3.4 & 5.3 & 1 & WHL J110730+285101\
920$^{a,c,d}$ & PSZ2 G203.32$+$08.91 & 5.15 & 07 05 56.53 & $+$12 30 33.66 & 4.25 & $-$ & $-$ & 0.27$\pm$0.03 & 9.1$\pm$3.0 & 4.6 & 2 & WHL J070556.5$+$123034\
921$^{a}$ & PSZ2 G203.71$+$50.82 & 4.65 & 09 55 15.56 & $+$26 19 37.70 & 2.03 & 0.082 ;$-$ & 22 & $-$ & $-$ & $-$ & 1 & (1)\
952 & PSZ2 G210.71$+$63.08 & 7.37 & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & ND &\
953 & PSZ2 G210.78$-$36.25 & 6.32 & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & ND &\
982 & PSZ2 G218.58$+$08.71 & 4.63 & 07 32 40.27 & $-$01 03 21.55 & 3.40 & $-$ & $-$ & 0.55$\pm$0.05 & 30.7$\pm$5.5 & 7.6 & 2 & 1RXS J073246.4$-$010205\
1018 & PSZ2 G226.15$+$09.02 & 4.66 & 07 47 58.81 & $-$07 29 22.70 & 2.96 & $-$ & $-$ & 0.30$\pm$0.05 & 28.9$\pm$5.4 & 10.2 & 2 &\
1023 & PSZ2 G227.30$+$09.00 & 4.62 & 07 50 15.74 & $-$08 24 32.56 & 1.85 & $-$ & $-$ & 0.24$\pm$0.03 & $-$ & $-$ & 2 & 1RXS J075020.3$-$082605\
1049$^{a,b}$ & PSZ2 G231.41$+$77.48 & 4.54 & 12 00 26.54 & $+$22 34 19.55 & 6.32 & 0.346 ; 0.3469 & 2 & 0.35$\pm$0.03 & $-$ & $-$ & 2 & WHL J120026.5$+$223420\
1054 & PSZ2 G232.27$+$12.59 & 4.52 & 08 12 39.17 & $-$10 52 02.90 & 2.35 & $-$ & $-$ & 0.28$\pm$0.04 & 35.3$\pm$5.9 & 10.3 & 2 &\
1062 & PSZ2 G233.46$+$25.46 & 4.79 & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & ND &\
1074$^{a,b}$ & PSZ2 G237.68$+$57.83 & 5.36 & 10 53 17.80 & $+$10 52 37.13 & 4.88 & $-$ & $-$ & 0.70$\pm$0.05 & $-$ & $-$ & 2 & (3)\
1095 & PSZ2 G241.98$+$19.56 & 4.51 & 08 58 04.54 & $-$14 43 01.87 & 3.95 & $-$ & $-$ & 0.27$\pm$0.04 & 33.1$\pm$5.7 & 8.3 & 2 &\
1151 & PSZ2 G252.45$+$73.44 & 5.57 & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & ND &\
1162 & PSZ2 G253.95$+$39.12 & 4.66 & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & ND &\
1168 & PSZ2 G254.52$+$62.52 & 4.85 & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & ND &\
1219 & PSZ2 G263.96$+$40.64 & 4.58 & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & ND &\
1262$^{a,b}$ & PSZ2 G271.53$+$36.41 & 5.19 & 11 05 19.71 & $-$19 59 15.61 & 4.73 & $-$ & $-$ & 0.50$\pm$0.03 & $-$ & $-$ & 2 & WHL J110519.6$-$195852\
1493 & PSZ2 G316.43$+$54.02 & 5.18 & 13 23 14.77 & $-$07 58 49.20 & 2.02 & 0.527 ; 0.5325 & 27 & 0.55$\pm$0.05 & 12.7$\pm$3.6 & 3.3 & 1 & Sub-structured\
1510$^{a,c}$ & PSZ2 G320.94$+$83.69 & 7.32 & 13 00 05.74 & $+$21 01 28.29 & 7.00 & 0.461 ; 0.4612 & 5 & 0.45$\pm$0.04 & $-$ & $-$ & 3 &\
1513 & PSZ2 G321.94$+$75.57 & 4.66 & $-$ & $-$ & & $-$ & $-$ & $-$ & $-$ & $-$ & ND &\
1532 & PSZ2 G325.19$+$49.12 & 4.62 & 13 49 55.18 & $-$11 15 24.44 & 3.61 & $-$ & $-$ & 0.20$\pm$0.03 & 28.1$\pm$5.3 & 6.5 & 2 & WHY J135000.5$-$111724\
1548$^{a,b}$ & PSZ2 G328.96$+$71.97 & 5.85 & 13 23 02.10 & $+$11 01 32.12 &18.03 & 0.090 ; 0.0937 & 94 & 0.09$\pm$0.01 & $-$ & $-$ & 3 &\
$^1$ SZ targets identified with the ID followed by an A or B label indicate the presence of multiple counterparts.
$^2$ References. (1) [@Burenin17], (2) [@Rykoff14], (3) [@Burenin18], (4) [@Zaznobin19], (5) [@Boada18]
$^a$ Photometric and/or spectroscopic redshift obtained from SDSS DR14 data.
$^b$ Already confirmed in [@Streblyanska18]
$^c$ Classified as “potentially associated” in [@Streblyanska18]
$^d$ Richness study from PAN-STARRS
Notes on individual objects {#sec:notes}
---------------------------
PSZ2 G058.31$+$41.96 This candidate is flagged as a non-detection due to a bright star located near the [*Planck* ]{}pointing. The star prevents photometric measurements of this region, and thus, we are unable to visually identify an over-density of galaxies. Despite this problem, we cannot identify visually any over-density of galaxies in the region.
PSZ2 G104.52$+$39.39 The distance from the optical centre and the [*Planck* ]{}nominal pointing is $5{\overset{'}{.}}62$. Nevertheless, we validate this cluster with [Flag]{} 2 because the MILCA $y$-map contours are elongated along the line that links both optical and [*Planck* ]{}centres. Besides this fact, the position error in the [*Planck* ]{}catalogue is too high ($5{\overset{'}{.}}40$) compared to the nominal one ($ 2{\overset{'}{.}}43 $).
PSZ2 G130.64$+$37.16 This candidate has two optical counterparts validated already in [@Streblyanska18] and one of them in [@Burenin17]. This is a tricky case as it is shown in Fig. \[Fig643\]. The nominal SZ pointing is clearly closer to the cluster named 644-A which presents 14 spectroscopic members and shows a velocity dispersion close to $1000$kms$^{-1}$; the cluster named 644-B is $6{\overset{'}{.}}07$ away from the SZ centre but the MILCA $y$-map shows that the contours are elongated along this counterpart location which is twice as rich as the 644-A. This is also a case where the position error in the [*Planck* ]{}catalogue is $5{\overset{'}{.}}38$ which is more than twice the mean position error.
![Compton $y$-map superimposed on the INT $r'$-band of the PSZ2 G130.64$+$37.16. Blue contours correspond to the $4$, $4.6$ and $5.2\times10^{-6}$ levels of the Compton $y$-map in this area. The nominal SZ pointing (red) is clearly closer to the cluster named 644-A (zoomed in the lower-right region) which presents 14 spectroscopic members and shows a velocity dispersion close to $1000$kms$^{-1}$; the cluster named 644-B (zoomed in the upper-left region) is $6{\overset{'}{.}}07$ away from the SZ centre but the MILCA $y$-map shows that the contours are elongated along this counterpart location which is twice as rich as the 644-A.[]{data-label="Fig643"}](Figures/PSZ2_643_final.jpeg){width="\columnwidth"}
PSZ2 G146.13$+$40.97 The optical centre of the proposed counterpart (its BCG) is at $7{\overset{'}{.}}18$ away from the Planck SZ pointing, which is affected by a position error of $5{\overset{'}{.}}89$. The y-map contours present a very irregular shape, maybe due to galactic dust contamination around this region. We estimate a richness of $\sigma_{R}=3.5$ for this system and we find four cluster members at $z_{\rm spec} = 0.342$ in the SDSS DR14 spectroscopic sample. So, we classify this counterpart with [Flag]{}$=2$. The ultimate confirmation will be obtained using MOS observations.
PSZ2 G152.47$+$42.11 We find a possible cluster counterpart at $z_{\rm spec} =
0.900$. The deepness of our images makes it impossible to estimate the richness at this redshift. However, we have observed this system spectroscopically and we have found six cluster members. From these six galaxies, we obtain a very low-velocity dispersion ($<400$kms$^{-1}$), revealing a low-mass galaxy system. Thus, in this case, we classify this optical counterpart with [ Flag]{}$=3$.
PSZ2 G156.24$+$22.32 This region encloses two very bright stars, making it very difficult to obtain accurate photometry or richness. However, by eye inspection, we identify a cluster showing a galaxy population with coherent colours. A detailed study of the photometry of some individual likely members and the BCG reveals a $z_{\rm phot}=0.30$. In addition, the $y$-map contours present a very regular profile centred on this system. For all these reasons, we classify this system with [Flag]{}$=2$.
PSZ2 G177.03$+$32.64 [@Burenin17] reports a counterpart for this candidate at $z\sim0.28$. We have analysed this over-density finding seven galaxies with spectroscopic redshifts in the SDSS archive. Four of these galaxies are more than 4 Mpc away from the [*Planck* ]{}centre and the velocity dispersion accounting for the seven galaxies is less than $300$kms$^{-1}$. For this reason, we present here only one counterpart at $z_{\rm spec} = 0.511$ whose velocity dispersion, calculated using 9 members is approximately $1000$kms$^{-1}$.
PSZ2 G183.92+16.36 The distance between the BCG of this cluster and the [*Planck* ]{}pointing is $6{\overset{'}{.}}61$, which is only 0.67Mpc at the redshift of the cluster, $z_{\rm spec} = 0.091$ (see Fig \[Figdz\]). We perform multi-object spectroscopy and retrieve 18 cluster members, showing a $\sigma_{v} \sim
650$kms$^{-1}$. In addition, this cluster is known as Abell 567, which is well consolidated by other observations in the past [@Abell89]. Therefore, we confirm Abell 567 as the counterpart of this SZ source, classifying it as [Flag]{}$=1$.
PSZ2 G202.61$-$26.26 and PSZ2 G203.32$+$08.91 Both candidates were analysed using PANSTARRS photometric data [@PANSTARRS]. Both systems are rich but the optical centre of the first one is more than $5 {'}$ away from the [*Planck* ]{}nominal pointing, so classified as [Flag 3]{}.
PSZ2 G227.30$+$09.00 This is an SZ source placed at very low galactic latitude, so a large number of stars are crowding this field. For this reason, we were not able to compute the richness: the galaxies of the background are partially masked by the foreground stars. However, this system presents X-ray emission and has been catalogued as 1RXS J075020.3-082605 in the ROSAT survey. So, we classify this source as [Flag]{}$=2$.
PSZ2 G237.68$+$57.83 This cluster has been already validated by [@Streblyanska18] using SDSS data. Here, we confirm this association using the INT images. Despite we are not able to perform a richness estimation at this redshift, some individual likely cluster members show photometry in agreement with a $z_{\rm phot} = 0.70\pm0.05$. We also find two additional over-densities at (RA=10:53:35.55, Dec=+10:43:45.71) and (RA=10:53:59.602, Dec=+10:46:38.23). However they are at over $> 10{'}$ distance from the SZ coordinates and thus, probably not contributing to the SZ signal. Therefore we validate PSZ2 G237.68+57.83 as a single counterpart at $z_{\rm phot} =0 .70$.
PSZ2 G271.53$+$36.41 This candidate was confirmed photometrically in [@Streblyanska18] as a double detection. However, only one cluster is visible in the INT images. This cluster is at $z_{\rm phot}=0.50\pm 0.03$. No more systems are associated with this SZ source.
PSZ2 G328.96$+$71.97 was validated by [@Streblyanska18] using SDSS DR12 data. Here, we use new spectroscopic information provided by SDSS DR14 in order to update the information there reported. We find 94 cluster members at $<z_{\rm
spec}>=0.090$. However, the BCG of this structure is at $34{\overset{'}{.}}6$ from the [*Planck* ]{}pointing. Fig. \[Fig1547\] shows the scenario around this region. SZ emission presents a very spread and irregular profile, with several peaks. The 94 cluster members present a $\sigma_{v} \sim 800$km$s^{-1}$ and a virial radius of 1.6Mpc, but the cluster seems to be placed completely off the SZ peak. Notice that the distance between the [*Planck* ]{}pointing and the optical cluster centre is larger (double) than the virial radius of the cluster. So, for all these reasons, we conclude that no optical counterpart is found for this SZ source, and the actual counterpart (if it exists) is still unknown.
![SZ emission as seen in the Compton $y$-maps for the source PSZ2 G328.96$+$71.97. The red dot represents the SZ coordinate as it appears in the PSZ2 catalogue. The BCG of the cluster is plotted in magenta, while the rest of the galaxies members are shown in green. The blue circle encloses the virial radius of this cluster. The 94 cluster members are placed completely off the SZ peak and the optical counterpart of this SZ source remains unknown.[]{data-label="Fig1547"}](Figures/PSZ2_1548.jpeg){width="\columnwidth"}
Observations of other PSZ2 candidates beyond the LP15 sample {#sec:wise}
============================================================
In the PSZ2 catalogue, there are 73 clusters validated using the AllWISE mid-infrared source catalogue [@Cutri13]. This catalogue is based on the observations from the Wide-field Infrared Survey Explorer mission [WISE, @Wright10]. Using the (W1-W2) colour they search for galaxy over-densities in the redshift range $0.3<z<1.5$. The details of this validation process can be found in Sect. 7.4 in [@PSZ2]. Those objects had [ validation]{} $= 16$ in the original PSZ2 catalogue, and thus, they were not included in the definition of the [LP15]{} sample.
Here, we present an update on 38 of those ALLWISE sources, providing their spectroscopic redshifts which were obtained using dedicated observations carried out with the telescope time within the [LP15]{} programme. Table \[tab:wise\] presents this information, and it is organised in the same way as Table \[tab:LP15\].
The double detections PSZ2 G086.28$+$74.76, PSZ2 G139.00$+$50.92 and PSZ2 G141.98$+$69.31 can be considered single detection even if secondary clusters are detected, because they are very low mass systems, and so not capable of contributing significantly to the SZ signal.
As noted in earlier by the Planck Collaboration, the number of double cluster detections is relatively high, as compared to other surveys, either in X-rays [@PcolintI] or in optical [@PcolintXXVI].
We have cross-checked our sample with two galaxy cluster catalogues WHL [@WHL] and WHY [@WHY], based on optical and infrared data, respectively. WHL catalogue was published using SDSS data, while WHY used 2MASS [@2MASS], WISE [@Wright10] and SuperCOSMOS [@SuperCOSMOS] data. We only find three matches with WHL. PSZ2 G076.55$+$60.29 and PSZ2 G141.98$+$69.31 will be discussed in the next sub-section, and PSZ2 G021.02$-$29.0 which is also part of the WHY catalogue. We find seven matches with the WHY catalogue, we are in a $1\sigma$ agreement in redshift except for PSZ2 G056.38$+$23.36. In this case, we estimate a photometric redshift of $z_{\rm phot} = 0.21 \pm 0.02$ while [@WHY] reports $z_{\rm phot} = 0.31 \pm 0.04$, compatible within $2\sigma$.
We also present in this section an update on four sources that were already confirmed in the PSZ2 original catalogue [@PSZ2] as they were matched with PSZ1 clusters but without an estimation of their redshifts. Here, we provide the photometric redshift for three of them and invalidate the already confirmed PSZ2 G198.73$+$13.34, for which we are unable to find any galaxy over-density. In a future publication, we will discuss this type of sources that we believe are false validations. These four sources can be found in table \[tab\_psz1\].
Discussion on special cases {#sec:notes_update}
---------------------------
We found that PSZ2 G076.55$+$60.29, which it was classified as an individual counterpart by [@Streblyanska18], is, in fact, a superposition of two clusters, at $z_{\rm spec}=0.287$ and $z_{\rm spec}=0.632$, respectively. The first one (327-A) was already proposed as potentially associated cluster. Here, we confirmed it with 5 spectroscopic members. The distance to the [*Planck* ]{}nominal pointing of the second counterpart (327-B) is slightly greater than $5{'}$ but the MILCA $y$-map contours superimposed on an INT image (Fig. \[Fig326\]) show that the SZ emission is clearly a superposition of both clusters. Both counterparts are two of the richest systems studied in this work, presenting a $\sigma_{R}$ of 23.1 and 8.4, respectively.
![Compton $y$-map superimposed to the INT $r'$-band of the PSZ2 G076.55$+$60.29. White contours correspond to the $3$, $4$ and $5\times10^{-6}$ levels of the Compton $y$-map in this area. The [*Planck* ]{} nominal pointing is represented in red. In the upper corners we show zoomed RGB images of the 327-A (left) and 327-B (right), respectively. Both systems, at $z_{\rm spec} = 0.632$ and $z_{\rm spec} = 0.287$, are associated with this SZ signal. This is a clear example of a multiple detection.[]{data-label="Fig326"}](Figures/PSZ2_326_final.jpeg){width="\columnwidth"}
PSZ2 G086.28$+$74.76 We find two clusters around the SZ emission at $z_{\rm
spec} = 0.246$ and $z_{\rm spec} = 0.701$ which we name 381-A and 381-B, respectively. Both present high-velocity dispersion. However, the centre of 381-A is $8{\overset{'}{.}}93$ away from the [*Planck* ]{}centre. So we conclude that this source has only one optical counterpart at $z_{\rm spec} = 0.701$.
PSZ2 G126.57$+$51.61 is one of the most distant cluster in our sample, at $z_{\rm spec}=0.816$. @Burenin18 published confirmation of one galaxy at $z_{\rm spec} = 0.815$. This cluster is at the detection limits of our deep optical images, with most of the members detected almost at the noise level of the $i'-$band image. No RS for this cluster could be constructed. However, inspection of the RGB image revealed an over-density of red sources close to the [*Planck* ]{}position supported by the contours extracted from the MILCA $y$-map (Fig. \[Fig624\]). Our spectroscopic data together with SDSS data confirm this cluster with 20 members and a $\sigma_{v} \sim 850$kms$^{-1}$.
![The zoomed RGB image of the central area of high-z cluster ($z_{\rm
spec} = 0.816$) associated with PSZ2 G126.57+51.61. We marked as a yellow circles cluster members confirmed spectroscopically. The small top panel shows the WFC/INT $i'-$band image with white contours corresponding to the 3 and 6$\times$10$^{-6}$ levels of the Compton $y$-map in this area. The red circle indicates the nominal PSZ2 position. The blue square shows the size of the area presented in the main RGB image.[]{data-label="Fig624"}](Figures/psz2_624_fin.jpeg){width="\columnwidth"}
PSZ2 G133.92$-$42.73 There is a potential counterpart but it is discarded due to its low $\sigma_{R}$. Looking at the RGB image (left panel, Fig. \[Fig653\]) it seems to be a high redshift cluster as seen in the WISE image (right panel, Fig. \[Fig653\]). In SDSS there are three galaxies with $z_{\rm spec} \sim
0.581$ but they are not associated with any galaxy over-density. Deeper imaging or spectroscopic observations would be needed in order to reject the possibility of a high-z ($z > 0.8$) cluster.
![Left: RGB image of the source PSZ2 G133.92$-$42.73. Right: WISE $W1$-band of the same region. In both images Compton $y$-maps are represented in white ($5$, $5.8$, $6.6$ and $7.4\times 10^{-6}$ levels). The blue region corresponds to $1$Mpc ($2{\overset{'}{.}}45$) at the mean redshift of the three galaxies represented in green. The [*Planck* ]{}nominal pointing is marked in red.[]{data-label="Fig653"}](Figures/PSZ2_653.jpeg){width="\columnwidth"}
PSZ2 G139.00$+$50.92 was already confirmed by [@Streblyanska18] at $z_{\rm
phot}=0.6$. We have performed spectroscopic observations for this cluster, named 681-A in this work, finding a velocity dispersion below the confirmation limit ($\sigma_{v} < 650$kms$^{-1}$). We conclude that it is not the main counterpart to the SZ emission. However, we find another cluster (681-B) at $z_{\rm spec} = 0.784$ showing a $\sigma_{v} > 800$kms$^{-1}$, so we conclude that this last counterpart is the responsible for the SZ emission.
PSZ2 G141.98$+$69.31 This is a case of double detection. We find two over-densities in this field but once we made spectroscopic observations and calculate the velocity dispersion, we realised that the object named 690-B presented a $\sigma_{v} < 400$kms$^{-1}$, very low to be associated with the SZ emission. Consequently, we only validate the object proposed in [@Streblyanska18] with a spectroscopic redshift of $z_{\rm spec} = 0.713$, here named as 690-A.
PSZ2 G270.78$+$36.83 This candidate was already validated by [@Streblyanska18] as a double detection. Here, we spectroscopically confirm one of these counterparts by detecting 25 cluster members at $z_{\rm spec} = 0.516$, showing a $\sigma_{v} \sim 900$kms$^{-1}$. The second counterpart remains unconfirmed spectroscopically.
[@p[1.1cm]{} c c c c c c c c c c c p[4.5cm]{} @]{} &\
ID$^{1}$& Planck Name & SZ SNR & R. A. & Decl.& Dist.(${'}$) & $<z_{\rm spec}>$ ; $z_{\rm spec,BCG}$&$N_{\rm spec}$&$z_{\rm phot}$&$R_{\rm cor}$&$\sigma_{\rm R}$&[Flag]{} &\
65 & PSZ2 G020.21$-$36.84 & 5.17 & 20 51 57.42 & $-$25 29 17.03 & 1.07 & $-$ & $-$ & 0.21$\pm$0.02 & 18.1$\pm$4.3 & 3.6 & 2 &\
68$^{a,c}$ & PSZ2 G021.02$-$29.04 & 4.89 & 20 20 28.21 & $-$22 25 14.78 & 1.83 & 0.300 ; 0.3017 & 25 & 0.32$\pm$0.04 & $-$ & $-$ & 1 & WHY/WHL J202028.2$-$222515\
92 & PSZ2 G027.77$-$49.72 & 4.60 & 21 53 03.42 & $-$23 34 13.08 & 1.45 & 0.165 ; 0.1641 & 15 & 0.12$\pm$0.02 & 20.6$\pm$4.5 & 4.4 & 1 & 2dF Survey, (1)\
93 & PSZ2 G027.81$-$45.93 & 5.24 & 21 37 16.11 & $-$22 32 19.80 & 2.23 & & $-$ & 0.45$\pm$0.05 & 15.8$\pm$4.0 & 5.2 & 2 &\
120 & PSZ2 G033.83$-$46.57 & 5.53 & 21 45 12.35 & $-$18 42 57.41 & 2.50 & $-$ & $-$ & 0.32$\pm$0.04 & 3.7$\pm$1.9 & 1.7 & 2 & WHY J214532.4$-$184130, (2)\
206 & PSZ2 G050.98$-$61.48 & 5.10 & 22 58 53.05 & $-$15 35 30.87 & 1.76 & $-$ & $-$ & 0.18$\pm$0.03 & 14.2$\pm$3.8 & 6.5 & 2 &\
227 & PSZ2 G056.38$+$23.36 & 4.85 & 18 01 16.53 & $+$30 23 20.71 & 3.43 & $-$ & $-$ & 0.21$\pm$0.02 & 11.4$\pm$3.4 & 7.0 & 2 & WHY J180116.5$+$302321\
277$^{a,c}$ & PSZ2 G066.34$+$26.14 & 5.63 & 18 01 06.52 & $+$39 52 06.73 & 1.64 & 0.622 ; 0.6167 & 52 & 0.63$\pm$0.06 & $-$ & $-$ & 1 &\
282 & PSZ2 G066.85$+$22.48 & 4.88 & 18 20 08.12 & $+$39 15 52.93 & 2.05 & $-$ & $-$ & 0.19$\pm$0.02 & 22.3$\pm$4.7 & 10.3 & 2 & WHY J182008.2$+$391553\
294$^{a,c}$ & PSZ2 G069.39$+$68.05 & 4.51 & 14 21 38.32 & $+$38 21 17.54 & 0.46 & 0.766 ; 0.7630 & 23 & 0.70$\pm$0.06 & $-$ & $-$ & 1 & CIG-J142138.3$+$382118, (3)\
304$^{b}$ & PSZ2 G071.82$-$56.55 & 4.78 & 23 09 37.35 & $-$04 09 52.17 & 1.20 & 0.822 ;$-$ & 38 & $-$ & $-$ & $-$ & 1 &\
323 & PSZ2 G075.85$+$15.53 & 4.64 & 19 10 50.62 & $+$44 54 48.99 & 3.30 & $-$ & $-$ & 0.30$\pm$0.03 & 7.8$\pm$2.8 & 7.3 & 2 &\
327-A$^{a,d}$ & PSZ2 G076.55$+$60.29 & 5.42 & 14 52 00.51 & $+$44 31 21.31 & 4.21 & 0.287 ;$-$ & 5 & 0.25$\pm$0.03 & 35.9$\pm$6.0 & 23.1 & 2 & WHL J145206.4$+$443235\
327-B$^{a}$ & & & 14 52 24.24 & $+$44 22 56.58 & 5.33 & 0.632 ;$-$ & 1 & 0.70$\pm$0.06 & 14.5$\pm$3.8 & 8.4 & 2 &\
333 & PSZ2 G078.10$-$83.83 & 4.87 & 00 32 30.94 & $-$22 42 10.95 & 2.00 & $-$ & $-$ & 0.28$\pm$0.03 & 18.4$\pm$4.3 & 8.7 & 2 &\
355 & PSZ2 G082.37$+$22.35 & 5.93 & 18 44 31.37 & $+$53 00 09.01 & 0.60 & $-$ & $-$ & 0.29$\pm$0.03 & 14.5$\pm$3.8 & 9.3 & 2 &\
359 & PSZ2 G083.56$+$24.90 & 6.13 & 18 29 28.55 & $+$54 43 08.80 & 1.67 & $-$ & $-$ & 0.32$\pm$0.04 & 7.0$\pm$2.6 & 2.7 & 2 &\
378 & PSZ2 G085.95$+$25.23 & 5.55 & 18 30 23.81 & $+$56 53 11.12 & 0.62 & $-$ & $-$ & 0.65$\pm$0.05 & 4.6$\pm$2.1 & 1.5 & 2 &\
381-A & PSZ2 G086.28$+$74.76 & 5.07 & 13 38 40.43 & $+$38 52 32.57 & 8.93 & 0.246 ;$-$ & 20 & $-$ & $-$ & $-$ & 3 &\
381-B$^{a,c}$ & & & 13 37 54.11 & $+$38 53 30.94 & 1.28 & 0.701 ;$-$ & 21 & 0.80$\pm$0.06 & $-$ & $-$ & 1 &\
394$^{a,c}$ & PSZ2 G087.39$-$34.58 & 4.62 & 22 49 09.53 & $+$19 44 30.50 & 1.93 & 0.772 ;$-$ & 31 & 0.70$\pm$0.07 & $-$ & $-$ & 1 & (4)\
468 & PSZ2 G098.75$-$28.63 & 4.74 & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & ND &\
475 & PSZ2 G099.55$+$34.23 & 5.34 & 17 10 33.34 & $+$68 44 43.60 & 1.01 & $-$ & $-$ & 0.31$\pm$0.03 & 15.4$\pm$3.9 & 6.0 & 2 & WHY J171033.4$+$684443\
548 & PSZ2 G113.27$+$48.39 & 5.30 & 13 58 59.49 & $+$67 25 50.29 & 0.75 & $-$ & $-$ & 0.32$\pm$0.03 & 7.9$\pm$2.8 & 4.5 & 2 &\
581 & PSZ2 G118.49$+$48.17 & 5.16 & 13 23 55.03 & $+$68 39 30.73 & 1.09 & $-$ & $-$ & 0.35$\pm$0.03 & 24.4$\pm$4.9 & 6.2 & 2 & WHY J132355.0$+$683931\
582 & PSZ2 G118.56$-$13.14 & 4.63 & 00 25 13.35 & $+$49 30 35.84 & 0.50 & $-$ & $-$ & 0.23$\pm$0.03 & 20.9$\pm$4.6 & 3.7 & 2 &\
590 & PSZ2 G120.30$+$44.47 & 5.31 & 13 16 38.50 & $+$72 32 15.60 & 1.09 & $-$ & $-$ & 0.23$\pm$0.03 & 23.3$\pm$4.8 & 7.9 & 2 & WHY J131638.6$+$723217\
623$^{b}$ & PSZ2 G126.28$+$65.62 & 4.67 & 12 42 23.33 & $+$51 26 20.98 & 1.67 & 0.819 ; 0.8201 & 16 & $-$ & $-$ & $-$ & 1 & (4)\
625$^{a,c}$ & PSZ2 G126.57$+$51.61 & 6.35 & 12 29 47.56 & $+$65 21 13.41 & 0.33 & 0.817 ;$-$ & 20 & 0.80$\pm$0.10 & $-$ & $-$ & 1 & (4)\
654$^{a}$ & PSZ2 G133.92$-$42.73 & 4.70 & 01 25 33.36 & $+$19 22 53.51 & 1.65 & 0.581 ;$-$ & 3 & 0.65$\pm$0.07 & 2.3$\pm$1.5 & 1.1 & ND &\
681-A & PSZ2 G139.00$+$50.92 & 4.98 & 11 20 22.76 & $+$63 14 38.35 & 1.63 & 0.636 ;$-$ & 13 & $-$ & $-$ & $-$ & 3 &\
681-B & & & 11 20 27.45 & $+$63 14 46.15 & 2.04 & 0.784 ;$-$ & 11 & $-$ & $-$ & $-$ & 1 &\
690-A$^{a,d}$ & PSZ2 G141.98$+$69.31 & 4.71 & 12 12 38.98 & $+$46 21 06.46 & 3.25 & 0.713 ;$-$ & 16 & 0.70$\pm$0.06 & $-$ & $-$ & 1 & WHL J121240.6$+$462123\
690-B & & & 12 12 42.63 & $+$46 21 04.59 & 2.65 & 0.796 ;$-$ & 9 & $-$ & $-$ & $-$ & 3 &\
701 & PSZ2 G144.23$-$18.19 & 5.22 & 02 38 55.30 & $+$40 11 11.57 & 0.41 & $-$ & $-$ & 0.31$\pm$0.03 & 13.4$\pm$3.7 & 2.2 & 2 &\
764 & PSZ2 G159.40$-$40.67 & 5.05 & 02 42 22.99 & $+$14 15 14.60 & 2.81 & $-$ & $-$ & 0.22$\pm$0.03 & 23.2$\pm$4.8 & 5.4 & 2 &\
768 & PSZ2 G160.83$-$70.63 & 6.30 & 01 39 20.72 & $-$11 22 19.49 & 1.20 & $-$ & $-$ & 0.24$\pm$0.03 & 26.2$\pm$5.1 & 7.7 & 2 &\
810$^{a,c}$ & PSZ2 G171.08$-$80.38 & 4.90 & 01 21 53.44 & $-$20 33 26.45 & 2.67 & 0.313 ; 0.3134 & 33 & 0.33$\pm$0.03 & $-$ & $-$ & 1 & WHL J012153.4$-$203327\
902$^{b}$ & PSZ2 G198.80$-$57.57 & 4.83 & 03 02 06.58 & $-$15 33 41.69 & 0.53 & 0.530 ; 0.5292 & 16 & $-$ & $-$ & $-$ & 1 &\
937$^{b}$ & PSZ2 G208.57$-$44.31 & 4.53 & 04 02 36.08 & $-$15 40 49.56 & 1.47 & 0.820 ; 0.8196 & 17 & $-$ & $-$ & $-$ & 1 &\
1254$^{a,c}$ & PSZ2 G270.78$+$36.83 & 4.99 & 11 04 21.06 & $-$19 14 18.34 & 2.55 & 0.516 ; 0.5146 & 25 & 0.52$\pm$0.05 & $-$ & $-$ & 1 & Double detection in (5)\
1606 & PSZ2 G343.46$+$52.65 & 4.89 & 14 24 23.15 & $-$02 43 49.34 & 0.88 & 0.711 ;$-$ & 20 & 0.70$\pm$0.07 & $-$ & $-$ & 1 & (4)\
$^1$ SZ targets identified with the ID followed by an A or B label indicate the presence of multiple counterparts.
$^2$ References. (1) [@2dF], (2) [@Amodeo18], (3) [@Buddendiek15], (4) [@Burenin18], (5) [@Streblyanska18]
$^a$ Photometric and/or spectroscopic redshift obtained from SDSS DR14 data.
$^b$ No imaging performed, private communication with Saclay group
$^c$ Already confirmed in [@Streblyanska18]
$^d$ Classified as “potentially associated” in [@Streblyanska18]
On PSZ2 statistics in the northern sky {#sec:fullpsz2}
======================================
In PSZ2 catalogue there are 1003 sources with $Dec.>-15\degree$. After the two years of [LP15]{} observations, a total of 226 sources have been observed; 184 of them were part of the [LP15]{} sample and thus were not validated at the time the PSZ2 catalogue was published. In addition, we updated the redshift for 42 additional sources. In this section, we will carry out the statistical analysis of this “northern sky” sub-sample of the PSZ2, such as the purity and effects that can influence the PSZ detection. For definiteness, we will refer to this sub-sample as PSZ2-North, which represents the $60\,\%$ of the complete PSZ2 sample.
We note that this PSZ2-North sample also includes some PSZ2 sources associated with PSZ1 objects, that were observed during the [ITP13]{} [@paper1; @paper2]. There are still five sources ($<0.5\,\%$) that could not be observed in order to validate the full PSZ2-North, so we exclude them of the sample for the computation of the statistics in this section.
![PSZ2 cluster counts as a function of the signal-to-noise ratio of the SZ detection $S/N$. The PSZ2-North sample is represented in light blue, the sources still not observed are represented in red ($<0.5\%$), the updated sources described in Sect. \[sec:wise\] are shown in dark blue ($3.4\%$) and the sources originally not confirmed that were observed during [LP15]{} and [ITP13]{} are shown in green ($18.4\%$ and $4.8\%$, respectively). The bin size is $0.25$.[]{data-label="FigSNR"}](Figures/N_SNR.jpeg){width="\columnwidth"}
Figure \[FigSNR\] shows the number of clusters as a function of the signal-to-noise ratio in the catalogue. The vast majority of the sources studied in this work present $S/N < 6$ and it is within this range where this optical follow-up has performed the largest contribution. In particular, we observed $37\%$ of the sources with $4.5 < S/N < 6$.
![Accumulated purity of the PSZ2-North sample (Dec.$>-15\degree$) studied as a function of the $S/N$, i.e. the percentage of the sources that are actual clusters and related to the SZ signal. In blue we represent the original purity of the catalogue and in red the same purity but after this work. []{data-label="Figpurity"}](Figures/purity.jpeg){width="\columnwidth"}
We define the purity as the ratio between confirmed clusters and the total number of SZ sources. It is important to take into account that we have explored the optical range in which the dust emission could be masking the possible counterpart, later in this section we quantify this effect. Figure \[Figpurity\] shows the accumulated purity of the PSZ2-North sample as a function of the $S/N$. While originally, it showed a purity of 76.7%, after every validation programme to date, the purity increases up to $86.2\,\%$ for $S/N > 4.5$. The feature in Fig. \[Figpurity\] showing a decrease of purity in the range $12 < S/N < 16$ is due to the existence of one non-detection listed in the PSZ2 as high $S/N$ source (PSZ2 G153.56+36.82), studied in detail in Paper I.
![Cluster counts as a function of redshift. Colour codes are the same as in Fig. \[FigSNR\]. Dark blue bars represent the ratio between clusters confirmed during our follow-ups and the total confirmed clusters. The size of the redshift bin is $0.05$.[]{data-label="FigNz"}](Figures/N_z.jpeg){width="\columnwidth"}
Figure \[FigNz\] shows the distribution of redshifts of the [*Planck* ]{}confirmed clusters. We note that $77\,\%$ of them present a redshift between $0.05 < z <
0.4$ which is the optimal range for the cluster detection of the [*Planck* ]{}mission. The median redshift of the PSZ2-North sample is 0.23 while the median redshift of the clusters confirmed during [LP15]{} is 0.29. While we confirm about $10\,\%$ of the clusters at $z < 0.4$, this rate is $\sim 20\,\%$ for $z >
0.4$. Moreover, we confirmed in [@paper2] the most distant [*Planck* ]{}SZ cluster in the northern hemisphere: the PSZ2 G123.35+25.39, at $z_{\rm phot}=0.95$.
[@Burenin17] presented an extension for the PSZ2 catalogue using SDSS and WISE. We find 28 matches between this catalogue and the [LP15]{} sample. Our results are in good agreement for the majority of the sources. For PSZ2 G069.47$-$29.06 and PSZ2 G130.64$+$37.16, the author only reports one counterpart, while we find two. PSZ2 G069.47-29.06 was discussed in Paper I, where both candidates where confirmed presenting 44 and 30 spectroscopic members, as mentioned in [@Zaznobin19]. PSZ2 G130.64$+$37.16 was discussed in Sect. \[sec:notes\]. On the contrary, for PSZ2 G066.59$-$58.51, we find only one counterpart, while [@Burenin17] reports more than one.
We compare our results with those of [@Zohren19 in prep., private communication]. They use the WHT to validate high-z clusters of the [*Planck* ]{}catalogues. They report the redshift, richness and mass for 23 candidates. Twenty of them were also observed during the [LP15]{} programme. We agree with their results but for three cases. They claim as well as [@Burenin18] that PSZ2 092.69$+$59.92 has two counterparts, at $z = 0.46$ and $z = 0.84$. Our spectroscopic observations reveal that the galaxy over-density at $z = 0.84$ is a low mass system as it presents $\sigma_{v} <
450$kms$^{-1}$. They found PSZ2 G139.00$+$50.92 presents a mass below their limit for validation. As discussed in Sect. \[sec:notes\_update\], we find two possible counterparts, one of them (681-B) showing a $\sigma_{v} >
800$kms$^{-1}$. PSZ2 G165.41$+$25.93 is also below their mass limit whereas in our richness analysis it shows a $\sigma_{R} = 1.8$, just above our validation limit of $\sigma_{R} = 1.5$.
We also compare our results with [@Zaznobin19] where the authors report 38 spectroscopic redshift for PSZ2 candidates. We find 20 matches between this catalogue and the [LP15]{} sample. We find discrepancies in only one case: PSZ2 G202.61$-$26.26. The authors report three spectroscopic redshifts at $z_{\rm spec} = 0.533$ while we find a galaxy over-density at $z_{\rm phot} = 0.23$ but further than $5{'}$ away from the [*Planck* ]{}centre, so not linked to the SZ emission.
![Ratio between unconfirmed and total number of sources for the PSZ2-North sample (Dec.$>-15\degree$) as a function of the $857$GHz signal in bins of $1$MJysr$^{-1}$. Error bars correspond to a Poisson error in the distribution of total number of sources. []{data-label="Figdust"}](Figures/dust.jpeg){width="\columnwidth"}
In order to study the galactic disturbance on the SZ [*Planck* ]{}detection, we compute the number of non-detections as a function of the $857\,GHz$ signal in the [*Planck* ]{}map. This map could be used as a tracer of the thermal dust emission . The signal is computed as the mean value within a region of $0.5\degree$ radius around the nominal pointing in the PSZ2 catalogue. Figure \[Figdust\] represents the ratio between unconfirmed and total number of sources for the PSZ2-North sample as a function of the $857$GHz signal in bins of $1$MJysr$^{-1}$. This figure shows a clear correlation between these two magnitudes. Below $7$MJysr$^{-1}$ the ratio of unconfirmed sources is under $20\%$. However, in zones with high dust emission (mainly places in the galactic plane), the false SZ clusters can be higher than $60$–$70\%$.
![Number of cluster-candidates versus the neural network quality flag value for the PSZ2-North sample. The confirmed candidates are represented in green while the still not confirmed are shown in red. The bin size is 0.05. []{data-label="FigQbad"}](Figures/Qbadall.jpeg){width="\columnwidth"}
Figure \[FigQbad\] shows the number of cluster-candidates versus the neural network quality flag value for the PSZ2-North sample ($Q_{bad}$). This value was defined in [@Aghanim15], it is an indicator of the reliability of a SZ source to be a real galaxy cluster. Candidates presenting values of $Q_{bad} >
0.6$ are considered low reliable sources. In the PSZ2-North sample, we observe that the vast majority ($> 93\,\%$) of the clusters with $Q_{bad} < 0.6$ are actual clusters while less than $33\,\%$ with $Q_{bad} > 0.6$ are confirmed.
We also compare the full validation results with [@Khatri16] where the author published a method of validation for the PSZ2 catalogue based on the combination of CO and $y$-distortion maps. He classifies the sources in five different groups depending on the value of his estimator: [*MOC, pMOC, CLG, pCLG*]{} and [*IND*]{}. The signal of the sources classified as [*MOC*]{} and [*pMOC*]{} is considered to come from molecular clouds while [*CLG*]{} and [ *pCLG*]{} come from galaxy clusters. [*IND*]{} is indeterminable. We find that $95.2\,\%$ of the [*IND*]{} and $94.6\,\%$ of the [*CLG*]{} $+$ [*pCLG*]{} correspond to actual validated clusters. On the other hand, $64.7\,\%$ of the [*MOC*]{} $+$ [*pMOC*]{} are also validated clusters. We expected that the sources with this classification present a lower rate of validation, however, it is not the case. A possible explanation for these results is that the threshold used by [@Khatri16] to distinguish between molecular clouds and clusters was shifted towards high values of $\Delta(\Sigma\chi^{2})_{CO-y}$. To illustrate this fact, for the 59 sources that the author classifies as [*pMOC*]{}, 48 ($81.3\,\%$) are actual clusters. A full study on this matter will be approached in a future publication where we will discuss the properties of the clusters, velocity dispersion, masses and their relations with the SZ signal.
Conclusions {#sec:conclusions}
===========
This work presents the final results of the observational programme [LP15]{} started in Paper I. We report here about the second year of observations carried out using INT, TNG and GTC at ORM Observatory, as part of the optical follow-up confirmation and characterization programme of [*Planck* ]{}SZ sources in the northern hemisphere.
During the second year of observations, 78 PSZ2 sources with no known optical counterpart have been observed. Thanks to a robust confirmation criterion based on velocity dispersion, when available, and richness estimations we were able to confirm 40 candidates, providing 18 spectroscopic and 22 photometric redshifts.
We update the information on 42 sources that were already validated in the original PSZ2 catalogue but presented no redshift estimation. We provide spectroscopic redshift for 20 of them and photometric redshift for 20. We also perform the richness study and apply the same criteria as for the candidates in order to check the associations to the SZ signal. We discover that three already confirmed counterparts were not present in the optical range studied here.
At the end of the whole observational programme [LP15]{}, we were able to confirm 81 new cluster candidates, with a median redshift of $0.29$ while the mean redshift of the catalogue is $0.23$. Our main contribution appears in the redshift interval $0.4 < z < 0.7$, where our confirmations correspond to $20\,\%$ of the total clusters confirmed in the PSZ2 in that range. The purity of the catalogue has been updated from $76.7\,\%$ to $86.2\,\%$.
Finally, we found a clear correlation between the number of unconfirmed sources and galactic thermal dust emission. This correlation suggests that there are spurious detections inside the PSZ2 catalogue. Some of these false detections have been discussed here. In particular, we find that more than $50\,\%$ of the sources with a mean signal in the $857$GHz maps greater than $7$MJy/sr remain unconfirmed after this work.
This article is based on observations made with a) the Gran Telescopio Canarias operated by the Instituto de Astrofisica de Canarias, b) the Isaac Newton Telescope, and the William Herschel Telescope operated by the Isaac Newton Group of Telescopes, and c) the Italian Telescopio Nazionale Galileo operated by the Fundacion Galileo Galilei of the INAF (Istituto Nazionale di Astrofisica). All these facilities are located at the Spanish Roque de los Muchachos Observatory of the Instituto de Astrofisica de Canarias on the island of La Palma. This research has been carried out with telescope time awarded for the programme 128-MULTIPLE-16/15B. Also, during our analysis, we used the following databases: the SZ-Cluster Database operated by the Integrated Data and Operation Center (IDOC) at the IAS under contract with CNES and CNRS and the Sloan Digital Sky Survey (SDSS) DR14 database. Funding for the SDSS has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Aeronautics and Space Administration, the National Science Foundation, the U.S. Department of Energy, the Japanese Monbukagakusho, and the Max Planck Society. This work has been partially funded by the Spanish Ministry of Economy and Competitiveness (MINECO) under the projects ESP2013-48362-C2-1-P, AYA2014-60438-P and AYA2017-84185-P. AS and RB acknowledge financial support from the Spanish Ministry of Economy and Competitiveness (MINECO) under the Severo Ochoa Programmes SEV-2011-0187 and SEV-2015-0548. HL is funded by PUT1627 grant from the Estonian Research Council and by the European Structural Funds grant for the Centre of Excellence “Dark Matter in (Astro)particle Physics and Cosmology” TK133. Some of the results in this paper have been derived using the [HEALPix]{} [@Healpix] package.
[^1]: [*Planck* ]{} <http://www.esa.int/Planck> is a project of the European Space Agency (ESA) with instruments provided by two scientific consortia funded by ESA member states and led by Principal Investigators from France and Italy, telescope reflectors provided through a collaboration between ESA and a scientific consortium led and funded by Denmark, and additional contributions from NASA (USA).
[^2]: The NASA/IPAC Extragalactic Database (NED) is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration
[^3]: [IRAF]{} (<http://iraf.noao.edu/>) is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc., under the cooperative agreement with the National Science Foundation.
[^4]: RVSAO was developed at the Smithsonian Astrophysical Observatory Telescope Data Center.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We present new [*Far Ultraviolet Spectroscopic Explorer*]{} (FUSE) observations of Mira A’s wind-accreting companion star, Mira B. We find that the strongest lines in the FUSE spectrum are H$_{2}$ lines fluoresced by H I Ly$\alpha$. A previously analyzed [*Hubble Space Telescope*]{} (HST) spectrum also shows numerous Ly$\alpha$-fluoresced H$_{2}$ lines. The HST lines are all Lyman band lines, while the FUSE H$_{2}$ lines are mostly Werner band lines, many of them never before identified in an astrophysical spectrum. We combine the FUSE and HST data to refine estimates of the physical properties of the emitting H$_{2}$ gas. We find that the emission can be reproduced by an H$_{2}$ layer with a temperature and column density of $T=3900$ K and $\log N(H_{2})=17.1$, respectively. Another similarity between the HST and FUSE data, besides the prevalence of H$_{2}$ emission, is the surprising weakness of the continuum and high temperature emission lines, suggesting that accretion onto Mira B has weakened dramatically. The UV fluxes observed by HST on 1999 August 2 were previously reported to be over an order of magnitude lower than those observed by HST and the [*International Ultraviolet Explorer*]{} (IUE) from 1979–1995. Analysis of the FUSE data reveals that Mira B was still in a similarly low state on 2001 November 22.'
author:
- 'Brian E. Wood and Margarita Karovska'
title: '**Molecular Hydrogen Emission Lines in [*Far Ultraviolet Spectroscopic Explorer*]{} Observations of Mira B**'
---
INTRODUCTION
============
Mira A (o Ceti, HD 14386) is the prototype for a class of pulsating giant stars on the asymptotic giant branch. The pulsations of Mira variables help drive very strong winds from the surfaces of these stars. Mass loss rate estimates for Mira A itself generally fall in the range $4\times 10^{-8}$ to $4\times 10^{-7}$ M$_{\odot}$ yr$^{-1}$ [@yy78; @grk85; @pfb88; @pp90; @grk98; @nr01]. Mira A has a companion star, Mira B, which is located $0.6^{\prime\prime}$ away [@mk97], corresponding to a projected distance of about 70 AU at its distance of $128\pm 18$ pc [@macp97]. Mira A’s wind is being accreted by Mira B, forming an accretion disk. The Mira system is attractive for studying accretion processes since it is one of the few wind accretion systems in which the components of the system are resolvable.
Mira B’s accretion disk emits broad, high temperature emission lines of C IV $\lambda$1550, Si III\] $\lambda$1892, and Mg II $\lambda$2800, among others, which were first observed by the [*International Ultraviolet Explorer*]{} (IUE) [@dr85]. The optical and UV continuum of Mira B appears to be dominated by the accretion based on its strong variability on many timescales, and based on accretion rate estimates of $(8-30)\times 10^{-10}$ M$_{\odot}$ yr$^{-1}$ [@bw72; @yy77; @mj84; @dr85]. Because of the complexities involved with the wind accretion onto Mira B, it is uncertain whether the star is a white dwarf or a red dwarf.
The UV spectrum of Mira B was observed many times by IUE between 1979 and 1995 [@dr85], and also by the Faint Object Camera (FOC) instrument on the [*Hubble Space Telescope*]{} (HST) on 1995 December 11 [@mk97]. The UV continuum and emission lines of Mira B show some modest variability within the 1979–1995 data, with fluxes varying by about a factor of 2. However, on 1999 August 2 the Space Telescope Imaging Spectrograph (STIS) instrument on HST obtained a UV spectrum from Mira B that was radically different from any previous observation [@bew01 hereafter Paper I]. The fluxes of the continuum and high temperature emission lines (e.g., C IV $\lambda$1550, Si III\] $\lambda$1892, Mg II $\lambda$2800, etc.) were over an order of magnitude lower than ever observed before. Furthermore, the character of the spectrum below 1700 Å had changed dramatically, with the spectrum dominated by many narrow H$_{2}$ lines rather than being dominated by the aforementioned broad, high temperature lines. These H$_{2}$ lines are pumped by the strong H I Ly$\alpha$ line, a fluorescence mechanism that has been found to produce detectable H$_{2}$ emission from the Sun and more recently from many other astrophysical sources [@cj77; @ab81; @adm99; @jav00; @dra02]. The surprising STIS data raise many new questions about the accretion onto Mira B. Why did the UV fluxes fall so dramatically? Where are all these H$_{2}$ lines coming from? Why were the H$_{2}$ lines not observed by IUE?
@bew02 [hereafter Paper II] analyzed the H$_{2}$ lines in detail. They found that the dominance of H$_{2}$ emission in the 1999 HST/STIS data is due at least in part to an H I Ly$\alpha$ line that is [*not*]{} weaker than during the IUE era, unlike the continuum and every other non-H$_{2}$ line in the spectrum. Therefore, the H$_{2}$ lines pumped by Ly$\alpha$ appear much stronger relative to other UV emission lines than before, when the H$_{2}$ lines were not even detectable by IUE. It was proposed in Papers I and II that the fundamental cause of the change in Mira B’s UV spectrum was an order of magnitude decrease in the accretion rate onto the star. This interpretation is supported by analysis of wind absorption in the Mg II h & k lines at 2803 and 2796 Å, respectively, which shows that the accretion-driven mass loss rate from Mira B at the time of the HST/STIS observations is lower by about an order of magnitude from what it was during the IUE era, consistent with the observed decrease in accretion luminosity. Exactly why the Ly$\alpha$ flux did [*not*]{} decrease with everything else in the spectrum remains somewhat of a mystery. In Paper II, we suggested that the Ly$\alpha$ emission may have indeed decreased, but the weaker wind opacity at the time of the HST/STIS observations allowed more Ly$\alpha$ emission to escape and compensated for this decrease.
As far as where the H$_{2}$ lines are coming from, several arguments were presented in Paper II against the H$_{2}$ emission being from the accretion disk. Instead, the H$_{2}$ lines are most likely coming from H$_{2}$ within Mira A’s wind, which is being heated and dissociated by H I Ly$\alpha$ as it approaches Mira B. The H$_{2}$ emission line ratios and the amount of H$_{2}$ absorption observed for the pumping transitions within Ly$\alpha$ are both consistent with an H$_{2}$ layer with $T\approx 3600$ K and $\log N(H_{2})\approx 17.3$. This temperature is close to the dissociation temperature of H$_{2}$, suggesting that the H$_{2}$ could be from an H$_{2}$ photodissociation front surrounding Mira B. A photodissociation front model presented in Paper II demonstrates that such a front can indeed reproduce the properties of the H$_{2}$ emission, although it was suggested that the collision of the winds of Mira A and B could also play a role in heating the H$_{2}$. The H$_{2}$ photodissociation rate estimated from the data is roughly consistent with Mira B’s $\sim 10^{-10}$ M$_{\odot}$ yr$^{-1}$ total accretion rate, meaning that the H$_{2}$ we are seeing being fluoresced and dissociated by Ly$\alpha$ is probably on its way to being ultimately accreted onto Mira B. Molecular hydrogen is the dominant constituent of Mira A’s wind by mass, so the accretion processes relating to H$_{2}$ are particularly important. The fluorescence, dissociation, and heating of the H$_{2}$ by Ly$\alpha$, which is what the UV H$_{2}$ lines are probing, is therefore an important step in the process of accretion onto Mira B.
In this paper, we report on new UV observations of Mira B from the [*Far Ultraviolet Spectroscopic Explorer*]{} (FUSE). The FUSE satellite observes the $905-1187$ Å wavelength range, which is almost entirely inaccessible to the HST. The Mira binary system has never been observed in this wavelength region, so the FUSE data allow us to search for new emission line diagnostics for Mira B.
FUSE OBSERVATIONS
=================
The FUSE satellite observed Mira B on 2001 November 27 starting at UT 9:04:57. The observation consisted of 11 separate exposures through the low resolution (LWRS) aperture, with a combined exposure time of 29,183 s. The spectra were processed using version 2.0.5 of the CALFUSE pipeline software. Note that while Mira A and Mira B are both within the $30^{\prime\prime}\times 30^{\prime\prime}$ LWRS aperture, experience with IUE data demonstrates that Miras do not produce any detectable emission below 2000 Å [@bew00], so any emission detected by FUSE should be from Mira B.
In order to fully cover its 905–1187 Å spectral range, FUSE has a multi-channel design — two channels (LiF1 and LiF2) use Al+LiF coatings, two channels (SiC1 and SiC2) use SiC coatings, and there are two different detectors (A and B). For a full description of the instrument, see @hwm00. With this design FUSE acquires spectra in 8 segments covering different, overlapping wavelength ranges. We coadded the individual exposures to produce spectra for each segment. We experimented with cross-correlating the exposures before coaddition, but there are not many spectral features strong enough to cross-correlate against, so the simple straight coaddition proved more reliable. Coaddition of the various segments with each other can potentially lead to degradation of spectral resolution, but we decided to do this anyway to increase the low signal-to-noise (S/N) of our data. We inspected the individual segments carefully to ensure they were reasonably well aligned before coadding them.
Figure 1 shows two regions of the FUSE spectrum. The top panel is a coaddition of the SiC1B and SiC2A segments and the bottom panel is a combination of the LiF1B and LiF2A segments. We have rebinned the two spectra by factors of 15 and 5, respectively, to increase S/N. The FUSE data highly oversample the line spread function, so this rebinning does not severely degrade the spectral resolution even for the factor-of-15 rebinning. The bottom spectrum in Figure 1 has also been smoothed slightly, but only for purposes of display. No stellar emission is detected between 990 and 1115 Å, so that region of the FUSE spectrum is not shown.
The top panel of Figure 1 shows mostly only airglow lines of H I and O I. There is a broad emission feature at 976 Å that may be C III 977.02 Å emission from the star, but the feature is shifted from its expected location by a suspiciously large amount ($\sim$ 200 km s$^{-1}$), making this identification tentative (see §5). The bottom panel shows numerous narrow lines that we identify as H$_{2}$ emission, as will be described in detail below. Many of these lines are blended with the C III $\lambda$1175 multiplet, making it unclear whether C III is really contributing any flux to the blend at all. Both C III lines shown in Figure 1 are strong lines frequently seen in stellar spectra, and we expected to be able to detect these lines from the Mira B accretion disk. We will discuss them further in §5.
We have fitted Gaussians to all of the stellar emission features shown in Figure 1, using a chi-squared minimization routine to determine the best fit [e.g., @prb92]. The first three columns of Table 1 list the fit parameters and their 1$\sigma$ uncertainties. The parameters are the central wavelength ($\lambda_{meas}$), flux ($f_{obs}$, in units of $10^{-15}$ ergs cm$^{-2}$ s$^{-1}$), and full-width-at-half-maximum (FWHM). The uncertainties are estimated using a Monte Carlo technique whereby the fluxes in the spectrum are varied within the bounds suggested by the flux error vector, and a large number of fits are performed on the altered spectra to see how much fit parameters vary. Our quoted 1$\sigma$ uncertainties are the 1$\sigma$ variations of the Monte Carlo trials. Figure 2 shows the fit to the complicated blend at 1176 Å. The narrow lines are all H$_{2}$ lines, and we assume the sum of the two broad components is representative of the C III emission. The parameters for both broad Gaussians are listed in Table 1 and tentatively identified as C III.
The line spread function of FUSE is not particularly well defined, so we have made no attempt to correct for instrumental broadening in our fits. The H$_{2}$ lines observed by HST/STIS have an average width of $FWHM=19.7\pm 0.4$ km s$^{-1}$, after correction for instrumental broadening (Paper II). The effective resolution of FUSE is $\sim 20$ km s$^{-1}$, so we expect the FUSE H$_{2}$ lines to have widths of $\sim 28$ km s$^{-1}$. This agrees reasonably well with the measured line widths in Table 1, with blending perhaps being responsible for the broadening of a few H$_{2}$ lines. The H$_{2}$ lines have average redshifts of $+62$ km s$^{-1}$ relative to their rest wavelengths. Considering the $\sim 5$ km s$^{-1}$ uncertainty in the FUSE wavelength calibration, this is consistent with the $+56.9\pm 0.2$ km s$^{-1}$ velocity found for the STIS H$_{2}$ lines, which in turn is consistent with the Mira system’s radial velocity of $\sim 56$ km s$^{-1}$ [@pfb88; @pp90; @ej00].
IDENTIFYING THE H$_{2}$ LINES
=============================
Most of the H$_{2}$ lines listed in Table 1 have never before been identified in any astrophysical spectrum. In Paper II, we found that all of the H$_{2}$ lines observed by HST/STIS could be associated with Lyman band fluorescence sequences pumped by the strong H I Ly$\alpha$ line. The Ly$\alpha$ emission excites H$_{2}$ from various rovibrational states within the ground electronic state ($X^{1}\Sigma_{g}^{+}$) to the excited $B^{1}\Sigma_{u}^{+}$ state. Radiative deexcitation back to various levels of the ground electronic state yields the Lyman band fluorescence sequences observed by STIS. We conclusively identified H$_{2}$ lines in the STIS spectrum by finding many lines in each fluorescence sequence, each having a Ly$\alpha$ pumping path, and noting that the line ratios are at least roughly consistent with the transition branching ratios. This is harder to do in our FUSE spectrum, because we have fewer lines to work with. The Lyman band sequences from Paper II only contribute to a couple of the H$_{2}$ lines in Figure 1 (at 1143 Å and 1162 Å). The other narrow lines are from previously unidentified sequences.
We used the Lyman band line list of @ha93a to search for possible matches to our narrow FUSE lines, but we quickly found that most of these lines are clearly not Lyman band H$_{2}$. However, we were much more successful when we searched the Werner band H$_{2}$ line list of @ha93b. Just as emission from Ly$\alpha$ can excite H$_{2}$ to the $B^{1}\Sigma_{u}^{+}$ state to produce Lyman band emission, Ly$\alpha$ can also excite H$_{2}$ to the excited $C^{1}\Pi_{u}$ state to produce Werner band emission, most of which falls within the FUSE bandpass rather than the higher wavelength range of STIS.
The Werner band fluorescence is of a slightly different character than the Lyman band fluorescence. Quantum selection rules require Lyman band transitions to have $\Delta J\equiv J''-J'=\pm 1$, where $J'$ and $J''$ are the upper and lower rotational quantum numbers, respectively. Transitions with $\Delta J=+1$ are defined as P-branch lines and those with $\Delta J=-1$ are R-branch lines. The $C^{1}\Pi_{u}$ state is actually degenerate and consists of a $C^{+}$ state and a $C^{-}$ state. The energy levels of the two states are practically identical, but Werner band transitions from $C^{+}$ must have $\Delta J=\pm 1$ (i.e., P and R branch transitions), while transitions from $C^{-}$ must have $\Delta J=0$ (Q branch transitions). The Werner band is therefore different from the Lyman band in having a Q branch, but because the $C^{+}$ and $C^{-}$ states are different the Q branch lines can only be pumped by Q branch transitions within Ly$\alpha$, and the P and R branch lines can only be pumped by P and R branch transitions.
Perusal of the @ha93a [@ha93b] line lists leads to possible H$_{2}$ identifications for all the narrow lines seen in the FUSE data, but conclusive identification is complicated due to most potential new fluorescence sequences having only 1 or 2 observed lines, and also due to many of the lines possibly being blends of more than one H$_{2}$ line. In order to clarify the situation and provide support for the detection of some of the weaker H$_{2}$ lines, we perform a forward modeling exercise using the results of Paper II. In Paper II, we used the STIS H$_{2}$ lines to estimate the temperature and column density of the H$_{2}$, and we also reconstructed the Ly$\alpha$ profile that the H$_{2}$ must see in order to account for the observed H$_{2}$ flux. Using these results, we create a synthetic H$_{2}$ spectrum, considering all new Lyman and Werner band fluorescence sequences that may be contributing to the FUSE H$_{2}$ lines. In Table 1, we list the line fluxes predicted by this forward modeling exercise ($f_{mod}$). In particular, Table 1 lists all lines that we believe contribute at least 10% of the flux of an observed FUSE H$_{2}$ feature based on the $f_{mod}$ values. Note that there is a broad, weak emission feature at 1134–1138 Å that our forward modeling exercise suggests may be partially H$_{2}$ emission. However, H$_{2}$ lines cannot seem to account for most of the flux of the line, so we do not consider it a clearly detected H$_{2}$ feature.
Another purpose of the forward modeling exercise, besides line identification, is to see if the reconstructed Ly$\alpha$ profile and H$_{2}$ parameters from Paper II can reproduce the fluxes in our FUSE data. Table 1 shows that the predicted fluxes generally agree with the measured fluxes rather well, with a few exceptions. The Ly$\alpha$ profile must therefore be about the same at the time of the FUSE observation (2001 November 27) as it was at the time of the STIS observation (1999 August 2). This means that we can consider the FUSE and STIS H$_{2}$ lines together to refine the H$_{2}$ analysis presented in Paper II, which we will do in §4.
While trying to identify the FUSE H$_{2}$ lines, we also revisited the STIS data and found one new STIS H$_{2}$ fluorescence sequence, which is also listed in Table 1. The five detected lines in this sequence are individually rather weak, but collectively they amount to a convincing detection. This new sequence is important because it is fluoresced at a lower wavelength than any other Mira B H$_{2}$ sequence, thereby providing a useful diagnostic for the H$_{2}$-observed Ly$\alpha$ flux at shorter wavelengths (see §4). We note that this sequence was also detected in observations of the T Tauri star TW Hya [@gjh02; @gjh03].
We emphasize that only a few of the H$_{2}$ lines listed in Table 1 have been identified in previous astrophysical spectra. While fluoresced Lyman band H$_{2}$ lines have been observed and analyzed from a number of different sources (see §1), the list of Werner band detections is very short. @jdfb79 noted a couple Werner lines in solar spectra, which are actually fluoresced by the O VI $\lambda$1032 line rather than Ly$\alpha$, and @ew02 report FUSE detections of a few Werner band H$_{2}$ lines from the prototype young accreting star system, T Tauri. Planetary aurora produce abundant H$_{2}$ emission in both the Lyman and Werner bands, but this is collisionally excited rather than fluoresced emission [e.g., @pfm97].
COMPLETE ANALYSIS OF THE H$_{2}$ LINES OF MIRA B
================================================
Table 2 lists all the pumping transitions within Ly$\alpha$ that are responsible for the H$_{2}$ lines observed by FUSE and STIS. Our analysis has expanded the list from the 13 Lyman band sequences analyzed in Paper II to 16 Lyman and 13 Werner sequences. This increase in H$_{2}$ data is sufficient to justify a reanalysis of the H$_{2}$ lines to see if conclusions from Paper II remain unchanged.
The ratios of H$_{2}$ fluxes within each fluorescence sequence are inconsistent with the line branching ratios and are therefore clearly affected by opacity effects (see Paper II). The opacity of the H$_{2}$ lines depends on the level populations of the H$_{2}$ molecules being fluoresced, which in turn depends on the temperature, T, and total column density of the H$_{2}$, $N(H_{2})$. The line ratios are therefore a diagnostic for T and $N(H_{2})$. In Paper II, we developed a plane parallel, Monte Carlo radiative transfer code to determine which T and $N(H_{2})$ values best fit the line ratios. We repeat this analysis including the new H$_{2}$ data presented here. For the blended lines in Table 1, we divide up the line flux according to the percent contributions to the line suggested by the $f_{mod}$ values.
For each transition, we need to know the absorption strength ($f$), and the energy ($E_{low}$) and statistical weight ($g_{low}$) of the lower level. The necessary atomic data are taken from @ha93a [@ha93b] and @id84. Table 2 lists the $f$, $E_{low}$, and $g_{low}$ values of the Ly$\alpha$ pumping transitions. Another factor considered in the radiative transfer model is dissociation. Fluorescence to the $B^{1}\Sigma_{u}^{+}$ and $C^{1}\Pi_{u}$ states can result in dissociation of the H$_{2}$ molecule rather than radiative deexcitation back to the ground electronic state. The dissociation probabilities listed in @ha00 are considered in the model. The $f_{dis}^{\prime}$ values in Table 2 indicate the fractional probability of dissociation for each excitation to the excited electronic state. The $f_{dis}$ values indicate the fraction of fluorescences within each sequence that ultimately lead to dissociation rather than the emergence of an H$_{2}$ line photon, based on our best-fit radiative transfer model (see below). Line opacity can cause multiple photoexcitations before an H$_{2}$ photon emerges or an H$_{2}$ molecule is destroyed, so $f_{dis}>f_{dis}^{\prime}$. In Paper II, the @ha00 tables were read incorrectly, resulting in the dissociation fractions being significantly overestimated. The $f_{dis}^{\prime}$ and $f_{dis}$ values in Table 2 of Paper II are therefore too large in general. Besides consideration of the new FUSE data, a secondary reason for reanalyzing the H$_{2}$ lines is to revise our analysis to see if conclusions made in Paper II regarding the importance of dissociation still stand.
The only free parameters of the model and $T$ and $N(H_{2})$, and Figure 3 shows $\chi^{2}_{\nu}$ contours measuring how well the H$_{2}$ line ratios are reproduced by models with different T and $\log N(H_{2})$ values. The H$_{2}$ line ratios are best explained by an H$_{2}$ layer with a temperature and column density of $T=3900$ K and $\log N(H_{2})=17.1$, respectively, similar to the $T=3600$ K and $\log N(H_{2})=17.3$ results from Paper II. Thus, neither the new FUSE data nor the corrected treatment of dissociation change these values much.
The third column of Table 2 lists the total Ly$\alpha$ flux absorbed that emerges as H$_{2}$ line flux, $F_{obs}$, not including the absorbed flux that leads to H$_{2}$ dissociation. These $F_{obs}$ values include corrections for H$_{2}$ lines in each sequence that are too weak to be detected, with the correction factors provided by the best-fit radiative transfer model. The $F_{obs}$ values in Table 2 are slightly higher than those from Paper II (for the 13 original Lyman band sequences) due to the lower dissociation fractions.
The $F_{obs}$ and $f_{dis}$ values in Table 2 can be used to compute the Ly$\alpha$ flux that must be overlying each H$_{2}$ pumping transition within Ly$\alpha$, thereby allowing us to reconstruct the Ly$\alpha$ profile seen by the H$_{2}$, $f_{0}(\lambda)$. Once again we refer the reader to Paper II for details of this computation, although here we assume our revised values of $T=3900$ K and $\log N(H_{2})=17.1$ in the calculations. Figure 4 shows the new reconstructed Ly$\alpha$ profile. The analysis assumes that the fluoresced H$_{2}$ completely surrounds the star and that the H$_{2}$ emission emerges isotropically. If these assumptions are inaccurate, the derived profile will be different from the actual profile by some geometric scaling factor, $\eta$, explaining why the y-axis of Figure 4 is labeled $\eta f_{0}$. The red and green boxes in Figure 4 show the inferred flux overlying Lyman and Werner band transitions, respectively, which pump observed fluorescence sequences. These data points collectively map out a self-consistent Ly$\alpha$ line profile, albeit with some scatter. The consistency of the Werner band data points, which are entirely from the FUSE data, and the Lyman band data points, which are mostly from the STIS data, provides further evidence that the Ly$\alpha$ profile is the same at the time of the FUSE and STIS observations.
Our best estimate for the reconstructed Ly$\alpha$ profile is indicated by the dashed line in Figure 4. This is not very different from the profile derived in Paper II (also shown in Fig. 4). The biggest difference between the two profiles is below 1215.5 Å, which is entirely due to the addition of the new 3–1 R(15) Lyman band sequence fluoresced at 1214.4648 Å. This is the only sequence fluoresced at wavelengths blueward of the stellar rest frame. Its existence demonstrates that the H$_{2}$ sees at least some Ly$\alpha$ flux on the blue side of the line. As described in Paper II, the reason the observed and reconstructed Ly$\alpha$ lines are highly redshifted is because of absorption from Mira B’s wind on the blue side of the line. The upper limits in Figure 4 are from Lyman and Werner band transitions within Ly$\alpha$ that do not fluoresce enough H$_{2}$ emission to be detectable. These upper limits are consistent with the Ly$\alpha$ profile derived from the detected fluorescence sequences.
Photodissociation rates were overestimated in Paper II due to the use of inaccurate dissociation fractions (see above). Thus, it is worthwhile to derive better estimates for H$_{2}$ photodissociation rates using the revised analysis presented here. We computed these rates for the detected sequences listed in Table 2, using the $F_{obs}$ and $f_{dis}$ values in that table, and we found that some of the weaker sequences that are newly detected in our FUSE data \[e.g., 5–0 P(20)\] are surprisingly strong contributors to the total H$_{2}$ dissociation rate. This suggests that undetected sequences might collectively contribute substantially to H$_{2}$ dissociation via Ly$\alpha$ fluorescence. Therefore, the best way to accurately estimate the H$_{2}$ photodissociation rate is to perform a forward modeling exercise computing contributions for [*all*]{} H$_{2}$ transitions within Ly$\alpha$, not just the ones that yield detected H$_{2}$ emission. The numerous upper limits in Figure 4 provide some idea for how many H$_{2}$ transitions must be considered.
We perform this calculation using the Monte Carlo radiative transfer routine described above and in Paper II, assuming the Ly$\alpha$ profile derived in Figure 4, and assuming the $T=3900$ K and $\log N(H_{2})=17.1$ values from the line ratio analysis to compute the H$_{2}$ level populations. Figure 5 shows a simulated UV H$_{2}$ spectrum computed from this calculation. We estimate that the total Ly$\alpha$ flux absorbed and reemitted by H$_{2}$ is $1.06\times 10^{-12}$ ergs cm$^{-2}$ s$^{-1}$, about 31% higher than the flux one would estimate from the detected sequences alone. About 6.8% of the fluorescences lead to H$_{2}$ dissociation rather than the emergence of an H$_{2}$ line photon, suggesting a dissociation rate of $8.6\times 10^{39}$ s$^{-1}$ (at Mira’s distance of 128 pc), or $4.5\times 10^{-10}$ M$_{\odot}$ yr$^{-1}$. Surprisingly, this is a factor of 2.1 times higher than the value one would derive from the detected sequences alone. Although the undetected sequences absorb less Ly$\alpha$ flux, they tend to have higher dissociation fractions and therefore contribute significantly to the dissociation rate. Nevertheless, the largest contributor to H$_{2}$ dissociation is the detected 4–0 P(19) Lyman band sequence, which accounts for 23.5% of the total dissociation rate.
Although the dissociation fractions assumed in Paper II were too high, this is mitigated by the legitimate increase in dissociation rate provided by consideration of undetected fluorescence sequences. Thus, the dissociation rate quoted above is only a factor of 2 lower than reported in Paper II. In Paper II, we proposed that the H$_{2}$ emission is arising within a photodissociation front within Mira A’s wind as it approaches Mira B, and we constructed a model of this front consistent with the observations. We experiment with new photodissociation front models assuming our revised estimates for photodissociation rates, and we find that the factor of 2 change is not enough to significantly alter the model. Thus, the photodissociation front model remains a very viable interpretation for the H$_{2}$ emission. As mentioned in §1, the implied H$_{2}$ dissociation rate is comparable to Mira B’s total accretion rate, consistent with the idea that the UV H$_{2}$ emission is showing us H$_{2}$ in Mira A’s wind being fluoresced, heated, and dissociated by Ly$\alpha$ as it approaches and is ultimately accreted by Mira B.
We now discuss some of the limitations of our H$_{2}$ modeling efforts, because our estimation of $T$ and $N(H_{2})$ and our reconstruction of the Ly$\alpha$ profile in Figure 4 rely on several assumptions. For example, our plane-parallel modeling essentially assumes that Mira B is surrounded by an isothermal H$_{2}$ layer of uniform $N(H_{2})$ and that H$_{2}$ level populations are precisely thermal. In reality, the H$_{2}$ may have a distribution of temperatures (as the photodissociation front models actually predict), it may be distributed inhomogenously around the star, and the fluorescence process itself could lead to some level of nonthermality in the level populations. The limitations introduced by our simplifying assumptions are presumably responsible for the significant scatter of Ly$\alpha$ flux data points in Figure 4 about the best-fit dashed line Ly$\alpha$ profile, beyond the measurement uncertainties. Since the reconstructed profile imprecisely reproduces the data points in Figure 4, the H$_{2}$ fluxes of the synthetic spectrum described above are naturally imprecise as well. For example, the synthetic spectrum predicts lines at 1130 Å and 1135 Å that are significantly stronger than observed and a line at 1162 Åsignificantly weaker than observed (see Fig. 5). Another source of uncertainty is line blending, which is not considered in our radiative transfer calculations. For example, there are 2 Lyman band data points at 1219 Å in Figure 4 that are too low, probably because these H$_{2}$ pumping transitions are blended enough that they are partially shielding each other from the full Ly$\alpha$ flux present at that wavelength. Emission line blends could also lead to H$_{2}$ photons shifting from one fluorescence sequence to another. All the effects and assumptions described above could lead to inaccuracies in our derived H$_{2}$ properties.
THE C III LINES AND MIRA B’S VARIABILITY
========================================
One goal of our FUSE observations was to try to detect broad, high temperature emission lines formed in the accretion disk of Mira B. We see no evidence of the O VI $\lambda\lambda$1032,1038 lines, which are typically very strong lines in coronal spectra. We estimate a 3$\sigma$ upper limit of $9\times 10^{-16}$ ergs cm$^{-2}$ s$^{-1}$ for these lines. The only possible accretion lines are the two C III lines shown in Figure 1. The widths of these two lines are roughly consistent with observations of high temperature lines from IUE [@dr85], but the identification with C III is still somewhat tentative. We identify the broad bump at 976.447 Å as C III $\lambda$977, but if it is C III it has a very large blueshift of $-232$ km s$^{-1}$, assuming a stellar velocity of $+56$ km s$^{-1}$ [@pfb88; @pp90; @ej00]. The scattered solar C III $\lambda$977 line is sometimes seen in FUSE spectra, but the observed feature is far too broad for that to be responsible. The spike seen on the red edge of the line could perhaps be the solar emission.
One possible explanation for the blueshift of the C III line is that we are seeing emission predominantly from the side of the accretion disk that is rotating towards us. There is no corresponding blueshift of the C III $\lambda$1175 line (see below), but the $\lambda$977 line is a high opacity resonance line, meaning intervening C III material could in principle scatter background $\lambda$977 emission, while the $\lambda$1175 lines are lower opacity intersystem lines, so the same process may not work for $\lambda$1175. Another possible explanation for the C III blueshift is absorption from overlying H$_{2}$. The density of H$_{2}$ transitions increases towards the shorter wavelengths accessible to FUSE. These transitions will have higher opacities than any of the observed H$_{2}$ lines, since they will be originating from lower, more populated levels of the ground electronic state. Below about 1110 Å there are Lyman and Werner band transitions from the lowest levels that will be populated even at very cool temperatures. The column densities of H$_{2}$ in these lowest levels could be much higher than for the $T=3900$ K population that we have detected via Ly$\alpha$ fluorescence. This is because most of the H$_{2}$ within Mira A’s wind, in which Mira B is embedded, will be at low $T\sim 100$ K temperatures, and only the lowest energy levels will be populated in this cold H$_{2}$. Because of the significant H$_{2}$ opacity in the neighborhood of the C III $\lambda$977 line, the red side of the line might be absorbed, making the resulting C III profile appear as blueshifted as seen in our FUSE data. Despite these plausible explanations for the blueshift, the detection of the $\lambda$977 line will remain tentative until a more definitive explanation of the blueshift is available.
The C III $\lambda$1175 multiplet is blended with many H$_{2}$ emission lines and it is therefore unclear that C III is actually contributing to the observed emission. At the end of §4 we described a forward modeling calculation designed to compute the total H$_{2}$ photodissociation rate from Ly$\alpha$ fluorescence, including contributions from undetected fluorescence sequences. This calculation also provides us with the synthetic H$_{2}$ emission spectrum shown in Figure 5. This figure compares the synthetic spectrum with the FUSE data, demonstrating a reasonably good fit to the observed H$_{2}$ lines. The bottom panel focuses on the 1176 Å region. We subtract the H$_{2}$ spectrum from the data to see if the H$_{2}$ lines can account for all of the line flux. The residual flux shown below the lowest panel of Figure 5 suggests that there is flux remaining after this subtraction that we can associate with the C III $\lambda$1175 multiplet. The integrated flux remaining is $1.4\times 10^{-14}$ ergs cm$^{-2}$ s$^{-1}$, in good agreement with the C III flux suggested by the two-Gaussian representation in Figure 2, providing support for this being a reasonably accurate C III flux measurement. The flux ratio of C III $\lambda$977 and C III $\lambda$1175 is often used as a density diagnostic, but in Mira B’s case the significant uncertainties involved in measuring both lines are too large to obtain a believable density measurement.
The 1176 Å emission feature is the only feature in our FUSE spectrum that is also detectable in the previous HST/STIS spectrum (see Fig. 2 in Paper I). The sensitivity of STIS drops dramatically below 1200 Å, so the S/N of the STIS spectrum at 1176 Å is very low. The data quality does not allow us to separate the H$_{2}$ emission from the C III emission, so we can only report a total STIS flux of $(2.7\pm 0.8)\times 10^{-14}$ ergs cm$^{-2}$ s$^{-1}$ for the 1176 Å line. This agrees well with the integrated FUSE flux of $(3.00\pm 0.13)\times 10^{-14}$ ergs cm$^{-2}$ s$^{-1}$. In §3 and §4, we demonstrated that the H I Ly$\alpha$ line had not changed between the times of the STIS and FUSE observations. The excellent flux agreement for the 1176 Å feature provides further evidence that the UV spectrum of Mira B is unchanged.
The C III $\lambda$1175 line has also been observed numerous times by IUE. We found 32 short-wavelength, low-resolution (SW-LO) observations of Mira B in the IUE archive. Only about half of them are long enough exposures with sufficient S/N to measure a C III flux. Note that while H$_{2}$ may be contaminating the C III line in the HST/STIS and FUSE data, we do not believe that this is the case for the IUE data. In Paper I it was argued that unblended H$_{2}$ features should have been seen in the IUE data if H$_{2}$ were contributing flux to other emission features such as C III $\lambda$1175.
In Figure 6, we plot the C III fluxes as a function of time, where for purposes of this figure we have made no attempt to correct the STIS and FUSE fluxes for H$_{2}$ contamination. The STIS and FUSE fluxes are about an order of magnitude lower than the IUE fluxes, consistent with the drop in flux seen in the continuum and many other emission lines (see Paper I). The drop in flux is presumably due to a substantial decrease in accretion rate, which also led to a similar decrease in mass loss rate for Mira B (see Paper II). The consistency of the 1999 STIS and 2001 FUSE data demonstrates that this decrease is an enduring phenomenon.
It is possible that similar changes in accretion rate have been detected from optical data. Mira B is difficult to observe from the ground due to its close proximity with the generally brighter Mira A, but during Mira A minimum it is possible to detect the presence of Mira B. @ahj54 and @yy77 report a possible periodic variation in Mira B’s optical light curve with a period of 14 years and with an intensity range of about a factor of 4. Considering that these optical intensities will be contaminated with Mira A emission even at Mira A minimum, these variations might be consistent with the factor of 20 variations seen in the UV.
However, if the period of these variations is 14 years, why did the long-lived IUE not see them? @yy77 report a Mira B optical minimum in 1971, so in Figure 6 we display a 14-year period sine curve consistent with this phasing. All the IUE data points occur near predicted maxima, while the STIS and FUSE data points are near a predicted minimum. There is an unfortunate time gap in the IUE data from 1981–1990 that means the predicted Mira B minimum period was not covered by the IUE SW-LO data set. Thus, it is possible that the substantial drop in flux seen by STIS and FUSE is simply the UV manifestation of the 14-year optical cycle detected by @ahj54 and @yy77. One argument against this is that while there are no SW-LO IUE observations from 1981–1990, there are long-wavelength, low-resolution (LW-LO) spectra from 1983 July 9 and 1988 January 8. Although Mg II $\lambda$2800 and continuum fluxes are somewhat lower than average in these spectra, neither spectrum shows flux levels dramatically below those of other LW-LO observations in the IUE data set. This leaves only about a 5-year time gap in the IUE coverage during which Mira B fluxes could have dropped tremendously like they did in 1999–2001. Nevertheless, it would be worthwhile to observe Mira B again with FUSE and/or HST in 2004–2007 to see if high UV fluxes return as Figure 6 predicts.
SUMMARY
=======
We have analyzed UV observations of the wind-accreting star Mira B, using new FUSE spectra combined with previous HST/STIS data. Our results are summarized as follows:
1.
: In the new FUSE data, we detect Ly$\alpha$-fluoresced H$_{2}$ lines that are mostly Werner band lines rather than the Lyman band lines previously detected by HST. Most of the FUSE H$_{2}$ lines have never before been identified in an astrophysical spectrum.
2.
: Using previously developed techniques, we analyze the Mira B H$_{2}$ emission, combining the old HST/STIS and the new FUSE H$_{2}$ data. We estimate a temperature and column density for the H$_{2}$ layer responsible for the emission of $T=3900$ K and $\log N(H_{2})=17.1$, respectively.
3.
: Our modeling efforts demonstrate that undetected H$_{2}$ fluorescence sequences actually produce more H$_{2}$ photodissociation from Ly$\alpha$ fluorescence than do the detected sequences. Considering both, we estimate a total photodissociation rate of $4.5\times 10^{-10}$ M$_{\odot}$ yr$^{-1}$, comparable to the $(8-30)\times 10^{-10}$ M$_{\odot}$ yr$^{-1}$ total accretion rate of Mira B. The FUSE and HST H$_{2}$ emission may be coming from a photodissociation front, as first proposed in the analysis of the HST data (see Paper II).
4.
: The only stellar lines detected in our FUSE spectrum other than H$_{2}$ are the broad C III $\lambda$977 and C III $\lambda$1175 lines, which originate from within Mira B’s accretion disk. The C III $\lambda$977 line is highly blueshifted from its expected location, which leads us to consider its identification as tentative. Overlying H$_{2}$ absorption or particularly bright emission from one side of the accretion disk could in principle be responsible for the blueshift. The C III $\lambda$1175 line is heavily blended with numerous H$_{2}$ emission lines, making an accurate flux measurement difficult.
5.
: Analysis of the H$_{2}$ and C III $\lambda$1175 lines in the 1999 HST/STIS and 2001 FUSE data demonstrates that the UV spectrum of Mira B is roughly the same at these two times. The UV fluxes of both data sets are dramatically lower than ever observed by IUE in 1979–1995. The presence of the lower fluxes over at least two years (1999–2001) demonstrates that they are a persistent phenomenon.
6.
: We hypothesize that the drop in UV flux in 1999–2001 is associated with a previously identified 14-year periodic variation in optical emission from Mira B, and that IUE missed the variation due to most IUE observations falling near the two maxima of the cycle during the IUE era.
Support for this work was provided by NASA through grant NAG5-11950 to the University of Colorado. M. K. is a member of the Chandra Science Center, which is operated under contract NAS8-39073, and is partially supported by NASA.
Abgrall, H. A., Roueff, E., & Drira, I. 2000, A&AS, 141, 297 Abgrall, H. A., Roueff, E., Launay, F., Roncin, J. -Y., & Subtil, J. -L. 1993a, A&AS, 101, 273 Abgrall, H. A., Roueff, E., Launay, F., Roncin, J. -Y., & Subtil, J. -L. 1993b, A&AS, 101, 323 Ardila, D. R., Basri, G., Walter, F. M., Valenti, J. A., & Johns-Krull, C. M. 2002, ApJ, 566, 1100 Bartoe, J. -D., F. Brueckner, G. E., Nicolas, K. R., Sandlin, G. D., VanHoosier, M. E., & Jordan, C. 1979, MNRAS, 187, 463 Bevington, P. R., & Robinson, D. K. 1992, Data Reduction and Error Analysis for the Physical Sciences (New York: McGraw-Hill) Bowers, P. F., & Knapp, G. R. 1988, ApJ, 332, 299 Brown, A., Jordan, C., Millar, T. J., Gondhalekar, P., & Wilson, R. 1981, Nature, 290, 34 Dabrowski, I. 1984, Can. J. Phys., 62, 1639 Herczeg, G. J., Linsky, J. L., Valenti, J. A., Johns-Krull, C. M., & Wood, B. E. 2002, ApJ, 572, 310 Herczeg, G. J., Wood, B. E., Linsky, J. L., Valenti, J. A., & Johns-Krull, C. M. 2003, ApJ, submitted Jordan, C., Brueckner, G. E., Bartoe, J.-D. F., Sandlin, G. D., & VanHoosier, M. E. 1977, Nature, 270, 326 Josselin, E., Mauron, N., Planesas, P., & Bachiller, R. 2000, A&A, 362, 255 Joy, A. H. 1954, ApJS, 1, 39 Jura, M., & Helfand, D. J. 1984, ApJ, 287, 785 Karovska, M., Hack, W., Raymond, J., & Guinan, E. 1997, ApJ, 482, L175 Knapp, G. R. 1985, ApJ, 293, 273 Knapp, G. R., Young, K., Lee, E., & Jorissen, A. 1998, ApJS, 117, 209 McMurry, A. D., Jordan, C., & Carpenter, K. G. 1999, MNRAS, 302, 48 Moos, H. W., et al. 2000, ApJ, 538, L1 Morrissey, P. F., Feldman, P. D., Clarke, J. T., Wolven, B. C., Strobel, D. F., Durrance, S. T., & Trauger, J. T. 1997, ApJ, 476, 918 Perryman, M. A. C., et al. 1997, A&A, 323, L49 Planesas, P., Bachiller, R., Martin-Pintado, J., & Bujarrabal, V. 1990, ApJ, 351, 263 Reimers, D., & Cassatella, A. 1985, ApJ, 297, 275 Ryde, N., & Schöier, F. L. 2001, ApJ, 547, 384 Valenti, J. A., Johns-Krull, C. M., & Linsky, J. L. 2000, ApJS, 129, 399 Warner, B. 1972, MNRAS, 159, 95 Wilkinson, E., Harper, G. M., Brown, A., & Herczeg, G. J. 2002, AJ, 124, 1077 Wood, B. E., & Karovska, M. 2000, ApJ, 535, 304 Wood, B. E., Karovska, M., & Hack, W. 2001, ApJ, 556, L51 (Paper I) Wood, B. E., Karovska, M., & Raymond, J. C. 2002, ApJ, 575, 1057 (Paper II) Yamashita, Y., & Maehara, H. 1977, PASJ, 29, 319 Yamashita, Y., & Maehara, H. 1978, PASJ, 30, 409
[ccclccl]{}\
$1127.433\pm 0.008$ & $2.87\pm 0.23$ & $50\pm 5$ & 0-2 Q(10) & 1127.160 & 1.60 & 0-4 Q(10) $\lambda$1217.263\
& & & 1-3 P(5) & 1127.245 & 0.78 & 1-5 P(5) $\lambda$1216.988\
$1130.571\pm 0.015$ & $0.66\pm 0.14$ & $31\pm 7$ & 1-3 Q(7) & 1130.365 & 1.65 & 1-5 Q(7) $\lambda$1218.507\
$1131.608\pm 0.015$ & $0.98\pm 0.18$ & $37\pm 8$ & 2-4 R(1) & 1131.309 & 0.22 & 2-6 R(1) $\lambda$1217.298\
& & & 1-3 R(9) & 1131.390 & 1.03 & 1-5 R(9) $\lambda$1216.997\
$1143.257\pm 0.021$ & $0.61\pm 0.16$ & $39\pm 9$ & 2-4 P(5) & 1142.956 & 0.38 & 2-6 R(3) $\lambda$1217.488\
& & &2-0 P(11)& 1143.112 & 0.07 & 2-2 R(9) $\lambda$1219.101\
$1155.154\pm 0.008$ & $1.53\pm 0.23$ & $24\pm 4$ & 1-3 P(11) & 1154.910 & 1.84 & 1-5 R(9) $\lambda$1216.997\
$1160.912\pm 0.010$ & $1.01\pm 0.20$ & $24\pm 5$ & 1-4 R(3) & 1160.647 & 1.38 & 1-5 P(5) $\lambda$1216.988\
$1162.051\pm 0.019$ & $2.18\pm 0.32$ & $61\pm 10$&0-1 R(0)& 1161.693 & 0.72 & 0-2 R(0) $\lambda$1217.205\
& & &1-1 P(5)& 1161.814 & 0.16 & 1-2 P(5) $\lambda$1216.070\
& & &1-1 R(6)& 1161.949 & 0.21 & 1-2 R(6) $\lambda$1215.726\
$1172.242\pm 0.006$ & $4.60\pm 0.40$ & $41\pm 4$ & 1-4 P(5) & 1171.953 & 1.62 & 1-5 P(5) $\lambda$1216.988\
& & & 0-3 Q(10) & 1172.031 & 2.67 & 0-4 Q(10) $\lambda$1217.263\
$1174.571\pm 0.007$ & $4.29\pm 0.50$ & $37\pm 5$ & 2-5 R(1) & 1174.298 & 1.41 & 2-6 R(1) $\lambda$1217.298\
& & & 1-4 Q(7) & 1174.347 & 4.04 & 1-5 Q(7) $\lambda$1218.507\
$1174.846\pm 0.013$ & $1.25\pm 0.35$ & $28\pm 9$ & 2-5 R(0) & 1174.574 & 0.29 & 2-6 R(0) $\lambda$1217.680\
& & & 2-5 R(2) & 1174.588 & 0.30 & 2-6 R(2) $\lambda$1217.440\
$1175.110\pm 0.007$ & $2.07\pm 0.35$ & $25\pm 4$ & 2-5 R(3) & 1174.869 & 1.64 & 2-6 R(3) $\lambda$1217.488\
$1176.060\pm 0.013$ & $0.89\pm 0.28$ & $25\pm 8$ & 2-5 Q(1) & 1175.826 & 0.78 & 2-6 Q(1) $\lambda$1218.940\
$1176.319\pm 0.010$ & $2.22\pm 0.36$ & $34\pm 6$&5-0 R(18)& 1176.066 & 0.57 & 5-0 P(20) $\lambda$1217.717\
& & & 0-2 Q(18) & 1176.082 & 1.58 & 0-3 Q(18) $\lambda$1216.692\
& & & 2-5 R(4) & 1176.085 & 0.33 & 2-6 R(4) $\lambda$1218.457\
$1176.978\pm 0.016$ & $0.60\pm 0.25$ & $20\pm 10$&6-0 R(19)& 1176.706 & 0.53 & 6-0 P(21) $\lambda$1218.841\
& & & 2-5 Q(2) & 1176.788 & 0.19 & 2-6 Q(2) $\lambda$1219.804\
$1177.518\pm 0.026$ & $0.95\pm 0.36$ & $39\pm 17$ & 1-3 Q(16) & 1177.269 & 1.75 & 1-4 Q(16) $\lambda$1216.930\
$1180.710\pm 0.013$ & $0.64\pm 0.18$ & $16\pm 7$ & 2-5 P(3) & 1180.457 & 1.38 & 2-6 R(1) $\lambda$1217.298\
\
$976.447\pm 0.121$ & $6.99\pm 1.60$ & $338\pm 84$ & C III? & 977.020 & &\
$1174.616\pm 0.083$ & $7.03\pm 1.82$ & $287\pm 47$ & C III? & 1175mult & &\
$1177.206\pm 0.115$ & $8.57\pm 1.32$ & $405\pm 68$ & C III? & 1175mult & &\
\
$1254.369\pm 0.013$ & $1.01\pm 0.31$ & $16\pm 4$&3-1 P(17)& 1254.125 & 0.26 & 3-1 R(15) $\lambda$1214.465\
$1265.421\pm 0.009$ & $0.99\pm 0.27$ & $17\pm 5$&3-2 R(15)& 1265.180 & 0.31 & 3-1 R(15) $\lambda$1214.465\
$1593.575\pm 0.016$ & $3.77\pm 0.83$ & $23\pm 6$&3-9 R(15)& 1593.258 & 0.74 & 3-1 R(15) $\lambda$1214.465\
$1621.385\pm 0.021$ & $3.54\pm 0.89$ & $26\pm 8$&3-10 R(15)& 1621.119 & 0.39 & 3-1 R(15) $\lambda$1214.465\
$1622.434\pm 0.011$ & $5.19\pm 1.01$ & $26\pm 6$&3-9 P(17)& 1622.133 & 0.87 & 3-1 R(15) $\lambda$1214.465\
[cccccccc]{}\
1-2 R(6) & 1215.7263 & 0.502 & 0.0349 & 10261.20 & 13 & 0.0 & 0.0\
1-2 P(5) & 1216.0696 & 0.914 & 0.0289 & 9654.15 & 33 & 0.0 & 0.0\
3-3 P(1) & 1217.0377 & 0.115 & 0.0013 & 11883.51 & 9 & 0.0 & 0.0\
0-2 R(0) & 1217.2045 & 0.858 & 0.0441 & 8086.93 & 1 & 0.0 & 0.0\
4-0 P(19) & 1217.4100 & 0.190 & 0.0093 & 17750.25 & 117 & 0.417 & 0.469\
0-2 R(1) & 1217.6426 & 1.355 & 0.0289 & 8193.81 & 9 & 0.0 & 0.0\
2-1 P(13) & 1217.9041 & 1.141 & 0.0192 & 13191.06 & 81 & 0.002 & 0.002\
2-1 R(14) & 1218.5205 & 0.324 & 0.0181 & 14399.08 & 29 & 0.006 & 0.006\
0-2 R(2) & 1219.0887 & 0.165 & 0.0256 & 8406.29 & 5 & 0.0 & 0.0\
2-2 R(9) & 1219.1005 & 0.320 & 0.0318 & 12584.80 & 57 & 0.0 & 0.0\
2-2 P(8) & 1219.1543 & 0.340 & 0.0214 & 11732.12 & 17 & 0.0 & 0.0\
0-2 P(1) & 1219.3676 & 0.318 & 0.0149 & 8193.81 & 9 & 0.0 & 0.0\
0-1 R(11) & 1219.7454 & 0.166 & 0.0037 & 10927.12 & 69 & 0.0 & 0.0\
\
3-1 R(15) & 1214.4648 & 0.284 & 0.0236 & 15649.58 & 93 & 0.031 & 0.037\
5-0 P(20) & 1217.7165 & 0.069 & 0.0115 & 19213.16 & 41 & 0.478 & 0.499\
6-0 P(21) & 1218.8403 & 0.051 & 0.0130 & 20688.04 & 129 & 0.520 & 0.555\
\
0-3 Q(18) & 1216.6926 & 0.036 & 0.0396 & 25499.74 & 37 & 0.0 & 0.0\
1-4 Q(16) & 1216.9302 & 0.034 & 0.0712 & 25929.42 & 33 & 0.0 & 0.0\
1-5 P(5) & 1216.9878 & 0.072 & 0.0071 & 19807.03 & 33 & 0.016 & 0.031\
1-5 R(9) & 1216.9969 & 0.099 & 0.0197 & 22251.21 & 57 & 0.017 & 0.031\
0-4 Q(10) & 1217.2628 & 0.065 & 0.0100 & 20074.45 & 21 & 0.0 & 0.0\
2-6 R(1) & 1217.2978 & 0.052 & 0.0558 & 21589.82 & 9 & 0.008 & 0.012\
2-6 R(2) & 1217.4400 & 0.047 & 0.0399 & 21756.98 & 5 & 0.066 & 0.089\
2-6 R(3) & 1217.4883 & 0.090 & 0.0364 & 22005.58 & 21 & 0.073 & 0.113\
2-6 R(0) & 1217.6800 & 0.027 & 0.1098 & 21505.78 & 1 & 0.002 & 0.002\
2-6 R(4) & 1218.4566 & 0.015 & 0.0399 & 22332.85 & 9 & 0.017 & 0.025\
1-5 Q(7) & 1218.5084 & 0.055 & 0.0303 & 20894.94 & 45 & 0.0 & 0.0\
2-6 Q(1) & 1218.9402 & 0.023 & 0.0532 & 21589.82 & 9 & 0.0 & 0.0\
2-6 Q(2) & 1219.8038 & 0.004 & 0.0532 & 21756.98 & 5 & 0.0 & 0.0\
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'MIMO systems are considered as most promising for wireless communications. However, with an increasing number of radio front ends the corresponding energy consumption and costs become an issue, which can be relieved by the utilization of low-resolution quantizers. In this study we propose an optimal precoding algorithm constrained to constant envelope signals and phase quantization that maximizes the minimum distance to the decision threshold at the receivers using a branch-and-bound strategy. The proposed algorithm is superior to the existing methods in terms of bit error rate. Numerical results show that the proposed approach has significantly lower complexity than exhaustive search.'
author:
- 'Erico S. P. Lopes, and Lukas T. N. Landau, [^1] [^2]'
bibliography:
- 'bib-refs.bib'
title: 'Optimal Precoding for Multiuser MIMO Systems With Phase Quantization and PSK Modulation via Branch-and-Bound'
---
Precoding, low-resolution quantization, MIMO systems, branch-and-bound methods.
[^1]: The authors are with Centro de Estudos em Telecomunicações Pontifícia Universidade Católica do Rio de Janeiro, Rio de Janeiro CEP 22453-900, Brazil, (email: {erico, lukas.landau}@cetuc.puc-rio.br).
[^2]: This work has been supported by the [ELIOT ANR18-CE40-0030 and FAPESP 2018/12579-7]{} project.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'L. Borgonovo'
- 'F. Frontera'
- 'C. Guidorzi'
- 'E. Montanari'
- 'L. Vetere'
- 'P. Soffitta'
date: 'Received ; accepted '
title: |
On the temporal variability classes found in long\
gamma-ray bursts with known redshift
---
[ We study the properties of the bursts belonging to each ACF class and look for significant differences between them. ]{} [ We complement previous ACF analysis studying the corresponding power density spectra (PDS). With the addition of [*Beppo-SAX* ]{} data and taken advantage of its broad-band capability, we not only increase the burst sample but we extend the analysis to X-ray energies. ]{} [ The rest-frame PDS analysis at $\gamma$-ray energies shows that the two ACF classes are not simply characterised by a different low frequency cut-off, but they have a distinct variability as a whole in the studied frequency range. Both classes exhibit average PDS with power-law behaviour at high frequencies ($f' \ge 0.1$ Hz) but significantly different slopes, with index values close to those of Brownian ($-2$) and Kolmogorov ($-5/3$) spectra for the [*narrow* ]{} and [*broad* ]{} classes respectively. The latter spectrum presents an additional PDS component, a low-frequency noise [*excess* ]{} with a sharp cut-off. At X-ray energies we find the power-law index unchanged for the broad class, but a significantly steeper slope in the narrow case ($\sim -3$). We interpret this as an indication that the broad class bursts have weaker spectral evolution than the narrow ones, as suggested also by our analysis of the ACF energy dependence. The low and high frequency PDS components may then arise from two radiating regions involving different emission mechanisms. We compare our GRB sample conditioned by afterglow detections with a complete, flux limited BATSE sample, finding a significant bias against narrow ACF bursts. ]{}
Introduction {#intro}
============
Determining the relevant timescales for any astronomical phenomenon is essential to understand its underlying physical processes. However, in spite of extensive research, temporal studies on the prompt emission phase of long gamma-ray bursts (GRBs) are not yet able to describe and explain their basic temporal properties. The main challenges encountered in the temporal analysis of GRBs are related to intrinsic characteristics of the emitted signal. Firstly, bursts are non-repetitive short-term events. Consequently, the total duration of the emission in a given observational energy window is the first timescale used to characterised them. Through out this paper we will only consider the class of [*long*]{} bursts, i.e., those with duration time $T_{90}>2$ s. Second, burst light curves (LCs) show a remarkable morphological diversity and they appear to have a composite structure. While a significant fraction of bright long bursts ($\sim
15\%$) exhibits a single smooth pulse structure, in most cases they appear to be the result of a complex, seemingly random distribution of several pulses. Burst pulses are commonly described as having fast-rise exponential-decay (FRED) shape, although the decay is not strictly exponential. Therefore, the second timescale that seems relevant for the description of a burst is a “typical” pulse duration. However, analysis of the pulse parameters has shown broad log-normal distributions not only among different bursts, but also within a single burst [see, e.g., @Norr96]. Consideration of other timescales might be relevant, e.g., for GRBs with [*precursors*]{} [@Kos95] or when long quiescent periods occur [@NP02], although these temporal features appear only in a small fraction of bursts.
Due to these characteristics, much of the GRB temporal analysis has been done directly on the LCs, i.e., modelling the pulses and studying their shape and distribution. Standard linear analysis tools must be used with some caution, since most of the inferences based on them would in principle require “long” stationary signals, i.e., the most suitable bursts are those where the duration is much longer than the typical pulse width. This is the case of the temporal analysis based on power density spectra (PDS). Individual PDS of GRB have very diverse shapes, and they do not seem to have common features, although the longest bursts show spectra with a consistent power-law behaviour. One way to overcome these limitations is to estimate an average PDS from a sample GRBs. This approach will only produce physically meaningful results if each burst can be considered a realisation of the same stochastic process, i.e., there are no subclasses in the sample.
Under this assumption @BSS98 [@BSS00] calculated with a large sample of bright GRBs, an average PDS showing a clear power-law behaviour extended over two frequency decades (approximately within the 0.01–1 Hz frequency range), and more remarkably with an exponent value approximately equal to that of the Kolmogorov spectrum found in fluid turbulence [@Kol41]. The significance of a thus obtained average PDS depends on two important additional factors. First, light curves have to be normalised to balance the weight between bursts of different [*brightness*]{}. It is not clear at this point which norm should be used to produce the most meaningful average. @BSS98 favoured the use of the peak flux normalisation, however they tested several other norms with qualitatively similar results. Thus, a good normalisation should give better convergence but the norm should not affect the final result for a sufficiently large sample. The second problem is the shift in frequencies produced by cosmic dilation effects. Having no redshifts $z$ determined for their GRB sample, @BSS00 did not correct for these effects. However, they argued that if the underlying PDS shape for every GRB is a featureless single power-law with a constant exponent, the frequency shifts will not affect the obtained average PDS. If this is the case, considering that the two decades range of the power-law is much larger than the standard deviation of the redshift distribution ($\sigma_z \sim 2$ based on the few known redshifts), indeed the shifts should just [*smear*]{} the cut-off frequencies.
The same statistical approach was used by @Fen95 [ hereafter F95] in their study of the average autocorrelation function (ACF) of a sample of bright GRBs. The ACF gives a measure of the correlation between different points in the light curve that are separated by a given time lag. Since it is the Fourier transform of the PDS, it contains in principle the same information that can be visualised in a different way. Therefore, the same caveats regarding the average PDS apply to the average ACF. It was only after the discovery of the afterglow emission [@Cos97] and the determination of their redshift for a significant number of bursts that we were able to address some of those issues. @Bor04 [ hereafter B04] showed for a sample of 16 bright GRB with known $z$ that when corrected for cosmic dilation effects the ACFs exhibit a clear bimodal distribution. Using as a measure the half-width at half-maximum, there is a highly significant gap between a narrow and a broad width class, the separation in standard deviations being $>\!7
\sigma$. The estimated local or intrinsic values (i.e., those calculated at the rest-frame of the source) for the average widths were 1.6 s and 7.5 s, and the relative dispersions were 32% and 4% for the narrow and broad classes, respectively. It is remarkable the low dispersion found in the last subset, which comprised $\sim 1/3$ of the total sample.
This article builds on the ACF analysis done in . In Sect. \[data\] we present our data samples and in Sect. \[methods\] we briefly describe the methods used in the subsequent temporal analysis. In Sect. \[gLC\] we strengthen previous findings based on the ACF, expanding the previous sample of GRB with known $z$ by the inclusion of proprietary data from the [*BeppoSAX*]{} mission, and we complement the temporal analysis estimating the [*intrinsic*]{} PDS for each of the subsets identified using the ACF. Furthermore, using the broadband capability of [*BeppoSAX*]{} combined instruments, in Sect. \[XLC\] we are able to extend the study to the X-ray energies. In Sect. \[class\], we investigate whether the typical values of several physical parameters commonly used to characterised GRBs differ significantly between the two temporal classes. We look into the problem of the energy dependence of the ACF and discuss possible biases in our sample of bursts with known $z$ in Sect. \[tau\_E\]. In Sect. \[discussion\] we discuss our main results.
$$\begin{array}{llclcccccccc}
\hline
\hline
\noalign{\smallskip}
\mathrm{GRB} &
\mathrm{Instrument} \; (\gamma) &
\mathrm{Inst.} \; (\mathrm{X}) &
z &
\mathrm{Ref.}^{b} &
w_{\gamma} &
w'_{\gamma} &
w_{\mathrm X} &
w'_{\mathrm X} &
w_{\mathrm X}/w_{\gamma} &
\xi_{\mathrm 2ch}^{c} &
\mathrm{Class}^{d} \\
& \textrm{(55--320 keV)}^{a} & \textrm{(2--28 keV)} & & & (s) & (s) & (s) & (s) & & & \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
970228 & \mathrm{ GRBM} & \mathrm{WFC} & 0.695 & (1) & 1.3\pm0.1 & 0.77\pm0.06 & 3.5\pm0.2 & 2.1\pm0.1 & 2.7 & & \mathrm{n} \\
970508 & \mathrm{ BATSE/GRBM } & \mathrm{WFC} & 0.835 & (2) & 2.7\pm0.1 & 1.47\pm0.05 & 10.3\pm0.4 & 5.6\pm0.2 & 3.8 & 0.06\pm0.05 & \mathrm{n} \\
970828 & \mathrm{ BATSE} & & 0.9578 & (3) & 15.33\pm0.06 & 7.83\pm0.03 & & & & 0.09\pm0.03 & \mathrm{b} \\
971214 & \mathrm{ BATSE/GRBM/Konus}& \mathrm{WFC} & 3.418 & (4) & 8.02\pm0.08 & 1.81\pm0.02 & 11.8\pm0.8 & 2.7\pm0.2 & 1.5 & 0.10\pm0.04 & \mathrm{n} \\
980326 & \mathrm{ GRBM } & \mathrm{WFC} & 1.2 & (5) & 1.34\pm0.1 & 0.61\pm0.04 & 2.35\pm0.4 & 1.05\pm0.2 & 1.7 & & \mathrm{n} \\
980329 & \mathrm{ BATSE/GRBM/Konus}& \mathrm{WFC} & 3\pm1 & (6) & 5.96\pm0.02 & 1.5\pm0.5 & 8.8\pm0.4 & 2.2\pm0.7 & 1.5 & 0.05\pm0.01 & \mathrm{n} \\
980425 & \mathrm{ BATSE/GRBM} & \mathrm{WFC} & 0.0085 & (7) & 7.62\pm0.08 & 7.56\pm0.08 & 14.8\pm0.9 & 14.7\pm0.9 & 1.9 & 0.24\pm0.06 & \mathrm{b} \\
980703 & \mathrm{ BATSE} & & 0.966 & (8) & 14.15\pm0.1 & 7.19\pm0.05 & & & & 0.11\pm0.01 & \mathrm{b} \\
990123 & \mathrm{ BATSE/GRBM/Konus}& \mathrm{WFC} & 1.600 & (9) & 19.81\pm0.03 & 7.62\pm0.01 & 35.6\pm0.2 & 13.70\pm0.08 & 1.8 & 0.30\pm0.05 & \mathrm{b} \\
990506 & \mathrm{ BATSE/GRBM/Konus}& & 1.3066 & (10) & 3.83\pm0.02 & 1.66\pm0.01 & & & & 0.20\pm0.05 & \mathrm{n} \\
990510 & \mathrm{ BATSE/GRBM/Konus}& & 1.619 & (11) & 2.54\pm0.03 & 0.97\pm0.01 & & & & 0.20\pm0.01 & \mathrm{n} \\
990705 & \mathrm{ GRBM } & \mathrm{WFC} & 0.86 & (12) & 14.3\pm0.2 & 7.7\pm0.1 & 22.3\pm0.4 & 12.0\pm0.2 & 1.6 & & \mathrm{b} \\
990712 & \mathrm{ GRBM } & \mathrm{WFC} & 0.433 & (13) & 4.1\pm0.2 & 2.85\pm0.1 & 4.8\pm0.2 & 3.35\pm0.1 & 1.2 & & \mathrm{n} \\
991208 & \mathrm{ Konus } & & 0.7055 & (14) & 3.67\pm0.04 & 2.15\pm0.02 & & & & & \mathrm{n} \\
991216 & \mathrm{ BATSE/GRBM/Konus}& & 1.02 & (15) & 3.80\pm0.02 & 1.88\pm0.01 & & & & 0.18\pm0.02 & \mathrm{n} \\
000131 & \mathrm{ BATSE} & & 4.500 & (16) & 5.77\pm0.08 & 1.05\pm0.01 & & & & 0.21\pm0.06 & \mathrm{n} \\
000210 & \mathrm{ GRBM/Konus } & \mathrm{WFC} & 0.846 & (17) & 2.4\pm0.2 & 1.3\pm0.1 & 5.35\pm0.3 & 2.9\pm0.2 & 2.2 & & \mathrm{n} \\
000214 & \mathrm{ GRBM} & \mathrm{WFC} & 0.47 & (18) & 2.5\pm0.4 & 1.7\pm0.3 & 6.8\pm0.3 & 4.65\pm0.2 & 2.7 & & \mathrm{n} \\
010222 & \mathrm{ GRBM/Konus } & \mathrm{WFC} & 1.477 & (19) & 3.68\pm0.07 & 1.48\pm0.03 & 42.2\pm0.4 & 17.0\pm0.2 & 11.5 & & \mathrm{n} \\
010921 & \mathrm{ GRBM } & & 0.451 & (20) & 9.8\pm0.3 & 6.75\pm0.2 & & & & & \mathrm{b} \\
011121 & \mathrm{ GRBM/Konus } & \mathrm{WFC} & 0.362 & (21) & 10.0\pm0.3 & 7.35\pm0.2 & 18.9\pm0.2 & 13.9\pm0.15 & 1.9 & & \mathrm{b} \\
030329 & \mathrm{ Konus} & & 0.1685 & (22) & 2.6\pm0.1 & 2.19\pm0.08 & & & & & \mathrm{n} \\
\noalign{\smallskip}
\hline
\end{array}$$
The energy range of BATSE data which was taken as reference instrument. For Konus and GRBM data the actual energy ranges are 50–200 keV and 40–700 keV respectively.
\(1) @bloom01; (2) @mrm97; (3) @djo01; (4) @kul98; (5) @bloom99; (6) @lcr99; (7) @tin98; (8) @djo98; (9) @kul99; (10) @bloom03; (11) @bea99; (12) @Ama00; (13) @vea01; (14) @dod99; (15) @vea99; (16) @And00; (17) @pir02; (18) @Ant00; (19) @jha01; (20) @Pri02; (21) @gar03; (22) @grei03.
Index $\xi$ has been estimated for an energy window width of two BATSE channels.
\(n) and (b) indicate [*narrow*]{} and [*broad*]{} width ACF class, respectively.
Data
====
This work is mainly based on the analysis of light curves from GRBs with known redshift. Given the scarce number of cases available for study, we combined data (in the $gamma$ energy band) from three instruments to improve our statistics. Increasing a sample in this way presents an obvious trade-off, since we use count time series and the difference between instrument responses introduce an additional dispersion that may counteract the benefits. For this reason, we initially analysed the data of the bursts that were observed by more than one instrument, evaluating whether the differences were acceptable for our purposes.
The comparisons were made taking BATSE as the reference instrument. Its data comprise half of our GRB sample (in the gamma energy band), showing the best signal-to-noise ratio (S/N) with relatively low directional dependence thanks to the large collecting area of its eight Large Area Detectors (LADs) placed on each corner of the [*Compton Gamma-Ray Observatory*]{} [CGRO; @Fish89], giving full sky coverage. It flew during the period 1991–2000 collecting the largest GRB catalog up to date. The [*CGRO*]{} Science Support Center (GROSSC) provides the so-called concatenated 64 ms burst data, which is a concatenation of the three standard BATSE data types DISCLA, PREB, and DISCSC. All three data types have four energy channels (approximately 25–55, 55–110, 110–320, and $>320$ keV). The DISCLA data is a continuous stream of 1.024 s and the PREB data covers the 2.048 s prior to the trigger time at 64 ms resolution, both types obtained from the 8 LADs. They have been scaled to overlap the DISCSC 64 ms burst data, that was gathered by the triggered LADs (usually the four closer to the line of sight). This combined data format was used when available, since the concatenated pre-burst data allows a better estimation of the background. In the case of GRB 970828 the DISCSC data are incomplete, and we used instead the 16-channel MER data type, binned up into 4 DISCSC-like energy channels. All BATSE bursts with known $z$ were considered for study, excluding two cases were the data are incomplete or were no recorded (i.e., GRB 980326 and GRB 980613), resulting in a total of 11 cases.
We also include the set of bursts selected in that were observed by Konus, which is a GRB detector on board the [*Wind*]{} mission [@Apt95]. Light curves of its bursts are publicly available at 64 ms resolution in the 50–200 keV energy band. The collecting area of this experiment is about 20 times smaller than the one of BATSE and consequently, in most cases, the signal is too weak for our temporal studies. However, in a comparative ACF analysis of the bursts observed by both instruments showed a good agreement for very [*bright*]{} bursts, and the selection criteria for Konus cases were set requiring peak count rates larger than 3000 counts s$^{-1}$ and the availability of post-burst data, resulting in the 5 Konus GRBs included in burst sample.
The [*BeppoSAX*]{} mission, that operated between the years 1996–2002, had broad energy band capabilities thanks to the combined operation of several instruments. The Gamma Ray Burst Monitor [GRBM; see @Fro97] covered the 40–700 keV energy range, roughly matching the range of Konus and BATSE (i.e., 55–320 keV using $2+3$ channels). Note that since the ACF and the PDS are quadratic functions of the number of counts, and generally there are more counts at lower energies, the agreement for these temporal analysis functions will depend mainly on having a similar lower-end energy limit. We made use of the standard high resolution 7.8125 ms LCs, but we binned them up into a 62.5 ms time resolution to improve the S/N ratio. This has a negligible effect on the measurement of the time scales that concern us here and it reduces some noise artifacts (i.e., the noise becomes more Poissonian). In addition, it approximately matches the standard 64 ms BATSE temporal resolution for better comparison. The LCs were dead-time corrected and background subtracted. In @Bor05 a comparative ACF analysis was presented, including 17 GRBS detected by the GRBM with known redshift from which 8 were also observed by BATSE. It was concluded there that although the measured dispersions of the local ACF widths were larger than in the BATSE case ($\sim 15\%$ at half-maximum), the average sample values for each ACF width class were equal to within uncertainties. Therefore, at least in the context of the present temporal variability analysis, we can consider that these data have more intrinsic dispersion, but the introduced errors are mainly stochastic.
The two Wide Field Cameras (WFCs) also on board [*BeppoSAX*]{} covered the 2–26 keV energy range [@Jag97]. During their operation time they detected 53 GRBs in conjunction with the GRBM allowing the first broad band studies of GRBs [see, e.g., @Ama02]. Furthermore, a considerable fraction of the WFC bursts ($\sim 36\%$) were also detected by BATSE. The LCs were extracted with a time resolution of 62.5 ms and discriminated in three energy channels (i.e., 2–5, 5–10, and 10–26 keV). The energy intervals were chosen in order to have a similar amount of counts in each channel for a typical GRB. However, except for the brightest GRBs, the signal in each channel is too weak for the purposes of our temporal analysis and the LCs had to be integrated into a single energy channel. Here we will focus on the analysis of the 13 GRBs for which we have redshift estimations. These bursts constitute a subset of our sample of GRBs with known redshift in the $\gamma$ energy band. Table \[tab:sample\] lists all these bursts indicating in its first columns their name, the source instruments in $\gamma$ and X-ray bands when available, the estimated redshift $z$, and the corresponding reference.
To further study the ACF width energy dependence we selected from the BATSE current catalog all long bursts (i.e., duration time $T_{90}>2$ s) with a peak flux measured on the 1.024 s timescale $F_{\rm 1s} \ge 4$ photons ${\rm cm}^{-2} {\rm s}^{-1}$ in the 50–300 keV band that have available concatenated 64 ms data. This resulted in a sample of 188 bright bursts for which, in most cases, the redshift is unknown.
Methods
=======
For the autocorrelation function analysis we follow the same method presented in , that was based on earlier works of @LRW93 and . Here, we will summarise the method, and we refer to for further details. Following the same notation, from a uniformly sampled count history with $\Delta T$ time resolution and $N$ time bins, let $m_i$ be the total observed counts at bin $i$. Also let $b_i$ be the corresponding background level and $c_i = m_i - b_i$ the net counts. The discrete ACF as a function of the time lag $\tau = k
\Delta T$ is $$A(\tau=k \Delta T) \equiv \sum_{i=0}^{N-1} \frac{c_{i} c_{i+k}-m_i \delta_{0k}}{A_0},\;\;\;\; k=0,
\ldots , N-1 \; ,
\label{acf}$$ where $\delta$ is the Kronecker function. Here the periodic boundary conditions ($c_i=c_{i+N}$) are assumed. The normalisation constant $A_0$ is defined as $$A_0 \equiv \sum_{i=0}^{N-1} (c_{i}^2 - m_i) \; ,
\label{A0}$$ such that $A(0)=1$ for $k=0$. The term $m_i$ in Eq. \[A0\] subtracts the contribution of the uncorrelated noise assuming that it follows the Poisson statistics.
For practical reasons, the actual calculation of Eq. \[acf\] was done using a Fast Fourier Transform (FFT) routine. Denoting by $C_f$ the Fourier transformed of the background subtracted light curve, then the definition of the power density spectrum (PDS) can be written as $P_f \equiv |C_f|^2$. The noise contribution is subtracted from the PDS assuming Poisson statistics. Since the PDS and the ACF are Fourier pairs (Wiener-Khinchin theorem), the latter is obtained by inverse transforming the first. Zero padding of the time series was used to avoid the artifacts produced by the periodic boundary condition. The normalisation used for the ACF, that gives to its central maximum unity value, is equivalent (aside from a noise correction term) to the scaling of the LC by the square root of total power $\sqrt{P_{\rm
tot}}$, where $P_{\rm tot} \equiv \sum_{i} c_i^2$. This normalisation is a natural choice for the ACF analysis and it makes the ACF of each burst independent (to first order) of its brightness. Note however that for our PDS analysis we found it more suitable to scale the LCs by their respective net count fluences ${{\cal F}} \equiv \sum_{i} c_i$ (or equivalently dividing the PDS of the original LCs by ${{\cal F}}^2$). Since the $a_0$ zero order Fourier coefficient is equal to the fluence ${{\cal F}}$, our normalisation makes the PDS converge to unity towards the low frequencies, therefore in Fourier space it can be interpreted also as normalising by the amount of [*power*]{} at the lowest frequency. In Sect. \[gpds\] we will further discuss this choice over other possible normalisations.
The background estimations were done by fitting with up to a second order polynomial the pre- and post-burst data, that were judged by visual inspection to be inactive. This is critical for weak bursts and particularly aggravated in the Konus case since the publicly available LCs have almost no pre-burst data. This is the main reason why dim bursts were excluded in our sample selection in Sect. \[data\].
In it was found empirically, considering the analysis of the average ACFs of the narrow and broad width sets and their dispersion, that at half-maximum the separation between the two sets is most significant. Therefore, the ACF half-width at half-maximum $w$ given by $A(\tau\!=\!w) \equiv 0.5$ (here we deviate from the notation for simplicity) was chosen as the measure that best characterises each class. Since it has been shown that the average ACF decreases approximately following a stretch exponential , the ACF width $w$ was calculated fitting the logarithm of the ACF in the range $0.4 \le A(\tau)\le 0.6$ with a second degree polynomial.
The uncertainties due to stochastic fluctuation were estimated using a Monte Carlo method. For any given LC a large number of realisations is generated assuming that the fluctuations on the gross number of counts in each bin $m_i$ follows a Poisson distribution. Then, the LCs are background subtracted and the ACF widths $w$ are calculated following the same method as for the original data. However, when subtracting the Poisson noise contribution in the ACF (Eq. \[A0\]) a factor of 2 must be introduced in the corresponding noise term to take into account that the procedure to generate the synthetic LCs doubles the noise variance, since they are in a sense second order realisations. In general, if one considers a Poisson process with expected value $\mu$ and iteratively assumes its counts as expected counts of another Poisson process, after $n$ iterations the expected value is still $\mu$, but the variance is $n \mu$ (with $n=2$ in this case). Finally, the standard deviation of the obtained widths is used as estimator of the uncertainty. Nevertheless, the main source of uncertainty in dim bursts is the background estimation, that introduces a [*systematic*]{} error in the determination of the ACF width. Therefore, the reported uncertainties should be regarded as lower limits. Another likely source of systematic errors is the non uniform response of the detectors, since we do not deconvolve the count data.
Analysis of the $\gamma$-ray LCs {#gLC}
================================
Some LC examples from our sample of GRBs with known redshift are shown in Fig. \[fig:LCs\]. No morphological differences are evident between the [*narrow*]{} and [*broad*]{} classes by simple inspection. However, standard linear analysis tools reveal clear variability differences between the two, as we report in this section.
ACFs {#gacf}
----
With the inclusion of GRBM bursts we were able to expand the sample presented in . The added ACFs (shown in Fig. \[acf\_g\] in [*dashed lines*]{}) also follows a bimodal distribution when corrected for time dilation effects, reinforcing the previously found pattern. In Fig. \[acf\_g\]a the observed ACFs $A(\tau)$ are shown first for comparison, while Fig. \[acf\_g\]b shows the rest-frame or [*local*]{} ACFs $A(\tau')$, where $\tau'=\tau/(1+z)$ is the time lag corrected for cosmic dilation. We obtained mean values for the rest-frame ACF half-width at half-maximum $\bar{w'}^{\rm (n)}_{\gamma}=(1.56\pm0.15)$ s and $\bar{w'}^{\rm (b)}_{\gamma}=(7.42\pm0.14)$ s, and also sample standard deviations $\sigma^{\rm (n)}_{\gamma}=0.6$ s and $\sigma^{\rm
(b)}_{\gamma}=0.5$ s for the narrow and broad subsets respectively. The mean values are equal to those reported in within uncertainties. The [*gap*]{} between the two subsets (defined as the mean difference) represents a $4.9 \sigma$ separation, smaller than the $7 \sigma$ separation reported in mainly due to the increased dispersion in the broad class. However, thanks to a larger sample, we have slightly increased the [*significance*]{} of this separation. The probability $p$ of a random occurrence of such gap was estimated using Monte Carlo methods. Assuming that there are no characteristic timescales and an underlying uniform probability distribution (the most favourable case) we estimated $p < 4 \times 10^{-7}$.
Hereafter our sample of bursts with known redshifts in the $\gamma$ band will be divided for analysis into two subsets of 15 narrow and 7 broad ACF width bursts respectively. In Table \[tab:sample\] we show the newly added ACFs widths together with those presented in for completeness.
PDS {#gpds}
---
We calculate the PDS for our sample of GRBs with known redshift, correcting the LCs for cosmic time dilation effects. We group bursts following the classes established in § \[gacf\] based on the ACF width. The upper panels of Figs. \[PDS\_narrow\] and \[PDS\_broad\] show for the narrow and broad classes all individual PDS overlaid for comparison. Given the chosen normalisation, all PDS must converge to unity at low frequencies (see Sect. \[methods\]). The frequencies have been equally binned on the logarithmic scale. The procedure not only smooths out the PDS stochastic variations but enable us to estimate an average PDS for each class. This is necessary because even if all the LCs had the same time duration, the redshift correction would make the Fourier frequencies differ for each burst (i.e., since the observed frequency $f$ transforms to the rest-frame as $f'=f
(1+z)$). The lower panels of Figs. \[PDS\_narrow\] and \[PDS\_broad\] show the corresponding central values $\tilde{P}_{f}$, where we have chosen the use of the [*median*]{} over the [*mean*]{} as a more robust estimator of the expected or “underlying” spectrum. Although qualitatively similar results are obtained in both cases, the median shows smaller fluctuations and it is less sensitive to the random exclusion of a few bursts from each class. The figures show also the quartile deviation ([*dashed lines*]{}) as a measure of the dispersion that was chosen to be consistent with the use of the median (i.e., the 25% and the 75% quartiles around the median value). Comparing the two median PDS it is evident that they have remarkably different shapes. The narrow class PDS as expected shows more power at high frequencies and it is well described by a single power-law model $\tilde{P}_f\propto 1/f^{\alpha}$ up to a low frequency cut-off at $\approx 0.1$ Hz. In Fig. \[PDS\_narrow\]b we show a fit to the median PDS in the frequency range $0.1 \le f' \le
10$ Hz for which we found a best-fit-parameter $\alpha^{\rm
(n)}_{\gamma}=1.97\pm0.04$ (i.e., $\gamma$ band - narrow class). The obtained exponent is consistent with that of [*Brownian or red noise*]{} (i.e., with $\alpha=2$), which suggests the alternative use of a single Lorentzian model to describe the entire PDS as in [*shot noise*]{} models [@Bel92]. However, as shown also in Fig. \[PDS\_narrow\]b, there seems to be a slight systematic deviation around the cut-off region. The broad class $\tilde{P}_{f}$ on the other hand appears to have two distinct components, i.e., a broad low frequency component with a sharp break and a power-law component (Fig. \[PDS\_broad\]b). If we assume that the two components are independent then the single power-law should have an independent cut-off at low frequencies which cannot be determined but only constrained by our data. The low frequency component is well fitted by a stretched exponential $P_1(f)=\exp[(-f/f_1)^\eta]$. The best-fit-parameters for this model will depend to some extent on the cut-off frequency parameter of the power-law component and vice versa, however this will affect mainly the estimation of the exponent $\eta$. Taking into account these uncertainties we found a characteristic frequency $f_1=(0.025\pm0.003)$ Hz and an index $\eta=1.3\pm0.2$. We considered other empirical models, concluding that functions with power-law asymptotic behaviour (e.g., a Lorentzian) do not decay fast enough to fit the data. For the high frequency component we found a best-fit power-law index $\alpha^{\rm
(b)}_{\gamma}=1.6\pm0.2$ (i.e., $\gamma$ band - broad class), consistent with the Kolmogorov spectral value $5/3$. However, note from Fig. \[PDS\_broad\]a that for different bursts the latter component appears to vary much more than the first.
As we discussed in Sect. \[intro\], the obtained mean PDS (as well as the median) may depend on the chosen normalisation since each burst will be [*weighted*]{} in a different way. If all LCs in the sample are realisations of the same stochastic process then the norm should simply improve the convergence to the mean value but should give identical results for a sufficiently large sample. Aside from the fluence norm, we tested several ways to scale the LCs, e.g., the peak flux used previously by @BSS98 [@BSS00], the @Miy92 normalisation (which expresses the PDS in fractional root-mean-square), and the root of the total power $\sqrt{P_{\rm
tot}}$ used for the ACF. We found in all cases qualitatively the same median PDS for each class but significantly larger spreads. Obviously, this does not prove that each class sample is “uniform” since it is a necessary but not sufficient condition. In this respect, it is worth mentioning that when combined into a single sample and without correcting for redshifts effects we obtained a median PDS (shown in Fig. \[mean\_all\]) that is well fitted in the 0.01–1 Hz frequency range by a single power-law with index $1.73\pm0.07$, consistent within uncertainties with the Kolmogorov spectrum and in full agreement with @BSS98. Therefore, our sample of GRBs with known $z$ is not different, at least in this regard, to the general sample of bright BATSE GRBs used in previous works. Note also that the overall dispersion about the median is significantly larger than in Figs. \[PDS\_narrow\] and \[PDS\_broad\] where, despite the smaller sample sizes, redshift corrections and ACF classes have been taken into account.
Analysis of the X-ray LCs {#XLC}
=========================
In this section we use the WFC sample of 13 GRBs with known redshifts (see Table \[tab:sample\]) to extend the previous temporal analysis into the X-ray energy range. These bursts represent a subset of the previously used sample in the $\gamma$-ray band (Sect. \[gLC\]). We find also at these energies significant differences between the two ACF classes although the results inevitable have larger associated uncertainties.
ACFs {#xacf}
----
Although the ACF widths are substantially broader at these lower energies (2–26 keV), the two ACF classes are still easily distinguishable as shown in Fig. \[wfc\_acf\_loc\]. However, the GRB 010222 classified as having a narrow ACF in the $\gamma$ band ([*gray lines*]{}) has broaded to such extent that it falls into the broad width range, although its ACF decays more slowly than any of the broad width cases. As shown in Table \[tab:sample\] where we list all the observed and the local ACF widths, this is the only [*outlier*]{} in the sample. In column 10 we list as a measure of the broadening the ACF width ratio for the $\gamma$ and X-ray bands. While the average width ratio is $\langle
w_{\mathrm X}/w_{\gamma} \rangle \simeq 2$, the broadening in the GRB 010222 case ($ w_{\mathrm X}/w_{\gamma} \simeq 11.5$) is much larger than any other burst in the sample and it is most likely caused by a systematic error, as we will discuss later in Sect. \[tau\_E\]. For this reason this case will be kept separate from the sample for the rest of our WFC data temporal analysis. Consequently, the obtained average local width ACF are $\bar{ w'}^{\rm (n)}_{\rm
X}=(3.1\pm0.5)$ s and $\bar{ w'}^{\rm (b)}_{\rm X}=(13.6\pm0.6)$ s, with sample standard deviations (relative dispersions) $\sigma^{\rm
(n)}_{\rm X}=1.9$ s (60%) and $\sigma^{\rm (b)}_{\rm X}=1.6$ s (12%) for the narrow and broad subsets respectively.
PDS {#xpds}
---
Once again using our sample of WFC bursts with known $z$, we estimate the [*local*]{} median PDS for each ACF width class following the same methods used in Sect. \[gpds\]. We consider separately the special case GRB 010222 due to the unique broadening of its ACF. Figure \[wfc\_PDS\_n\] shows our results for the narrow class. As expected from the ACF analysis the break appears now at approximately half the corresponding frequency at $\gamma$ energies (Fig. \[PDS\_narrow\]), but most noticeably the behaviour towards the high frequencies now shows a much faster decay. A best-fit for frequencies $f' \gtrsim 0.08$ Hz using a $P_f\propto 1/f^{\alpha}$ model gives an exponent $\alpha^{\rm (n)}_{X}=3.0\pm0.2$, a significantly steeper power-law than in the $\gamma$ band. The PDS of GRB 010222 ([*gray line*]{}) is also shown in Fig. \[wfc\_PDS\_n\] for comparison. Although with much less power at low frequencies than the other bursts in this class, at high frequencies ($f' \ge 0.2$ Hz) it follows a very similar asymptotic behaviour.
Our estimation of the median PDS $\tilde{P}_{f}$ for the broad class is shown in Fig. \[wfc\_PDS\_broad\]. In spite of the large uncertainties associated with such a small sample, the two components found in Fig. \[PDS\_broad\] are recognisable. The low frequency component is not as prominent and broad as before, and consequently the data are not enough to properly constrain the fit parameters of a stretched exponential as it was previously done. However, we estimated that the high frequency component follows a power-law with index $\alpha^{\rm (b)}_{X}=1.7\pm0.2$ consistent within uncertainties with the Kolmogorov index found in Sect. \[gpds\] for the $\gamma$-ray LCs. When compared with the broad class $\tilde{P}_{f}$ in Fig. \[wfc\_PDS\_broad\], the GRB 010222 spectrum matches neither the $\tilde{P}_{f}$ general shape nor any of its components. Power-law fits using the indices and the frequency ranges shown in Figs. \[wfc\_PDS\_n\] and \[wfc\_PDS\_broad\] for the respective high-frequency components resulted in squared residual totals per degrees of freedom $\chi^2/\nu = 16.2/11$ and $\chi^2/\nu = 179./17$ for the narrow and broad PDS respectively. Based on these fits we conclude that the first model is acceptable while the second model can be rejected at a very high confidence level. Uncertainties were estimated using synthetic shot noise and calculating standard deviations on a sample of PDS following the same procedures and frequency binnings applied to the burst data.
Comparing ACF classes {#class}
=====================
In order to understand the origin of the two classes found based on our temporal analysis we looked for any additional GRB characteristic or physical parameter that might differ between them. As point out in Sect. \[gLC\], the visual inspection of the LCs reveals no trivial morphological differences between the classes (see Fig. \[fig:LCs\]). Both classes present cases with simple and complex structure, showing from a few smooth pulses to many heavily overlapped sharp pulses. To cover a broad range of physical parameters, we made use of the database of GRB redshifts and other burst parameters compiled in @FB05 and references therein, that contains much additional information about the 22 bursts in Table \[tab:sample\], although not all parameter estimations are available for every burst. In particular, we considered the following parameters (primed quantities are calculated at the source rest-frame): the peak energy ${E_{\rm pk}}'$, the isotropic equivalent gamma-ray energy $E_{\rm iso}$, the duration time $T'_{90}$, the time of the observed break in the afterglow light curves $T'_{\rm b}$, and the redshift $z$. In all cases we compared the distributions within the narrow and broad ACF classes using the standard Kolmogorov-Smirnov (K-S) test [see, e.g., @Press], which is particularly sensitive to median deviations, but we found no significant difference for the cumulative distributions of any of those parameters. No significant $R$ correlation coefficients with $w'$ were found for ${E_{\rm pk}}'$ and $E_{\rm iso}$ (logarithms taken in all cases), although marginal results were obtained for the temporal parameters $T'_{90}$ ($R=0.44$ with significance $p<0.04$) and $T'_{\rm b}$ ($R=0.46$ with significance $p<0.05$ for only 18 bursts with determined values). Note that a correlation between $w$ and $T_{90}$ has been well established ($R=0.58$), and although in the observer-frame is partially due to the cosmic temporal dilations, it should be found also in the rest-frame [@Bor06]. However, our present sample is too small to determine it at a high confidence level. More surprising is the finding of an indication of some correlation between a prompt ($w'$) and an afterglow ($T'_{\rm b}$) parameter, but this would need better statistics to be confirmed.
Other subdivisions of the long/soft GRB class have been suggested in the literature but we found them to be unrelated to with the one under consideration. Evaluation of the distributions of the observed $T_{90}$ and ${E_{\rm pk}}$ for the broad ACF class excludes an association between these bursts and the “intermediate” duration GRB group reported by @Hor98 and @Muk98. Likewise, we found no correspondence between the ACF bimodality and the bimodality of the afterglow optical luminosity distribution reported by @Nar06 and @LZ06.
ACF energy dependence {#tau_E}
=====================
In our analysis we have only corrected the LCs for cosmic time dilation. However, the data are collected over a finite energy band and therefore additional corrections might be needed to account for the shift in energy, since in principle we should compare LCs emitted in the same energy band. As LC pulses are sharper at higher energies [@Norr96], correspondingly the ACF is narrower. For a sample of 45 bright bursts, found that the width of the mean ACF depends on the energy $E$ as $w(E) \propto E^{-0.43}$. Since for large redshifts the instrument will see photons emitted at higher mean energies, the anti-correlation of the ACF width with the energy should partially counteract the time dilation effect. In possible additional redshift corrections to the accounted cosmic dilation were studied using the transformation $$w'=w/(1+z)^{1+a}
\label{w'}$$ for the observed widths, where $a$ represents a smaller than unity correction. The relative dispersion of each width class was calculated as a function of the parameter $a$ looking for minimum values since, under the assumption that the $w'$ intrinsic distribution does not depend on $z$ (i.e., evolution effects are neglected), redshift dependencies should [*increase*]{} the observed dispersion. We obtained minimal dispersions at $a^{(n)}_{\rm min}=-0.35\pm0.2$ and $a^{(b)}_{\rm min}=0.04\pm0.04$ for the narrow and broad classes respectively, equal within uncertainties to previous results. The uncertainties were estimated numerically using the [*bootstrap*]{} method [see, e.g., @Press]. If the dispersion is calculated over the whole sample (i.e., without separating the two width classes) no well-defined minimum is found, with an approximately constant value within the free parameter range $|a|<1$, which further supports the separation into two classes. However, while the narrow class minimum agrees within errors with the expected effect from the energy shift, the broad class dispersion does not improve much with any additional redshift correction. Actually, in both cases a perfect cancellation (i.e., $a_{\rm min}=0$) cannot be ruled out, although it would appear coincidental that all other possible redshift dependencies would exactly compensate the energy shift. Another possibility is that for the broad class the ACF dependence with energy is weaker. At face value the low dispersion already suggests this and in this case a nearly null $a$ correction would be much more likely. In what follows we will show evidence that supports the latter alternative.
Using 4 energy channel data from a sample of 188 bright BATSE bursts we analyse the ACF width dependence with energy for each individual case using a single power-law model $w(E)~\propto~E^{-\xi}$ (the channels are approximately equally spaced on a logarithmic scale). We found that in most cases the relation $w(E)$ in individual bursts is not well fitted by a power-law (less than 33% of the bursts satisfied for the sum of the reduced residuals $\chi^2_{\rm r}<2$), but for most cases it can be considered a good approximation within the BATSE energy range with typical total relative errors $<10\%$. The obtained distribution of $\xi$ indices is shown in Fig. \[index\_histo\]. Since we found for our fits poorer figures-of-merit (i.e., $\chi^2_{\rm r}$) for larger $\xi$ values, the largest values are related to the cases that strongly deviate from the power-law model for $w(E)$ and are therefore ill determined. For this same sample, a power-law fit of the energy dependence of the width of the [*mean*]{} ACF gives an index $-0.40\pm0.03$ in agreement with . However, note in Fig. \[index\_histo\] that for such a long-tailed distribution most individual bursts typically have a weaker energy dependence and therefore the median $\tilde{\xi}_{\rm
1ch}=0.29$ is a more representative value of individual indices.
In addition, since we are particularly interested in quantifying the effect of the energy dependence in our ACF analysis where the observation energy window has the width of two BATSE channels, we studied the ACF width energy dependence $w(E)$ combining the 4-channel data into three overlapped 2-channel width energy bands. The distribution of the thus obtained indices $\xi_{\rm 2ch}$ is shown in Fig. \[index\_histo\] ([*dashed lines*]{}). The median $\tilde{\xi}_{\rm 2ch}=0.21$ is significantly smaller, as expected since less correction will be needed the wider the energy window. If we consider the median value of the ACF broadening $w_{\mathrm
X}/w_{\gamma}$ for the sample in Table \[tab:sample\] (excluding GRB 010222) we derive a median index $\tilde{\xi} \approx 0.20$, in good agreement with the 2-channel energy window distribution in Fig. \[index\_histo\], either taking the lower edge of each energy window or their geometric means. In Table \[tab:sample\] we list our estimations of the index ${\xi}_{\rm 2ch}$ for the 11 BATSE GRBs with known $z$. Both ACF width classes show typical values that seem consistent with the general distribution in Fig. \[index\_histo\], but unfortunately the small sample size only allowed us to distinguish very substantial differences.
Since we do not have redshifts for most of the BATSE GRBs, the estimated distribution of index $\xi$ was derived mixing both ACF width classes; nevertheless it is probably fairly representative of the narrow class that dominates the sample. In principle, we do not know how representative that general distribution is of the broad class. For our sample of 11 BATSE GRBs with known $z$ the fraction of broad cases is $f_{\rm b}=4/11$ ($f_{\rm b}=7/22$ for our whole sample). However, since the afterglow localisations needed for the redshift determinations always involved another instruments with different trigger responses, that fraction is most likely not representative of the general BATSE catalog. An analysis of the ACF widths of the 188 brightest long BATSE bursts sample shows that the distribution is well described by a log-normal distribution with a mean $10^{\langle \log w \rangle}=3.0$ s and a standard deviation representing a factor of $3.1$, with no significant hint of an underlying bimodality [see also @Bor06]. Since all 11 BATSE GRB with known redshift have $w>2.5$ s almost half of the BATSE $w$ distribution without redshifts seems truncated. Using the K-S test that compares cumulative distributions we calculate a chance probability $p<0.02$ that they are drawn from the same distribution. We conclude that the sample with known $z$ is significantly biased towards large ACF widths, mainly affecting the representation of the narrow class. Taking into account this truncation and the fact that for a burst to belong to the broad class its observed ACF width must be $w\gtrsim 7$ s independently of the redshift, we estimate that the unbiased BATSE fraction of broad cases should lie approximately between $0.08 \lesssim f_{\rm b} \lesssim 0.22$ to be consistent with the observations.
Having characterised the energy dependence of the timescale $w$ with the index $\xi$, we investigated possible correlations with other GRB parameters, but bearing in mind that the largest $\xi$ values are most likely meaningless. We tested a broad range of temporal and spectral parameters (e.g., duration time $T_{90}$, emission time, ACF width $w$, time lag between energy channels, $\nu F_\nu$ peak energy $E_{\rm
pk}$, low and high spectral indices $\alpha$ and $\beta$) following definitions and methods described in @Bor06 for a multivariate correlation analysis. In particular, we used a measure of the overall spectral evolution of a burst introduced there, where the LC is divided in two fluence-halves and the peak energies of the corresponding integrated spectra are estimated and compared. In that way a ratio of peak energies is defined as ${{\cal R}E_{\rm pk}}\equiv
{{E_{\rm pk}}}^{(1)}/{{E_{\rm pk}}}^{(2)}$, so that ${{\cal R}E_{\rm pk}}>1$ implies an overall hard to soft evolution. Although we found no significant global correlations with the index $\xi$, there are indications of a different behaviour at low $\xi$ values in two of the studied parameters, i.e., $w$ and ${{\cal R}E_{\rm pk}}$.
In Fig. \[xi\_corrs\] we show scatter plots of the ACF width $w$ and the ratio of peak energies ${{\cal R}E_{\rm pk}}$ versus the index $\xi$. The $\xi$ observed range was divided into four bins with equal number of data points, and the corresponding geometric means (more appropriate for log-normal distributions) and uncertainties are shown. Due to the large spread the variables seem uncorrelated based on their correlation coefficients $R$, nevertheless the binning reveals a trend where lower $\xi$ values correspond on average to broader ACF widths $w$ and lower values of the ${{\cal R}E_{\rm pk}}$ ratio. In addition, a weak but significant anti-correlation exists between $w$ and ${{\cal R}E_{\rm pk}}$ with coefficient $R=-0.25$ taken their logarithms. For $\xi\lesssim0.3$ there is noticeably less dispersion in $\log
{{\cal R}E_{\rm pk}}$ (i.e., less relative dispersion in ${{\cal R}E_{\rm pk}}$) and a clear positive correlation that comprises half of the sample (with coefficient $R=0.45$). Comparing a linear fit to those data to the no correlation null hypothesis gives for the F-test (the standard test for a ratio of $\chi^2$ values) a level of rejection $p<0.001$. However, since we do not have an objective criterion to truncate the index $\xi$ distribution, the real significance of this found correlation is hard to evaluate; therefore we approach the problem using Monte Carlo methods. We draw random samples for both parameters using the bootstrap method so that they are uncorrelated. Then sorting the data based on their $\xi$ values, we look for the $\xi$-truncate that would give the maximum chance correlation with ${{\cal R}E_{\rm pk}}$. Assuming conservatively that the correlation has to comprise at least 1/3 of the sample and $|R|>0.3$ then the trial is judged as successful. In this way we found a probability of $p<0.008$ of finding a similar or stronger correlation by chance. Although the energy dependence of the ACF is clearly a consequence of the spectral variability, the spectral evolution of a burst is as complex as its LC, shifting from hard to soft over each individual pulse [@BR01]. For this reason it is only under certain limited conditions that we are able to actually observe a correlation between $\xi$ and ${{\cal R}E_{\rm pk}}$. In summary, from the analysis of Fig. \[xi\_corrs\] there are clear indications that bursts with broad ACF on average show weaker ACF energy dependence and less spectral evolution.
Finally, we attempted to extend the analysis of the ACF energy dependence to X-ray energies with the use of 3-channel data from the WFC bursts which were simultaneously observed by BATSE. However, only the signal of the very brightest of the WFC bursts were found suitable for this purpose and the resultant sample was considered too small for a systematic broad-band analysis. Furthermore, although for most of these bursts $w(E)$ shows a smooth, approximately power-law behaviour, we observed in a significant number of bursts a function discontinuity or “shift” between the WFC and BATSE data, a systematic error most likely due to the background estimation. The observed fraction of cases that present this artifact ($\lesssim 20 \%$) is consistent with the occurrence of one such case within the 13 WFC bursts with known $z$ in Table \[tab:sample\]. Therefore, we conclude that the unusual ACF broadening of the [*outlier*]{} case GRB 010222 discussed in Sect. \[xacf\] is most likely an artifact. The PDS analysis also supports this since an incorrect background subtraction would mainly affect the power at low frequencies, and we have shown in Fig. \[wfc\_PDS\_n\] that GRB 010222 deviates from the narrow class typical behaviour only at those frequencies.
Discussion
==========
The reported ACF bimodality in was based on the analysis of a small sample of GRBs with available $z$. Nevertheless, the gap between the two ACF classes was found to be highly significant (with $p<6 \times 10^{-7}$). Low-number statistics can often suggest nonexistent features and should be regarded with caution. As always, one should consider the possibility of a selection effect not accounted for in the probability estimation. However, while it is easy to conceive of upper and lower biases in the observed ACF width distribution (e.g., due to trigger criteria), it is harder to think of a reason why this would happen for middle range $w'$ values. We have shown in Sect \[tau\_E\] that, when compared to a complete, flux limited large sample of BATSE bursts, our sample with known $z$ is likely biased against very narrow ACFs. The addition of GRBM data with their larger intrinsic dispersion only slightly increased the statistical significance of the found ACF bimodality. Nevertheless, with three sets of $\gamma$-ray data showing the same bimodal pattern (Fig. \[acf\_g\]) we can safely rule out an instrumental effect as the cause of it.
It is clear from the PDS analysis presented in Sects. \[gLC\] and \[XLC\] that narrow and broad ACF bursts have very different overall variability, with the latter showing a prominent low-frequency PDS component with a very sharp cut-off, which explains in part the associated low dispersion observed in this ACF width class, despite the typical fluctuations in individual burst PDS and the observed dispersion in the high-frequency power-law component. Another possible difference between the two ACF classes is hinted at by the trends shown in Fig. \[xi\_corrs\], which could be interpreted as an indication that broad ACF bursts show less spectral evolution than the narrow ones. This view is indirectly supported by the comparison of the respective median PDS $\tilde{P}_{f}$ at different energies estimated in Sects. \[gpds\] and \[xpds\]. The narrow class exhibits a single power-law PDS with index $\alpha^{(n)}_{\gamma}\approx2$ at high energies and a steeper $\alpha^{(n)}_{X}\approx3$ low energies. In [*shot noise*]{} models where LCs are generated by random pulses, an index 2 commonly appears when the pulses have sharp rising phase, like the typical FRED-like pulses in many GRBs. But due to the hard-to-soft evolution [@BR01] the same pulses at X-ray energies look smoother with slower rise phases giving steeper indices. On the other hand, the high-frequency power-law component in the broad class PDS shows approximately the same index in both energy bands (close to the $5/3$ Kolmogorov index). Visual inspection in the $\gamma$ and X-ray bands of the bright broad ACF case GRB 990123 [see, e.g., @Cor05], which shows well-separated pulses, indicates that the broadening of the ACF at low energies is not as much due to a corresponding smoothing of individual short pulses as to a decrease in their relative importance with respect to a long smooth LC component that seems to lie underneath them. If this case were representative of the whole class then that would explain in terms of LC components the observed $\tilde{P}_{f}$ behaviour at different energies.
One remarkable feature of the found ACF bimodality is the low relative dispersion of the broad class ($\sim 6\%$ in the $\gamma$ band) around the $\bar{w'}^{\rm (b)}_{\gamma}=(7.42\pm0.14)$ s timescale. If confirmed such a temporal feature would have very important implications, both for our physical understanding of GRBs and because of its potential use in cosmological studies. A crucial issue then is to properly account for redshift dependencies. The reported ACF widths were derived taking into account only the cosmic time dilation, since other effects are of a lesser order. The latter were quantified by means of the free parameter $a$ included in Eq. \[w’\] width transformation. We found well-defined minima at $a^{(n)}_{\rm
min}=-0.35\pm0.2$ and $a^{(b)}_{\rm min}=0.04\pm0.04$ for the narrow and broad ACF class respectively . Since reported a considerable energy dependence for the average ACF ($w(E) \propto E^{-0.43}$) then the cosmic energy shift would be able to account within errors for the found $a_{\rm min}$ in the narrow class. However, for the broad class either we would need to assume that there are other dependencies at work that coincidentally cancel out almost exactly the energy shift effect or that simply the ACF energy dependence is weaker in this case. From the analysis of this alternative presented in Sect. \[tau\_E\] we conclude that in the context of our ACF study the power-law energy dependence is better characterised by a median index $\tilde{\xi}_{\rm 2ch}=0.21$, still consistent with the narrow class within errors and a considerably smaller effect to account for. We evaluated by Monte Carlo simulations that the estimated $w'$ dispersion (for a broad class sample of only 7 bursts) is smaller but not significantly different to the one expected assuming the ${\xi}_{\rm 2ch}$ distribution found in Fig. \[index\_histo\]. Moreover, we found indications that the broad ACF class may have lower typical $\xi$ indices than the general sample. This would be an expected consequence of the weaker spectral evolution already mentioned above when comparing the two classes.
Based on the analysis of WFC X-ray data, @Vet06 reported that in a fraction of bursts the existence of two LC components is apparent. They selected 10 bursts by visual inspection and found for their [*slow/fast*]{} (low/high frequency variability) LC decomposition that the slow components were in general softer than the corresponding fast ones, suggesting different underlying emission processes. There are only 3 GRBs that they claim to have slow components among our sample with known $z$, but they all show broad ACFs in the source rest-frame (including the outlier GRB 010222), which suggests a connection to our own results that will need further analysis to be established. The discussion about the nature of the slow/fast components in @Vet06 would apply also to the low/high frequency components found in broad ACF bursts. Within the framework of the internal shock model developed by @MR00, there would be two major radiating regions in the relativistic outflow and the PDS components would arise from the different expected variability of the emission, i.e., from a thermal photospheric component and from the non-thermal flux of optically thin dissipative regions above the photospheric radius. In this context, ACF class differences could be due to the conditions controlling the relative strength of those two emission components [@Ryde06].
We thank S. Larsson and C.-I. Björnsson for useful comments and careful reading of the manuscript. We are grateful to the anonymous referee for constructive criticism. This research has made use of BATSE and Konus data obtained from the High Energy Astrophysics Science Archive Research Center (HEASARC), provided by NASA’s Goddard Space Flight Center.
Amati, L., et al. 2000, Science, 290, 953
Amati, L., et al. 2002, , 390, 81
Andersen, M. I. et al. 2000, , 364, L54
Antonelli, L. A., et al. 2000, , 545, L39
Aptekar, R. L. et al. 1995, Space Science Reviews, 71, 265
Belli, B. M. 1992, , 393, 266
Beloborodov, A. M., Stern, B. E., & Svensson, R. 1998, , 508, L25
Beloborodov, A. M., Stern, B. E., & Svensson, R. 2000, , 535, 158
Beuermann, K., et al. 1999, A&A, 352, L26
Bloom, J. S., et al. 1999, , 401, 453
Bloom, J. S., Djorgovski, S. G., & Kulkarni, S. R. 2001, , 554, 678
Bloom, J. S., Berger, E., Kulkarni, S. R., Djorgovski, S. G., & Frail, D. A. 2003, , 125, 999
Borgonovo, L. 2004, , 418, 487 (B04)
Borgonovo, L., & Ryde, F. 2001, , 548, 770
Borgonovo, L., & Björnsson, C.-I. 2006, ApJ, 652, 1423
Borgonovo, L., Frontera, F., Guidorzi, C., Montanari, E., & Soffitta, P. 2005, Il Nuovo Cimento, 28 C, 275
Corsi, A., et al. 2005, , 438, 829
Costa, E., et al. 1997, , 387, 783
Djorgovski, S. G., et al. 1998, ApJ, 508, L17
Djorgovski, S. G., Frail, D. A., Kulkarni, S. R., Bloom, J. S., Odewahn, S. C., & Diercks, A. 2001, , 562, 654
Dodonov, S. N., Afanasiev, V. L., Sokolov, V. V., Moiseev, A. V., & Castro-Tirado, A. J. 1999, GCN Report 475
Fenimore, E. E., int’t Zand, J. J. M., Norris, J. P., Bonnell, J. T., & Nemiroff, R. J. 1995, , 448, L101 (F95)
Fishman, G. J., et al. 1989, in Proc. of the GRO Science Workshop, ed. W. N. Johnson, 2
Friedman, A. S., & Bloom, J. S. 2005, , 627, 1
Frontera, F., Costa, E., dal Fiume, D., Feroci, M., Nicastro, L., Orlandini, M., Palazzi, E., & Zavattini, G. 1997, , 122, 357
Garnavich, P. M. et al. 2003, , 582, 924
Greiner, J., et al. 2003, GCN Report 2020
Horv[á]{}th, I. 1998, , 508, 757
Jager, R., et al. 1997, , 125, 557
Jha, S. et al. 2001, , 554, L155
Kolmogorov, A. 1941, Dokl. Akad. Nauk SSSR, 31, 538
Koshut, T. M., Kouveliotou, C., Paciesas, W. S., van Paradijs, J., Pendleton, G. N., Briggs, M. S., Fishman, G. J., & Meegan, C. A. 1995, , 452, 145
Kulkarni, S. R., et al. 1998, Nature, 393, 35
Kulkarni, S. R., et al. 1999, Nature, 398, 389
Lamb, D. Q., Castander, F. J., & Reichart, D. E. 1999, , 138, 479
Lazzati, D. 2002, , 337, 1426
Liang, E., & Zhang, B. 2006, , 638, L67
Link, B., Epstein, R. I., & Priedhorsky, W. C. 1993, , 408, L81
M[é]{}sz[á]{}ros, P., & Rees, M. J. 2000, , 530, 292
Metzger, M. R., et al. 1997, Nature, 387, 878
Miyamoto, S., Kitamoto, S., Iga, S., Negoro, H., & Terada, K. 1992, , 391, L21
Mukherjee, S., Feigelson, E. D., Jogesh Babu, G., Murtagh, F., Fraley, C., & Raftery, A. 1998, , 508, 314
Nardini, M., Ghisellini, G., Ghirlanda, G., Tavecchio, F., Firmani, C., & Lazzati, D. 2006, , 451, 821
Nakar, E., & Piran, T. 2002, , 331, 40
Norris, J. P., Nemiroff, R. J., Bonnell, J. T., Scargle, J. D., Kouveliotou, C., Paciesas, W. S., Meegan, C. A., & Fishman, G. J. 1996, , 459, 393
Piro, L. et al. 2002, , 577, 680
Press, W. H., Teukolsky, S. A., Vetterling, W. T., & Flannery, B. P. 1992, Numerical Recipes in Fortran (2d ed.; Cambridge: Cambridge Univ. Press )
Price, P. A., et al. 2002, , 571, L121
Reichart, D. E., Lamb, D. Q., Fenimore, E. E., Ramirez-Ruiz, E., Cline, T. L., & Hurley, K. 2001, , 552, 57
Ryde, F., Bj[ö]{}rnsson, C.-I., Kaneko, Y., M[é]{}sz[á]{}ros, P., Preece, R., & Battelino, M. 2006, , 652, 1400
Tinney, C., et al. 1998, IAU Circular 6896
Vetere, L., Massaro, E., Costa, E., Soffitta, P., & Ventura, G. 2006, , 447, 499
Vreeswijk, P. M., et al. 1999, GCN Report 496
Vreeswijk, P. M., et al. 2001, , 546, 672
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'During the execution of large scale construction projects performed by Virtual Organizations (VO), relatively complex technical models have to be exchanged between the VO members. For linking the trade and transfer of these models, a so-called multi-model container format was developed. Considering the different skills and tasks of the involved partners, it is not necessary for them to know all the models in every technical detailing. Furthermore, the model size can lead to a delay in communication. In this paper an approach is presented for defining model cut-outs according to the current project context. Dynamic dependencies to the project context as well as static dependencies on the organizational structure are mapped in a context-sensitive rule. As a result, an approach for dynamic filtering of multi-models is obtained which ensures, together with a filtering service, that the involved VO members get a simplified view of complex multi-models as well as sufficient permissions depending on their tasks.'
author:
- Frank Hilbert
- 'Raimar J. Scherer'
- Larissa Araujo
bibliography:
- 'generic.bib'
nocite: '[@*]'
title: 'Multi-model-based Access Control in Construction Projects'
---
Introduction
============
For the processing of large and complex construction projects, single independent organizations work together as Virtual Organizations (VO) \[1\], combining their core competencies \[2, 3\]. To manage the complexity in construction in which the costs incurred within a project vary, both physical and functional information split into various domain-specific application models (e.g. building model, cost model, time model, etc.) on the basis of which the work is running, typically discipline-specific \[4\]. The use of Building Information Modelling (BIM) is becoming increasingly important for collaborative processing \[5\]. The elements of the various specialized models are implicitly linked together (e.g. the component inside wall “xy” from the model building with the process “concrete first floor” from the process model). To transfer application models together with their links, in the German Project MEFISTO [^1] and in the international Project HESMOS [^2] , a multi-model container format (MMC) was developed which summarizes application models with their dependencies \[6\]. When using this MMC approach, however, the following aspects must be considered:
- (A1) The MMC, based on the number and granularity of the application models, becomes relatively large. Because of distributed processing in Virtual Organizations, these multi-models are shared very often and the size of the model can lead to the obstruction of communication (e.g. the size of the IFC building model of a simple high-rise building is about 40 MB).
- (A2) For the processing of individual project tasks only a relatively small part of the subject model is usually interesting. According to their current project responsibilities the agents are only interested, in most cases, in the specific aspects (e.g. building services, building spatial structures, structural analysis, etc.).
- (A3) It is not necessary, and for reasons of data protection, may also not be desirable that all partners involved must know all the technical details of the project model.
From these considerations, it is useful to facilitate the working with multi-models, the transmitting of model cut-outs or the generating of model views. For this purpose, a central filter and mapping service are currently being developed at the TU Dresden, which generate partial application models while maintaining the link structure between the application models \[7\]. This filtering service is used within this approach to enforce the definition of model views and covers the aspect of access and privacy when accessing multi models within virtual organizations. In this manner, an initiated approach, as to the integration of these filter services in a context-sensitive fine-grained access control, is described. The next section briefly describes the structure of the multi-model used. On this basis, in Chapter 3, different multi-model filter designs are considered; first the uses of multi-model templates for the generation of model views, and second, the filtering of elementary models for creating cut-outs. Then, in chapter 4, the idea of an extended context-sensitive role model for virtual organizations is pursued, and in chapter 5, a proposed methodology for using this model for the filtered access to multi-models follows. The last chapter outlines practical scenarios and gives an outlook into further development.
Basic Concepts
==============
The multi-model container format
--------------------------------
In order to distribute project information together from the various semantically and structurally inhomogeneous domain expert models and, on the other side, externalize the implicit relationship between the model elements, a generic link model was developed, which will be shown together with the associated application models within a multi-model. Then the link model in the form of an XML document is stored, the format of the application models is not specified. The multi-model was changed in a so-called multi-model container (MMC) which is within the multi-model container and separated in each model, described by metadata (see Figure 1). These multi-model containers are read into the specialized applications of the VO partners, modified and shared with other partners. More information about the structure and the use of MMC is given by Fuchs in \[6\].
Filter Concepts for Multi-Models
--------------------------------
In the filtering of multi-models it is necessary to distinguish between the pure selection of application models as they are defined for model-views (domain-views) with any setting of the desired granularity (see Figure 2) and of the filter model to generate model cut-outs (see Figure 3).
![Example of a Multi-Model-Cutout[]{data-label="Bild_B"}](figures/figure02)
![Example of a Multi-Model-Cutout[]{data-label="Bild_B"}](figures/figure03)
For handling different tasks, it is often not necessary for all application models to be available. The use of model views serves not only for the simplified representation of complex issues but also to increase the security against unauthorized data access and is achieved by the simplest form of filtering through the use of multi-model templates.
In the MEFISTO Project seven exemplary different multi-model templates have been defined which are based on the phases of the project and selected in the first step of each different subset of application models (Building, Operation, Performance, Costs and Construction Site Model). In a second step, the associated link model can be reduced to the remaining links of the tray model elements. Then, the templates demand not only the existence of application models in the multi-model, but claim additionally the compliance model-dependent levels of detail and processing status. It is conceivable that in a production model, the application building model (BM) and operation model (OM) are structured much more comprehensively than within the model range. In a third step, these additional conditions are checked using the metadata within the container. In Table 1, examples of granularities are outlined (e.g. in the application model the process model-column templates of the frame schedule (0.1), coarse schedule (0.2)) to detail schedule (1)). Other various project plans are discussed in more detail in \[4\].
[|c|c|c|c|c|c|c|c|]{} **Project Phase** & **Task** & **MM-Template** & **BM** & **OM** & **PM** & **CM** & **CSM**\
Offer phase & Offer request & Tender Model & 0,5 & 0,1 & 0,1 & 0 & 0\
Offer phase & Offer delivery & Offer Model & 0,5 & 0,2 & 1 & 0,5 & 0\
Contract phase & Negotiation & Negotiation Model & 0,5 & 0,2 & 1 & 0,7 & 0\
Contract phase & Commissioning & Contract Model & 1 & 1 & 1 & 1 & 0\
Execution phase & Scheduling& Tender Model & 1 & 1 & 1 & 1 & 0,7\
For the selection of individual object fields for further processing or to facilitate semantic multi-model filter, the use of specialized model cut-outs is desirable. Then these excerpts can turn into various, in part dynamic, criteria to be defined, for example as geometric part-models, time periods or (external) object properties of the model elements. The generation of task-specific ad-hoc multi-model views, which are combined to link models of BIM data with information from other application models is relatively complex and is a topic of current research. Katranuschkov describes in \[7\] an approach for the generation of BIM-based multi-model views, and a reference implementation, based on the use of open IFC toolset from the University of Weimar and HOCHTIEF that, with the use of the generic Model Subset Definition Schema (GMSD), focuses on these goals. Another challenge is the generation of partial geometric models, here under certain circumstances broken up into objects, because it must be generated at different levels of the object’s detail. This aspect is the subject of current research within the project HESMOS and is not described in this work.
Context-dependent Access Control
=================================
Camarinha-Matos differentiated in \[8\], exogenous and endogenous VO roles, with the former relationships reflecting outwards, such as interaction with the environment, customers, competitors and potential partners, and on the other side endogenous roles which are represented by the relationships within the VO. These endogenous roles handle the VO participants during the exercising of their cooperation and are defined in the VO-initiation, as well as other descriptive attributes, in the Organization Model. The classic role-based access control (RBAC) \[9\] defines the relationship in a simple way:
*WHO (project role) may WHAT (authorizations)?*
Hence, VO participants may bring different talents or skills into a VO, so that the VO actors are assigned with so-called potential roles. Both in \[10\] and in \[11\], it was found that the role assignment of the VO members is not rigidly fixed in practice, but contextual and dynamic role assignments are subject to restrictions and must therefore be context-dependently modelled and dynamically evaluated. Only with this combination, the observance of the principle of Separation of Duty \[9\] and the Principle of Least Privilege \[12\] can be ensured. We use an expanded role model for which the access control including the access context offers and determines illustrated in simplified form:
*WHO (current role) may WHAT (authorizations) with WHOM (object) in which SITUATION (context)?*
An example scenario: A player is entrusted with the tasks within the VO plan creation and plan evaluation. To ensure the Separation of Duty, this actor should not be allowed to control himself, so both roles can only be selected when accessing different objects.
The use of context information
-------------------------------
To add context-sensitive abilities to security mechanisms, context attributes must be described in a formal context model. The context attributes are not individually described within this context model, but they are assigned real world objects. This so-called deputy approach describes a basic concept for the representation of contextual information \[13\] and link distributed context attributes of the subject and the object model together with the environment context attributes. Using a context model now we can combine previously individually held context information so that situations can be precisely described. The relevant relationships for access decisions can be complex and involve various types and combinations of context information. For the modelling of the knowledge about their relationships, the impact of this information has to be formally described. For the definition of the access control policies, we use the Web Rule Language (SWRL) (Horrocks et al., 2003) to achieve flexibility and interoperability as well as easy administration. Some examples of SWRL rules:
- Role assignment:\
(Actor(?a), Role(?e), Resource(?r), hasTarget(?a,?r), hasRole(?a,?e), hasOwner(?r,?a)) ->PermittedRole(?e)
- Ownership:\
(Action(?a), Resource(?r), hasTarget(?a,?r), hasActor(?a,?x), hasOwner(?r,?x)) ->PermittedAction(?a)
- Sstate dependence:\
(Action(?a), Resource(?r), hasTarget(?a,?r), hasObjectRestriction(?a,?x), hasPassRestriction(?r,?x)) ->PermittedAction(?a)
Structure of the VO model
-------------------------
We have modelled the RBAC-Model in our VO-Ontology model with the classes Actor, Role, Permission and Restriction. Objects of the Class Restriction define the context sensitive Assignment-Restriction in the form of SWRL-rules. The Relation sr is modelled by the bijective mapping of the ObjectPropertys hasActorRoleAssignment and hasRoleActorAssignment, which connect the classes Actor and Role.
Both, the ObjectPropertys hasPermissionRoleAssignment and hasRolePermissionAssignment connect the classes Role and Permission and reflect the Relation sr. The SubPropertys hasRoleRestriction and hasPermissionRestriction link the Class Restriction to the ObjectPropertys hasActorRoleAssignment and hasPermissionRoleAssignment and realize the functions fc(sr) and gc(pr) (in Figure 3 dotted). The Restriction-Objects contain the restrictions to be evaluated, where the „Open World Assumption“ is valid, i.e. the modelled Relations sr and pr are true, if there are no Restrictions or the Restrictions can be evaluated to be true.
Dynamic Authorization Decisions
--------------------------------
Scholz describes in \[15\] mutual trust as the basis for VOs and the transparent handling of explicit rules and standards as the basis for this confidence. For the rules of a context-based access control for multi-models, we use the Fine Grain Access Control (FGAC) rules described by Franzoni in \[16\] with a context-sensitive expansion, which additionally accesses the context in which access decisions are integrated. The actual access role when accessing an object instance from a dynamic access control according to the access context, consists of various determined static (potential roles with permissions) and dynamic conditions (object status and subject status, subject-object relationships and project status). This consists mainly of an empty MMC with pre-populated metadata descriptions. Using these templates it is easy to determine whether a padded MMC corresponds to a template. In MEFISTO a VO-based service platform is used which registers all created MMC templates and the metadata of all transmitted application models.
We propose a context-sensitive access control in 5 steps:
1. The role of a requesting actor is determined by evaluating its potential VO roles and the access context.
2. The User ID is checked in the VO registry and connected access rules are evaluated.
3. With the Task description in the core ontology, the necessary permissions are evaluated
4. The task-dependent Template is searched on the VO-platform. If there is no template, then it will return an unfiltered Multi-Model, according to the preferences, or the access will not be permitted. If there is a template, it is created on the basis of a container template a Multi-Model-Container through the base service. Then, the registered trade models use the service from the platform and search in the model registry for an application model in the desired resolution.
5. If an application model does not exist in the required granularity, the subject-specific model type of filter methods and services are requested and an application model with higher granularity and the desired level of detail are given as an argument. The application model produced is inserted into the Multi-Model-Container.
6. This Multi-Model-Container is transferred to the actor.
Related work
============
The common standardised access model role-based access control model (RBAC) \[9\] is based on the concept of the user role as an intermediary between subjects and permissions and is purely subject-based. The properties of the objects have no effect on the access decision. Therefore, role assignments may be wrong after a state change of the entities (objects as well as subjects). Various proposals have extended the RBAC model with attribute-based access control capabilities \[17\] which also consider environment attributes (e.g. system status, time, date), subject attributes (e.g. age, location, proofs of identity) as well as object attributes (e.g. size, value, location, state of objects). Priebe et al. \[18\] proposes an attribute-based access control approach, but roles are subject- dependently modelled. Kumar et al. use in \[19\] context filters for role assignment, based on hash tables describing user context and object context-attributes. Franzoni et al. presents in \[16\] a fine-grained access control model, with focus on Semantic Databases, using ontologies to describe the subject-object relationships. Last but not least, Goslar describes a context model with dynamic real-world object linking in an economic focus \[13\]. Common to all these approaches is an attribute based assignment of subjects to roles while the roles-permissions assignment remains static.
Outlook
=======
It has outlined an approach which covers the aspects mentioned in chapter 1: (A1) Due to one filter of the Multi-Model-Container, the communication between the partners was simplified and accelerated. Example, it is now possible to use Multi-Model-Container also on mobile devices. (A2) Only those application models necessary for the current processing situation were transferred to the Multi-Model-Container. (A3) The transferred application models have the exact resolution (size) needed for further processing.
Both the major components of the filtering services as well as most of the required multi-model templates are still under development, making the approach only roughly sketched. Currently in the focus of the research at the TU Dresden are the methods for the generation of geometric partial models and scheduling time spans.
Another development is focused on the filtering of MMC for transitive object properties of model elements. An example scenario for this comes from subsequent processing. Here, the complete reduction of MMC to a supplementary multi-model is desired which contains only subsequent positions in the application model specifications, corresponding to a reduced geometry and link model. Such application model sections, due to the dynamic component, cannot be generated with template models.
[^1]: http://www.mefisto-bau.de
[^2]: http://hesmos.eu
| {
"pile_set_name": "ArXiv"
} | ArXiv |